xref: /dflybsd-src/sys/vm/vm_fault.c (revision d638c6eedc81671c3ceddd06ef20463940cb6a43)
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 
160 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int);
161 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *,
162 			vpte_t, int, int);
163 #if 0
164 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
165 #endif
166 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry);
167 static void vm_prefault(pmap_t pmap, vm_offset_t addra,
168 			vm_map_entry_t entry, int prot, int fault_flags);
169 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
170 			vm_map_entry_t entry, int prot, int fault_flags);
171 
172 static __inline void
173 release_page(struct faultstate *fs)
174 {
175 	vm_page_deactivate(fs->m);
176 	vm_page_wakeup(fs->m);
177 	fs->m = NULL;
178 }
179 
180 /*
181  * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse
182  *	 requires relocking and then checking the timestamp.
183  *
184  * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do
185  *	 not have to update fs->map_generation here.
186  *
187  * NOTE: This function can fail due to a deadlock against the caller's
188  *	 holding of a vm_page BUSY.
189  */
190 static __inline int
191 relock_map(struct faultstate *fs)
192 {
193 	int error;
194 
195 	if (fs->lookup_still_valid == FALSE && fs->map) {
196 		error = vm_map_lock_read_to(fs->map);
197 		if (error == 0)
198 			fs->lookup_still_valid = TRUE;
199 	} else {
200 		error = 0;
201 	}
202 	return error;
203 }
204 
205 static __inline void
206 unlock_map(struct faultstate *fs)
207 {
208 	if (fs->lookup_still_valid && fs->map) {
209 		vm_map_lookup_done(fs->map, fs->entry, 0);
210 		fs->lookup_still_valid = FALSE;
211 	}
212 }
213 
214 /*
215  * Clean up after a successful call to vm_fault_object() so another call
216  * to vm_fault_object() can be made.
217  */
218 static void
219 _cleanup_successful_fault(struct faultstate *fs, int relock)
220 {
221 	/*
222 	 * We allocated a junk page for a COW operation that did
223 	 * not occur, the page must be freed.
224 	 */
225 	if (fs->object != fs->first_object) {
226 		KKASSERT(fs->first_shared == 0);
227 		vm_page_free(fs->first_m);
228 		vm_object_pip_wakeup(fs->object);
229 		fs->first_m = NULL;
230 	}
231 
232 	/*
233 	 * Reset fs->object.
234 	 */
235 	fs->object = fs->first_object;
236 	if (relock && fs->lookup_still_valid == FALSE) {
237 		if (fs->map)
238 			vm_map_lock_read(fs->map);
239 		fs->lookup_still_valid = TRUE;
240 	}
241 }
242 
243 static void
244 _unlock_things(struct faultstate *fs, int dealloc)
245 {
246 	_cleanup_successful_fault(fs, 0);
247 	if (dealloc) {
248 		/*vm_object_deallocate(fs->first_object);*/
249 		/*fs->first_object = NULL; drop used later on */
250 	}
251 	unlock_map(fs);
252 	if (fs->vp != NULL) {
253 		vput(fs->vp);
254 		fs->vp = NULL;
255 	}
256 }
257 
258 #define unlock_things(fs) _unlock_things(fs, 0)
259 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
260 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
261 
262 /*
263  * TRYPAGER
264  *
265  * Determine if the pager for the current object *might* contain the page.
266  *
267  * We only need to try the pager if this is not a default object (default
268  * objects are zero-fill and have no real pager), and if we are not taking
269  * a wiring fault or if the FS entry is wired.
270  */
271 #define TRYPAGER(fs)	\
272 		(fs->object->type != OBJT_DEFAULT && \
273 		(((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
274 
275 /*
276  * vm_fault:
277  *
278  * Handle a page fault occuring at the given address, requiring the given
279  * permissions, in the map specified.  If successful, the page is inserted
280  * into the associated physical map.
281  *
282  * NOTE: The given address should be truncated to the proper page address.
283  *
284  * KERN_SUCCESS is returned if the page fault is handled; otherwise,
285  * a standard error specifying why the fault is fatal is returned.
286  *
287  * The map in question must be referenced, and remains so.
288  * The caller may hold no locks.
289  * No other requirements.
290  */
291 int
292 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
293 {
294 	int result;
295 	vm_pindex_t first_pindex;
296 	struct faultstate fs;
297 	struct lwp *lp;
298 	struct proc *p;
299 	thread_t td;
300 	struct vm_map_ilock ilock;
301 	int didilock;
302 	int growstack;
303 	int retry = 0;
304 	int inherit_prot;
305 
306 	inherit_prot = fault_type & VM_PROT_NOSYNC;
307 	fs.hardfault = 0;
308 	fs.fault_flags = fault_flags;
309 	fs.vp = NULL;
310 	fs.shared = vm_shared_fault;
311 	fs.first_shared = vm_shared_fault;
312 	growstack = 1;
313 
314 	/*
315 	 * vm_map interactions
316 	 */
317 	td = curthread;
318 	if ((lp = td->td_lwp) != NULL)
319 		lp->lwp_flags |= LWP_PAGING;
320 
321 RetryFault:
322 	/*
323 	 * Find the vm_map_entry representing the backing store and resolve
324 	 * the top level object and page index.  This may have the side
325 	 * effect of executing a copy-on-write on the map entry and/or
326 	 * creating a shadow object, but will not COW any actual VM pages.
327 	 *
328 	 * On success fs.map is left read-locked and various other fields
329 	 * are initialized but not otherwise referenced or locked.
330 	 *
331 	 * NOTE!  vm_map_lookup will try to upgrade the fault_type to
332 	 *	  VM_FAULT_WRITE if the map entry is a virtual page table
333 	 *	  and also writable, so we can set the 'A'accessed bit in
334 	 *	  the virtual page table entry.
335 	 */
336 	fs.map = map;
337 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
338 			       &fs.entry, &fs.first_object,
339 			       &first_pindex, &fs.first_prot, &fs.wired);
340 
341 	/*
342 	 * If the lookup failed or the map protections are incompatible,
343 	 * the fault generally fails.
344 	 *
345 	 * The failure could be due to TDF_NOFAULT if vm_map_lookup()
346 	 * tried to do a COW fault.
347 	 *
348 	 * If the caller is trying to do a user wiring we have more work
349 	 * to do.
350 	 */
351 	if (result != KERN_SUCCESS) {
352 		if (result == KERN_FAILURE_NOFAULT) {
353 			result = KERN_FAILURE;
354 			goto done;
355 		}
356 		if (result != KERN_PROTECTION_FAILURE ||
357 		    (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
358 		{
359 			if (result == KERN_INVALID_ADDRESS && growstack &&
360 			    map != &kernel_map && curproc != NULL) {
361 				result = vm_map_growstack(map, vaddr);
362 				if (result == KERN_SUCCESS) {
363 					growstack = 0;
364 					++retry;
365 					goto RetryFault;
366 				}
367 				result = KERN_FAILURE;
368 			}
369 			goto done;
370 		}
371 
372 		/*
373 		 * If we are user-wiring a r/w segment, and it is COW, then
374 		 * we need to do the COW operation.  Note that we don't
375 		 * currently COW RO sections now, because it is NOT desirable
376 		 * to COW .text.  We simply keep .text from ever being COW'ed
377 		 * and take the heat that one cannot debug wired .text sections.
378 		 */
379 		result = vm_map_lookup(&fs.map, vaddr,
380 				       VM_PROT_READ|VM_PROT_WRITE|
381 				        VM_PROT_OVERRIDE_WRITE,
382 				       &fs.entry, &fs.first_object,
383 				       &first_pindex, &fs.first_prot,
384 				       &fs.wired);
385 		if (result != KERN_SUCCESS) {
386 			/* could also be KERN_FAILURE_NOFAULT */
387 			result = KERN_FAILURE;
388 			goto done;
389 		}
390 
391 		/*
392 		 * If we don't COW now, on a user wire, the user will never
393 		 * be able to write to the mapping.  If we don't make this
394 		 * restriction, the bookkeeping would be nearly impossible.
395 		 *
396 		 * XXX We have a shared lock, this will have a MP race but
397 		 * I don't see how it can hurt anything.
398 		 */
399 		if ((fs.entry->protection & VM_PROT_WRITE) == 0) {
400 			atomic_clear_char(&fs.entry->max_protection,
401 					  VM_PROT_WRITE);
402 		}
403 	}
404 
405 	/*
406 	 * fs.map is read-locked
407 	 *
408 	 * Misc checks.  Save the map generation number to detect races.
409 	 */
410 	fs.map_generation = fs.map->timestamp;
411 	fs.lookup_still_valid = TRUE;
412 	fs.first_m = NULL;
413 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
414 	fs.prot = fs.first_prot;	/* default (used by uksmap) */
415 
416 	if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) {
417 		if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
418 			panic("vm_fault: fault on nofault entry, addr: %p",
419 			      (void *)vaddr);
420 		}
421 		if ((fs.entry->eflags & MAP_ENTRY_KSTACK) &&
422 		    vaddr >= fs.entry->start &&
423 		    vaddr < fs.entry->start + PAGE_SIZE) {
424 			panic("vm_fault: fault on stack guard, addr: %p",
425 			      (void *)vaddr);
426 		}
427 	}
428 
429 	/*
430 	 * A user-kernel shared map has no VM object and bypasses
431 	 * everything.  We execute the uksmap function with a temporary
432 	 * fictitious vm_page.  The address is directly mapped with no
433 	 * management.
434 	 */
435 	if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) {
436 		struct vm_page fakem;
437 
438 		bzero(&fakem, sizeof(fakem));
439 		fakem.pindex = first_pindex;
440 		fakem.flags = PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED;
441 		fakem.valid = VM_PAGE_BITS_ALL;
442 		fakem.pat_mode = VM_MEMATTR_DEFAULT;
443 		if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) {
444 			result = KERN_FAILURE;
445 			unlock_things(&fs);
446 			goto done2;
447 		}
448 		pmap_enter(fs.map->pmap, vaddr, &fakem, fs.prot | inherit_prot,
449 			   fs.wired, fs.entry);
450 		goto done_success;
451 	}
452 
453 	/*
454 	 * A system map entry may return a NULL object.  No object means
455 	 * no pager means an unrecoverable kernel fault.
456 	 */
457 	if (fs.first_object == NULL) {
458 		panic("vm_fault: unrecoverable fault at %p in entry %p",
459 			(void *)vaddr, fs.entry);
460 	}
461 
462 	/*
463 	 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
464 	 * is set.
465 	 *
466 	 * Unfortunately a deadlock can occur if we are forced to page-in
467 	 * from swap, but diving all the way into the vm_pager_get_page()
468 	 * function to find out is too much.  Just check the object type.
469 	 *
470 	 * The deadlock is a CAM deadlock on a busy VM page when trying
471 	 * to finish an I/O if another process gets stuck in
472 	 * vop_helper_read_shortcut() due to a swap fault.
473 	 */
474 	if ((td->td_flags & TDF_NOFAULT) &&
475 	    (retry ||
476 	     fs.first_object->type == OBJT_VNODE ||
477 	     fs.first_object->type == OBJT_SWAP ||
478 	     fs.first_object->backing_object)) {
479 		result = KERN_FAILURE;
480 		unlock_things(&fs);
481 		goto done2;
482 	}
483 
484 	/*
485 	 * If the entry is wired we cannot change the page protection.
486 	 */
487 	if (fs.wired)
488 		fault_type = fs.first_prot;
489 
490 	/*
491 	 * We generally want to avoid unnecessary exclusive modes on backing
492 	 * and terminal objects because this can seriously interfere with
493 	 * heavily fork()'d processes (particularly /bin/sh scripts).
494 	 *
495 	 * However, we also want to avoid unnecessary retries due to needed
496 	 * shared->exclusive promotion for common faults.  Exclusive mode is
497 	 * always needed if any page insertion, rename, or free occurs in an
498 	 * object (and also indirectly if any I/O is done).
499 	 *
500 	 * The main issue here is going to be fs.first_shared.  If the
501 	 * first_object has a backing object which isn't shadowed and the
502 	 * process is single-threaded we might as well use an exclusive
503 	 * lock/chain right off the bat.
504 	 */
505 	if (fs.first_shared && fs.first_object->backing_object &&
506 	    LIST_EMPTY(&fs.first_object->shadow_head) &&
507 	    td->td_proc && td->td_proc->p_nthreads == 1) {
508 		fs.first_shared = 0;
509 	}
510 
511 	/*
512 	 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
513 	 * VM_FAULT_DIRTY  - may require swap_pager_unswapped() later, but
514 	 *		     we can try shared first.
515 	 */
516 	if (fault_flags & VM_FAULT_UNSWAP) {
517 		fs.first_shared = 0;
518 	}
519 
520 	/*
521 	 * Obtain a top-level object lock, shared or exclusive depending
522 	 * on fs.first_shared.  If a shared lock winds up being insufficient
523 	 * we will retry with an exclusive lock.
524 	 *
525 	 * The vnode pager lock is always shared.
526 	 */
527 	if (fs.first_shared)
528 		vm_object_hold_shared(fs.first_object);
529 	else
530 		vm_object_hold(fs.first_object);
531 	if (fs.vp == NULL)
532 		fs.vp = vnode_pager_lock(fs.first_object);
533 
534 	/*
535 	 * The page we want is at (first_object, first_pindex), but if the
536 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
537 	 * page table to figure out the actual pindex.
538 	 *
539 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
540 	 * ONLY
541 	 */
542 	didilock = 0;
543 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
544 		vm_map_interlock(fs.map, &ilock, vaddr, vaddr + PAGE_SIZE);
545 		didilock = 1;
546 		result = vm_fault_vpagetable(&fs, &first_pindex,
547 					     fs.entry->aux.master_pde,
548 					     fault_type, 1);
549 		if (result == KERN_TRY_AGAIN) {
550 			vm_map_deinterlock(fs.map, &ilock);
551 			vm_object_drop(fs.first_object);
552 			++retry;
553 			goto RetryFault;
554 		}
555 		if (result != KERN_SUCCESS) {
556 			vm_map_deinterlock(fs.map, &ilock);
557 			goto done;
558 		}
559 	}
560 
561 	/*
562 	 * Now we have the actual (object, pindex), fault in the page.  If
563 	 * vm_fault_object() fails it will unlock and deallocate the FS
564 	 * data.   If it succeeds everything remains locked and fs->object
565 	 * will have an additional PIP count if it is not equal to
566 	 * fs->first_object
567 	 *
568 	 * vm_fault_object will set fs->prot for the pmap operation.  It is
569 	 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
570 	 * page can be safely written.  However, it will force a read-only
571 	 * mapping for a read fault if the memory is managed by a virtual
572 	 * page table.
573 	 *
574 	 * If the fault code uses the shared object lock shortcut
575 	 * we must not try to burst (we can't allocate VM pages).
576 	 */
577 	result = vm_fault_object(&fs, first_pindex, fault_type, 1);
578 
579 	if (debug_fault > 0) {
580 		--debug_fault;
581 		kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x "
582 			"fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n",
583 			result, (intmax_t)vaddr, fault_type, fault_flags,
584 			fs.m, fs.prot, fs.wired, fs.entry);
585 	}
586 
587 	if (result == KERN_TRY_AGAIN) {
588 		if (didilock)
589 			vm_map_deinterlock(fs.map, &ilock);
590 		vm_object_drop(fs.first_object);
591 		++retry;
592 		goto RetryFault;
593 	}
594 	if (result != KERN_SUCCESS) {
595 		if (didilock)
596 			vm_map_deinterlock(fs.map, &ilock);
597 		goto done;
598 	}
599 
600 	/*
601 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
602 	 * will contain a busied page.
603 	 *
604 	 * Enter the page into the pmap and do pmap-related adjustments.
605 	 */
606 	KKASSERT(fs.lookup_still_valid == TRUE);
607 	vm_page_flag_set(fs.m, PG_REFERENCED);
608 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot | inherit_prot,
609 		   fs.wired, fs.entry);
610 
611 	if (didilock)
612 		vm_map_deinterlock(fs.map, &ilock);
613 
614 	/*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
615 	KKASSERT(fs.m->flags & PG_BUSY);
616 
617 	/*
618 	 * If the page is not wired down, then put it where the pageout daemon
619 	 * can find it.
620 	 */
621 	if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
622 		if (fs.wired)
623 			vm_page_wire(fs.m);
624 		else
625 			vm_page_unwire(fs.m, 1);
626 	} else {
627 		vm_page_activate(fs.m);
628 	}
629 	vm_page_wakeup(fs.m);
630 
631 	/*
632 	 * Burst in a few more pages if possible.  The fs.map should still
633 	 * be locked.  To avoid interlocking against a vnode->getblk
634 	 * operation we had to be sure to unbusy our primary vm_page above
635 	 * first.
636 	 *
637 	 * A normal burst can continue down backing store, only execute
638 	 * if we are holding an exclusive lock, otherwise the exclusive
639 	 * locks the burst code gets might cause excessive SMP collisions.
640 	 *
641 	 * A quick burst can be utilized when there is no backing object
642 	 * (i.e. a shared file mmap).
643 	 */
644 	if ((fault_flags & VM_FAULT_BURST) &&
645 	    (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
646 	    fs.wired == 0) {
647 		if (fs.first_shared == 0 && fs.shared == 0) {
648 			vm_prefault(fs.map->pmap, vaddr,
649 				    fs.entry, fs.prot, fault_flags);
650 		} else {
651 			vm_prefault_quick(fs.map->pmap, vaddr,
652 					  fs.entry, fs.prot, fault_flags);
653 		}
654 	}
655 
656 done_success:
657 	mycpu->gd_cnt.v_vm_faults++;
658 	if (td->td_lwp)
659 		++td->td_lwp->lwp_ru.ru_minflt;
660 
661 	/*
662 	 * Unlock everything, and return
663 	 */
664 	unlock_things(&fs);
665 
666 	if (td->td_lwp) {
667 		if (fs.hardfault) {
668 			td->td_lwp->lwp_ru.ru_majflt++;
669 		} else {
670 			td->td_lwp->lwp_ru.ru_minflt++;
671 		}
672 	}
673 
674 	/*vm_object_deallocate(fs.first_object);*/
675 	/*fs.m = NULL; */
676 	/*fs.first_object = NULL; must still drop later */
677 
678 	result = KERN_SUCCESS;
679 done:
680 	if (fs.first_object)
681 		vm_object_drop(fs.first_object);
682 done2:
683 	if (lp)
684 		lp->lwp_flags &= ~LWP_PAGING;
685 
686 #if !defined(NO_SWAPPING)
687 	/*
688 	 * Check the process RSS limit and force deactivation and
689 	 * (asynchronous) paging if necessary.  This is a complex operation,
690 	 * only do it for direct user-mode faults, for now.
691 	 *
692 	 * To reduce overhead implement approximately a ~16MB hysteresis.
693 	 */
694 	p = td->td_proc;
695 	if ((fault_flags & VM_FAULT_USERMODE) && lp &&
696 	    p->p_limit && map->pmap && vm_pageout_memuse_mode >= 1 &&
697 	    map != &kernel_map) {
698 		vm_pindex_t limit;
699 		vm_pindex_t size;
700 
701 		limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur,
702 					p->p_rlimit[RLIMIT_RSS].rlim_max));
703 		size = pmap_resident_tlnw_count(map->pmap);
704 		if (limit >= 0 && size > 4096 && size - 4096 >= limit) {
705 			vm_pageout_map_deactivate_pages(map, limit);
706 		}
707 	}
708 #endif
709 
710 	return (result);
711 }
712 
713 /*
714  * Fault in the specified virtual address in the current process map,
715  * returning a held VM page or NULL.  See vm_fault_page() for more
716  * information.
717  *
718  * No requirements.
719  */
720 vm_page_t
721 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type,
722 		    int *errorp, int *busyp)
723 {
724 	struct lwp *lp = curthread->td_lwp;
725 	vm_page_t m;
726 
727 	m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
728 			  fault_type, VM_FAULT_NORMAL,
729 			  errorp, busyp);
730 	return(m);
731 }
732 
733 /*
734  * Fault in the specified virtual address in the specified map, doing all
735  * necessary manipulation of the object store and all necessary I/O.  Return
736  * a held VM page or NULL, and set *errorp.  The related pmap is not
737  * updated.
738  *
739  * If busyp is not NULL then *busyp will be set to TRUE if this routine
740  * decides to return a busied page (aka VM_PROT_WRITE), or FALSE if it
741  * does not (VM_PROT_WRITE not specified or busyp is NULL).  If busyp is
742  * NULL the returned page is only held.
743  *
744  * If the caller has no intention of writing to the page's contents, busyp
745  * can be passed as NULL along with VM_PROT_WRITE to force a COW operation
746  * without busying the page.
747  *
748  * The returned page will also be marked PG_REFERENCED.
749  *
750  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
751  * error will be returned.
752  *
753  * No requirements.
754  */
755 vm_page_t
756 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
757 	      int fault_flags, int *errorp, int *busyp)
758 {
759 	vm_pindex_t first_pindex;
760 	struct faultstate fs;
761 	int result;
762 	int retry;
763 	int growstack;
764 	vm_prot_t orig_fault_type = fault_type;
765 
766 	retry = 0;
767 	fs.hardfault = 0;
768 	fs.fault_flags = fault_flags;
769 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
770 
771 	/*
772 	 * Dive the pmap (concurrency possible).  If we find the
773 	 * appropriate page we can terminate early and quickly.
774 	 *
775 	 * This works great for normal programs but will always return
776 	 * NULL for host lookups of vkernel maps in VMM mode.
777 	 */
778 	fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type, busyp);
779 	if (fs.m) {
780 		if (fault_type & (VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE))
781 			vm_page_dirty(fs.m);
782 		*errorp = 0;
783 		return(fs.m);
784 	}
785 
786 	/*
787 	 * Otherwise take a concurrency hit and do a formal page
788 	 * fault.
789 	 */
790 	fs.vp = NULL;
791 	fs.shared = vm_shared_fault;
792 	fs.first_shared = vm_shared_fault;
793 	growstack = 1;
794 
795 	/*
796 	 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
797 	 * VM_FAULT_DIRTY  - may require swap_pager_unswapped() later, but
798 	 *		     we can try shared first.
799 	 */
800 	if (fault_flags & VM_FAULT_UNSWAP) {
801 		fs.first_shared = 0;
802 	}
803 
804 RetryFault:
805 	/*
806 	 * Find the vm_map_entry representing the backing store and resolve
807 	 * the top level object and page index.  This may have the side
808 	 * effect of executing a copy-on-write on the map entry and/or
809 	 * creating a shadow object, but will not COW any actual VM pages.
810 	 *
811 	 * On success fs.map is left read-locked and various other fields
812 	 * are initialized but not otherwise referenced or locked.
813 	 *
814 	 * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
815 	 *	  if the map entry is a virtual page table and also writable,
816 	 *	  so we can set the 'A'accessed bit in the virtual page table
817 	 *	  entry.
818 	 */
819 	fs.map = map;
820 	result = vm_map_lookup(&fs.map, vaddr, fault_type,
821 			       &fs.entry, &fs.first_object,
822 			       &first_pindex, &fs.first_prot, &fs.wired);
823 
824 	if (result != KERN_SUCCESS) {
825 		if (result == KERN_FAILURE_NOFAULT) {
826 			*errorp = KERN_FAILURE;
827 			fs.m = NULL;
828 			goto done;
829 		}
830 		if (result != KERN_PROTECTION_FAILURE ||
831 		    (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
832 		{
833 			if (result == KERN_INVALID_ADDRESS && growstack &&
834 			    map != &kernel_map && curproc != NULL) {
835 				result = vm_map_growstack(map, vaddr);
836 				if (result == KERN_SUCCESS) {
837 					growstack = 0;
838 					++retry;
839 					goto RetryFault;
840 				}
841 				result = KERN_FAILURE;
842 			}
843 			fs.m = NULL;
844 			*errorp = result;
845 			goto done;
846 		}
847 
848 		/*
849 		 * If we are user-wiring a r/w segment, and it is COW, then
850 		 * we need to do the COW operation.  Note that we don't
851 		 * currently COW RO sections now, because it is NOT desirable
852 		 * to COW .text.  We simply keep .text from ever being COW'ed
853 		 * and take the heat that one cannot debug wired .text sections.
854 		 */
855 		result = vm_map_lookup(&fs.map, vaddr,
856 				       VM_PROT_READ|VM_PROT_WRITE|
857 				        VM_PROT_OVERRIDE_WRITE,
858 				       &fs.entry, &fs.first_object,
859 				       &first_pindex, &fs.first_prot,
860 				       &fs.wired);
861 		if (result != KERN_SUCCESS) {
862 			/* could also be KERN_FAILURE_NOFAULT */
863 			*errorp = KERN_FAILURE;
864 			fs.m = NULL;
865 			goto done;
866 		}
867 
868 		/*
869 		 * If we don't COW now, on a user wire, the user will never
870 		 * be able to write to the mapping.  If we don't make this
871 		 * restriction, the bookkeeping would be nearly impossible.
872 		 *
873 		 * XXX We have a shared lock, this will have a MP race but
874 		 * I don't see how it can hurt anything.
875 		 */
876 		if ((fs.entry->protection & VM_PROT_WRITE) == 0) {
877 			atomic_clear_char(&fs.entry->max_protection,
878 					  VM_PROT_WRITE);
879 		}
880 	}
881 
882 	/*
883 	 * fs.map is read-locked
884 	 *
885 	 * Misc checks.  Save the map generation number to detect races.
886 	 */
887 	fs.map_generation = fs.map->timestamp;
888 	fs.lookup_still_valid = TRUE;
889 	fs.first_m = NULL;
890 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
891 
892 	if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
893 		panic("vm_fault: fault on nofault entry, addr: %lx",
894 		    (u_long)vaddr);
895 	}
896 
897 	/*
898 	 * A user-kernel shared map has no VM object and bypasses
899 	 * everything.  We execute the uksmap function with a temporary
900 	 * fictitious vm_page.  The address is directly mapped with no
901 	 * management.
902 	 */
903 	if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) {
904 		struct vm_page fakem;
905 
906 		bzero(&fakem, sizeof(fakem));
907 		fakem.pindex = first_pindex;
908 		fakem.flags = PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED;
909 		fakem.valid = VM_PAGE_BITS_ALL;
910 		fakem.pat_mode = VM_MEMATTR_DEFAULT;
911 		if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) {
912 			*errorp = KERN_FAILURE;
913 			fs.m = NULL;
914 			unlock_things(&fs);
915 			goto done2;
916 		}
917 		fs.m = PHYS_TO_VM_PAGE(fakem.phys_addr);
918 		vm_page_hold(fs.m);
919 		if (busyp)
920 			*busyp = 0;	/* don't need to busy R or W */
921 		unlock_things(&fs);
922 		*errorp = 0;
923 		goto done;
924 	}
925 
926 
927 	/*
928 	 * A system map entry may return a NULL object.  No object means
929 	 * no pager means an unrecoverable kernel fault.
930 	 */
931 	if (fs.first_object == NULL) {
932 		panic("vm_fault: unrecoverable fault at %p in entry %p",
933 			(void *)vaddr, fs.entry);
934 	}
935 
936 	/*
937 	 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
938 	 * is set.
939 	 *
940 	 * Unfortunately a deadlock can occur if we are forced to page-in
941 	 * from swap, but diving all the way into the vm_pager_get_page()
942 	 * function to find out is too much.  Just check the object type.
943 	 */
944 	if ((curthread->td_flags & TDF_NOFAULT) &&
945 	    (retry ||
946 	     fs.first_object->type == OBJT_VNODE ||
947 	     fs.first_object->type == OBJT_SWAP ||
948 	     fs.first_object->backing_object)) {
949 		*errorp = KERN_FAILURE;
950 		unlock_things(&fs);
951 		fs.m = NULL;
952 		goto done2;
953 	}
954 
955 	/*
956 	 * If the entry is wired we cannot change the page protection.
957 	 */
958 	if (fs.wired)
959 		fault_type = fs.first_prot;
960 
961 	/*
962 	 * Make a reference to this object to prevent its disposal while we
963 	 * are messing with it.  Once we have the reference, the map is free
964 	 * to be diddled.  Since objects reference their shadows (and copies),
965 	 * they will stay around as well.
966 	 *
967 	 * The reference should also prevent an unexpected collapse of the
968 	 * parent that might move pages from the current object into the
969 	 * parent unexpectedly, resulting in corruption.
970 	 *
971 	 * Bump the paging-in-progress count to prevent size changes (e.g.
972 	 * truncation operations) during I/O.  This must be done after
973 	 * obtaining the vnode lock in order to avoid possible deadlocks.
974 	 */
975 	if (fs.first_shared)
976 		vm_object_hold_shared(fs.first_object);
977 	else
978 		vm_object_hold(fs.first_object);
979 	if (fs.vp == NULL)
980 		fs.vp = vnode_pager_lock(fs.first_object);	/* shared */
981 
982 	/*
983 	 * The page we want is at (first_object, first_pindex), but if the
984 	 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
985 	 * page table to figure out the actual pindex.
986 	 *
987 	 * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
988 	 * ONLY
989 	 */
990 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
991 		result = vm_fault_vpagetable(&fs, &first_pindex,
992 					     fs.entry->aux.master_pde,
993 					     fault_type, 1);
994 		if (result == KERN_TRY_AGAIN) {
995 			vm_object_drop(fs.first_object);
996 			++retry;
997 			goto RetryFault;
998 		}
999 		if (result != KERN_SUCCESS) {
1000 			*errorp = result;
1001 			fs.m = NULL;
1002 			goto done;
1003 		}
1004 	}
1005 
1006 	/*
1007 	 * Now we have the actual (object, pindex), fault in the page.  If
1008 	 * vm_fault_object() fails it will unlock and deallocate the FS
1009 	 * data.   If it succeeds everything remains locked and fs->object
1010 	 * will have an additinal PIP count if it is not equal to
1011 	 * fs->first_object
1012 	 */
1013 	fs.m = NULL;
1014 	result = vm_fault_object(&fs, first_pindex, fault_type, 1);
1015 
1016 	if (result == KERN_TRY_AGAIN) {
1017 		vm_object_drop(fs.first_object);
1018 		++retry;
1019 		goto RetryFault;
1020 	}
1021 	if (result != KERN_SUCCESS) {
1022 		*errorp = result;
1023 		fs.m = NULL;
1024 		goto done;
1025 	}
1026 
1027 	if ((orig_fault_type & VM_PROT_WRITE) &&
1028 	    (fs.prot & VM_PROT_WRITE) == 0) {
1029 		*errorp = KERN_PROTECTION_FAILURE;
1030 		unlock_and_deallocate(&fs);
1031 		fs.m = NULL;
1032 		goto done;
1033 	}
1034 
1035 	/*
1036 	 * DO NOT UPDATE THE PMAP!!!  This function may be called for
1037 	 * a pmap unrelated to the current process pmap, in which case
1038 	 * the current cpu core will not be listed in the pmap's pm_active
1039 	 * mask.  Thus invalidation interlocks will fail to work properly.
1040 	 *
1041 	 * (for example, 'ps' uses procfs to read program arguments from
1042 	 * each process's stack).
1043 	 *
1044 	 * In addition to the above this function will be called to acquire
1045 	 * a page that might already be faulted in, re-faulting it
1046 	 * continuously is a waste of time.
1047 	 *
1048 	 * XXX could this have been the cause of our random seg-fault
1049 	 *     issues?  procfs accesses user stacks.
1050 	 */
1051 	vm_page_flag_set(fs.m, PG_REFERENCED);
1052 #if 0
1053 	pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL);
1054 	mycpu->gd_cnt.v_vm_faults++;
1055 	if (curthread->td_lwp)
1056 		++curthread->td_lwp->lwp_ru.ru_minflt;
1057 #endif
1058 
1059 	/*
1060 	 * On success vm_fault_object() does not unlock or deallocate, and fs.m
1061 	 * will contain a busied page.  So we must unlock here after having
1062 	 * messed with the pmap.
1063 	 */
1064 	unlock_things(&fs);
1065 
1066 	/*
1067 	 * Return a held page.  We are not doing any pmap manipulation so do
1068 	 * not set PG_MAPPED.  However, adjust the page flags according to
1069 	 * the fault type because the caller may not use a managed pmapping
1070 	 * (so we don't want to lose the fact that the page will be dirtied
1071 	 * if a write fault was specified).
1072 	 */
1073 	if (fault_type & VM_PROT_WRITE)
1074 		vm_page_dirty(fs.m);
1075 	vm_page_activate(fs.m);
1076 
1077 	if (curthread->td_lwp) {
1078 		if (fs.hardfault) {
1079 			curthread->td_lwp->lwp_ru.ru_majflt++;
1080 		} else {
1081 			curthread->td_lwp->lwp_ru.ru_minflt++;
1082 		}
1083 	}
1084 
1085 	/*
1086 	 * Unlock everything, and return the held or busied page.
1087 	 */
1088 	if (busyp) {
1089 		if (fault_type & (VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE)) {
1090 			vm_page_dirty(fs.m);
1091 			*busyp = 1;
1092 		} else {
1093 			*busyp = 0;
1094 			vm_page_hold(fs.m);
1095 			vm_page_wakeup(fs.m);
1096 		}
1097 	} else {
1098 		vm_page_hold(fs.m);
1099 		vm_page_wakeup(fs.m);
1100 	}
1101 	/*vm_object_deallocate(fs.first_object);*/
1102 	/*fs.first_object = NULL; */
1103 	*errorp = 0;
1104 
1105 done:
1106 	if (fs.first_object)
1107 		vm_object_drop(fs.first_object);
1108 done2:
1109 	return(fs.m);
1110 }
1111 
1112 /*
1113  * Fault in the specified (object,offset), dirty the returned page as
1114  * needed.  If the requested fault_type cannot be done NULL and an
1115  * error is returned.
1116  *
1117  * A held (but not busied) page is returned.
1118  *
1119  * The passed in object must be held as specified by the shared
1120  * argument.
1121  */
1122 vm_page_t
1123 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
1124 		     vm_prot_t fault_type, int fault_flags,
1125 		     int *sharedp, int *errorp)
1126 {
1127 	int result;
1128 	vm_pindex_t first_pindex;
1129 	struct faultstate fs;
1130 	struct vm_map_entry entry;
1131 
1132 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1133 	bzero(&entry, sizeof(entry));
1134 	entry.object.vm_object = object;
1135 	entry.maptype = VM_MAPTYPE_NORMAL;
1136 	entry.protection = entry.max_protection = fault_type;
1137 
1138 	fs.hardfault = 0;
1139 	fs.fault_flags = fault_flags;
1140 	fs.map = NULL;
1141 	fs.shared = vm_shared_fault;
1142 	fs.first_shared = *sharedp;
1143 	fs.vp = NULL;
1144 	KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
1145 
1146 	/*
1147 	 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
1148 	 * VM_FAULT_DIRTY  - may require swap_pager_unswapped() later, but
1149 	 *		     we can try shared first.
1150 	 */
1151 	if (fs.first_shared && (fault_flags & VM_FAULT_UNSWAP)) {
1152 		fs.first_shared = 0;
1153 		vm_object_upgrade(object);
1154 	}
1155 
1156 	/*
1157 	 * Retry loop as needed (typically for shared->exclusive transitions)
1158 	 */
1159 RetryFault:
1160 	*sharedp = fs.first_shared;
1161 	first_pindex = OFF_TO_IDX(offset);
1162 	fs.first_object = object;
1163 	fs.entry = &entry;
1164 	fs.first_prot = fault_type;
1165 	fs.wired = 0;
1166 	/*fs.map_generation = 0; unused */
1167 
1168 	/*
1169 	 * Make a reference to this object to prevent its disposal while we
1170 	 * are messing with it.  Once we have the reference, the map is free
1171 	 * to be diddled.  Since objects reference their shadows (and copies),
1172 	 * they will stay around as well.
1173 	 *
1174 	 * The reference should also prevent an unexpected collapse of the
1175 	 * parent that might move pages from the current object into the
1176 	 * parent unexpectedly, resulting in corruption.
1177 	 *
1178 	 * Bump the paging-in-progress count to prevent size changes (e.g.
1179 	 * truncation operations) during I/O.  This must be done after
1180 	 * obtaining the vnode lock in order to avoid possible deadlocks.
1181 	 */
1182 	if (fs.vp == NULL)
1183 		fs.vp = vnode_pager_lock(fs.first_object);
1184 
1185 	fs.lookup_still_valid = TRUE;
1186 	fs.first_m = NULL;
1187 	fs.object = fs.first_object;	/* so unlock_and_deallocate works */
1188 
1189 #if 0
1190 	/* XXX future - ability to operate on VM object using vpagetable */
1191 	if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1192 		result = vm_fault_vpagetable(&fs, &first_pindex,
1193 					     fs.entry->aux.master_pde,
1194 					     fault_type, 0);
1195 		if (result == KERN_TRY_AGAIN) {
1196 			if (fs.first_shared == 0 && *sharedp)
1197 				vm_object_upgrade(object);
1198 			goto RetryFault;
1199 		}
1200 		if (result != KERN_SUCCESS) {
1201 			*errorp = result;
1202 			return (NULL);
1203 		}
1204 	}
1205 #endif
1206 
1207 	/*
1208 	 * Now we have the actual (object, pindex), fault in the page.  If
1209 	 * vm_fault_object() fails it will unlock and deallocate the FS
1210 	 * data.   If it succeeds everything remains locked and fs->object
1211 	 * will have an additinal PIP count if it is not equal to
1212 	 * fs->first_object
1213 	 *
1214 	 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact.
1215 	 * We may have to upgrade its lock to handle the requested fault.
1216 	 */
1217 	result = vm_fault_object(&fs, first_pindex, fault_type, 0);
1218 
1219 	if (result == KERN_TRY_AGAIN) {
1220 		if (fs.first_shared == 0 && *sharedp)
1221 			vm_object_upgrade(object);
1222 		goto RetryFault;
1223 	}
1224 	if (result != KERN_SUCCESS) {
1225 		*errorp = result;
1226 		return(NULL);
1227 	}
1228 
1229 	if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
1230 		*errorp = KERN_PROTECTION_FAILURE;
1231 		unlock_and_deallocate(&fs);
1232 		return(NULL);
1233 	}
1234 
1235 	/*
1236 	 * On success vm_fault_object() does not unlock or deallocate, so we
1237 	 * do it here.  Note that the returned fs.m will be busied.
1238 	 */
1239 	unlock_things(&fs);
1240 
1241 	/*
1242 	 * Return a held page.  We are not doing any pmap manipulation so do
1243 	 * not set PG_MAPPED.  However, adjust the page flags according to
1244 	 * the fault type because the caller may not use a managed pmapping
1245 	 * (so we don't want to lose the fact that the page will be dirtied
1246 	 * if a write fault was specified).
1247 	 */
1248 	vm_page_hold(fs.m);
1249 	vm_page_activate(fs.m);
1250 	if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY))
1251 		vm_page_dirty(fs.m);
1252 	if (fault_flags & VM_FAULT_UNSWAP)
1253 		swap_pager_unswapped(fs.m);
1254 
1255 	/*
1256 	 * Indicate that the page was accessed.
1257 	 */
1258 	vm_page_flag_set(fs.m, PG_REFERENCED);
1259 
1260 	if (curthread->td_lwp) {
1261 		if (fs.hardfault) {
1262 			curthread->td_lwp->lwp_ru.ru_majflt++;
1263 		} else {
1264 			curthread->td_lwp->lwp_ru.ru_minflt++;
1265 		}
1266 	}
1267 
1268 	/*
1269 	 * Unlock everything, and return the held page.
1270 	 */
1271 	vm_page_wakeup(fs.m);
1272 	/*vm_object_deallocate(fs.first_object);*/
1273 	/*fs.first_object = NULL; */
1274 
1275 	*errorp = 0;
1276 	return(fs.m);
1277 }
1278 
1279 /*
1280  * Translate the virtual page number (first_pindex) that is relative
1281  * to the address space into a logical page number that is relative to the
1282  * backing object.  Use the virtual page table pointed to by (vpte).
1283  *
1284  * Possibly downgrade the protection based on the vpte bits.
1285  *
1286  * This implements an N-level page table.  Any level can terminate the
1287  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
1288  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
1289  */
1290 static
1291 int
1292 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
1293 		    vpte_t vpte, int fault_type, int allow_nofault)
1294 {
1295 	struct lwbuf *lwb;
1296 	struct lwbuf lwb_cache;
1297 	int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
1298 	int result;
1299 	vpte_t *ptep;
1300 
1301 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1302 	for (;;) {
1303 		/*
1304 		 * We cannot proceed if the vpte is not valid, not readable
1305 		 * for a read fault, or not writable for a write fault.
1306 		 */
1307 		if ((vpte & VPTE_V) == 0) {
1308 			unlock_and_deallocate(fs);
1309 			return (KERN_FAILURE);
1310 		}
1311 		if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) {
1312 			unlock_and_deallocate(fs);
1313 			return (KERN_FAILURE);
1314 		}
1315 		if ((vpte & VPTE_PS) || vshift == 0)
1316 			break;
1317 
1318 		/*
1319 		 * Get the page table page.  Nominally we only read the page
1320 		 * table, but since we are actively setting VPTE_M and VPTE_A,
1321 		 * tell vm_fault_object() that we are writing it.
1322 		 *
1323 		 * There is currently no real need to optimize this.
1324 		 */
1325 		result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
1326 					 VM_PROT_READ|VM_PROT_WRITE,
1327 					 allow_nofault);
1328 		if (result != KERN_SUCCESS)
1329 			return (result);
1330 
1331 		/*
1332 		 * Process the returned fs.m and look up the page table
1333 		 * entry in the page table page.
1334 		 */
1335 		vshift -= VPTE_PAGE_BITS;
1336 		lwb = lwbuf_alloc(fs->m, &lwb_cache);
1337 		ptep = ((vpte_t *)lwbuf_kva(lwb) +
1338 		        ((*pindex >> vshift) & VPTE_PAGE_MASK));
1339 		vm_page_activate(fs->m);
1340 
1341 		/*
1342 		 * Page table write-back - entire operation including
1343 		 * validation of the pte must be atomic to avoid races
1344 		 * against the vkernel changing the pte.
1345 		 *
1346 		 * If the vpte is valid for the* requested operation, do
1347 		 * a write-back to the page table.
1348 		 *
1349 		 * XXX VPTE_M is not set properly for page directory pages.
1350 		 * It doesn't get set in the page directory if the page table
1351 		 * is modified during a read access.
1352 		 */
1353 		for (;;) {
1354 			vpte_t nvpte;
1355 
1356 			/*
1357 			 * Reload for the cmpset, but make sure the pte is
1358 			 * still valid.
1359 			 */
1360 			vpte = *ptep;
1361 			cpu_ccfence();
1362 			nvpte = vpte;
1363 
1364 			if ((vpte & VPTE_V) == 0)
1365 				break;
1366 
1367 			if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW))
1368 				nvpte |= VPTE_M | VPTE_A;
1369 			if (fault_type & VM_PROT_READ)
1370 				nvpte |= VPTE_A;
1371 			if (vpte == nvpte)
1372 				break;
1373 			if (atomic_cmpset_long(ptep, vpte, nvpte)) {
1374 				vm_page_dirty(fs->m);
1375 				break;
1376 			}
1377 		}
1378 		lwbuf_free(lwb);
1379 		vm_page_flag_set(fs->m, PG_REFERENCED);
1380 		vm_page_wakeup(fs->m);
1381 		fs->m = NULL;
1382 		cleanup_successful_fault(fs);
1383 	}
1384 
1385 	/*
1386 	 * When the vkernel sets VPTE_RW it expects the real kernel to
1387 	 * reflect VPTE_M back when the page is modified via the mapping.
1388 	 * In order to accomplish this the real kernel must map the page
1389 	 * read-only for read faults and use write faults to reflect VPTE_M
1390 	 * back.
1391 	 *
1392 	 * Once VPTE_M has been set, the real kernel's pte allows writing.
1393 	 * If the vkernel clears VPTE_M the vkernel must be sure to
1394 	 * MADV_INVAL the real kernel's mappings to force the real kernel
1395 	 * to re-fault on the next write so oit can set VPTE_M again.
1396 	 */
1397 	if ((fault_type & VM_PROT_WRITE) == 0 &&
1398 	    (vpte & (VPTE_RW | VPTE_M)) != (VPTE_RW | VPTE_M)) {
1399 		fs->first_prot &= ~VM_PROT_WRITE;
1400 	}
1401 
1402 	/*
1403 	 * Combine remaining address bits with the vpte.
1404 	 */
1405 	*pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
1406 		  (*pindex & ((1L << vshift) - 1));
1407 	return (KERN_SUCCESS);
1408 }
1409 
1410 
1411 /*
1412  * This is the core of the vm_fault code.
1413  *
1414  * Do all operations required to fault-in (fs.first_object, pindex).  Run
1415  * through the shadow chain as necessary and do required COW or virtual
1416  * copy operations.  The caller has already fully resolved the vm_map_entry
1417  * and, if appropriate, has created a copy-on-write layer.  All we need to
1418  * do is iterate the object chain.
1419  *
1420  * On failure (fs) is unlocked and deallocated and the caller may return or
1421  * retry depending on the failure code.  On success (fs) is NOT unlocked or
1422  * deallocated, fs.m will contained a resolved, busied page, and fs.object
1423  * will have an additional PIP count if it is not equal to fs.first_object.
1424  *
1425  * If locks based on fs->first_shared or fs->shared are insufficient,
1426  * clear the appropriate field(s) and return RETRY.  COWs require that
1427  * first_shared be 0, while page allocations (or frees) require that
1428  * shared be 0.  Renames require that both be 0.
1429  *
1430  * NOTE! fs->[first_]shared might be set with VM_FAULT_DIRTY also set.
1431  *	 we will have to retry with it exclusive if the vm_page is
1432  *	 PG_SWAPPED.
1433  *
1434  * fs->first_object must be held on call.
1435  */
1436 static
1437 int
1438 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex,
1439 		vm_prot_t fault_type, int allow_nofault)
1440 {
1441 	vm_object_t next_object;
1442 	vm_pindex_t pindex;
1443 	int error;
1444 
1445 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1446 	fs->prot = fs->first_prot;
1447 	fs->object = fs->first_object;
1448 	pindex = first_pindex;
1449 
1450 	vm_object_chain_acquire(fs->first_object, fs->shared);
1451 	vm_object_pip_add(fs->first_object, 1);
1452 
1453 	/*
1454 	 * If a read fault occurs we try to upgrade the page protection
1455 	 * and make it also writable if possible.  There are three cases
1456 	 * where we cannot make the page mapping writable:
1457 	 *
1458 	 * (1) The mapping is read-only or the VM object is read-only,
1459 	 *     fs->prot above will simply not have VM_PROT_WRITE set.
1460 	 *
1461 	 * (2) If the mapping is a virtual page table fs->first_prot will
1462 	 *     have already been properly adjusted by vm_fault_vpagetable().
1463 	 *     to detect writes so we can set VPTE_M in the virtual page
1464 	 *     table.  Used by vkernels.
1465 	 *
1466 	 * (3) If the VM page is read-only or copy-on-write, upgrading would
1467 	 *     just result in an unnecessary COW fault.
1468 	 *
1469 	 * (4) If the pmap specifically requests A/M bit emulation, downgrade
1470 	 *     here.
1471 	 */
1472 #if 0
1473 	/* see vpagetable code */
1474 	if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1475 		if ((fault_type & VM_PROT_WRITE) == 0)
1476 			fs->prot &= ~VM_PROT_WRITE;
1477 	}
1478 #endif
1479 
1480 	if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
1481 	    pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
1482 		if ((fault_type & VM_PROT_WRITE) == 0)
1483 			fs->prot &= ~VM_PROT_WRITE;
1484 	}
1485 
1486 	/* vm_object_hold(fs->object); implied b/c object == first_object */
1487 
1488 	for (;;) {
1489 		/*
1490 		 * The entire backing chain from first_object to object
1491 		 * inclusive is chainlocked.
1492 		 *
1493 		 * If the object is dead, we stop here
1494 		 */
1495 		if (fs->object->flags & OBJ_DEAD) {
1496 			vm_object_pip_wakeup(fs->first_object);
1497 			vm_object_chain_release_all(fs->first_object,
1498 						    fs->object);
1499 			if (fs->object != fs->first_object)
1500 				vm_object_drop(fs->object);
1501 			unlock_and_deallocate(fs);
1502 			return (KERN_PROTECTION_FAILURE);
1503 		}
1504 
1505 		/*
1506 		 * See if the page is resident.  Wait/Retry if the page is
1507 		 * busy (lots of stuff may have changed so we can't continue
1508 		 * in that case).
1509 		 *
1510 		 * We can theoretically allow the soft-busy case on a read
1511 		 * fault if the page is marked valid, but since such
1512 		 * pages are typically already pmap'd, putting that
1513 		 * special case in might be more effort then it is
1514 		 * worth.  We cannot under any circumstances mess
1515 		 * around with a vm_page_t->busy page except, perhaps,
1516 		 * to pmap it.
1517 		 */
1518 		fs->m = vm_page_lookup_busy_try(fs->object, pindex,
1519 						TRUE, &error);
1520 		if (error) {
1521 			vm_object_pip_wakeup(fs->first_object);
1522 			vm_object_chain_release_all(fs->first_object,
1523 						    fs->object);
1524 			if (fs->object != fs->first_object)
1525 				vm_object_drop(fs->object);
1526 			unlock_things(fs);
1527 			vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1528 			mycpu->gd_cnt.v_intrans++;
1529 			/*vm_object_deallocate(fs->first_object);*/
1530 			/*fs->first_object = NULL;*/
1531 			fs->m = NULL;
1532 			return (KERN_TRY_AGAIN);
1533 		}
1534 		if (fs->m) {
1535 			/*
1536 			 * The page is busied for us.
1537 			 *
1538 			 * If reactivating a page from PQ_CACHE we may have
1539 			 * to rate-limit.
1540 			 */
1541 			int queue = fs->m->queue;
1542 			vm_page_unqueue_nowakeup(fs->m);
1543 
1544 			if ((queue - fs->m->pc) == PQ_CACHE &&
1545 			    vm_page_count_severe()) {
1546 				vm_page_activate(fs->m);
1547 				vm_page_wakeup(fs->m);
1548 				fs->m = NULL;
1549 				vm_object_pip_wakeup(fs->first_object);
1550 				vm_object_chain_release_all(fs->first_object,
1551 							    fs->object);
1552 				if (fs->object != fs->first_object)
1553 					vm_object_drop(fs->object);
1554 				unlock_and_deallocate(fs);
1555 				if (allow_nofault == 0 ||
1556 				    (curthread->td_flags & TDF_NOFAULT) == 0) {
1557 					thread_t td;
1558 
1559 					vm_wait_pfault();
1560 					td = curthread;
1561 					if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
1562 						return (KERN_PROTECTION_FAILURE);
1563 				}
1564 				return (KERN_TRY_AGAIN);
1565 			}
1566 
1567 			/*
1568 			 * If it still isn't completely valid (readable),
1569 			 * or if a read-ahead-mark is set on the VM page,
1570 			 * jump to readrest, else we found the page and
1571 			 * can return.
1572 			 *
1573 			 * We can release the spl once we have marked the
1574 			 * page busy.
1575 			 */
1576 			if (fs->m->object != &kernel_object) {
1577 				if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1578 				    VM_PAGE_BITS_ALL) {
1579 					goto readrest;
1580 				}
1581 				if (fs->m->flags & PG_RAM) {
1582 					if (debug_cluster)
1583 						kprintf("R");
1584 					vm_page_flag_clear(fs->m, PG_RAM);
1585 					goto readrest;
1586 				}
1587 			}
1588 			break; /* break to PAGE HAS BEEN FOUND */
1589 		}
1590 
1591 		/*
1592 		 * Page is not resident, If this is the search termination
1593 		 * or the pager might contain the page, allocate a new page.
1594 		 */
1595 		if (TRYPAGER(fs) || fs->object == fs->first_object) {
1596 			/*
1597 			 * Allocating, must be exclusive.
1598 			 */
1599 			if (fs->object == fs->first_object &&
1600 			    fs->first_shared) {
1601 				fs->first_shared = 0;
1602 				vm_object_pip_wakeup(fs->first_object);
1603 				vm_object_chain_release_all(fs->first_object,
1604 							    fs->object);
1605 				if (fs->object != fs->first_object)
1606 					vm_object_drop(fs->object);
1607 				unlock_and_deallocate(fs);
1608 				return (KERN_TRY_AGAIN);
1609 			}
1610 			if (fs->object != fs->first_object &&
1611 			    fs->shared) {
1612 				fs->first_shared = 0;
1613 				fs->shared = 0;
1614 				vm_object_pip_wakeup(fs->first_object);
1615 				vm_object_chain_release_all(fs->first_object,
1616 							    fs->object);
1617 				if (fs->object != fs->first_object)
1618 					vm_object_drop(fs->object);
1619 				unlock_and_deallocate(fs);
1620 				return (KERN_TRY_AGAIN);
1621 			}
1622 
1623 			/*
1624 			 * If the page is beyond the object size we fail
1625 			 */
1626 			if (pindex >= fs->object->size) {
1627 				vm_object_pip_wakeup(fs->first_object);
1628 				vm_object_chain_release_all(fs->first_object,
1629 							    fs->object);
1630 				if (fs->object != fs->first_object)
1631 					vm_object_drop(fs->object);
1632 				unlock_and_deallocate(fs);
1633 				return (KERN_PROTECTION_FAILURE);
1634 			}
1635 
1636 			/*
1637 			 * Allocate a new page for this object/offset pair.
1638 			 *
1639 			 * It is possible for the allocation to race, so
1640 			 * handle the case.
1641 			 */
1642 			fs->m = NULL;
1643 			if (!vm_page_count_severe()) {
1644 				fs->m = vm_page_alloc(fs->object, pindex,
1645 				    ((fs->vp || fs->object->backing_object) ?
1646 					VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL :
1647 					VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1648 					VM_ALLOC_USE_GD | VM_ALLOC_ZERO));
1649 			}
1650 			if (fs->m == NULL) {
1651 				vm_object_pip_wakeup(fs->first_object);
1652 				vm_object_chain_release_all(fs->first_object,
1653 							    fs->object);
1654 				if (fs->object != fs->first_object)
1655 					vm_object_drop(fs->object);
1656 				unlock_and_deallocate(fs);
1657 				if (allow_nofault == 0 ||
1658 				    (curthread->td_flags & TDF_NOFAULT) == 0) {
1659 					thread_t td;
1660 
1661 					vm_wait_pfault();
1662 					td = curthread;
1663 					if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
1664 						return (KERN_PROTECTION_FAILURE);
1665 				}
1666 				return (KERN_TRY_AGAIN);
1667 			}
1668 
1669 			/*
1670 			 * Fall through to readrest.  We have a new page which
1671 			 * will have to be paged (since m->valid will be 0).
1672 			 */
1673 		}
1674 
1675 readrest:
1676 		/*
1677 		 * We have found an invalid or partially valid page, a
1678 		 * page with a read-ahead mark which might be partially or
1679 		 * fully valid (and maybe dirty too), or we have allocated
1680 		 * a new page.
1681 		 *
1682 		 * Attempt to fault-in the page if there is a chance that the
1683 		 * pager has it, and potentially fault in additional pages
1684 		 * at the same time.
1685 		 *
1686 		 * If TRYPAGER is true then fs.m will be non-NULL and busied
1687 		 * for us.
1688 		 */
1689 		if (TRYPAGER(fs)) {
1690 			int rv;
1691 			int seqaccess;
1692 			u_char behavior = vm_map_entry_behavior(fs->entry);
1693 
1694 			if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1695 				seqaccess = 0;
1696 			else
1697 				seqaccess = -1;
1698 
1699 			/*
1700 			 * Doing I/O may synchronously insert additional
1701 			 * pages so we can't be shared at this point either.
1702 			 *
1703 			 * NOTE: We can't free fs->m here in the allocated
1704 			 *	 case (fs->object != fs->first_object) as
1705 			 *	 this would require an exclusively locked
1706 			 *	 VM object.
1707 			 */
1708 			if (fs->object == fs->first_object &&
1709 			    fs->first_shared) {
1710 				vm_page_deactivate(fs->m);
1711 				vm_page_wakeup(fs->m);
1712 				fs->m = NULL;
1713 				fs->first_shared = 0;
1714 				vm_object_pip_wakeup(fs->first_object);
1715 				vm_object_chain_release_all(fs->first_object,
1716 							    fs->object);
1717 				if (fs->object != fs->first_object)
1718 					vm_object_drop(fs->object);
1719 				unlock_and_deallocate(fs);
1720 				return (KERN_TRY_AGAIN);
1721 			}
1722 			if (fs->object != fs->first_object &&
1723 			    fs->shared) {
1724 				vm_page_deactivate(fs->m);
1725 				vm_page_wakeup(fs->m);
1726 				fs->m = NULL;
1727 				fs->first_shared = 0;
1728 				fs->shared = 0;
1729 				vm_object_pip_wakeup(fs->first_object);
1730 				vm_object_chain_release_all(fs->first_object,
1731 							    fs->object);
1732 				if (fs->object != fs->first_object)
1733 					vm_object_drop(fs->object);
1734 				unlock_and_deallocate(fs);
1735 				return (KERN_TRY_AGAIN);
1736 			}
1737 
1738 			/*
1739 			 * Avoid deadlocking against the map when doing I/O.
1740 			 * fs.object and the page is PG_BUSY'd.
1741 			 *
1742 			 * NOTE: Once unlocked, fs->entry can become stale
1743 			 *	 so this will NULL it out.
1744 			 *
1745 			 * NOTE: fs->entry is invalid until we relock the
1746 			 *	 map and verify that the timestamp has not
1747 			 *	 changed.
1748 			 */
1749 			unlock_map(fs);
1750 
1751 			/*
1752 			 * Acquire the page data.  We still hold a ref on
1753 			 * fs.object and the page has been PG_BUSY's.
1754 			 *
1755 			 * The pager may replace the page (for example, in
1756 			 * order to enter a fictitious page into the
1757 			 * object).  If it does so it is responsible for
1758 			 * cleaning up the passed page and properly setting
1759 			 * the new page PG_BUSY.
1760 			 *
1761 			 * If we got here through a PG_RAM read-ahead
1762 			 * mark the page may be partially dirty and thus
1763 			 * not freeable.  Don't bother checking to see
1764 			 * if the pager has the page because we can't free
1765 			 * it anyway.  We have to depend on the get_page
1766 			 * operation filling in any gaps whether there is
1767 			 * backing store or not.
1768 			 */
1769 			rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1770 
1771 			if (rv == VM_PAGER_OK) {
1772 				/*
1773 				 * Relookup in case pager changed page. Pager
1774 				 * is responsible for disposition of old page
1775 				 * if moved.
1776 				 *
1777 				 * XXX other code segments do relookups too.
1778 				 * It's a bad abstraction that needs to be
1779 				 * fixed/removed.
1780 				 */
1781 				fs->m = vm_page_lookup(fs->object, pindex);
1782 				if (fs->m == NULL) {
1783 					vm_object_pip_wakeup(fs->first_object);
1784 					vm_object_chain_release_all(
1785 						fs->first_object, fs->object);
1786 					if (fs->object != fs->first_object)
1787 						vm_object_drop(fs->object);
1788 					unlock_and_deallocate(fs);
1789 					return (KERN_TRY_AGAIN);
1790 				}
1791 				++fs->hardfault;
1792 				break; /* break to PAGE HAS BEEN FOUND */
1793 			}
1794 
1795 			/*
1796 			 * Remove the bogus page (which does not exist at this
1797 			 * object/offset); before doing so, we must get back
1798 			 * our object lock to preserve our invariant.
1799 			 *
1800 			 * Also wake up any other process that may want to bring
1801 			 * in this page.
1802 			 *
1803 			 * If this is the top-level object, we must leave the
1804 			 * busy page to prevent another process from rushing
1805 			 * past us, and inserting the page in that object at
1806 			 * the same time that we are.
1807 			 */
1808 			if (rv == VM_PAGER_ERROR) {
1809 				if (curproc) {
1810 					kprintf("vm_fault: pager read error, "
1811 						"pid %d (%s)\n",
1812 						curproc->p_pid,
1813 						curproc->p_comm);
1814 				} else {
1815 					kprintf("vm_fault: pager read error, "
1816 						"thread %p (%s)\n",
1817 						curthread,
1818 						curproc->p_comm);
1819 				}
1820 			}
1821 
1822 			/*
1823 			 * Data outside the range of the pager or an I/O error
1824 			 *
1825 			 * The page may have been wired during the pagein,
1826 			 * e.g. by the buffer cache, and cannot simply be
1827 			 * freed.  Call vnode_pager_freepage() to deal with it.
1828 			 *
1829 			 * Also note that we cannot free the page if we are
1830 			 * holding the related object shared. XXX not sure
1831 			 * what to do in that case.
1832 			 */
1833 			if (fs->object != fs->first_object) {
1834 				vnode_pager_freepage(fs->m);
1835 				fs->m = NULL;
1836 				/*
1837 				 * XXX - we cannot just fall out at this
1838 				 * point, m has been freed and is invalid!
1839 				 */
1840 			}
1841 			/*
1842 			 * XXX - the check for kernel_map is a kludge to work
1843 			 * around having the machine panic on a kernel space
1844 			 * fault w/ I/O error.
1845 			 */
1846 			if (((fs->map != &kernel_map) &&
1847 			    (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1848 				if (fs->m) {
1849 					if (fs->first_shared) {
1850 						vm_page_deactivate(fs->m);
1851 						vm_page_wakeup(fs->m);
1852 					} else {
1853 						vnode_pager_freepage(fs->m);
1854 					}
1855 					fs->m = NULL;
1856 				}
1857 				vm_object_pip_wakeup(fs->first_object);
1858 				vm_object_chain_release_all(fs->first_object,
1859 							    fs->object);
1860 				if (fs->object != fs->first_object)
1861 					vm_object_drop(fs->object);
1862 				unlock_and_deallocate(fs);
1863 				if (rv == VM_PAGER_ERROR)
1864 					return (KERN_FAILURE);
1865 				else
1866 					return (KERN_PROTECTION_FAILURE);
1867 				/* NOT REACHED */
1868 			}
1869 		}
1870 
1871 		/*
1872 		 * We get here if the object has a default pager (or unwiring)
1873 		 * or the pager doesn't have the page.
1874 		 *
1875 		 * fs->first_m will be used for the COW unless we find a
1876 		 * deeper page to be mapped read-only, in which case the
1877 		 * unlock*(fs) will free first_m.
1878 		 */
1879 		if (fs->object == fs->first_object)
1880 			fs->first_m = fs->m;
1881 
1882 		/*
1883 		 * Move on to the next object.  The chain lock should prevent
1884 		 * the backing_object from getting ripped out from under us.
1885 		 *
1886 		 * The object lock for the next object is governed by
1887 		 * fs->shared.
1888 		 */
1889 		if ((next_object = fs->object->backing_object) != NULL) {
1890 			if (fs->shared)
1891 				vm_object_hold_shared(next_object);
1892 			else
1893 				vm_object_hold(next_object);
1894 			vm_object_chain_acquire(next_object, fs->shared);
1895 			KKASSERT(next_object == fs->object->backing_object);
1896 			pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1897 		}
1898 
1899 		if (next_object == NULL) {
1900 			/*
1901 			 * If there's no object left, fill the page in the top
1902 			 * object with zeros.
1903 			 */
1904 			if (fs->object != fs->first_object) {
1905 #if 0
1906 				if (fs->first_object->backing_object !=
1907 				    fs->object) {
1908 					vm_object_hold(fs->first_object->backing_object);
1909 				}
1910 #endif
1911 				vm_object_chain_release_all(
1912 					fs->first_object->backing_object,
1913 					fs->object);
1914 #if 0
1915 				if (fs->first_object->backing_object !=
1916 				    fs->object) {
1917 					vm_object_drop(fs->first_object->backing_object);
1918 				}
1919 #endif
1920 				vm_object_pip_wakeup(fs->object);
1921 				vm_object_drop(fs->object);
1922 				fs->object = fs->first_object;
1923 				pindex = first_pindex;
1924 				fs->m = fs->first_m;
1925 			}
1926 			fs->first_m = NULL;
1927 
1928 			/*
1929 			 * Zero the page and mark it valid.
1930 			 */
1931 			vm_page_zero_fill(fs->m);
1932 			mycpu->gd_cnt.v_zfod++;
1933 			fs->m->valid = VM_PAGE_BITS_ALL;
1934 			break;	/* break to PAGE HAS BEEN FOUND */
1935 		}
1936 		if (fs->object != fs->first_object) {
1937 			vm_object_pip_wakeup(fs->object);
1938 			vm_object_lock_swap();
1939 			vm_object_drop(fs->object);
1940 		}
1941 		KASSERT(fs->object != next_object,
1942 			("object loop %p", next_object));
1943 		fs->object = next_object;
1944 		vm_object_pip_add(fs->object, 1);
1945 	}
1946 
1947 	/*
1948 	 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1949 	 * is held.]
1950 	 *
1951 	 * object still held.
1952 	 *
1953 	 * local shared variable may be different from fs->shared.
1954 	 *
1955 	 * If the page is being written, but isn't already owned by the
1956 	 * top-level object, we have to copy it into a new page owned by the
1957 	 * top-level object.
1958 	 */
1959 	KASSERT((fs->m->flags & PG_BUSY) != 0,
1960 		("vm_fault: not busy after main loop"));
1961 
1962 	if (fs->object != fs->first_object) {
1963 		/*
1964 		 * We only really need to copy if we want to write it.
1965 		 */
1966 		if (fault_type & VM_PROT_WRITE) {
1967 			/*
1968 			 * This allows pages to be virtually copied from a
1969 			 * backing_object into the first_object, where the
1970 			 * backing object has no other refs to it, and cannot
1971 			 * gain any more refs.  Instead of a bcopy, we just
1972 			 * move the page from the backing object to the
1973 			 * first object.  Note that we must mark the page
1974 			 * dirty in the first object so that it will go out
1975 			 * to swap when needed.
1976 			 */
1977 			if (
1978 				/*
1979 				 * Must be holding exclusive locks
1980 				 */
1981 				fs->first_shared == 0 &&
1982 				fs->shared == 0 &&
1983 				/*
1984 				 * Map, if present, has not changed
1985 				 */
1986 				(fs->map == NULL ||
1987 				fs->map_generation == fs->map->timestamp) &&
1988 				/*
1989 				 * Only one shadow object
1990 				 */
1991 				(fs->object->shadow_count == 1) &&
1992 				/*
1993 				 * No COW refs, except us
1994 				 */
1995 				(fs->object->ref_count == 1) &&
1996 				/*
1997 				 * No one else can look this object up
1998 				 */
1999 				(fs->object->handle == NULL) &&
2000 				/*
2001 				 * No other ways to look the object up
2002 				 */
2003 				((fs->object->type == OBJT_DEFAULT) ||
2004 				 (fs->object->type == OBJT_SWAP)) &&
2005 				/*
2006 				 * We don't chase down the shadow chain
2007 				 */
2008 				(fs->object == fs->first_object->backing_object) &&
2009 
2010 				/*
2011 				 * grab the lock if we need to
2012 				 */
2013 				(fs->lookup_still_valid ||
2014 				 fs->map == NULL ||
2015 				 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
2016 			    ) {
2017 				/*
2018 				 * (first_m) and (m) are both busied.  We have
2019 				 * move (m) into (first_m)'s object/pindex
2020 				 * in an atomic fashion, then free (first_m).
2021 				 *
2022 				 * first_object is held so second remove
2023 				 * followed by the rename should wind
2024 				 * up being atomic.  vm_page_free() might
2025 				 * block so we don't do it until after the
2026 				 * rename.
2027 				 */
2028 				fs->lookup_still_valid = 1;
2029 				vm_page_protect(fs->first_m, VM_PROT_NONE);
2030 				vm_page_remove(fs->first_m);
2031 				vm_page_rename(fs->m, fs->first_object,
2032 					       first_pindex);
2033 				vm_page_free(fs->first_m);
2034 				fs->first_m = fs->m;
2035 				fs->m = NULL;
2036 				mycpu->gd_cnt.v_cow_optim++;
2037 			} else {
2038 				/*
2039 				 * Oh, well, lets copy it.
2040 				 *
2041 				 * Why are we unmapping the original page
2042 				 * here?  Well, in short, not all accessors
2043 				 * of user memory go through the pmap.  The
2044 				 * procfs code doesn't have access user memory
2045 				 * via a local pmap, so vm_fault_page*()
2046 				 * can't call pmap_enter().  And the umtx*()
2047 				 * code may modify the COW'd page via a DMAP
2048 				 * or kernel mapping and not via the pmap,
2049 				 * leaving the original page still mapped
2050 				 * read-only into the pmap.
2051 				 *
2052 				 * So we have to remove the page from at
2053 				 * least the current pmap if it is in it.
2054 				 * Just remove it from all pmaps.
2055 				 */
2056 				KKASSERT(fs->first_shared == 0);
2057 				vm_page_copy(fs->m, fs->first_m);
2058 				vm_page_protect(fs->m, VM_PROT_NONE);
2059 				vm_page_event(fs->m, VMEVENT_COW);
2060 			}
2061 
2062 			/*
2063 			 * We no longer need the old page or object.
2064 			 */
2065 			if (fs->m)
2066 				release_page(fs);
2067 
2068 			/*
2069 			 * We intend to revert to first_object, undo the
2070 			 * chain lock through to that.
2071 			 */
2072 #if 0
2073 			if (fs->first_object->backing_object != fs->object)
2074 				vm_object_hold(fs->first_object->backing_object);
2075 #endif
2076 			vm_object_chain_release_all(
2077 					fs->first_object->backing_object,
2078 					fs->object);
2079 #if 0
2080 			if (fs->first_object->backing_object != fs->object)
2081 				vm_object_drop(fs->first_object->backing_object);
2082 #endif
2083 
2084 			/*
2085 			 * fs->object != fs->first_object due to above
2086 			 * conditional
2087 			 */
2088 			vm_object_pip_wakeup(fs->object);
2089 			vm_object_drop(fs->object);
2090 
2091 			/*
2092 			 * Only use the new page below...
2093 			 */
2094 			mycpu->gd_cnt.v_cow_faults++;
2095 			fs->m = fs->first_m;
2096 			fs->object = fs->first_object;
2097 			pindex = first_pindex;
2098 		} else {
2099 			/*
2100 			 * If it wasn't a write fault avoid having to copy
2101 			 * the page by mapping it read-only.
2102 			 */
2103 			fs->prot &= ~VM_PROT_WRITE;
2104 		}
2105 	}
2106 
2107 	/*
2108 	 * Relock the map if necessary, then check the generation count.
2109 	 * relock_map() will update fs->timestamp to account for the
2110 	 * relocking if necessary.
2111 	 *
2112 	 * If the count has changed after relocking then all sorts of
2113 	 * crap may have happened and we have to retry.
2114 	 *
2115 	 * NOTE: The relock_map() can fail due to a deadlock against
2116 	 *	 the vm_page we are holding BUSY.
2117 	 */
2118 	if (fs->lookup_still_valid == FALSE && fs->map) {
2119 		if (relock_map(fs) ||
2120 		    fs->map->timestamp != fs->map_generation) {
2121 			release_page(fs);
2122 			vm_object_pip_wakeup(fs->first_object);
2123 			vm_object_chain_release_all(fs->first_object,
2124 						    fs->object);
2125 			if (fs->object != fs->first_object)
2126 				vm_object_drop(fs->object);
2127 			unlock_and_deallocate(fs);
2128 			return (KERN_TRY_AGAIN);
2129 		}
2130 	}
2131 
2132 	/*
2133 	 * If the fault is a write, we know that this page is being
2134 	 * written NOW so dirty it explicitly to save on pmap_is_modified()
2135 	 * calls later.
2136 	 *
2137 	 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
2138 	 * if the page is already dirty to prevent data written with
2139 	 * the expectation of being synced from not being synced.
2140 	 * Likewise if this entry does not request NOSYNC then make
2141 	 * sure the page isn't marked NOSYNC.  Applications sharing
2142 	 * data should use the same flags to avoid ping ponging.
2143 	 *
2144 	 * Also tell the backing pager, if any, that it should remove
2145 	 * any swap backing since the page is now dirty.
2146 	 */
2147 	vm_page_activate(fs->m);
2148 	if (fs->prot & VM_PROT_WRITE) {
2149 		vm_object_set_writeable_dirty(fs->m->object);
2150 		vm_set_nosync(fs->m, fs->entry);
2151 		if (fs->fault_flags & VM_FAULT_DIRTY) {
2152 			vm_page_dirty(fs->m);
2153 			if (fs->m->flags & PG_SWAPPED) {
2154 				/*
2155 				 * If the page is swapped out we have to call
2156 				 * swap_pager_unswapped() which requires an
2157 				 * exclusive object lock.  If we are shared,
2158 				 * we must clear the shared flag and retry.
2159 				 */
2160 				if ((fs->object == fs->first_object &&
2161 				     fs->first_shared) ||
2162 				    (fs->object != fs->first_object &&
2163 				     fs->shared)) {
2164 					vm_page_wakeup(fs->m);
2165 					fs->m = NULL;
2166 					if (fs->object == fs->first_object)
2167 						fs->first_shared = 0;
2168 					else
2169 						fs->shared = 0;
2170 					vm_object_pip_wakeup(fs->first_object);
2171 					vm_object_chain_release_all(
2172 						fs->first_object, fs->object);
2173 					if (fs->object != fs->first_object)
2174 						vm_object_drop(fs->object);
2175 					unlock_and_deallocate(fs);
2176 					return (KERN_TRY_AGAIN);
2177 				}
2178 				swap_pager_unswapped(fs->m);
2179 			}
2180 		}
2181 	}
2182 
2183 	vm_object_pip_wakeup(fs->first_object);
2184 	vm_object_chain_release_all(fs->first_object, fs->object);
2185 	if (fs->object != fs->first_object)
2186 		vm_object_drop(fs->object);
2187 
2188 	/*
2189 	 * Page had better still be busy.  We are still locked up and
2190 	 * fs->object will have another PIP reference if it is not equal
2191 	 * to fs->first_object.
2192 	 */
2193 	KASSERT(fs->m->flags & PG_BUSY,
2194 		("vm_fault: page %p not busy!", fs->m));
2195 
2196 	/*
2197 	 * Sanity check: page must be completely valid or it is not fit to
2198 	 * map into user space.  vm_pager_get_pages() ensures this.
2199 	 */
2200 	if (fs->m->valid != VM_PAGE_BITS_ALL) {
2201 		vm_page_zero_invalid(fs->m, TRUE);
2202 		kprintf("Warning: page %p partially invalid on fault\n", fs->m);
2203 	}
2204 
2205 	return (KERN_SUCCESS);
2206 }
2207 
2208 /*
2209  * Wire down a range of virtual addresses in a map.  The entry in question
2210  * should be marked in-transition and the map must be locked.  We must
2211  * release the map temporarily while faulting-in the page to avoid a
2212  * deadlock.  Note that the entry may be clipped while we are blocked but
2213  * will never be freed.
2214  *
2215  * No requirements.
2216  */
2217 int
2218 vm_fault_wire(vm_map_t map, vm_map_entry_t entry,
2219 	      boolean_t user_wire, int kmflags)
2220 {
2221 	boolean_t fictitious;
2222 	vm_offset_t start;
2223 	vm_offset_t end;
2224 	vm_offset_t va;
2225 	pmap_t pmap;
2226 	int rv;
2227 	int wire_prot;
2228 	int fault_flags;
2229 	vm_page_t m;
2230 
2231 	if (user_wire) {
2232 		wire_prot = VM_PROT_READ;
2233 		fault_flags = VM_FAULT_USER_WIRE;
2234 	} else {
2235 		wire_prot = VM_PROT_READ | VM_PROT_WRITE;
2236 		fault_flags = VM_FAULT_CHANGE_WIRING;
2237 	}
2238 	if (kmflags & KM_NOTLBSYNC)
2239 		wire_prot |= VM_PROT_NOSYNC;
2240 
2241 	pmap = vm_map_pmap(map);
2242 	start = entry->start;
2243 	end = entry->end;
2244 
2245 	switch(entry->maptype) {
2246 	case VM_MAPTYPE_NORMAL:
2247 	case VM_MAPTYPE_VPAGETABLE:
2248 		fictitious = entry->object.vm_object &&
2249 			    ((entry->object.vm_object->type == OBJT_DEVICE) ||
2250 			     (entry->object.vm_object->type == OBJT_MGTDEVICE));
2251 		break;
2252 	case VM_MAPTYPE_UKSMAP:
2253 		fictitious = TRUE;
2254 		break;
2255 	default:
2256 		fictitious = FALSE;
2257 		break;
2258 	}
2259 
2260 	if (entry->eflags & MAP_ENTRY_KSTACK)
2261 		start += PAGE_SIZE;
2262 	map->timestamp++;
2263 	vm_map_unlock(map);
2264 
2265 	/*
2266 	 * We simulate a fault to get the page and enter it in the physical
2267 	 * map.
2268 	 */
2269 	for (va = start; va < end; va += PAGE_SIZE) {
2270 		rv = vm_fault(map, va, wire_prot, fault_flags);
2271 		if (rv) {
2272 			while (va > start) {
2273 				va -= PAGE_SIZE;
2274 				m = pmap_unwire(pmap, va);
2275 				if (m && !fictitious) {
2276 					vm_page_busy_wait(m, FALSE, "vmwrpg");
2277 					vm_page_unwire(m, 1);
2278 					vm_page_wakeup(m);
2279 				}
2280 			}
2281 			goto done;
2282 		}
2283 	}
2284 	rv = KERN_SUCCESS;
2285 done:
2286 	vm_map_lock(map);
2287 
2288 	return (rv);
2289 }
2290 
2291 /*
2292  * Unwire a range of virtual addresses in a map.  The map should be
2293  * locked.
2294  */
2295 void
2296 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
2297 {
2298 	boolean_t fictitious;
2299 	vm_offset_t start;
2300 	vm_offset_t end;
2301 	vm_offset_t va;
2302 	pmap_t pmap;
2303 	vm_page_t m;
2304 
2305 	pmap = vm_map_pmap(map);
2306 	start = entry->start;
2307 	end = entry->end;
2308 	fictitious = entry->object.vm_object &&
2309 			((entry->object.vm_object->type == OBJT_DEVICE) ||
2310 			 (entry->object.vm_object->type == OBJT_MGTDEVICE));
2311 	if (entry->eflags & MAP_ENTRY_KSTACK)
2312 		start += PAGE_SIZE;
2313 
2314 	/*
2315 	 * Since the pages are wired down, we must be able to get their
2316 	 * mappings from the physical map system.
2317 	 */
2318 	for (va = start; va < end; va += PAGE_SIZE) {
2319 		m = pmap_unwire(pmap, va);
2320 		if (m && !fictitious) {
2321 			vm_page_busy_wait(m, FALSE, "vmwrpg");
2322 			vm_page_unwire(m, 1);
2323 			vm_page_wakeup(m);
2324 		}
2325 	}
2326 }
2327 
2328 /*
2329  * Copy all of the pages from a wired-down map entry to another.
2330  *
2331  * The source and destination maps must be locked for write.
2332  * The source and destination maps token must be held
2333  * The source map entry must be wired down (or be a sharing map
2334  * entry corresponding to a main map entry that is wired down).
2335  *
2336  * No other requirements.
2337  *
2338  * XXX do segment optimization
2339  */
2340 void
2341 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
2342 		    vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
2343 {
2344 	vm_object_t dst_object;
2345 	vm_object_t src_object;
2346 	vm_ooffset_t dst_offset;
2347 	vm_ooffset_t src_offset;
2348 	vm_prot_t prot;
2349 	vm_offset_t vaddr;
2350 	vm_page_t dst_m;
2351 	vm_page_t src_m;
2352 
2353 	src_object = src_entry->object.vm_object;
2354 	src_offset = src_entry->offset;
2355 
2356 	/*
2357 	 * Create the top-level object for the destination entry. (Doesn't
2358 	 * actually shadow anything - we copy the pages directly.)
2359 	 */
2360 	vm_map_entry_allocate_object(dst_entry);
2361 	dst_object = dst_entry->object.vm_object;
2362 
2363 	prot = dst_entry->max_protection;
2364 
2365 	/*
2366 	 * Loop through all of the pages in the entry's range, copying each
2367 	 * one from the source object (it should be there) to the destination
2368 	 * object.
2369 	 */
2370 	vm_object_hold(src_object);
2371 	vm_object_hold(dst_object);
2372 	for (vaddr = dst_entry->start, dst_offset = 0;
2373 	     vaddr < dst_entry->end;
2374 	     vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
2375 
2376 		/*
2377 		 * Allocate a page in the destination object
2378 		 */
2379 		do {
2380 			dst_m = vm_page_alloc(dst_object,
2381 					      OFF_TO_IDX(dst_offset),
2382 					      VM_ALLOC_NORMAL);
2383 			if (dst_m == NULL) {
2384 				vm_wait(0);
2385 			}
2386 		} while (dst_m == NULL);
2387 
2388 		/*
2389 		 * Find the page in the source object, and copy it in.
2390 		 * (Because the source is wired down, the page will be in
2391 		 * memory.)
2392 		 */
2393 		src_m = vm_page_lookup(src_object,
2394 				       OFF_TO_IDX(dst_offset + src_offset));
2395 		if (src_m == NULL)
2396 			panic("vm_fault_copy_wired: page missing");
2397 
2398 		vm_page_copy(src_m, dst_m);
2399 		vm_page_event(src_m, VMEVENT_COW);
2400 
2401 		/*
2402 		 * Enter it in the pmap...
2403 		 */
2404 		pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry);
2405 
2406 		/*
2407 		 * Mark it no longer busy, and put it on the active list.
2408 		 */
2409 		vm_page_activate(dst_m);
2410 		vm_page_wakeup(dst_m);
2411 	}
2412 	vm_object_drop(dst_object);
2413 	vm_object_drop(src_object);
2414 }
2415 
2416 #if 0
2417 
2418 /*
2419  * This routine checks around the requested page for other pages that
2420  * might be able to be faulted in.  This routine brackets the viable
2421  * pages for the pages to be paged in.
2422  *
2423  * Inputs:
2424  *	m, rbehind, rahead
2425  *
2426  * Outputs:
2427  *  marray (array of vm_page_t), reqpage (index of requested page)
2428  *
2429  * Return value:
2430  *  number of pages in marray
2431  */
2432 static int
2433 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
2434 			  vm_page_t *marray, int *reqpage)
2435 {
2436 	int i,j;
2437 	vm_object_t object;
2438 	vm_pindex_t pindex, startpindex, endpindex, tpindex;
2439 	vm_page_t rtm;
2440 	int cbehind, cahead;
2441 
2442 	object = m->object;
2443 	pindex = m->pindex;
2444 
2445 	/*
2446 	 * we don't fault-ahead for device pager
2447 	 */
2448 	if ((object->type == OBJT_DEVICE) ||
2449 	    (object->type == OBJT_MGTDEVICE)) {
2450 		*reqpage = 0;
2451 		marray[0] = m;
2452 		return 1;
2453 	}
2454 
2455 	/*
2456 	 * if the requested page is not available, then give up now
2457 	 */
2458 	if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
2459 		*reqpage = 0;	/* not used by caller, fix compiler warn */
2460 		return 0;
2461 	}
2462 
2463 	if ((cbehind == 0) && (cahead == 0)) {
2464 		*reqpage = 0;
2465 		marray[0] = m;
2466 		return 1;
2467 	}
2468 
2469 	if (rahead > cahead) {
2470 		rahead = cahead;
2471 	}
2472 
2473 	if (rbehind > cbehind) {
2474 		rbehind = cbehind;
2475 	}
2476 
2477 	/*
2478 	 * Do not do any readahead if we have insufficient free memory.
2479 	 *
2480 	 * XXX code was broken disabled before and has instability
2481 	 * with this conditonal fixed, so shortcut for now.
2482 	 */
2483 	if (burst_fault == 0 || vm_page_count_severe()) {
2484 		marray[0] = m;
2485 		*reqpage = 0;
2486 		return 1;
2487 	}
2488 
2489 	/*
2490 	 * scan backward for the read behind pages -- in memory
2491 	 *
2492 	 * Assume that if the page is not found an interrupt will not
2493 	 * create it.  Theoretically interrupts can only remove (busy)
2494 	 * pages, not create new associations.
2495 	 */
2496 	if (pindex > 0) {
2497 		if (rbehind > pindex) {
2498 			rbehind = pindex;
2499 			startpindex = 0;
2500 		} else {
2501 			startpindex = pindex - rbehind;
2502 		}
2503 
2504 		vm_object_hold(object);
2505 		for (tpindex = pindex; tpindex > startpindex; --tpindex) {
2506 			if (vm_page_lookup(object, tpindex - 1))
2507 				break;
2508 		}
2509 
2510 		i = 0;
2511 		while (tpindex < pindex) {
2512 			rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2513 							     VM_ALLOC_NULL_OK);
2514 			if (rtm == NULL) {
2515 				for (j = 0; j < i; j++) {
2516 					vm_page_free(marray[j]);
2517 				}
2518 				vm_object_drop(object);
2519 				marray[0] = m;
2520 				*reqpage = 0;
2521 				return 1;
2522 			}
2523 			marray[i] = rtm;
2524 			++i;
2525 			++tpindex;
2526 		}
2527 		vm_object_drop(object);
2528 	} else {
2529 		i = 0;
2530 	}
2531 
2532 	/*
2533 	 * Assign requested page
2534 	 */
2535 	marray[i] = m;
2536 	*reqpage = i;
2537 	++i;
2538 
2539 	/*
2540 	 * Scan forwards for read-ahead pages
2541 	 */
2542 	tpindex = pindex + 1;
2543 	endpindex = tpindex + rahead;
2544 	if (endpindex > object->size)
2545 		endpindex = object->size;
2546 
2547 	vm_object_hold(object);
2548 	while (tpindex < endpindex) {
2549 		if (vm_page_lookup(object, tpindex))
2550 			break;
2551 		rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2552 						     VM_ALLOC_NULL_OK);
2553 		if (rtm == NULL)
2554 			break;
2555 		marray[i] = rtm;
2556 		++i;
2557 		++tpindex;
2558 	}
2559 	vm_object_drop(object);
2560 
2561 	return (i);
2562 }
2563 
2564 #endif
2565 
2566 /*
2567  * vm_prefault() provides a quick way of clustering pagefaults into a
2568  * processes address space.  It is a "cousin" of pmap_object_init_pt,
2569  * except it runs at page fault time instead of mmap time.
2570  *
2571  * vm.fast_fault	Enables pre-faulting zero-fill pages
2572  *
2573  * vm.prefault_pages	Number of pages (1/2 negative, 1/2 positive) to
2574  *			prefault.  Scan stops in either direction when
2575  *			a page is found to already exist.
2576  *
2577  * This code used to be per-platform pmap_prefault().  It is now
2578  * machine-independent and enhanced to also pre-fault zero-fill pages
2579  * (see vm.fast_fault) as well as make them writable, which greatly
2580  * reduces the number of page faults programs incur.
2581  *
2582  * Application performance when pre-faulting zero-fill pages is heavily
2583  * dependent on the application.  Very tiny applications like /bin/echo
2584  * lose a little performance while applications of any appreciable size
2585  * gain performance.  Prefaulting multiple pages also reduces SMP
2586  * congestion and can improve SMP performance significantly.
2587  *
2588  * NOTE!  prot may allow writing but this only applies to the top level
2589  *	  object.  If we wind up mapping a page extracted from a backing
2590  *	  object we have to make sure it is read-only.
2591  *
2592  * NOTE!  The caller has already handled any COW operations on the
2593  *	  vm_map_entry via the normal fault code.  Do NOT call this
2594  *	  shortcut unless the normal fault code has run on this entry.
2595  *
2596  * The related map must be locked.
2597  * No other requirements.
2598  */
2599 static int vm_prefault_pages = 8;
2600 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0,
2601 	   "Maximum number of pages to pre-fault");
2602 static int vm_fast_fault = 1;
2603 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
2604 	   "Burst fault zero-fill regions");
2605 
2606 /*
2607  * Set PG_NOSYNC if the map entry indicates so, but only if the page
2608  * is not already dirty by other means.  This will prevent passive
2609  * filesystem syncing as well as 'sync' from writing out the page.
2610  */
2611 static void
2612 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
2613 {
2614 	if (entry->eflags & MAP_ENTRY_NOSYNC) {
2615 		if (m->dirty == 0)
2616 			vm_page_flag_set(m, PG_NOSYNC);
2617 	} else {
2618 		vm_page_flag_clear(m, PG_NOSYNC);
2619 	}
2620 }
2621 
2622 static void
2623 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot,
2624 	    int fault_flags)
2625 {
2626 	struct lwp *lp;
2627 	vm_page_t m;
2628 	vm_offset_t addr;
2629 	vm_pindex_t index;
2630 	vm_pindex_t pindex;
2631 	vm_object_t object;
2632 	int pprot;
2633 	int i;
2634 	int noneg;
2635 	int nopos;
2636 	int maxpages;
2637 
2638 	/*
2639 	 * Get stable max count value, disabled if set to 0
2640 	 */
2641 	maxpages = vm_prefault_pages;
2642 	cpu_ccfence();
2643 	if (maxpages <= 0)
2644 		return;
2645 
2646 	/*
2647 	 * We do not currently prefault mappings that use virtual page
2648 	 * tables.  We do not prefault foreign pmaps.
2649 	 */
2650 	if (entry->maptype != VM_MAPTYPE_NORMAL)
2651 		return;
2652 	lp = curthread->td_lwp;
2653 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2654 		return;
2655 
2656 	/*
2657 	 * Limit pre-fault count to 1024 pages.
2658 	 */
2659 	if (maxpages > 1024)
2660 		maxpages = 1024;
2661 
2662 	object = entry->object.vm_object;
2663 	KKASSERT(object != NULL);
2664 	KKASSERT(object == entry->object.vm_object);
2665 
2666 	/*
2667 	 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively
2668 	 *	 now (or do something more complex XXX).
2669 	 */
2670 	vm_object_hold(object);
2671 	vm_object_chain_acquire(object, 0);
2672 
2673 	noneg = 0;
2674 	nopos = 0;
2675 	for (i = 0; i < maxpages; ++i) {
2676 		vm_object_t lobject;
2677 		vm_object_t nobject;
2678 		int allocated = 0;
2679 		int error;
2680 
2681 		/*
2682 		 * This can eat a lot of time on a heavily contended
2683 		 * machine so yield on the tick if needed.
2684 		 */
2685 		if ((i & 7) == 7)
2686 			lwkt_yield();
2687 
2688 		/*
2689 		 * Calculate the page to pre-fault, stopping the scan in
2690 		 * each direction separately if the limit is reached.
2691 		 */
2692 		if (i & 1) {
2693 			if (noneg)
2694 				continue;
2695 			addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2696 		} else {
2697 			if (nopos)
2698 				continue;
2699 			addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2700 		}
2701 		if (addr < entry->start) {
2702 			noneg = 1;
2703 			if (noneg && nopos)
2704 				break;
2705 			continue;
2706 		}
2707 		if (addr >= entry->end) {
2708 			nopos = 1;
2709 			if (noneg && nopos)
2710 				break;
2711 			continue;
2712 		}
2713 
2714 		/*
2715 		 * Skip pages already mapped, and stop scanning in that
2716 		 * direction.  When the scan terminates in both directions
2717 		 * we are done.
2718 		 */
2719 		if (pmap_prefault_ok(pmap, addr) == 0) {
2720 			if (i & 1)
2721 				noneg = 1;
2722 			else
2723 				nopos = 1;
2724 			if (noneg && nopos)
2725 				break;
2726 			continue;
2727 		}
2728 
2729 		/*
2730 		 * Follow the VM object chain to obtain the page to be mapped
2731 		 * into the pmap.
2732 		 *
2733 		 * If we reach the terminal object without finding a page
2734 		 * and we determine it would be advantageous, then allocate
2735 		 * a zero-fill page for the base object.  The base object
2736 		 * is guaranteed to be OBJT_DEFAULT for this case.
2737 		 *
2738 		 * In order to not have to check the pager via *haspage*()
2739 		 * we stop if any non-default object is encountered.  e.g.
2740 		 * a vnode or swap object would stop the loop.
2741 		 */
2742 		index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2743 		lobject = object;
2744 		pindex = index;
2745 		pprot = prot;
2746 
2747 		KKASSERT(lobject == entry->object.vm_object);
2748 		/*vm_object_hold(lobject); implied */
2749 
2750 		while ((m = vm_page_lookup_busy_try(lobject, pindex,
2751 						    TRUE, &error)) == NULL) {
2752 			if (lobject->type != OBJT_DEFAULT)
2753 				break;
2754 			if (lobject->backing_object == NULL) {
2755 				if (vm_fast_fault == 0)
2756 					break;
2757 				if ((prot & VM_PROT_WRITE) == 0 ||
2758 				    vm_page_count_min(0)) {
2759 					break;
2760 				}
2761 
2762 				/*
2763 				 * NOTE: Allocated from base object
2764 				 */
2765 				m = vm_page_alloc(object, index,
2766 						  VM_ALLOC_NORMAL |
2767 						  VM_ALLOC_ZERO |
2768 						  VM_ALLOC_USE_GD |
2769 						  VM_ALLOC_NULL_OK);
2770 				if (m == NULL)
2771 					break;
2772 				allocated = 1;
2773 				pprot = prot;
2774 				/* lobject = object .. not needed */
2775 				break;
2776 			}
2777 			if (lobject->backing_object_offset & PAGE_MASK)
2778 				break;
2779 			nobject = lobject->backing_object;
2780 			vm_object_hold(nobject);
2781 			KKASSERT(nobject == lobject->backing_object);
2782 			pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2783 			if (lobject != object) {
2784 				vm_object_lock_swap();
2785 				vm_object_drop(lobject);
2786 			}
2787 			lobject = nobject;
2788 			pprot &= ~VM_PROT_WRITE;
2789 			vm_object_chain_acquire(lobject, 0);
2790 		}
2791 
2792 		/*
2793 		 * NOTE: A non-NULL (m) will be associated with lobject if
2794 		 *	 it was found there, otherwise it is probably a
2795 		 *	 zero-fill page associated with the base object.
2796 		 *
2797 		 * Give-up if no page is available.
2798 		 */
2799 		if (m == NULL) {
2800 			if (lobject != object) {
2801 #if 0
2802 				if (object->backing_object != lobject)
2803 					vm_object_hold(object->backing_object);
2804 #endif
2805 				vm_object_chain_release_all(
2806 					object->backing_object, lobject);
2807 #if 0
2808 				if (object->backing_object != lobject)
2809 					vm_object_drop(object->backing_object);
2810 #endif
2811 				vm_object_drop(lobject);
2812 			}
2813 			break;
2814 		}
2815 
2816 		/*
2817 		 * The object must be marked dirty if we are mapping a
2818 		 * writable page.  m->object is either lobject or object,
2819 		 * both of which are still held.  Do this before we
2820 		 * potentially drop the object.
2821 		 */
2822 		if (pprot & VM_PROT_WRITE)
2823 			vm_object_set_writeable_dirty(m->object);
2824 
2825 		/*
2826 		 * Do not conditionalize on PG_RAM.  If pages are present in
2827 		 * the VM system we assume optimal caching.  If caching is
2828 		 * not optimal the I/O gravy train will be restarted when we
2829 		 * hit an unavailable page.  We do not want to try to restart
2830 		 * the gravy train now because we really don't know how much
2831 		 * of the object has been cached.  The cost for restarting
2832 		 * the gravy train should be low (since accesses will likely
2833 		 * be I/O bound anyway).
2834 		 */
2835 		if (lobject != object) {
2836 #if 0
2837 			if (object->backing_object != lobject)
2838 				vm_object_hold(object->backing_object);
2839 #endif
2840 			vm_object_chain_release_all(object->backing_object,
2841 						    lobject);
2842 #if 0
2843 			if (object->backing_object != lobject)
2844 				vm_object_drop(object->backing_object);
2845 #endif
2846 			vm_object_drop(lobject);
2847 		}
2848 
2849 		/*
2850 		 * Enter the page into the pmap if appropriate.  If we had
2851 		 * allocated the page we have to place it on a queue.  If not
2852 		 * we just have to make sure it isn't on the cache queue
2853 		 * (pages on the cache queue are not allowed to be mapped).
2854 		 */
2855 		if (allocated) {
2856 			/*
2857 			 * Page must be zerod.
2858 			 */
2859 			vm_page_zero_fill(m);
2860 			mycpu->gd_cnt.v_zfod++;
2861 			m->valid = VM_PAGE_BITS_ALL;
2862 
2863 			/*
2864 			 * Handle dirty page case
2865 			 */
2866 			if (pprot & VM_PROT_WRITE)
2867 				vm_set_nosync(m, entry);
2868 			pmap_enter(pmap, addr, m, pprot, 0, entry);
2869 			mycpu->gd_cnt.v_vm_faults++;
2870 			if (curthread->td_lwp)
2871 				++curthread->td_lwp->lwp_ru.ru_minflt;
2872 			vm_page_deactivate(m);
2873 			if (pprot & VM_PROT_WRITE) {
2874 				/*vm_object_set_writeable_dirty(m->object);*/
2875 				vm_set_nosync(m, entry);
2876 				if (fault_flags & VM_FAULT_DIRTY) {
2877 					vm_page_dirty(m);
2878 					/*XXX*/
2879 					swap_pager_unswapped(m);
2880 				}
2881 			}
2882 			vm_page_wakeup(m);
2883 		} else if (error) {
2884 			/* couldn't busy page, no wakeup */
2885 		} else if (
2886 		    ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2887 		    (m->flags & PG_FICTITIOUS) == 0) {
2888 			/*
2889 			 * A fully valid page not undergoing soft I/O can
2890 			 * be immediately entered into the pmap.
2891 			 */
2892 			if ((m->queue - m->pc) == PQ_CACHE)
2893 				vm_page_deactivate(m);
2894 			if (pprot & VM_PROT_WRITE) {
2895 				/*vm_object_set_writeable_dirty(m->object);*/
2896 				vm_set_nosync(m, entry);
2897 				if (fault_flags & VM_FAULT_DIRTY) {
2898 					vm_page_dirty(m);
2899 					/*XXX*/
2900 					swap_pager_unswapped(m);
2901 				}
2902 			}
2903 			if (pprot & VM_PROT_WRITE)
2904 				vm_set_nosync(m, entry);
2905 			pmap_enter(pmap, addr, m, pprot, 0, entry);
2906 			mycpu->gd_cnt.v_vm_faults++;
2907 			if (curthread->td_lwp)
2908 				++curthread->td_lwp->lwp_ru.ru_minflt;
2909 			vm_page_wakeup(m);
2910 		} else {
2911 			vm_page_wakeup(m);
2912 		}
2913 	}
2914 	vm_object_chain_release(object);
2915 	vm_object_drop(object);
2916 }
2917 
2918 /*
2919  * Object can be held shared
2920  */
2921 static void
2922 vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
2923 		  vm_map_entry_t entry, int prot, int fault_flags)
2924 {
2925 	struct lwp *lp;
2926 	vm_page_t m;
2927 	vm_offset_t addr;
2928 	vm_pindex_t pindex;
2929 	vm_object_t object;
2930 	int i;
2931 	int noneg;
2932 	int nopos;
2933 	int maxpages;
2934 
2935 	/*
2936 	 * Get stable max count value, disabled if set to 0
2937 	 */
2938 	maxpages = vm_prefault_pages;
2939 	cpu_ccfence();
2940 	if (maxpages <= 0)
2941 		return;
2942 
2943 	/*
2944 	 * We do not currently prefault mappings that use virtual page
2945 	 * tables.  We do not prefault foreign pmaps.
2946 	 */
2947 	if (entry->maptype != VM_MAPTYPE_NORMAL)
2948 		return;
2949 	lp = curthread->td_lwp;
2950 	if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2951 		return;
2952 	object = entry->object.vm_object;
2953 	if (object->backing_object != NULL)
2954 		return;
2955 	ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2956 
2957 	/*
2958 	 * Limit pre-fault count to 1024 pages.
2959 	 */
2960 	if (maxpages > 1024)
2961 		maxpages = 1024;
2962 
2963 	noneg = 0;
2964 	nopos = 0;
2965 	for (i = 0; i < maxpages; ++i) {
2966 		int error;
2967 
2968 		/*
2969 		 * Calculate the page to pre-fault, stopping the scan in
2970 		 * each direction separately if the limit is reached.
2971 		 */
2972 		if (i & 1) {
2973 			if (noneg)
2974 				continue;
2975 			addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2976 		} else {
2977 			if (nopos)
2978 				continue;
2979 			addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2980 		}
2981 		if (addr < entry->start) {
2982 			noneg = 1;
2983 			if (noneg && nopos)
2984 				break;
2985 			continue;
2986 		}
2987 		if (addr >= entry->end) {
2988 			nopos = 1;
2989 			if (noneg && nopos)
2990 				break;
2991 			continue;
2992 		}
2993 
2994 		/*
2995 		 * Follow the VM object chain to obtain the page to be mapped
2996 		 * into the pmap.  This version of the prefault code only
2997 		 * works with terminal objects.
2998 		 *
2999 		 * The page must already exist.  If we encounter a problem
3000 		 * we stop here.
3001 		 *
3002 		 * WARNING!  We cannot call swap_pager_unswapped() or insert
3003 		 *	     a new vm_page with a shared token.
3004 		 */
3005 		pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
3006 
3007 		m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
3008 		if (m == NULL || error)
3009 			break;
3010 
3011 		/*
3012 		 * Skip pages already mapped, and stop scanning in that
3013 		 * direction.  When the scan terminates in both directions
3014 		 * we are done.
3015 		 */
3016 		if (pmap_prefault_ok(pmap, addr) == 0) {
3017 			vm_page_wakeup(m);
3018 			if (i & 1)
3019 				noneg = 1;
3020 			else
3021 				nopos = 1;
3022 			if (noneg && nopos)
3023 				break;
3024 			continue;
3025 		}
3026 
3027 		/*
3028 		 * Stop if the page cannot be trivially entered into the
3029 		 * pmap.
3030 		 */
3031 		if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) ||
3032 		    (m->flags & PG_FICTITIOUS) ||
3033 		    ((m->flags & PG_SWAPPED) &&
3034 		     (prot & VM_PROT_WRITE) &&
3035 		     (fault_flags & VM_FAULT_DIRTY))) {
3036 			vm_page_wakeup(m);
3037 			break;
3038 		}
3039 
3040 		/*
3041 		 * Enter the page into the pmap.  The object might be held
3042 		 * shared so we can't do any (serious) modifying operation
3043 		 * on it.
3044 		 */
3045 		if ((m->queue - m->pc) == PQ_CACHE)
3046 			vm_page_deactivate(m);
3047 		if (prot & VM_PROT_WRITE) {
3048 			vm_object_set_writeable_dirty(m->object);
3049 			vm_set_nosync(m, entry);
3050 			if (fault_flags & VM_FAULT_DIRTY) {
3051 				vm_page_dirty(m);
3052 				/* can't happeen due to conditional above */
3053 				/* swap_pager_unswapped(m); */
3054 			}
3055 		}
3056 		pmap_enter(pmap, addr, m, prot, 0, entry);
3057 		mycpu->gd_cnt.v_vm_faults++;
3058 		if (curthread->td_lwp)
3059 			++curthread->td_lwp->lwp_ru.ru_minflt;
3060 		vm_page_wakeup(m);
3061 	}
3062 }
3063