xref: /dflybsd-src/sys/vm/vm_fault.c (revision e5e174adcc2678e90a7ffbe469dfe34d9989aa09)
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 	if ((curthread->td_flags & TDF_NOFAULT) &&
472 	    (retry ||
473 	     fs.first_object->type == OBJT_VNODE ||
474 	     fs.first_object->backing_object)) {
475 		result = KERN_FAILURE;
476 		unlock_things(&fs);
477 		goto done2;
478 	}
479 
480 	/*
481 	 * If the entry is wired we cannot change the page protection.
482 	 */
483 	if (fs.wired)
484 		fault_type = fs.first_prot;
485 
486 	/*
487 	 * We generally want to avoid unnecessary exclusive modes on backing
488 	 * and terminal objects because this can seriously interfere with
489 	 * heavily fork()'d processes (particularly /bin/sh scripts).
490 	 *
491 	 * However, we also want to avoid unnecessary retries due to needed
492 	 * shared->exclusive promotion for common faults.  Exclusive mode is
493 	 * always needed if any page insertion, rename, or free occurs in an
494 	 * object (and also indirectly if any I/O is done).
495 	 *
496 	 * The main issue here is going to be fs.first_shared.  If the
497 	 * first_object has a backing object which isn't shadowed and the
498 	 * process is single-threaded we might as well use an exclusive
499 	 * lock/chain right off the bat.
500 	 */
501 	if (fs.first_shared && fs.first_object->backing_object &&
502 	    LIST_EMPTY(&fs.first_object->shadow_head) &&
503 	    curthread->td_proc && curthread->td_proc->p_nthreads == 1) {
504 		fs.first_shared = 0;
505 	}
506 
507 	/*
508 	 * swap_pager_unswapped() needs an exclusive object
509 	 */
510 	if (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY)) {
511 		fs.first_shared = 0;
512 	}
513 
514 	/*
515 	 * Obtain a top-level object lock, shared or exclusive depending
516 	 * on fs.first_shared.  If a shared lock winds up being insufficient
517 	 * we will retry with an exclusive lock.
518 	 *
519 	 * The vnode pager lock is always shared.
520 	 */
521 	if (fs.first_shared)
522 		vm_object_hold_shared(fs.first_object);
523 	else
524 		vm_object_hold(fs.first_object);
525 	if (fs.vp == NULL)
526 		fs.vp = vnode_pager_lock(fs.first_object);
527 
528 	/*
529 	 * The page we want is at (first_object, first_pindex), but if the
530 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
531 	 * page table to figure out the actual pindex.
532 	 *
533 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
534 	 * ONLY
535 	 */
536 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
537 		result = vm_fault_vpagetable(&fs, &first_pindex,
538 					     fs.entry->aux.master_pde,
539 					     fault_type, 1);
540 		if (result == KERN_TRY_AGAIN) {
541 			vm_object_drop(fs.first_object);
542 			++retry;
543 			goto RetryFault;
544 		}
545 		if (result != KERN_SUCCESS)
546 			goto done;
547 	}
548 
549 	/*
550 	 * Now we have the actual (object, pindex), fault in the page.  If
551 	 * vm_fault_object() fails it will unlock and deallocate the FS
552 	 * data.   If it succeeds everything remains locked and fs->object
553 	 * will have an additional PIP count if it is not equal to
554 	 * fs->first_object
555 	 *
556 	 * vm_fault_object will set fs->prot for the pmap operation.  It is
557 	 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
558 	 * page can be safely written.  However, it will force a read-only
559 	 * mapping for a read fault if the memory is managed by a virtual
560 	 * page table.
561 	 *
562 	 * If the fault code uses the shared object lock shortcut
563 	 * we must not try to burst (we can't allocate VM pages).
564 	 */
565 	result = vm_fault_object(&fs, first_pindex, fault_type, 1);
566 
567 	if (debug_fault > 0) {
568 		--debug_fault;
569 		kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x "
570 			"fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n",
571 			result, (intmax_t)vaddr, fault_type, fault_flags,
572 			fs.m, fs.prot, fs.wired, fs.entry);
573 	}
574 
575 	if (result == KERN_TRY_AGAIN) {
576 		vm_object_drop(fs.first_object);
577 		++retry;
578 		goto RetryFault;
579 	}
580 	if (result != KERN_SUCCESS)
581 		goto done;
582 
583 	/*
584 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
585 	 * will contain a busied page.
586 	 *
587 	 * Enter the page into the pmap and do pmap-related adjustments.
588 	 */
589 	KKASSERT(fs.lookup_still_valid == TRUE);
590 	vm_page_flag_set(fs.m, PG_REFERENCED);
591 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot | inherit_prot,
592 		   fs.wired, fs.entry);
593 
594 	/*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
595 	KKASSERT(fs.m->flags & PG_BUSY);
596 
597 	/*
598 	 * If the page is not wired down, then put it where the pageout daemon
599 	 * can find it.
600 	 */
601 	if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
602 		if (fs.wired)
603 			vm_page_wire(fs.m);
604 		else
605 			vm_page_unwire(fs.m, 1);
606 	} else {
607 		vm_page_activate(fs.m);
608 	}
609 	vm_page_wakeup(fs.m);
610 
611 	/*
612 	 * Burst in a few more pages if possible.  The fs.map should still
613 	 * be locked.  To avoid interlocking against a vnode->getblk
614 	 * operation we had to be sure to unbusy our primary vm_page above
615 	 * first.
616 	 *
617 	 * A normal burst can continue down backing store, only execute
618 	 * if we are holding an exclusive lock, otherwise the exclusive
619 	 * locks the burst code gets might cause excessive SMP collisions.
620 	 *
621 	 * A quick burst can be utilized when there is no backing object
622 	 * (i.e. a shared file mmap).
623 	 */
624 	if ((fault_flags & VM_FAULT_BURST) &&
625 	    (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
626 	    fs.wired == 0) {
627 		if (fs.first_shared == 0 && fs.shared == 0) {
628 			vm_prefault(fs.map->pmap, vaddr,
629 				    fs.entry, fs.prot, fault_flags);
630 		} else {
631 			vm_prefault_quick(fs.map->pmap, vaddr,
632 					  fs.entry, fs.prot, fault_flags);
633 		}
634 	}
635 
636 done_success:
637 	mycpu->gd_cnt.v_vm_faults++;
638 	if (curthread->td_lwp)
639 		++curthread->td_lwp->lwp_ru.ru_minflt;
640 
641 	/*
642 	 * Unlock everything, and return
643 	 */
644 	unlock_things(&fs);
645 
646 	if (curthread->td_lwp) {
647 		if (fs.hardfault) {
648 			curthread->td_lwp->lwp_ru.ru_majflt++;
649 		} else {
650 			curthread->td_lwp->lwp_ru.ru_minflt++;
651 		}
652 	}
653 
654 	/*vm_object_deallocate(fs.first_object);*/
655 	/*fs.m = NULL; */
656 	/*fs.first_object = NULL; must still drop later */
657 
658 	result = KERN_SUCCESS;
659 done:
660 	if (fs.first_object)
661 		vm_object_drop(fs.first_object);
662 done2:
663 	lwkt_reltoken(&map->token);
664 	if (lp)
665 		lp->lwp_flags &= ~LWP_PAGING;
666 	if (vm_shared_fault && fs.shared == 0)
667 		++vm_shared_miss;
668 	return (result);
669 }
670 
671 /*
672  * Fault in the specified virtual address in the current process map,
673  * returning a held VM page or NULL.  See vm_fault_page() for more
674  * information.
675  *
676  * No requirements.
677  */
678 vm_page_t
679 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
680 {
681 	struct lwp *lp = curthread->td_lwp;
682 	vm_page_t m;
683 
684 	m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
685 			  fault_type, VM_FAULT_NORMAL, errorp);
686 	return(m);
687 }
688 
689 /*
690  * Fault in the specified virtual address in the specified map, doing all
691  * necessary manipulation of the object store and all necessary I/O.  Return
692  * a held VM page or NULL, and set *errorp.  The related pmap is not
693  * updated.
694  *
695  * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
696  * and marked PG_REFERENCED as well.
697  *
698  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
699  * error will be returned.
700  *
701  * No requirements.
702  */
703 vm_page_t
704 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
705 	      int fault_flags, int *errorp)
706 {
707 	vm_pindex_t first_pindex;
708 	struct faultstate fs;
709 	int result;
710 	int retry = 0;
711 	vm_prot_t orig_fault_type = fault_type;
712 
713 	fs.hardfault = 0;
714 	fs.fault_flags = fault_flags;
715 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
716 
717 	/*
718 	 * Dive the pmap (concurrency possible).  If we find the
719 	 * appropriate page we can terminate early and quickly.
720 	 */
721 	fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type);
722 	if (fs.m) {
723 		*errorp = 0;
724 		return(fs.m);
725 	}
726 
727 	/*
728 	 * Otherwise take a concurrency hit and do a formal page
729 	 * fault.
730 	 */
731 	fs.shared = vm_shared_fault;
732 	fs.first_shared = vm_shared_fault;
733 	fs.vp = NULL;
734 	lwkt_gettoken(&map->token);
735 
736 	/*
737 	 * swap_pager_unswapped() needs an exclusive object
738 	 */
739 	if (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY)) {
740 		fs.first_shared = 0;
741 	}
742 
743 RetryFault:
744 	/*
745 	 * Find the vm_map_entry representing the backing store and resolve
746 	 * the top level object and page index.  This may have the side
747 	 * effect of executing a copy-on-write on the map entry and/or
748 	 * creating a shadow object, but will not COW any actual VM pages.
749 	 *
750 	 * On success fs.map is left read-locked and various other fields
751 	 * are initialized but not otherwise referenced or locked.
752 	 *
753 	 * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
754 	 * if the map entry is a virtual page table and also writable,
755 	 * so we can set the 'A'accessed bit in the virtual page table entry.
756 	 */
757 	fs.map = map;
758 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
759 			       &fs.entry, &fs.first_object,
760 			       &first_pindex, &fs.first_prot, &fs.wired);
761 
762 	if (result != KERN_SUCCESS) {
763 		*errorp = result;
764 		fs.m = NULL;
765 		goto done;
766 	}
767 
768 	/*
769 	 * fs.map is read-locked
770 	 *
771 	 * Misc checks.  Save the map generation number to detect races.
772 	 */
773 	fs.map_generation = fs.map->timestamp;
774 	fs.lookup_still_valid = TRUE;
775 	fs.first_m = NULL;
776 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
777 
778 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
779 		panic("vm_fault: fault on nofault entry, addr: %lx",
780 		    (u_long)vaddr);
781 	}
782 
783 	/*
784 	 * A user-kernel shared map has no VM object and bypasses
785 	 * everything.  We execute the uksmap function with a temporary
786 	 * fictitious vm_page.  The address is directly mapped with no
787 	 * management.
788 	 */
789 	if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) {
790 		struct vm_page fakem;
791 
792 		bzero(&fakem, sizeof(fakem));
793 		fakem.pindex = first_pindex;
794 		fakem.flags = PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED;
795 		fakem.valid = VM_PAGE_BITS_ALL;
796 		fakem.pat_mode = VM_MEMATTR_DEFAULT;
797 		if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) {
798 			*errorp = KERN_FAILURE;
799 			fs.m = NULL;
800 			unlock_things(&fs);
801 			goto done2;
802 		}
803 		fs.m = PHYS_TO_VM_PAGE(fakem.phys_addr);
804 		vm_page_hold(fs.m);
805 
806 		unlock_things(&fs);
807 		*errorp = 0;
808 		goto done;
809 	}
810 
811 
812 	/*
813 	 * A system map entry may return a NULL object.  No object means
814 	 * no pager means an unrecoverable kernel fault.
815 	 */
816 	if (fs.first_object == NULL) {
817 		panic("vm_fault: unrecoverable fault at %p in entry %p",
818 			(void *)vaddr, fs.entry);
819 	}
820 
821 	/*
822 	 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
823 	 * is set.
824 	 */
825 	if ((curthread->td_flags & TDF_NOFAULT) &&
826 	    (retry ||
827 	     fs.first_object->type == OBJT_VNODE ||
828 	     fs.first_object->backing_object)) {
829 		*errorp = KERN_FAILURE;
830 		unlock_things(&fs);
831 		goto done2;
832 	}
833 
834 	/*
835 	 * If the entry is wired we cannot change the page protection.
836 	 */
837 	if (fs.wired)
838 		fault_type = fs.first_prot;
839 
840 	/*
841 	 * Make a reference to this object to prevent its disposal while we
842 	 * are messing with it.  Once we have the reference, the map is free
843 	 * to be diddled.  Since objects reference their shadows (and copies),
844 	 * they will stay around as well.
845 	 *
846 	 * The reference should also prevent an unexpected collapse of the
847 	 * parent that might move pages from the current object into the
848 	 * parent unexpectedly, resulting in corruption.
849 	 *
850 	 * Bump the paging-in-progress count to prevent size changes (e.g.
851 	 * truncation operations) during I/O.  This must be done after
852 	 * obtaining the vnode lock in order to avoid possible deadlocks.
853 	 */
854 	if (fs.first_shared)
855 		vm_object_hold_shared(fs.first_object);
856 	else
857 		vm_object_hold(fs.first_object);
858 	if (fs.vp == NULL)
859 		fs.vp = vnode_pager_lock(fs.first_object);	/* shared */
860 
861 	/*
862 	 * The page we want is at (first_object, first_pindex), but if the
863 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
864 	 * page table to figure out the actual pindex.
865 	 *
866 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
867 	 * ONLY
868 	 */
869 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
870 		result = vm_fault_vpagetable(&fs, &first_pindex,
871 					     fs.entry->aux.master_pde,
872 					     fault_type, 1);
873 		if (result == KERN_TRY_AGAIN) {
874 			vm_object_drop(fs.first_object);
875 			++retry;
876 			goto RetryFault;
877 		}
878 		if (result != KERN_SUCCESS) {
879 			*errorp = result;
880 			fs.m = NULL;
881 			goto done;
882 		}
883 	}
884 
885 	/*
886 	 * Now we have the actual (object, pindex), fault in the page.  If
887 	 * vm_fault_object() fails it will unlock and deallocate the FS
888 	 * data.   If it succeeds everything remains locked and fs->object
889 	 * will have an additinal PIP count if it is not equal to
890 	 * fs->first_object
891 	 */
892 	fs.m = NULL;
893 	result = vm_fault_object(&fs, first_pindex, fault_type, 1);
894 
895 	if (result == KERN_TRY_AGAIN) {
896 		vm_object_drop(fs.first_object);
897 		++retry;
898 		goto RetryFault;
899 	}
900 	if (result != KERN_SUCCESS) {
901 		*errorp = result;
902 		fs.m = NULL;
903 		goto done;
904 	}
905 
906 	if ((orig_fault_type & VM_PROT_WRITE) &&
907 	    (fs.prot & VM_PROT_WRITE) == 0) {
908 		*errorp = KERN_PROTECTION_FAILURE;
909 		unlock_and_deallocate(&fs);
910 		fs.m = NULL;
911 		goto done;
912 	}
913 
914 	/*
915 	 * DO NOT UPDATE THE PMAP!!!  This function may be called for
916 	 * a pmap unrelated to the current process pmap, in which case
917 	 * the current cpu core will not be listed in the pmap's pm_active
918 	 * mask.  Thus invalidation interlocks will fail to work properly.
919 	 *
920 	 * (for example, 'ps' uses procfs to read program arguments from
921 	 * each process's stack).
922 	 *
923 	 * In addition to the above this function will be called to acquire
924 	 * a page that might already be faulted in, re-faulting it
925 	 * continuously is a waste of time.
926 	 *
927 	 * XXX could this have been the cause of our random seg-fault
928 	 *     issues?  procfs accesses user stacks.
929 	 */
930 	vm_page_flag_set(fs.m, PG_REFERENCED);
931 #if 0
932 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL);
933 	mycpu->gd_cnt.v_vm_faults++;
934 	if (curthread->td_lwp)
935 		++curthread->td_lwp->lwp_ru.ru_minflt;
936 #endif
937 
938 	/*
939 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
940 	 * will contain a busied page.  So we must unlock here after having
941 	 * messed with the pmap.
942 	 */
943 	unlock_things(&fs);
944 
945 	/*
946 	 * Return a held page.  We are not doing any pmap manipulation so do
947 	 * not set PG_MAPPED.  However, adjust the page flags according to
948 	 * the fault type because the caller may not use a managed pmapping
949 	 * (so we don't want to lose the fact that the page will be dirtied
950 	 * if a write fault was specified).
951 	 */
952 	vm_page_hold(fs.m);
953 	vm_page_activate(fs.m);
954 	if (fault_type & VM_PROT_WRITE)
955 		vm_page_dirty(fs.m);
956 
957 	if (curthread->td_lwp) {
958 		if (fs.hardfault) {
959 			curthread->td_lwp->lwp_ru.ru_majflt++;
960 		} else {
961 			curthread->td_lwp->lwp_ru.ru_minflt++;
962 		}
963 	}
964 
965 	/*
966 	 * Unlock everything, and return the held page.
967 	 */
968 	vm_page_wakeup(fs.m);
969 	/*vm_object_deallocate(fs.first_object);*/
970 	/*fs.first_object = NULL; */
971 	*errorp = 0;
972 
973 done:
974 	if (fs.first_object)
975 		vm_object_drop(fs.first_object);
976 done2:
977 	lwkt_reltoken(&map->token);
978 	return(fs.m);
979 }
980 
981 /*
982  * Fault in the specified (object,offset), dirty the returned page as
983  * needed.  If the requested fault_type cannot be done NULL and an
984  * error is returned.
985  *
986  * A held (but not busied) page is returned.
987  *
988  * The passed in object must be held as specified by the shared
989  * argument.
990  */
991 vm_page_t
992 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
993 		     vm_prot_t fault_type, int fault_flags,
994 		     int *sharedp, int *errorp)
995 {
996 	int result;
997 	vm_pindex_t first_pindex;
998 	struct faultstate fs;
999 	struct vm_map_entry entry;
1000 
1001 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1002 	bzero(&entry, sizeof(entry));
1003 	entry.object.vm_object = object;
1004 	entry.maptype = VM_MAPTYPE_NORMAL;
1005 	entry.protection = entry.max_protection = fault_type;
1006 
1007 	fs.hardfault = 0;
1008 	fs.fault_flags = fault_flags;
1009 	fs.map = NULL;
1010 	fs.shared = vm_shared_fault;
1011 	fs.first_shared = *sharedp;
1012 	fs.vp = NULL;
1013 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
1014 
1015 	/*
1016 	 * Might require swap block adjustments
1017 	 */
1018 	if (fs.first_shared && (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY))) {
1019 		fs.first_shared = 0;
1020 		vm_object_upgrade(object);
1021 	}
1022 
1023 	/*
1024 	 * Retry loop as needed (typically for shared->exclusive transitions)
1025 	 */
1026 RetryFault:
1027 	*sharedp = fs.first_shared;
1028 	first_pindex = OFF_TO_IDX(offset);
1029 	fs.first_object = object;
1030 	fs.entry = &entry;
1031 	fs.first_prot = fault_type;
1032 	fs.wired = 0;
1033 	/*fs.map_generation = 0; unused */
1034 
1035 	/*
1036 	 * Make a reference to this object to prevent its disposal while we
1037 	 * are messing with it.  Once we have the reference, the map is free
1038 	 * to be diddled.  Since objects reference their shadows (and copies),
1039 	 * they will stay around as well.
1040 	 *
1041 	 * The reference should also prevent an unexpected collapse of the
1042 	 * parent that might move pages from the current object into the
1043 	 * parent unexpectedly, resulting in corruption.
1044 	 *
1045 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1046 	 * truncation operations) during I/O.  This must be done after
1047 	 * obtaining the vnode lock in order to avoid possible deadlocks.
1048 	 */
1049 	if (fs.vp == NULL)
1050 		fs.vp = vnode_pager_lock(fs.first_object);
1051 
1052 	fs.lookup_still_valid = TRUE;
1053 	fs.first_m = NULL;
1054 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
1055 
1056 #if 0
1057 	/* XXX future - ability to operate on VM object using vpagetable */
1058 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1059 		result = vm_fault_vpagetable(&fs, &first_pindex,
1060 					     fs.entry->aux.master_pde,
1061 					     fault_type, 0);
1062 		if (result == KERN_TRY_AGAIN) {
1063 			if (fs.first_shared == 0 && *sharedp)
1064 				vm_object_upgrade(object);
1065 			goto RetryFault;
1066 		}
1067 		if (result != KERN_SUCCESS) {
1068 			*errorp = result;
1069 			return (NULL);
1070 		}
1071 	}
1072 #endif
1073 
1074 	/*
1075 	 * Now we have the actual (object, pindex), fault in the page.  If
1076 	 * vm_fault_object() fails it will unlock and deallocate the FS
1077 	 * data.   If it succeeds everything remains locked and fs->object
1078 	 * will have an additinal PIP count if it is not equal to
1079 	 * fs->first_object
1080 	 *
1081 	 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact.
1082 	 * We may have to upgrade its lock to handle the requested fault.
1083 	 */
1084 	result = vm_fault_object(&fs, first_pindex, fault_type, 0);
1085 
1086 	if (result == KERN_TRY_AGAIN) {
1087 		if (fs.first_shared == 0 && *sharedp)
1088 			vm_object_upgrade(object);
1089 		goto RetryFault;
1090 	}
1091 	if (result != KERN_SUCCESS) {
1092 		*errorp = result;
1093 		return(NULL);
1094 	}
1095 
1096 	if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
1097 		*errorp = KERN_PROTECTION_FAILURE;
1098 		unlock_and_deallocate(&fs);
1099 		return(NULL);
1100 	}
1101 
1102 	/*
1103 	 * On success vm_fault_object() does not unlock or deallocate, so we
1104 	 * do it here.  Note that the returned fs.m will be busied.
1105 	 */
1106 	unlock_things(&fs);
1107 
1108 	/*
1109 	 * Return a held page.  We are not doing any pmap manipulation so do
1110 	 * not set PG_MAPPED.  However, adjust the page flags according to
1111 	 * the fault type because the caller may not use a managed pmapping
1112 	 * (so we don't want to lose the fact that the page will be dirtied
1113 	 * if a write fault was specified).
1114 	 */
1115 	vm_page_hold(fs.m);
1116 	vm_page_activate(fs.m);
1117 	if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY))
1118 		vm_page_dirty(fs.m);
1119 	if (fault_flags & VM_FAULT_UNSWAP)
1120 		swap_pager_unswapped(fs.m);
1121 
1122 	/*
1123 	 * Indicate that the page was accessed.
1124 	 */
1125 	vm_page_flag_set(fs.m, PG_REFERENCED);
1126 
1127 	if (curthread->td_lwp) {
1128 		if (fs.hardfault) {
1129 			curthread->td_lwp->lwp_ru.ru_majflt++;
1130 		} else {
1131 			curthread->td_lwp->lwp_ru.ru_minflt++;
1132 		}
1133 	}
1134 
1135 	/*
1136 	 * Unlock everything, and return the held page.
1137 	 */
1138 	vm_page_wakeup(fs.m);
1139 	/*vm_object_deallocate(fs.first_object);*/
1140 	/*fs.first_object = NULL; */
1141 
1142 	*errorp = 0;
1143 	return(fs.m);
1144 }
1145 
1146 /*
1147  * Translate the virtual page number (first_pindex) that is relative
1148  * to the address space into a logical page number that is relative to the
1149  * backing object.  Use the virtual page table pointed to by (vpte).
1150  *
1151  * This implements an N-level page table.  Any level can terminate the
1152  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
1153  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
1154  */
1155 static
1156 int
1157 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
1158 		    vpte_t vpte, int fault_type, int allow_nofault)
1159 {
1160 	struct lwbuf *lwb;
1161 	struct lwbuf lwb_cache;
1162 	int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
1163 	int result = KERN_SUCCESS;
1164 	vpte_t *ptep;
1165 
1166 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1167 	for (;;) {
1168 		/*
1169 		 * We cannot proceed if the vpte is not valid, not readable
1170 		 * for a read fault, or not writable for a write fault.
1171 		 */
1172 		if ((vpte & VPTE_V) == 0) {
1173 			unlock_and_deallocate(fs);
1174 			return (KERN_FAILURE);
1175 		}
1176 		if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) {
1177 			unlock_and_deallocate(fs);
1178 			return (KERN_FAILURE);
1179 		}
1180 		if ((vpte & VPTE_PS) || vshift == 0)
1181 			break;
1182 		KKASSERT(vshift >= VPTE_PAGE_BITS);
1183 
1184 		/*
1185 		 * Get the page table page.  Nominally we only read the page
1186 		 * table, but since we are actively setting VPTE_M and VPTE_A,
1187 		 * tell vm_fault_object() that we are writing it.
1188 		 *
1189 		 * There is currently no real need to optimize this.
1190 		 */
1191 		result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
1192 					 VM_PROT_READ|VM_PROT_WRITE,
1193 					 allow_nofault);
1194 		if (result != KERN_SUCCESS)
1195 			return (result);
1196 
1197 		/*
1198 		 * Process the returned fs.m and look up the page table
1199 		 * entry in the page table page.
1200 		 */
1201 		vshift -= VPTE_PAGE_BITS;
1202 		lwb = lwbuf_alloc(fs->m, &lwb_cache);
1203 		ptep = ((vpte_t *)lwbuf_kva(lwb) +
1204 		        ((*pindex >> vshift) & VPTE_PAGE_MASK));
1205 		vpte = *ptep;
1206 
1207 		/*
1208 		 * Page table write-back.  If the vpte is valid for the
1209 		 * requested operation, do a write-back to the page table.
1210 		 *
1211 		 * XXX VPTE_M is not set properly for page directory pages.
1212 		 * It doesn't get set in the page directory if the page table
1213 		 * is modified during a read access.
1214 		 */
1215 		vm_page_activate(fs->m);
1216 		if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
1217 		    (vpte & VPTE_RW)) {
1218 			if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
1219 				atomic_set_long(ptep, VPTE_M | VPTE_A);
1220 				vm_page_dirty(fs->m);
1221 			}
1222 		}
1223 		if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V)) {
1224 			if ((vpte & VPTE_A) == 0) {
1225 				atomic_set_long(ptep, VPTE_A);
1226 				vm_page_dirty(fs->m);
1227 			}
1228 		}
1229 		lwbuf_free(lwb);
1230 		vm_page_flag_set(fs->m, PG_REFERENCED);
1231 		vm_page_wakeup(fs->m);
1232 		fs->m = NULL;
1233 		cleanup_successful_fault(fs);
1234 	}
1235 	/*
1236 	 * Combine remaining address bits with the vpte.
1237 	 */
1238 	/* JG how many bits from each? */
1239 	*pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
1240 		  (*pindex & ((1L << vshift) - 1));
1241 	return (KERN_SUCCESS);
1242 }
1243 
1244 
1245 /*
1246  * This is the core of the vm_fault code.
1247  *
1248  * Do all operations required to fault-in (fs.first_object, pindex).  Run
1249  * through the shadow chain as necessary and do required COW or virtual
1250  * copy operations.  The caller has already fully resolved the vm_map_entry
1251  * and, if appropriate, has created a copy-on-write layer.  All we need to
1252  * do is iterate the object chain.
1253  *
1254  * On failure (fs) is unlocked and deallocated and the caller may return or
1255  * retry depending on the failure code.  On success (fs) is NOT unlocked or
1256  * deallocated, fs.m will contained a resolved, busied page, and fs.object
1257  * will have an additional PIP count if it is not equal to fs.first_object.
1258  *
1259  * If locks based on fs->first_shared or fs->shared are insufficient,
1260  * clear the appropriate field(s) and return RETRY.  COWs require that
1261  * first_shared be 0, while page allocations (or frees) require that
1262  * shared be 0.  Renames require that both be 0.
1263  *
1264  * fs->first_object must be held on call.
1265  */
1266 static
1267 int
1268 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex,
1269 		vm_prot_t fault_type, int allow_nofault)
1270 {
1271 	vm_object_t next_object;
1272 	vm_pindex_t pindex;
1273 	int error;
1274 
1275 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1276 	fs->prot = fs->first_prot;
1277 	fs->object = fs->first_object;
1278 	pindex = first_pindex;
1279 
1280 	vm_object_chain_acquire(fs->first_object, fs->shared);
1281 	vm_object_pip_add(fs->first_object, 1);
1282 
1283 	/*
1284 	 * If a read fault occurs we try to make the page writable if
1285 	 * possible.  There are three cases where we cannot make the
1286 	 * page mapping writable:
1287 	 *
1288 	 * (1) The mapping is read-only or the VM object is read-only,
1289 	 *     fs->prot above will simply not have VM_PROT_WRITE set.
1290 	 *
1291 	 * (2) If the mapping is a virtual page table we need to be able
1292 	 *     to detect writes so we can set VPTE_M in the virtual page
1293 	 *     table.
1294 	 *
1295 	 * (3) If the VM page is read-only or copy-on-write, upgrading would
1296 	 *     just result in an unnecessary COW fault.
1297 	 *
1298 	 * VM_PROT_VPAGED is set if faulting via a virtual page table and
1299 	 * causes adjustments to the 'M'odify bit to also turn off write
1300 	 * access to force a re-fault.
1301 	 */
1302 	if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1303 		if ((fault_type & VM_PROT_WRITE) == 0)
1304 			fs->prot &= ~VM_PROT_WRITE;
1305 	}
1306 
1307 	if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
1308 	    pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
1309 		if ((fault_type & VM_PROT_WRITE) == 0)
1310 			fs->prot &= ~VM_PROT_WRITE;
1311 	}
1312 
1313 	/* vm_object_hold(fs->object); implied b/c object == first_object */
1314 
1315 	for (;;) {
1316 		/*
1317 		 * The entire backing chain from first_object to object
1318 		 * inclusive is chainlocked.
1319 		 *
1320 		 * If the object is dead, we stop here
1321 		 */
1322 		if (fs->object->flags & OBJ_DEAD) {
1323 			vm_object_pip_wakeup(fs->first_object);
1324 			vm_object_chain_release_all(fs->first_object,
1325 						    fs->object);
1326 			if (fs->object != fs->first_object)
1327 				vm_object_drop(fs->object);
1328 			unlock_and_deallocate(fs);
1329 			return (KERN_PROTECTION_FAILURE);
1330 		}
1331 
1332 		/*
1333 		 * See if the page is resident.  Wait/Retry if the page is
1334 		 * busy (lots of stuff may have changed so we can't continue
1335 		 * in that case).
1336 		 *
1337 		 * We can theoretically allow the soft-busy case on a read
1338 		 * fault if the page is marked valid, but since such
1339 		 * pages are typically already pmap'd, putting that
1340 		 * special case in might be more effort then it is
1341 		 * worth.  We cannot under any circumstances mess
1342 		 * around with a vm_page_t->busy page except, perhaps,
1343 		 * to pmap it.
1344 		 */
1345 		fs->m = vm_page_lookup_busy_try(fs->object, pindex,
1346 						TRUE, &error);
1347 		if (error) {
1348 			vm_object_pip_wakeup(fs->first_object);
1349 			vm_object_chain_release_all(fs->first_object,
1350 						    fs->object);
1351 			if (fs->object != fs->first_object)
1352 				vm_object_drop(fs->object);
1353 			unlock_things(fs);
1354 			vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1355 			mycpu->gd_cnt.v_intrans++;
1356 			/*vm_object_deallocate(fs->first_object);*/
1357 			/*fs->first_object = NULL;*/
1358 			fs->m = NULL;
1359 			return (KERN_TRY_AGAIN);
1360 		}
1361 		if (fs->m) {
1362 			/*
1363 			 * The page is busied for us.
1364 			 *
1365 			 * If reactivating a page from PQ_CACHE we may have
1366 			 * to rate-limit.
1367 			 */
1368 			int queue = fs->m->queue;
1369 			vm_page_unqueue_nowakeup(fs->m);
1370 
1371 			if ((queue - fs->m->pc) == PQ_CACHE &&
1372 			    vm_page_count_severe()) {
1373 				vm_page_activate(fs->m);
1374 				vm_page_wakeup(fs->m);
1375 				fs->m = NULL;
1376 				vm_object_pip_wakeup(fs->first_object);
1377 				vm_object_chain_release_all(fs->first_object,
1378 							    fs->object);
1379 				if (fs->object != fs->first_object)
1380 					vm_object_drop(fs->object);
1381 				unlock_and_deallocate(fs);
1382 				if (allow_nofault == 0 ||
1383 				    (curthread->td_flags & TDF_NOFAULT) == 0) {
1384 					vm_wait_pfault();
1385 				}
1386 				return (KERN_TRY_AGAIN);
1387 			}
1388 
1389 			/*
1390 			 * If it still isn't completely valid (readable),
1391 			 * or if a read-ahead-mark is set on the VM page,
1392 			 * jump to readrest, else we found the page and
1393 			 * can return.
1394 			 *
1395 			 * We can release the spl once we have marked the
1396 			 * page busy.
1397 			 */
1398 			if (fs->m->object != &kernel_object) {
1399 				if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1400 				    VM_PAGE_BITS_ALL) {
1401 					goto readrest;
1402 				}
1403 				if (fs->m->flags & PG_RAM) {
1404 					if (debug_cluster)
1405 						kprintf("R");
1406 					vm_page_flag_clear(fs->m, PG_RAM);
1407 					goto readrest;
1408 				}
1409 			}
1410 			break; /* break to PAGE HAS BEEN FOUND */
1411 		}
1412 
1413 		/*
1414 		 * Page is not resident, If this is the search termination
1415 		 * or the pager might contain the page, allocate a new page.
1416 		 */
1417 		if (TRYPAGER(fs) || fs->object == fs->first_object) {
1418 			/*
1419 			 * Allocating, must be exclusive.
1420 			 */
1421 			if (fs->object == fs->first_object &&
1422 			    fs->first_shared) {
1423 				fs->first_shared = 0;
1424 				vm_object_pip_wakeup(fs->first_object);
1425 				vm_object_chain_release_all(fs->first_object,
1426 							    fs->object);
1427 				if (fs->object != fs->first_object)
1428 					vm_object_drop(fs->object);
1429 				unlock_and_deallocate(fs);
1430 				return (KERN_TRY_AGAIN);
1431 			}
1432 			if (fs->object != fs->first_object &&
1433 			    fs->shared) {
1434 				fs->first_shared = 0;
1435 				fs->shared = 0;
1436 				vm_object_pip_wakeup(fs->first_object);
1437 				vm_object_chain_release_all(fs->first_object,
1438 							    fs->object);
1439 				if (fs->object != fs->first_object)
1440 					vm_object_drop(fs->object);
1441 				unlock_and_deallocate(fs);
1442 				return (KERN_TRY_AGAIN);
1443 			}
1444 
1445 			/*
1446 			 * If the page is beyond the object size we fail
1447 			 */
1448 			if (pindex >= fs->object->size) {
1449 				vm_object_pip_wakeup(fs->first_object);
1450 				vm_object_chain_release_all(fs->first_object,
1451 							    fs->object);
1452 				if (fs->object != fs->first_object)
1453 					vm_object_drop(fs->object);
1454 				unlock_and_deallocate(fs);
1455 				return (KERN_PROTECTION_FAILURE);
1456 			}
1457 
1458 			/*
1459 			 * Allocate a new page for this object/offset pair.
1460 			 *
1461 			 * It is possible for the allocation to race, so
1462 			 * handle the case.
1463 			 */
1464 			fs->m = NULL;
1465 			if (!vm_page_count_severe()) {
1466 				fs->m = vm_page_alloc(fs->object, pindex,
1467 				    ((fs->vp || fs->object->backing_object) ?
1468 					VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL :
1469 					VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1470 					VM_ALLOC_USE_GD | VM_ALLOC_ZERO));
1471 			}
1472 			if (fs->m == NULL) {
1473 				vm_object_pip_wakeup(fs->first_object);
1474 				vm_object_chain_release_all(fs->first_object,
1475 							    fs->object);
1476 				if (fs->object != fs->first_object)
1477 					vm_object_drop(fs->object);
1478 				unlock_and_deallocate(fs);
1479 				if (allow_nofault == 0 ||
1480 				    (curthread->td_flags & TDF_NOFAULT) == 0) {
1481 					vm_wait_pfault();
1482 				}
1483 				return (KERN_TRY_AGAIN);
1484 			}
1485 
1486 			/*
1487 			 * Fall through to readrest.  We have a new page which
1488 			 * will have to be paged (since m->valid will be 0).
1489 			 */
1490 		}
1491 
1492 readrest:
1493 		/*
1494 		 * We have found an invalid or partially valid page, a
1495 		 * page with a read-ahead mark which might be partially or
1496 		 * fully valid (and maybe dirty too), or we have allocated
1497 		 * a new page.
1498 		 *
1499 		 * Attempt to fault-in the page if there is a chance that the
1500 		 * pager has it, and potentially fault in additional pages
1501 		 * at the same time.
1502 		 *
1503 		 * If TRYPAGER is true then fs.m will be non-NULL and busied
1504 		 * for us.
1505 		 */
1506 		if (TRYPAGER(fs)) {
1507 			int rv;
1508 			int seqaccess;
1509 			u_char behavior = vm_map_entry_behavior(fs->entry);
1510 
1511 			if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1512 				seqaccess = 0;
1513 			else
1514 				seqaccess = -1;
1515 
1516 			/*
1517 			 * Doing I/O may synchronously insert additional
1518 			 * pages so we can't be shared at this point either.
1519 			 *
1520 			 * NOTE: We can't free fs->m here in the allocated
1521 			 *	 case (fs->object != fs->first_object) as
1522 			 *	 this would require an exclusively locked
1523 			 *	 VM object.
1524 			 */
1525 			if (fs->object == fs->first_object &&
1526 			    fs->first_shared) {
1527 				vm_page_deactivate(fs->m);
1528 				vm_page_wakeup(fs->m);
1529 				fs->m = NULL;
1530 				fs->first_shared = 0;
1531 				vm_object_pip_wakeup(fs->first_object);
1532 				vm_object_chain_release_all(fs->first_object,
1533 							    fs->object);
1534 				if (fs->object != fs->first_object)
1535 					vm_object_drop(fs->object);
1536 				unlock_and_deallocate(fs);
1537 				return (KERN_TRY_AGAIN);
1538 			}
1539 			if (fs->object != fs->first_object &&
1540 			    fs->shared) {
1541 				vm_page_deactivate(fs->m);
1542 				vm_page_wakeup(fs->m);
1543 				fs->m = NULL;
1544 				fs->first_shared = 0;
1545 				fs->shared = 0;
1546 				vm_object_pip_wakeup(fs->first_object);
1547 				vm_object_chain_release_all(fs->first_object,
1548 							    fs->object);
1549 				if (fs->object != fs->first_object)
1550 					vm_object_drop(fs->object);
1551 				unlock_and_deallocate(fs);
1552 				return (KERN_TRY_AGAIN);
1553 			}
1554 
1555 			/*
1556 			 * Avoid deadlocking against the map when doing I/O.
1557 			 * fs.object and the page is PG_BUSY'd.
1558 			 *
1559 			 * NOTE: Once unlocked, fs->entry can become stale
1560 			 *	 so this will NULL it out.
1561 			 *
1562 			 * NOTE: fs->entry is invalid until we relock the
1563 			 *	 map and verify that the timestamp has not
1564 			 *	 changed.
1565 			 */
1566 			unlock_map(fs);
1567 
1568 			/*
1569 			 * Acquire the page data.  We still hold a ref on
1570 			 * fs.object and the page has been PG_BUSY's.
1571 			 *
1572 			 * The pager may replace the page (for example, in
1573 			 * order to enter a fictitious page into the
1574 			 * object).  If it does so it is responsible for
1575 			 * cleaning up the passed page and properly setting
1576 			 * the new page PG_BUSY.
1577 			 *
1578 			 * If we got here through a PG_RAM read-ahead
1579 			 * mark the page may be partially dirty and thus
1580 			 * not freeable.  Don't bother checking to see
1581 			 * if the pager has the page because we can't free
1582 			 * it anyway.  We have to depend on the get_page
1583 			 * operation filling in any gaps whether there is
1584 			 * backing store or not.
1585 			 */
1586 			rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1587 
1588 			if (rv == VM_PAGER_OK) {
1589 				/*
1590 				 * Relookup in case pager changed page. Pager
1591 				 * is responsible for disposition of old page
1592 				 * if moved.
1593 				 *
1594 				 * XXX other code segments do relookups too.
1595 				 * It's a bad abstraction that needs to be
1596 				 * fixed/removed.
1597 				 */
1598 				fs->m = vm_page_lookup(fs->object, pindex);
1599 				if (fs->m == NULL) {
1600 					vm_object_pip_wakeup(fs->first_object);
1601 					vm_object_chain_release_all(
1602 						fs->first_object, fs->object);
1603 					if (fs->object != fs->first_object)
1604 						vm_object_drop(fs->object);
1605 					unlock_and_deallocate(fs);
1606 					return (KERN_TRY_AGAIN);
1607 				}
1608 				++fs->hardfault;
1609 				break; /* break to PAGE HAS BEEN FOUND */
1610 			}
1611 
1612 			/*
1613 			 * Remove the bogus page (which does not exist at this
1614 			 * object/offset); before doing so, we must get back
1615 			 * our object lock to preserve our invariant.
1616 			 *
1617 			 * Also wake up any other process that may want to bring
1618 			 * in this page.
1619 			 *
1620 			 * If this is the top-level object, we must leave the
1621 			 * busy page to prevent another process from rushing
1622 			 * past us, and inserting the page in that object at
1623 			 * the same time that we are.
1624 			 */
1625 			if (rv == VM_PAGER_ERROR) {
1626 				if (curproc) {
1627 					kprintf("vm_fault: pager read error, "
1628 						"pid %d (%s)\n",
1629 						curproc->p_pid,
1630 						curproc->p_comm);
1631 				} else {
1632 					kprintf("vm_fault: pager read error, "
1633 						"thread %p (%s)\n",
1634 						curthread,
1635 						curproc->p_comm);
1636 				}
1637 			}
1638 
1639 			/*
1640 			 * Data outside the range of the pager or an I/O error
1641 			 *
1642 			 * The page may have been wired during the pagein,
1643 			 * e.g. by the buffer cache, and cannot simply be
1644 			 * freed.  Call vnode_pager_freepage() to deal with it.
1645 			 *
1646 			 * Also note that we cannot free the page if we are
1647 			 * holding the related object shared. XXX not sure
1648 			 * what to do in that case.
1649 			 */
1650 			if (fs->object != fs->first_object) {
1651 				vnode_pager_freepage(fs->m);
1652 				fs->m = NULL;
1653 				/*
1654 				 * XXX - we cannot just fall out at this
1655 				 * point, m has been freed and is invalid!
1656 				 */
1657 			}
1658 			/*
1659 			 * XXX - the check for kernel_map is a kludge to work
1660 			 * around having the machine panic on a kernel space
1661 			 * fault w/ I/O error.
1662 			 */
1663 			if (((fs->map != &kernel_map) &&
1664 			    (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1665 				if (fs->m) {
1666 					if (fs->first_shared) {
1667 						vm_page_deactivate(fs->m);
1668 						vm_page_wakeup(fs->m);
1669 					} else {
1670 						vnode_pager_freepage(fs->m);
1671 					}
1672 					fs->m = NULL;
1673 				}
1674 				vm_object_pip_wakeup(fs->first_object);
1675 				vm_object_chain_release_all(fs->first_object,
1676 							    fs->object);
1677 				if (fs->object != fs->first_object)
1678 					vm_object_drop(fs->object);
1679 				unlock_and_deallocate(fs);
1680 				if (rv == VM_PAGER_ERROR)
1681 					return (KERN_FAILURE);
1682 				else
1683 					return (KERN_PROTECTION_FAILURE);
1684 				/* NOT REACHED */
1685 			}
1686 		}
1687 
1688 		/*
1689 		 * We get here if the object has a default pager (or unwiring)
1690 		 * or the pager doesn't have the page.
1691 		 *
1692 		 * fs->first_m will be used for the COW unless we find a
1693 		 * deeper page to be mapped read-only, in which case the
1694 		 * unlock*(fs) will free first_m.
1695 		 */
1696 		if (fs->object == fs->first_object)
1697 			fs->first_m = fs->m;
1698 
1699 		/*
1700 		 * Move on to the next object.  The chain lock should prevent
1701 		 * the backing_object from getting ripped out from under us.
1702 		 *
1703 		 * The object lock for the next object is governed by
1704 		 * fs->shared.
1705 		 */
1706 		if ((next_object = fs->object->backing_object) != NULL) {
1707 			if (fs->shared)
1708 				vm_object_hold_shared(next_object);
1709 			else
1710 				vm_object_hold(next_object);
1711 			vm_object_chain_acquire(next_object, fs->shared);
1712 			KKASSERT(next_object == fs->object->backing_object);
1713 			pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1714 		}
1715 
1716 		if (next_object == NULL) {
1717 			/*
1718 			 * If there's no object left, fill the page in the top
1719 			 * object with zeros.
1720 			 */
1721 			if (fs->object != fs->first_object) {
1722 #if 0
1723 				if (fs->first_object->backing_object !=
1724 				    fs->object) {
1725 					vm_object_hold(fs->first_object->backing_object);
1726 				}
1727 #endif
1728 				vm_object_chain_release_all(
1729 					fs->first_object->backing_object,
1730 					fs->object);
1731 #if 0
1732 				if (fs->first_object->backing_object !=
1733 				    fs->object) {
1734 					vm_object_drop(fs->first_object->backing_object);
1735 				}
1736 #endif
1737 				vm_object_pip_wakeup(fs->object);
1738 				vm_object_drop(fs->object);
1739 				fs->object = fs->first_object;
1740 				pindex = first_pindex;
1741 				fs->m = fs->first_m;
1742 			}
1743 			fs->first_m = NULL;
1744 
1745 			/*
1746 			 * Zero the page and mark it valid.
1747 			 */
1748 			vm_page_zero_fill(fs->m);
1749 			mycpu->gd_cnt.v_zfod++;
1750 			fs->m->valid = VM_PAGE_BITS_ALL;
1751 			break;	/* break to PAGE HAS BEEN FOUND */
1752 		}
1753 		if (fs->object != fs->first_object) {
1754 			vm_object_pip_wakeup(fs->object);
1755 			vm_object_lock_swap();
1756 			vm_object_drop(fs->object);
1757 		}
1758 		KASSERT(fs->object != next_object,
1759 			("object loop %p", next_object));
1760 		fs->object = next_object;
1761 		vm_object_pip_add(fs->object, 1);
1762 	}
1763 
1764 	/*
1765 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1766 	 * is held.]
1767 	 *
1768 	 * object still held.
1769 	 *
1770 	 * local shared variable may be different from fs->shared.
1771 	 *
1772 	 * If the page is being written, but isn't already owned by the
1773 	 * top-level object, we have to copy it into a new page owned by the
1774 	 * top-level object.
1775 	 */
1776 	KASSERT((fs->m->flags & PG_BUSY) != 0,
1777 		("vm_fault: not busy after main loop"));
1778 
1779 	if (fs->object != fs->first_object) {
1780 		/*
1781 		 * We only really need to copy if we want to write it.
1782 		 */
1783 		if (fault_type & VM_PROT_WRITE) {
1784 			/*
1785 			 * This allows pages to be virtually copied from a
1786 			 * backing_object into the first_object, where the
1787 			 * backing object has no other refs to it, and cannot
1788 			 * gain any more refs.  Instead of a bcopy, we just
1789 			 * move the page from the backing object to the
1790 			 * first object.  Note that we must mark the page
1791 			 * dirty in the first object so that it will go out
1792 			 * to swap when needed.
1793 			 */
1794 			if (
1795 				/*
1796 				 * Must be holding exclusive locks
1797 				 */
1798 				fs->first_shared == 0 &&
1799 				fs->shared == 0 &&
1800 				/*
1801 				 * Map, if present, has not changed
1802 				 */
1803 				(fs->map == NULL ||
1804 				fs->map_generation == fs->map->timestamp) &&
1805 				/*
1806 				 * Only one shadow object
1807 				 */
1808 				(fs->object->shadow_count == 1) &&
1809 				/*
1810 				 * No COW refs, except us
1811 				 */
1812 				(fs->object->ref_count == 1) &&
1813 				/*
1814 				 * No one else can look this object up
1815 				 */
1816 				(fs->object->handle == NULL) &&
1817 				/*
1818 				 * No other ways to look the object up
1819 				 */
1820 				((fs->object->type == OBJT_DEFAULT) ||
1821 				 (fs->object->type == OBJT_SWAP)) &&
1822 				/*
1823 				 * We don't chase down the shadow chain
1824 				 */
1825 				(fs->object == fs->first_object->backing_object) &&
1826 
1827 				/*
1828 				 * grab the lock if we need to
1829 				 */
1830 				(fs->lookup_still_valid ||
1831 				 fs->map == NULL ||
1832 				 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1833 			    ) {
1834 				/*
1835 				 * (first_m) and (m) are both busied.  We have
1836 				 * move (m) into (first_m)'s object/pindex
1837 				 * in an atomic fashion, then free (first_m).
1838 				 *
1839 				 * first_object is held so second remove
1840 				 * followed by the rename should wind
1841 				 * up being atomic.  vm_page_free() might
1842 				 * block so we don't do it until after the
1843 				 * rename.
1844 				 */
1845 				fs->lookup_still_valid = 1;
1846 				vm_page_protect(fs->first_m, VM_PROT_NONE);
1847 				vm_page_remove(fs->first_m);
1848 				vm_page_rename(fs->m, fs->first_object,
1849 					       first_pindex);
1850 				vm_page_free(fs->first_m);
1851 				fs->first_m = fs->m;
1852 				fs->m = NULL;
1853 				mycpu->gd_cnt.v_cow_optim++;
1854 			} else {
1855 				/*
1856 				 * Oh, well, lets copy it.
1857 				 *
1858 				 * Why are we unmapping the original page
1859 				 * here?  Well, in short, not all accessors
1860 				 * of user memory go through the pmap.  The
1861 				 * procfs code doesn't have access user memory
1862 				 * via a local pmap, so vm_fault_page*()
1863 				 * can't call pmap_enter().  And the umtx*()
1864 				 * code may modify the COW'd page via a DMAP
1865 				 * or kernel mapping and not via the pmap,
1866 				 * leaving the original page still mapped
1867 				 * read-only into the pmap.
1868 				 *
1869 				 * So we have to remove the page from at
1870 				 * least the current pmap if it is in it.
1871 				 * Just remove it from all pmaps.
1872 				 */
1873 				KKASSERT(fs->first_shared == 0);
1874 				vm_page_copy(fs->m, fs->first_m);
1875 				vm_page_protect(fs->m, VM_PROT_NONE);
1876 				vm_page_event(fs->m, VMEVENT_COW);
1877 			}
1878 
1879 			/*
1880 			 * We no longer need the old page or object.
1881 			 */
1882 			if (fs->m)
1883 				release_page(fs);
1884 
1885 			/*
1886 			 * We intend to revert to first_object, undo the
1887 			 * chain lock through to that.
1888 			 */
1889 #if 0
1890 			if (fs->first_object->backing_object != fs->object)
1891 				vm_object_hold(fs->first_object->backing_object);
1892 #endif
1893 			vm_object_chain_release_all(
1894 					fs->first_object->backing_object,
1895 					fs->object);
1896 #if 0
1897 			if (fs->first_object->backing_object != fs->object)
1898 				vm_object_drop(fs->first_object->backing_object);
1899 #endif
1900 
1901 			/*
1902 			 * fs->object != fs->first_object due to above
1903 			 * conditional
1904 			 */
1905 			vm_object_pip_wakeup(fs->object);
1906 			vm_object_drop(fs->object);
1907 
1908 			/*
1909 			 * Only use the new page below...
1910 			 */
1911 			mycpu->gd_cnt.v_cow_faults++;
1912 			fs->m = fs->first_m;
1913 			fs->object = fs->first_object;
1914 			pindex = first_pindex;
1915 		} else {
1916 			/*
1917 			 * If it wasn't a write fault avoid having to copy
1918 			 * the page by mapping it read-only.
1919 			 */
1920 			fs->prot &= ~VM_PROT_WRITE;
1921 		}
1922 	}
1923 
1924 	/*
1925 	 * Relock the map if necessary, then check the generation count.
1926 	 * relock_map() will update fs->timestamp to account for the
1927 	 * relocking if necessary.
1928 	 *
1929 	 * If the count has changed after relocking then all sorts of
1930 	 * crap may have happened and we have to retry.
1931 	 *
1932 	 * NOTE: The relock_map() can fail due to a deadlock against
1933 	 *	 the vm_page we are holding BUSY.
1934 	 */
1935 	if (fs->lookup_still_valid == FALSE && fs->map) {
1936 		if (relock_map(fs) ||
1937 		    fs->map->timestamp != fs->map_generation) {
1938 			release_page(fs);
1939 			vm_object_pip_wakeup(fs->first_object);
1940 			vm_object_chain_release_all(fs->first_object,
1941 						    fs->object);
1942 			if (fs->object != fs->first_object)
1943 				vm_object_drop(fs->object);
1944 			unlock_and_deallocate(fs);
1945 			return (KERN_TRY_AGAIN);
1946 		}
1947 	}
1948 
1949 	/*
1950 	 * If the fault is a write, we know that this page is being
1951 	 * written NOW so dirty it explicitly to save on pmap_is_modified()
1952 	 * calls later.
1953 	 *
1954 	 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1955 	 * if the page is already dirty to prevent data written with
1956 	 * the expectation of being synced from not being synced.
1957 	 * Likewise if this entry does not request NOSYNC then make
1958 	 * sure the page isn't marked NOSYNC.  Applications sharing
1959 	 * data should use the same flags to avoid ping ponging.
1960 	 *
1961 	 * Also tell the backing pager, if any, that it should remove
1962 	 * any swap backing since the page is now dirty.
1963 	 */
1964 	vm_page_activate(fs->m);
1965 	if (fs->prot & VM_PROT_WRITE) {
1966 		vm_object_set_writeable_dirty(fs->m->object);
1967 		vm_set_nosync(fs->m, fs->entry);
1968 		if (fs->fault_flags & VM_FAULT_DIRTY) {
1969 			vm_page_dirty(fs->m);
1970 			swap_pager_unswapped(fs->m);
1971 		}
1972 	}
1973 
1974 	vm_object_pip_wakeup(fs->first_object);
1975 	vm_object_chain_release_all(fs->first_object, fs->object);
1976 	if (fs->object != fs->first_object)
1977 		vm_object_drop(fs->object);
1978 
1979 	/*
1980 	 * Page had better still be busy.  We are still locked up and
1981 	 * fs->object will have another PIP reference if it is not equal
1982 	 * to fs->first_object.
1983 	 */
1984 	KASSERT(fs->m->flags & PG_BUSY,
1985 		("vm_fault: page %p not busy!", fs->m));
1986 
1987 	/*
1988 	 * Sanity check: page must be completely valid or it is not fit to
1989 	 * map into user space.  vm_pager_get_pages() ensures this.
1990 	 */
1991 	if (fs->m->valid != VM_PAGE_BITS_ALL) {
1992 		vm_page_zero_invalid(fs->m, TRUE);
1993 		kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1994 	}
1995 
1996 	return (KERN_SUCCESS);
1997 }
1998 
1999 /*
2000  * Hold each of the physical pages that are mapped by the specified range of
2001  * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
2002  * and allow the specified types of access, "prot".  If all of the implied
2003  * pages are successfully held, then the number of held pages is returned
2004  * together with pointers to those pages in the array "ma".  However, if any
2005  * of the pages cannot be held, -1 is returned.
2006  */
2007 int
2008 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
2009     vm_prot_t prot, vm_page_t *ma, int max_count)
2010 {
2011 	vm_offset_t start, end;
2012 	int i, npages, error;
2013 
2014 	start = trunc_page(addr);
2015 	end = round_page(addr + len);
2016 
2017 	npages = howmany(end - start, PAGE_SIZE);
2018 
2019 	if (npages > max_count)
2020 		return -1;
2021 
2022 	for (i = 0; i < npages; i++) {
2023 		// XXX error handling
2024 		ma[i] = vm_fault_page_quick(start + (i * PAGE_SIZE),
2025 			prot,
2026 			&error);
2027 	}
2028 
2029 	return npages;
2030 }
2031 
2032 /*
2033  * Wire down a range of virtual addresses in a map.  The entry in question
2034  * should be marked in-transition and the map must be locked.  We must
2035  * release the map temporarily while faulting-in the page to avoid a
2036  * deadlock.  Note that the entry may be clipped while we are blocked but
2037  * will never be freed.
2038  *
2039  * No requirements.
2040  */
2041 int
2042 vm_fault_wire(vm_map_t map, vm_map_entry_t entry,
2043 	      boolean_t user_wire, int kmflags)
2044 {
2045 	boolean_t fictitious;
2046 	vm_offset_t start;
2047 	vm_offset_t end;
2048 	vm_offset_t va;
2049 	vm_paddr_t pa;
2050 	vm_page_t m;
2051 	pmap_t pmap;
2052 	int rv;
2053 	int wire_prot;
2054 	int fault_flags;
2055 
2056 	lwkt_gettoken(&map->token);
2057 
2058 	if (user_wire) {
2059 		wire_prot = VM_PROT_READ;
2060 		fault_flags = VM_FAULT_USER_WIRE;
2061 	} else {
2062 		wire_prot = VM_PROT_READ | VM_PROT_WRITE;
2063 		fault_flags = VM_FAULT_CHANGE_WIRING;
2064 	}
2065 	if (kmflags & KM_NOTLBSYNC)
2066 		wire_prot |= VM_PROT_NOSYNC;
2067 
2068 	pmap = vm_map_pmap(map);
2069 	start = entry->start;
2070 	end = entry->end;
2071 	switch(entry->maptype) {
2072 	case VM_MAPTYPE_NORMAL:
2073 	case VM_MAPTYPE_VPAGETABLE:
2074 		fictitious = entry->object.vm_object &&
2075 			    ((entry->object.vm_object->type == OBJT_DEVICE) ||
2076 			     (entry->object.vm_object->type == OBJT_MGTDEVICE));
2077 		break;
2078 	case VM_MAPTYPE_UKSMAP:
2079 		fictitious = TRUE;
2080 		break;
2081 	default:
2082 		fictitious = FALSE;
2083 		break;
2084 	}
2085 
2086 	if (entry->eflags & MAP_ENTRY_KSTACK)
2087 		start += PAGE_SIZE;
2088 	map->timestamp++;
2089 	vm_map_unlock(map);
2090 
2091 	/*
2092 	 * We simulate a fault to get the page and enter it in the physical
2093 	 * map.
2094 	 */
2095 	for (va = start; va < end; va += PAGE_SIZE) {
2096 		rv = vm_fault(map, va, wire_prot, fault_flags);
2097 		if (rv) {
2098 			while (va > start) {
2099 				va -= PAGE_SIZE;
2100 				if ((pa = pmap_extract(pmap, va)) == 0)
2101 					continue;
2102 				pmap_change_wiring(pmap, va, FALSE, entry);
2103 				if (!fictitious) {
2104 					m = PHYS_TO_VM_PAGE(pa);
2105 					vm_page_busy_wait(m, FALSE, "vmwrpg");
2106 					vm_page_unwire(m, 1);
2107 					vm_page_wakeup(m);
2108 				}
2109 			}
2110 			goto done;
2111 		}
2112 	}
2113 	rv = KERN_SUCCESS;
2114 done:
2115 	vm_map_lock(map);
2116 	lwkt_reltoken(&map->token);
2117 	return (rv);
2118 }
2119 
2120 /*
2121  * Unwire a range of virtual addresses in a map.  The map should be
2122  * locked.
2123  */
2124 void
2125 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
2126 {
2127 	boolean_t fictitious;
2128 	vm_offset_t start;
2129 	vm_offset_t end;
2130 	vm_offset_t va;
2131 	vm_paddr_t pa;
2132 	vm_page_t m;
2133 	pmap_t pmap;
2134 
2135 	lwkt_gettoken(&map->token);
2136 
2137 	pmap = vm_map_pmap(map);
2138 	start = entry->start;
2139 	end = entry->end;
2140 	fictitious = entry->object.vm_object &&
2141 			((entry->object.vm_object->type == OBJT_DEVICE) ||
2142 			 (entry->object.vm_object->type == OBJT_MGTDEVICE));
2143 	if (entry->eflags & MAP_ENTRY_KSTACK)
2144 		start += PAGE_SIZE;
2145 
2146 	/*
2147 	 * Since the pages are wired down, we must be able to get their
2148 	 * mappings from the physical map system.
2149 	 */
2150 	for (va = start; va < end; va += PAGE_SIZE) {
2151 		pa = pmap_extract(pmap, va);
2152 		if (pa != 0) {
2153 			pmap_change_wiring(pmap, va, FALSE, entry);
2154 			if (!fictitious) {
2155 				m = PHYS_TO_VM_PAGE(pa);
2156 				vm_page_busy_wait(m, FALSE, "vmwupg");
2157 				vm_page_unwire(m, 1);
2158 				vm_page_wakeup(m);
2159 			}
2160 		}
2161 	}
2162 	lwkt_reltoken(&map->token);
2163 }
2164 
2165 /*
2166  * Copy all of the pages from a wired-down map entry to another.
2167  *
2168  * The source and destination maps must be locked for write.
2169  * The source and destination maps token must be held
2170  * The source map entry must be wired down (or be a sharing map
2171  * entry corresponding to a main map entry that is wired down).
2172  *
2173  * No other requirements.
2174  *
2175  * XXX do segment optimization
2176  */
2177 void
2178 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
2179 		    vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
2180 {
2181 	vm_object_t dst_object;
2182 	vm_object_t src_object;
2183 	vm_ooffset_t dst_offset;
2184 	vm_ooffset_t src_offset;
2185 	vm_prot_t prot;
2186 	vm_offset_t vaddr;
2187 	vm_page_t dst_m;
2188 	vm_page_t src_m;
2189 
2190 	src_object = src_entry->object.vm_object;
2191 	src_offset = src_entry->offset;
2192 
2193 	/*
2194 	 * Create the top-level object for the destination entry. (Doesn't
2195 	 * actually shadow anything - we copy the pages directly.)
2196 	 */
2197 	vm_map_entry_allocate_object(dst_entry);
2198 	dst_object = dst_entry->object.vm_object;
2199 
2200 	prot = dst_entry->max_protection;
2201 
2202 	/*
2203 	 * Loop through all of the pages in the entry's range, copying each
2204 	 * one from the source object (it should be there) to the destination
2205 	 * object.
2206 	 */
2207 	vm_object_hold(src_object);
2208 	vm_object_hold(dst_object);
2209 	for (vaddr = dst_entry->start, dst_offset = 0;
2210 	    vaddr < dst_entry->end;
2211 	    vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
2212 
2213 		/*
2214 		 * Allocate a page in the destination object
2215 		 */
2216 		do {
2217 			dst_m = vm_page_alloc(dst_object,
2218 					      OFF_TO_IDX(dst_offset),
2219 					      VM_ALLOC_NORMAL);
2220 			if (dst_m == NULL) {
2221 				vm_wait(0);
2222 			}
2223 		} while (dst_m == NULL);
2224 
2225 		/*
2226 		 * Find the page in the source object, and copy it in.
2227 		 * (Because the source is wired down, the page will be in
2228 		 * memory.)
2229 		 */
2230 		src_m = vm_page_lookup(src_object,
2231 				       OFF_TO_IDX(dst_offset + src_offset));
2232 		if (src_m == NULL)
2233 			panic("vm_fault_copy_wired: page missing");
2234 
2235 		vm_page_copy(src_m, dst_m);
2236 		vm_page_event(src_m, VMEVENT_COW);
2237 
2238 		/*
2239 		 * Enter it in the pmap...
2240 		 */
2241 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry);
2242 
2243 		/*
2244 		 * Mark it no longer busy, and put it on the active list.
2245 		 */
2246 		vm_page_activate(dst_m);
2247 		vm_page_wakeup(dst_m);
2248 	}
2249 	vm_object_drop(dst_object);
2250 	vm_object_drop(src_object);
2251 }
2252 
2253 #if 0
2254 
2255 /*
2256  * This routine checks around the requested page for other pages that
2257  * might be able to be faulted in.  This routine brackets the viable
2258  * pages for the pages to be paged in.
2259  *
2260  * Inputs:
2261  *	m, rbehind, rahead
2262  *
2263  * Outputs:
2264  *  marray (array of vm_page_t), reqpage (index of requested page)
2265  *
2266  * Return value:
2267  *  number of pages in marray
2268  */
2269 static int
2270 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
2271 			  vm_page_t *marray, int *reqpage)
2272 {
2273 	int i,j;
2274 	vm_object_t object;
2275 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
2276 	vm_page_t rtm;
2277 	int cbehind, cahead;
2278 
2279 	object = m->object;
2280 	pindex = m->pindex;
2281 
2282 	/*
2283 	 * we don't fault-ahead for device pager
2284 	 */
2285 	if ((object->type == OBJT_DEVICE) ||
2286 	    (object->type == OBJT_MGTDEVICE)) {
2287 		*reqpage = 0;
2288 		marray[0] = m;
2289 		return 1;
2290 	}
2291 
2292 	/*
2293 	 * if the requested page is not available, then give up now
2294 	 */
2295 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
2296 		*reqpage = 0;	/* not used by caller, fix compiler warn */
2297 		return 0;
2298 	}
2299 
2300 	if ((cbehind == 0) && (cahead == 0)) {
2301 		*reqpage = 0;
2302 		marray[0] = m;
2303 		return 1;
2304 	}
2305 
2306 	if (rahead > cahead) {
2307 		rahead = cahead;
2308 	}
2309 
2310 	if (rbehind > cbehind) {
2311 		rbehind = cbehind;
2312 	}
2313 
2314 	/*
2315 	 * Do not do any readahead if we have insufficient free memory.
2316 	 *
2317 	 * XXX code was broken disabled before and has instability
2318 	 * with this conditonal fixed, so shortcut for now.
2319 	 */
2320 	if (burst_fault == 0 || vm_page_count_severe()) {
2321 		marray[0] = m;
2322 		*reqpage = 0;
2323 		return 1;
2324 	}
2325 
2326 	/*
2327 	 * scan backward for the read behind pages -- in memory
2328 	 *
2329 	 * Assume that if the page is not found an interrupt will not
2330 	 * create it.  Theoretically interrupts can only remove (busy)
2331 	 * pages, not create new associations.
2332 	 */
2333 	if (pindex > 0) {
2334 		if (rbehind > pindex) {
2335 			rbehind = pindex;
2336 			startpindex = 0;
2337 		} else {
2338 			startpindex = pindex - rbehind;
2339 		}
2340 
2341 		vm_object_hold(object);
2342 		for (tpindex = pindex; tpindex > startpindex; --tpindex) {
2343 			if (vm_page_lookup(object, tpindex - 1))
2344 				break;
2345 		}
2346 
2347 		i = 0;
2348 		while (tpindex < pindex) {
2349 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2350 							     VM_ALLOC_NULL_OK);
2351 			if (rtm == NULL) {
2352 				for (j = 0; j < i; j++) {
2353 					vm_page_free(marray[j]);
2354 				}
2355 				vm_object_drop(object);
2356 				marray[0] = m;
2357 				*reqpage = 0;
2358 				return 1;
2359 			}
2360 			marray[i] = rtm;
2361 			++i;
2362 			++tpindex;
2363 		}
2364 		vm_object_drop(object);
2365 	} else {
2366 		i = 0;
2367 	}
2368 
2369 	/*
2370 	 * Assign requested page
2371 	 */
2372 	marray[i] = m;
2373 	*reqpage = i;
2374 	++i;
2375 
2376 	/*
2377 	 * Scan forwards for read-ahead pages
2378 	 */
2379 	tpindex = pindex + 1;
2380 	endpindex = tpindex + rahead;
2381 	if (endpindex > object->size)
2382 		endpindex = object->size;
2383 
2384 	vm_object_hold(object);
2385 	while (tpindex < endpindex) {
2386 		if (vm_page_lookup(object, tpindex))
2387 			break;
2388 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2389 						     VM_ALLOC_NULL_OK);
2390 		if (rtm == NULL)
2391 			break;
2392 		marray[i] = rtm;
2393 		++i;
2394 		++tpindex;
2395 	}
2396 	vm_object_drop(object);
2397 
2398 	return (i);
2399 }
2400 
2401 #endif
2402 
2403 /*
2404  * vm_prefault() provides a quick way of clustering pagefaults into a
2405  * processes address space.  It is a "cousin" of pmap_object_init_pt,
2406  * except it runs at page fault time instead of mmap time.
2407  *
2408  * vm.fast_fault	Enables pre-faulting zero-fill pages
2409  *
2410  * vm.prefault_pages	Number of pages (1/2 negative, 1/2 positive) to
2411  *			prefault.  Scan stops in either direction when
2412  *			a page is found to already exist.
2413  *
2414  * This code used to be per-platform pmap_prefault().  It is now
2415  * machine-independent and enhanced to also pre-fault zero-fill pages
2416  * (see vm.fast_fault) as well as make them writable, which greatly
2417  * reduces the number of page faults programs incur.
2418  *
2419  * Application performance when pre-faulting zero-fill pages is heavily
2420  * dependent on the application.  Very tiny applications like /bin/echo
2421  * lose a little performance while applications of any appreciable size
2422  * gain performance.  Prefaulting multiple pages also reduces SMP
2423  * congestion and can improve SMP performance significantly.
2424  *
2425  * NOTE!  prot may allow writing but this only applies to the top level
2426  *	  object.  If we wind up mapping a page extracted from a backing
2427  *	  object we have to make sure it is read-only.
2428  *
2429  * NOTE!  The caller has already handled any COW operations on the
2430  *	  vm_map_entry via the normal fault code.  Do NOT call this
2431  *	  shortcut unless the normal fault code has run on this entry.
2432  *
2433  * The related map must be locked.
2434  * No other requirements.
2435  */
2436 static int vm_prefault_pages = 8;
2437 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0,
2438 	   "Maximum number of pages to pre-fault");
2439 static int vm_fast_fault = 1;
2440 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
2441 	   "Burst fault zero-fill regions");
2442 
2443 /*
2444  * Set PG_NOSYNC if the map entry indicates so, but only if the page
2445  * is not already dirty by other means.  This will prevent passive
2446  * filesystem syncing as well as 'sync' from writing out the page.
2447  */
2448 static void
2449 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
2450 {
2451 	if (entry->eflags & MAP_ENTRY_NOSYNC) {
2452 		if (m->dirty == 0)
2453 			vm_page_flag_set(m, PG_NOSYNC);
2454 	} else {
2455 		vm_page_flag_clear(m, PG_NOSYNC);
2456 	}
2457 }
2458 
2459 static void
2460 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot,
2461 	    int fault_flags)
2462 {
2463 	struct lwp *lp;
2464 	vm_page_t m;
2465 	vm_offset_t addr;
2466 	vm_pindex_t index;
2467 	vm_pindex_t pindex;
2468 	vm_object_t object;
2469 	int pprot;
2470 	int i;
2471 	int noneg;
2472 	int nopos;
2473 	int maxpages;
2474 
2475 	/*
2476 	 * Get stable max count value, disabled if set to 0
2477 	 */
2478 	maxpages = vm_prefault_pages;
2479 	cpu_ccfence();
2480 	if (maxpages <= 0)
2481 		return;
2482 
2483 	/*
2484 	 * We do not currently prefault mappings that use virtual page
2485 	 * tables.  We do not prefault foreign pmaps.
2486 	 */
2487 	if (entry->maptype != VM_MAPTYPE_NORMAL)
2488 		return;
2489 	lp = curthread->td_lwp;
2490 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2491 		return;
2492 
2493 	/*
2494 	 * Limit pre-fault count to 1024 pages.
2495 	 */
2496 	if (maxpages > 1024)
2497 		maxpages = 1024;
2498 
2499 	object = entry->object.vm_object;
2500 	KKASSERT(object != NULL);
2501 	KKASSERT(object == entry->object.vm_object);
2502 	vm_object_hold(object);
2503 	vm_object_chain_acquire(object, 0);
2504 
2505 	noneg = 0;
2506 	nopos = 0;
2507 	for (i = 0; i < maxpages; ++i) {
2508 		vm_object_t lobject;
2509 		vm_object_t nobject;
2510 		int allocated = 0;
2511 		int error;
2512 
2513 		/*
2514 		 * This can eat a lot of time on a heavily contended
2515 		 * machine so yield on the tick if needed.
2516 		 */
2517 		if ((i & 7) == 7)
2518 			lwkt_yield();
2519 
2520 		/*
2521 		 * Calculate the page to pre-fault, stopping the scan in
2522 		 * each direction separately if the limit is reached.
2523 		 */
2524 		if (i & 1) {
2525 			if (noneg)
2526 				continue;
2527 			addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2528 		} else {
2529 			if (nopos)
2530 				continue;
2531 			addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2532 		}
2533 		if (addr < entry->start) {
2534 			noneg = 1;
2535 			if (noneg && nopos)
2536 				break;
2537 			continue;
2538 		}
2539 		if (addr >= entry->end) {
2540 			nopos = 1;
2541 			if (noneg && nopos)
2542 				break;
2543 			continue;
2544 		}
2545 
2546 		/*
2547 		 * Skip pages already mapped, and stop scanning in that
2548 		 * direction.  When the scan terminates in both directions
2549 		 * we are done.
2550 		 */
2551 		if (pmap_prefault_ok(pmap, addr) == 0) {
2552 			if (i & 1)
2553 				noneg = 1;
2554 			else
2555 				nopos = 1;
2556 			if (noneg && nopos)
2557 				break;
2558 			continue;
2559 		}
2560 
2561 		/*
2562 		 * Follow the VM object chain to obtain the page to be mapped
2563 		 * into the pmap.
2564 		 *
2565 		 * If we reach the terminal object without finding a page
2566 		 * and we determine it would be advantageous, then allocate
2567 		 * a zero-fill page for the base object.  The base object
2568 		 * is guaranteed to be OBJT_DEFAULT for this case.
2569 		 *
2570 		 * In order to not have to check the pager via *haspage*()
2571 		 * we stop if any non-default object is encountered.  e.g.
2572 		 * a vnode or swap object would stop the loop.
2573 		 */
2574 		index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2575 		lobject = object;
2576 		pindex = index;
2577 		pprot = prot;
2578 
2579 		KKASSERT(lobject == entry->object.vm_object);
2580 		/*vm_object_hold(lobject); implied */
2581 
2582 		while ((m = vm_page_lookup_busy_try(lobject, pindex,
2583 						    TRUE, &error)) == NULL) {
2584 			if (lobject->type != OBJT_DEFAULT)
2585 				break;
2586 			if (lobject->backing_object == NULL) {
2587 				if (vm_fast_fault == 0)
2588 					break;
2589 				if ((prot & VM_PROT_WRITE) == 0 ||
2590 				    vm_page_count_min(0)) {
2591 					break;
2592 				}
2593 
2594 				/*
2595 				 * NOTE: Allocated from base object
2596 				 */
2597 				m = vm_page_alloc(object, index,
2598 						  VM_ALLOC_NORMAL |
2599 						  VM_ALLOC_ZERO |
2600 						  VM_ALLOC_USE_GD |
2601 						  VM_ALLOC_NULL_OK);
2602 				if (m == NULL)
2603 					break;
2604 				allocated = 1;
2605 				pprot = prot;
2606 				/* lobject = object .. not needed */
2607 				break;
2608 			}
2609 			if (lobject->backing_object_offset & PAGE_MASK)
2610 				break;
2611 			nobject = lobject->backing_object;
2612 			vm_object_hold(nobject);
2613 			KKASSERT(nobject == lobject->backing_object);
2614 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2615 			if (lobject != object) {
2616 				vm_object_lock_swap();
2617 				vm_object_drop(lobject);
2618 			}
2619 			lobject = nobject;
2620 			pprot &= ~VM_PROT_WRITE;
2621 			vm_object_chain_acquire(lobject, 0);
2622 		}
2623 
2624 		/*
2625 		 * NOTE: A non-NULL (m) will be associated with lobject if
2626 		 *	 it was found there, otherwise it is probably a
2627 		 *	 zero-fill page associated with the base object.
2628 		 *
2629 		 * Give-up if no page is available.
2630 		 */
2631 		if (m == NULL) {
2632 			if (lobject != object) {
2633 #if 0
2634 				if (object->backing_object != lobject)
2635 					vm_object_hold(object->backing_object);
2636 #endif
2637 				vm_object_chain_release_all(
2638 					object->backing_object, lobject);
2639 #if 0
2640 				if (object->backing_object != lobject)
2641 					vm_object_drop(object->backing_object);
2642 #endif
2643 				vm_object_drop(lobject);
2644 			}
2645 			break;
2646 		}
2647 
2648 		/*
2649 		 * The object must be marked dirty if we are mapping a
2650 		 * writable page.  m->object is either lobject or object,
2651 		 * both of which are still held.  Do this before we
2652 		 * potentially drop the object.
2653 		 */
2654 		if (pprot & VM_PROT_WRITE)
2655 			vm_object_set_writeable_dirty(m->object);
2656 
2657 		/*
2658 		 * Do not conditionalize on PG_RAM.  If pages are present in
2659 		 * the VM system we assume optimal caching.  If caching is
2660 		 * not optimal the I/O gravy train will be restarted when we
2661 		 * hit an unavailable page.  We do not want to try to restart
2662 		 * the gravy train now because we really don't know how much
2663 		 * of the object has been cached.  The cost for restarting
2664 		 * the gravy train should be low (since accesses will likely
2665 		 * be I/O bound anyway).
2666 		 */
2667 		if (lobject != object) {
2668 #if 0
2669 			if (object->backing_object != lobject)
2670 				vm_object_hold(object->backing_object);
2671 #endif
2672 			vm_object_chain_release_all(object->backing_object,
2673 						    lobject);
2674 #if 0
2675 			if (object->backing_object != lobject)
2676 				vm_object_drop(object->backing_object);
2677 #endif
2678 			vm_object_drop(lobject);
2679 		}
2680 
2681 		/*
2682 		 * Enter the page into the pmap if appropriate.  If we had
2683 		 * allocated the page we have to place it on a queue.  If not
2684 		 * we just have to make sure it isn't on the cache queue
2685 		 * (pages on the cache queue are not allowed to be mapped).
2686 		 */
2687 		if (allocated) {
2688 			/*
2689 			 * Page must be zerod.
2690 			 */
2691 			vm_page_zero_fill(m);
2692 			mycpu->gd_cnt.v_zfod++;
2693 			m->valid = VM_PAGE_BITS_ALL;
2694 
2695 			/*
2696 			 * Handle dirty page case
2697 			 */
2698 			if (pprot & VM_PROT_WRITE)
2699 				vm_set_nosync(m, entry);
2700 			pmap_enter(pmap, addr, m, pprot, 0, entry);
2701 			mycpu->gd_cnt.v_vm_faults++;
2702 			if (curthread->td_lwp)
2703 				++curthread->td_lwp->lwp_ru.ru_minflt;
2704 			vm_page_deactivate(m);
2705 			if (pprot & VM_PROT_WRITE) {
2706 				/*vm_object_set_writeable_dirty(m->object);*/
2707 				vm_set_nosync(m, entry);
2708 				if (fault_flags & VM_FAULT_DIRTY) {
2709 					vm_page_dirty(m);
2710 					/*XXX*/
2711 					swap_pager_unswapped(m);
2712 				}
2713 			}
2714 			vm_page_wakeup(m);
2715 		} else if (error) {
2716 			/* couldn't busy page, no wakeup */
2717 		} else if (
2718 		    ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2719 		    (m->flags & PG_FICTITIOUS) == 0) {
2720 			/*
2721 			 * A fully valid page not undergoing soft I/O can
2722 			 * be immediately entered into the pmap.
2723 			 */
2724 			if ((m->queue - m->pc) == PQ_CACHE)
2725 				vm_page_deactivate(m);
2726 			if (pprot & VM_PROT_WRITE) {
2727 				/*vm_object_set_writeable_dirty(m->object);*/
2728 				vm_set_nosync(m, entry);
2729 				if (fault_flags & VM_FAULT_DIRTY) {
2730 					vm_page_dirty(m);
2731 					/*XXX*/
2732 					swap_pager_unswapped(m);
2733 				}
2734 			}
2735 			if (pprot & VM_PROT_WRITE)
2736 				vm_set_nosync(m, entry);
2737 			pmap_enter(pmap, addr, m, pprot, 0, entry);
2738 			mycpu->gd_cnt.v_vm_faults++;
2739 			if (curthread->td_lwp)
2740 				++curthread->td_lwp->lwp_ru.ru_minflt;
2741 			vm_page_wakeup(m);
2742 		} else {
2743 			vm_page_wakeup(m);
2744 		}
2745 	}
2746 	vm_object_chain_release(object);
2747 	vm_object_drop(object);
2748 }
2749 
2750 /*
2751  * Object can be held shared
2752  */
2753 static void
2754 vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
2755 		  vm_map_entry_t entry, int prot, int fault_flags)
2756 {
2757 	struct lwp *lp;
2758 	vm_page_t m;
2759 	vm_offset_t addr;
2760 	vm_pindex_t pindex;
2761 	vm_object_t object;
2762 	int i;
2763 	int noneg;
2764 	int nopos;
2765 	int maxpages;
2766 
2767 	/*
2768 	 * Get stable max count value, disabled if set to 0
2769 	 */
2770 	maxpages = vm_prefault_pages;
2771 	cpu_ccfence();
2772 	if (maxpages <= 0)
2773 		return;
2774 
2775 	/*
2776 	 * We do not currently prefault mappings that use virtual page
2777 	 * tables.  We do not prefault foreign pmaps.
2778 	 */
2779 	if (entry->maptype != VM_MAPTYPE_NORMAL)
2780 		return;
2781 	lp = curthread->td_lwp;
2782 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2783 		return;
2784 	object = entry->object.vm_object;
2785 	if (object->backing_object != NULL)
2786 		return;
2787 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2788 
2789 	/*
2790 	 * Limit pre-fault count to 1024 pages.
2791 	 */
2792 	if (maxpages > 1024)
2793 		maxpages = 1024;
2794 
2795 	noneg = 0;
2796 	nopos = 0;
2797 	for (i = 0; i < maxpages; ++i) {
2798 		int error;
2799 
2800 		/*
2801 		 * Calculate the page to pre-fault, stopping the scan in
2802 		 * each direction separately if the limit is reached.
2803 		 */
2804 		if (i & 1) {
2805 			if (noneg)
2806 				continue;
2807 			addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2808 		} else {
2809 			if (nopos)
2810 				continue;
2811 			addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2812 		}
2813 		if (addr < entry->start) {
2814 			noneg = 1;
2815 			if (noneg && nopos)
2816 				break;
2817 			continue;
2818 		}
2819 		if (addr >= entry->end) {
2820 			nopos = 1;
2821 			if (noneg && nopos)
2822 				break;
2823 			continue;
2824 		}
2825 
2826 		/*
2827 		 * Follow the VM object chain to obtain the page to be mapped
2828 		 * into the pmap.  This version of the prefault code only
2829 		 * works with terminal objects.
2830 		 *
2831 		 * The page must already exist.  If we encounter a problem
2832 		 * we stop here.
2833 		 *
2834 		 * WARNING!  We cannot call swap_pager_unswapped() or insert
2835 		 *	     a new vm_page with a shared token.
2836 		 */
2837 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2838 
2839 		m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2840 		if (m == NULL || error)
2841 			break;
2842 
2843 		/*
2844 		 * Skip pages already mapped, and stop scanning in that
2845 		 * direction.  When the scan terminates in both directions
2846 		 * we are done.
2847 		 */
2848 		if (pmap_prefault_ok(pmap, addr) == 0) {
2849 			vm_page_wakeup(m);
2850 			if (i & 1)
2851 				noneg = 1;
2852 			else
2853 				nopos = 1;
2854 			if (noneg && nopos)
2855 				break;
2856 			continue;
2857 		}
2858 
2859 		/*
2860 		 * Stop if the page cannot be trivially entered into the
2861 		 * pmap.
2862 		 */
2863 		if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) ||
2864 		    (m->flags & PG_FICTITIOUS) ||
2865 		    ((m->flags & PG_SWAPPED) &&
2866 		     (prot & VM_PROT_WRITE) &&
2867 		     (fault_flags & VM_FAULT_DIRTY))) {
2868 			vm_page_wakeup(m);
2869 			break;
2870 		}
2871 
2872 		/*
2873 		 * Enter the page into the pmap.  The object might be held
2874 		 * shared so we can't do any (serious) modifying operation
2875 		 * on it.
2876 		 */
2877 		if ((m->queue - m->pc) == PQ_CACHE)
2878 			vm_page_deactivate(m);
2879 		if (prot & VM_PROT_WRITE) {
2880 			vm_object_set_writeable_dirty(m->object);
2881 			vm_set_nosync(m, entry);
2882 			if (fault_flags & VM_FAULT_DIRTY) {
2883 				vm_page_dirty(m);
2884 				/* can't happeen due to conditional above */
2885 				/* swap_pager_unswapped(m); */
2886 			}
2887 		}
2888 		pmap_enter(pmap, addr, m, prot, 0, entry);
2889 		mycpu->gd_cnt.v_vm_faults++;
2890 		if (curthread->td_lwp)
2891 			++curthread->td_lwp->lwp_ru.ru_minflt;
2892 		vm_page_wakeup(m);
2893 	}
2894 }
2895