1b228cb1b1613f8d9b5d5247eab122991ce9a863
[dragonfly.git] / sys / vm / vm_fault.c
1 /*
2  * (MPSAFE)
3  *
4  * Copyright (c) 1991, 1993
5  *      The Regents of the University of California.  All rights reserved.
6  * Copyright (c) 1994 John S. Dyson
7  * All rights reserved.
8  * Copyright (c) 1994 David Greenman
9  * All rights reserved.
10  *
11  *
12  * This code is derived from software contributed to Berkeley by
13  * The Mach Operating System project at Carnegie-Mellon University.
14  *
15  * Redistribution and use in source and binary forms, with or without
16  * modification, are permitted provided that the following conditions
17  * are met:
18  * 1. Redistributions of source code must retain the above copyright
19  *    notice, this list of conditions and the following disclaimer.
20  * 2. Redistributions in binary form must reproduce the above copyright
21  *    notice, this list of conditions and the following disclaimer in the
22  *    documentation and/or other materials provided with the distribution.
23  * 3. All advertising materials mentioning features or use of this software
24  *    must display the following acknowledgement:
25  *      This product includes software developed by the University of
26  *      California, Berkeley and its contributors.
27  * 4. Neither the name of the University nor the names of its contributors
28  *    may be used to endorse or promote products derived from this software
29  *    without specific prior written permission.
30  *
31  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41  * SUCH DAMAGE.
42  *
43  *      from: @(#)vm_fault.c    8.4 (Berkeley) 1/12/94
44  *
45  *
46  * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47  * All rights reserved.
48  *
49  * Authors: Avadis Tevanian, Jr., Michael Wayne Young
50  *
51  * Permission to use, copy, modify and distribute this software and
52  * its documentation is hereby granted, provided that both the copyright
53  * notice and this permission notice appear in all copies of the
54  * software, derivative works or modified versions, and any portions
55  * thereof, and that both notices appear in supporting documentation.
56  *
57  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
58  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
59  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
60  *
61  * Carnegie Mellon requests users of this software to return to
62  *
63  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
64  *  School of Computer Science
65  *  Carnegie Mellon University
66  *  Pittsburgh PA 15213-3890
67  *
68  * any improvements or extensions that they make and grant Carnegie the
69  * rights to redistribute these changes.
70  *
71  * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
72  * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $
73  */
74
75 /*
76  *      Page fault handling module.
77  */
78
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/proc.h>
83 #include <sys/vnode.h>
84 #include <sys/resourcevar.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vkernel.h>
87 #include <sys/lock.h>
88 #include <sys/sysctl.h>
89
90 #include <cpu/lwbuf.h>
91
92 #include <vm/vm.h>
93 #include <vm/vm_param.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_kern.h>
100 #include <vm/vm_pager.h>
101 #include <vm/vnode_pager.h>
102 #include <vm/vm_extern.h>
103
104 #include <sys/thread2.h>
105 #include <vm/vm_page2.h>
106
107 struct faultstate {
108         vm_page_t m;
109         vm_object_t object;
110         vm_pindex_t pindex;
111         vm_prot_t prot;
112         vm_page_t first_m;
113         vm_object_t first_object;
114         vm_prot_t first_prot;
115         vm_map_t map;
116         vm_map_entry_t entry;
117         int lookup_still_valid;
118         int didlimit;
119         int hardfault;
120         int fault_flags;
121         int map_generation;
122         boolean_t wired;
123         struct vnode *vp;
124 };
125
126 static int vm_fast_fault = 1;
127 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, "");
128 static int debug_cluster = 0;
129 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
130
131 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
132 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
133 #if 0
134 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
135 #endif
136 static int vm_fault_ratelimit(struct vmspace *);
137 static void vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry,
138                         int prot);
139
140 /*
141  * The caller must hold vm_token.
142  */
143 static __inline void
144 release_page(struct faultstate *fs)
145 {
146         vm_page_deactivate(fs->m);
147         vm_page_wakeup(fs->m);
148         fs->m = NULL;
149 }
150
151 /*
152  * The caller must hold vm_token.
153  */
154 static __inline void
155 unlock_map(struct faultstate *fs)
156 {
157         if (fs->lookup_still_valid && fs->map) {
158                 vm_map_lookup_done(fs->map, fs->entry, 0);
159                 fs->lookup_still_valid = FALSE;
160         }
161 }
162
163 /*
164  * Clean up after a successful call to vm_fault_object() so another call
165  * to vm_fault_object() can be made.
166  *
167  * The caller must hold vm_token.
168  */
169 static void
170 _cleanup_successful_fault(struct faultstate *fs, int relock)
171 {
172         if (fs->object != fs->first_object) {
173                 vm_page_free(fs->first_m);
174                 vm_object_pip_wakeup(fs->object);
175                 fs->first_m = NULL;
176         }
177         fs->object = fs->first_object;
178         if (relock && fs->lookup_still_valid == FALSE) {
179                 if (fs->map)
180                         vm_map_lock_read(fs->map);
181                 fs->lookup_still_valid = TRUE;
182         }
183 }
184
185 /*
186  * The caller must hold vm_token.
187  */
188 static void
189 _unlock_things(struct faultstate *fs, int dealloc)
190 {
191         vm_object_pip_wakeup(fs->first_object);
192         _cleanup_successful_fault(fs, 0);
193         if (dealloc) {
194                 vm_object_deallocate(fs->first_object);
195                 fs->first_object = NULL;
196         }
197         unlock_map(fs); 
198         if (fs->vp != NULL) { 
199                 vput(fs->vp);
200                 fs->vp = NULL;
201         }
202 }
203
204 #define unlock_things(fs) _unlock_things(fs, 0)
205 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
206 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
207
208 /*
209  * TRYPAGER 
210  *
211  * Determine if the pager for the current object *might* contain the page.
212  *
213  * We only need to try the pager if this is not a default object (default
214  * objects are zero-fill and have no real pager), and if we are not taking
215  * a wiring fault or if the FS entry is wired.
216  */
217 #define TRYPAGER(fs)    \
218                 (fs->object->type != OBJT_DEFAULT && \
219                 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
220
221 /*
222  * vm_fault:
223  *
224  * Handle a page fault occuring at the given address, requiring the given
225  * permissions, in the map specified.  If successful, the page is inserted
226  * into the associated physical map.
227  *
228  * NOTE: The given address should be truncated to the proper page address.
229  *
230  * KERN_SUCCESS is returned if the page fault is handled; otherwise,
231  * a standard error specifying why the fault is fatal is returned.
232  *
233  * The map in question must be referenced, and remains so.
234  * The caller may hold no locks.
235  * No other requirements.
236  */
237 int
238 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
239 {
240         int result;
241         vm_pindex_t first_pindex;
242         struct faultstate fs;
243         int growstack;
244
245         mycpu->gd_cnt.v_vm_faults++;
246
247         fs.didlimit = 0;
248         fs.hardfault = 0;
249         fs.fault_flags = fault_flags;
250         growstack = 1;
251
252 RetryFault:
253         /*
254          * Find the vm_map_entry representing the backing store and resolve
255          * the top level object and page index.  This may have the side
256          * effect of executing a copy-on-write on the map entry and/or
257          * creating a shadow object, but will not COW any actual VM pages.
258          *
259          * On success fs.map is left read-locked and various other fields 
260          * are initialized but not otherwise referenced or locked.
261          *
262          * NOTE!  vm_map_lookup will try to upgrade the fault_type to
263          * VM_FAULT_WRITE if the map entry is a virtual page table and also
264          * writable, so we can set the 'A'accessed bit in the virtual page
265          * table entry.
266          */
267         fs.map = map;
268         result = vm_map_lookup(&fs.map, vaddr, fault_type,
269                                &fs.entry, &fs.first_object,
270                                &first_pindex, &fs.first_prot, &fs.wired);
271
272         /*
273          * If the lookup failed or the map protections are incompatible,
274          * the fault generally fails.  However, if the caller is trying
275          * to do a user wiring we have more work to do.
276          */
277         if (result != KERN_SUCCESS) {
278                 if (result != KERN_PROTECTION_FAILURE ||
279                     (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
280                 {
281                         if (result == KERN_INVALID_ADDRESS && growstack &&
282                             map != &kernel_map && curproc != NULL) {
283                                 result = vm_map_growstack(curproc, vaddr);
284                                 if (result != KERN_SUCCESS)
285                                         return (KERN_FAILURE);
286                                 growstack = 0;
287                                 goto RetryFault;
288                         }
289                         return (result);
290                 }
291
292                 /*
293                  * If we are user-wiring a r/w segment, and it is COW, then
294                  * we need to do the COW operation.  Note that we don't
295                  * currently COW RO sections now, because it is NOT desirable
296                  * to COW .text.  We simply keep .text from ever being COW'ed
297                  * and take the heat that one cannot debug wired .text sections.
298                  */
299                 result = vm_map_lookup(&fs.map, vaddr,
300                                        VM_PROT_READ|VM_PROT_WRITE|
301                                         VM_PROT_OVERRIDE_WRITE,
302                                        &fs.entry, &fs.first_object,
303                                        &first_pindex, &fs.first_prot,
304                                        &fs.wired);
305                 if (result != KERN_SUCCESS)
306                         return result;
307
308                 /*
309                  * If we don't COW now, on a user wire, the user will never
310                  * be able to write to the mapping.  If we don't make this
311                  * restriction, the bookkeeping would be nearly impossible.
312                  */
313                 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
314                         fs.entry->max_protection &= ~VM_PROT_WRITE;
315         }
316
317         /*
318          * fs.map is read-locked
319          *
320          * Misc checks.  Save the map generation number to detect races.
321          */
322         fs.map_generation = fs.map->timestamp;
323
324         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
325                 panic("vm_fault: fault on nofault entry, addr: %lx",
326                     (u_long)vaddr);
327         }
328
329         /*
330          * A system map entry may return a NULL object.  No object means
331          * no pager means an unrecoverable kernel fault.
332          */
333         if (fs.first_object == NULL) {
334                 panic("vm_fault: unrecoverable fault at %p in entry %p",
335                         (void *)vaddr, fs.entry);
336         }
337
338         /*
339          * Make a reference to this object to prevent its disposal while we
340          * are messing with it.  Once we have the reference, the map is free
341          * to be diddled.  Since objects reference their shadows (and copies),
342          * they will stay around as well.
343          *
344          * Bump the paging-in-progress count to prevent size changes (e.g.
345          * truncation operations) during I/O.  This must be done after
346          * obtaining the vnode lock in order to avoid possible deadlocks.
347          *
348          * The vm_token is needed to manipulate the vm_object
349          */
350         lwkt_gettoken(&vm_token);
351         vm_object_reference(fs.first_object);
352         fs.vp = vnode_pager_lock(fs.first_object);
353         vm_object_pip_add(fs.first_object, 1);
354         lwkt_reltoken(&vm_token);
355
356         fs.lookup_still_valid = TRUE;
357         fs.first_m = NULL;
358         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
359
360         /*
361          * If the entry is wired we cannot change the page protection.
362          */
363         if (fs.wired)
364                 fault_type = fs.first_prot;
365
366         /*
367          * The page we want is at (first_object, first_pindex), but if the
368          * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
369          * page table to figure out the actual pindex.
370          *
371          * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
372          * ONLY
373          */
374         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
375                 result = vm_fault_vpagetable(&fs, &first_pindex,
376                                              fs.entry->aux.master_pde,
377                                              fault_type);
378                 if (result == KERN_TRY_AGAIN)
379                         goto RetryFault;
380                 if (result != KERN_SUCCESS)
381                         return (result);
382         }
383
384         /*
385          * Now we have the actual (object, pindex), fault in the page.  If
386          * vm_fault_object() fails it will unlock and deallocate the FS
387          * data.   If it succeeds everything remains locked and fs->object
388          * will have an additional PIP count if it is not equal to
389          * fs->first_object
390          *
391          * vm_fault_object will set fs->prot for the pmap operation.  It is
392          * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
393          * page can be safely written.  However, it will force a read-only
394          * mapping for a read fault if the memory is managed by a virtual
395          * page table.
396          */
397         result = vm_fault_object(&fs, first_pindex, fault_type);
398
399         if (result == KERN_TRY_AGAIN)
400                 goto RetryFault;
401         if (result != KERN_SUCCESS)
402                 return (result);
403
404         /*
405          * On success vm_fault_object() does not unlock or deallocate, and fs.m
406          * will contain a busied page.
407          *
408          * Enter the page into the pmap and do pmap-related adjustments.
409          */
410         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
411
412         /*
413          * Burst in a few more pages if possible.  The fs.map should still
414          * be locked.
415          */
416         if (fault_flags & VM_FAULT_BURST) {
417                 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
418                     fs.wired == 0) {
419                         vm_prefault(fs.map->pmap, vaddr, fs.entry, fs.prot);
420                 }
421         }
422         unlock_things(&fs);
423
424         vm_page_flag_clear(fs.m, PG_ZERO);
425         vm_page_flag_set(fs.m, PG_REFERENCED);
426
427         /*
428          * If the page is not wired down, then put it where the pageout daemon
429          * can find it.
430          *
431          * We do not really need to get vm_token here but since all the
432          * vm_*() calls have to doing it here improves efficiency.
433          */
434         lwkt_gettoken(&vm_token);
435         if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
436                 if (fs.wired)
437                         vm_page_wire(fs.m);
438                 else
439                         vm_page_unwire(fs.m, 1);
440         } else {
441                 vm_page_activate(fs.m);
442         }
443
444         if (curthread->td_lwp) {
445                 if (fs.hardfault) {
446                         curthread->td_lwp->lwp_ru.ru_majflt++;
447                 } else {
448                         curthread->td_lwp->lwp_ru.ru_minflt++;
449                 }
450         }
451
452         /*
453          * Unlock everything, and return
454          */
455         vm_page_wakeup(fs.m);
456         vm_object_deallocate(fs.first_object);
457         lwkt_reltoken(&vm_token);
458
459         return (KERN_SUCCESS);
460 }
461
462 /*
463  * Fault in the specified virtual address in the current process map, 
464  * returning a held VM page or NULL.  See vm_fault_page() for more 
465  * information.
466  *
467  * No requirements.
468  */
469 vm_page_t
470 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
471 {
472         struct lwp *lp = curthread->td_lwp;
473         vm_page_t m;
474
475         m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 
476                           fault_type, VM_FAULT_NORMAL, errorp);
477         return(m);
478 }
479
480 /*
481  * Fault in the specified virtual address in the specified map, doing all
482  * necessary manipulation of the object store and all necessary I/O.  Return
483  * a held VM page or NULL, and set *errorp.  The related pmap is not
484  * updated.
485  *
486  * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
487  * and marked PG_REFERENCED as well.
488  *
489  * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
490  * error will be returned.
491  *
492  * No requirements.
493  */
494 vm_page_t
495 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
496               int fault_flags, int *errorp)
497 {
498         vm_pindex_t first_pindex;
499         struct faultstate fs;
500         int result;
501         vm_prot_t orig_fault_type = fault_type;
502
503         mycpu->gd_cnt.v_vm_faults++;
504
505         fs.didlimit = 0;
506         fs.hardfault = 0;
507         fs.fault_flags = fault_flags;
508         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
509
510 RetryFault:
511         /*
512          * Find the vm_map_entry representing the backing store and resolve
513          * the top level object and page index.  This may have the side
514          * effect of executing a copy-on-write on the map entry and/or
515          * creating a shadow object, but will not COW any actual VM pages.
516          *
517          * On success fs.map is left read-locked and various other fields 
518          * are initialized but not otherwise referenced or locked.
519          *
520          * NOTE!  vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
521          * if the map entry is a virtual page table and also writable,
522          * so we can set the 'A'accessed bit in the virtual page table entry.
523          */
524         fs.map = map;
525         result = vm_map_lookup(&fs.map, vaddr, fault_type,
526                                &fs.entry, &fs.first_object,
527                                &first_pindex, &fs.first_prot, &fs.wired);
528
529         if (result != KERN_SUCCESS) {
530                 *errorp = result;
531                 return (NULL);
532         }
533
534         /*
535          * fs.map is read-locked
536          *
537          * Misc checks.  Save the map generation number to detect races.
538          */
539         fs.map_generation = fs.map->timestamp;
540
541         if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
542                 panic("vm_fault: fault on nofault entry, addr: %lx",
543                     (u_long)vaddr);
544         }
545
546         /*
547          * A system map entry may return a NULL object.  No object means
548          * no pager means an unrecoverable kernel fault.
549          */
550         if (fs.first_object == NULL) {
551                 panic("vm_fault: unrecoverable fault at %p in entry %p",
552                         (void *)vaddr, fs.entry);
553         }
554
555         /*
556          * Make a reference to this object to prevent its disposal while we
557          * are messing with it.  Once we have the reference, the map is free
558          * to be diddled.  Since objects reference their shadows (and copies),
559          * they will stay around as well.
560          *
561          * Bump the paging-in-progress count to prevent size changes (e.g.
562          * truncation operations) during I/O.  This must be done after
563          * obtaining the vnode lock in order to avoid possible deadlocks.
564          *
565          * The vm_token is needed to manipulate the vm_object
566          */
567         lwkt_gettoken(&vm_token);
568         vm_object_reference(fs.first_object);
569         fs.vp = vnode_pager_lock(fs.first_object);
570         vm_object_pip_add(fs.first_object, 1);
571         lwkt_reltoken(&vm_token);
572
573         fs.lookup_still_valid = TRUE;
574         fs.first_m = NULL;
575         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
576
577         /*
578          * If the entry is wired we cannot change the page protection.
579          */
580         if (fs.wired)
581                 fault_type = fs.first_prot;
582
583         /*
584          * The page we want is at (first_object, first_pindex), but if the
585          * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
586          * page table to figure out the actual pindex.
587          *
588          * NOTE!  DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
589          * ONLY
590          */
591         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
592                 result = vm_fault_vpagetable(&fs, &first_pindex,
593                                              fs.entry->aux.master_pde,
594                                              fault_type);
595                 if (result == KERN_TRY_AGAIN)
596                         goto RetryFault;
597                 if (result != KERN_SUCCESS) {
598                         *errorp = result;
599                         return (NULL);
600                 }
601         }
602
603         /*
604          * Now we have the actual (object, pindex), fault in the page.  If
605          * vm_fault_object() fails it will unlock and deallocate the FS
606          * data.   If it succeeds everything remains locked and fs->object
607          * will have an additinal PIP count if it is not equal to
608          * fs->first_object
609          */
610         result = vm_fault_object(&fs, first_pindex, fault_type);
611
612         if (result == KERN_TRY_AGAIN)
613                 goto RetryFault;
614         if (result != KERN_SUCCESS) {
615                 *errorp = result;
616                 return(NULL);
617         }
618
619         if ((orig_fault_type & VM_PROT_WRITE) &&
620             (fs.prot & VM_PROT_WRITE) == 0) {
621                 *errorp = KERN_PROTECTION_FAILURE;
622                 unlock_and_deallocate(&fs);
623                 return(NULL);
624         }
625
626         /*
627          * On success vm_fault_object() does not unlock or deallocate, and fs.m
628          * will contain a busied page.
629          */
630         unlock_things(&fs);
631
632         /*
633          * Return a held page.  We are not doing any pmap manipulation so do
634          * not set PG_MAPPED.  However, adjust the page flags according to
635          * the fault type because the caller may not use a managed pmapping
636          * (so we don't want to lose the fact that the page will be dirtied
637          * if a write fault was specified).
638          */
639         vm_page_hold(fs.m);
640         vm_page_flag_clear(fs.m, PG_ZERO);
641         if (fault_type & VM_PROT_WRITE)
642                 vm_page_dirty(fs.m);
643
644         /*
645          * Update the pmap.  We really only have to do this if a COW
646          * occured to replace the read-only page with the new page.  For
647          * now just do it unconditionally. XXX
648          */
649         pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
650         vm_page_flag_set(fs.m, PG_REFERENCED);
651
652         /*
653          * Unbusy the page by activating it.  It remains held and will not
654          * be reclaimed.
655          */
656         lwkt_gettoken(&vm_token);
657         vm_page_activate(fs.m);
658
659         if (curthread->td_lwp) {
660                 if (fs.hardfault) {
661                         curthread->td_lwp->lwp_ru.ru_majflt++;
662                 } else {
663                         curthread->td_lwp->lwp_ru.ru_minflt++;
664                 }
665         }
666
667         /*
668          * Unlock everything, and return the held page.
669          */
670         vm_page_wakeup(fs.m);
671         vm_object_deallocate(fs.first_object);
672         lwkt_reltoken(&vm_token);
673
674         *errorp = 0;
675         return(fs.m);
676 }
677
678 /*
679  * Fault in the specified (object,offset), dirty the returned page as
680  * needed.  If the requested fault_type cannot be done NULL and an
681  * error is returned.
682  *
683  * A held (but not busied) page is returned.
684  *
685  * No requirements.
686  */
687 vm_page_t
688 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
689                      vm_prot_t fault_type, int fault_flags, int *errorp)
690 {
691         int result;
692         vm_pindex_t first_pindex;
693         struct faultstate fs;
694         struct vm_map_entry entry;
695
696         bzero(&entry, sizeof(entry));
697         entry.object.vm_object = object;
698         entry.maptype = VM_MAPTYPE_NORMAL;
699         entry.protection = entry.max_protection = fault_type;
700
701         fs.didlimit = 0;
702         fs.hardfault = 0;
703         fs.fault_flags = fault_flags;
704         fs.map = NULL;
705         KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
706
707 RetryFault:
708         
709         fs.first_object = object;
710         first_pindex = OFF_TO_IDX(offset);
711         fs.entry = &entry;
712         fs.first_prot = fault_type;
713         fs.wired = 0;
714         /*fs.map_generation = 0; unused */
715
716         /*
717          * Make a reference to this object to prevent its disposal while we
718          * are messing with it.  Once we have the reference, the map is free
719          * to be diddled.  Since objects reference their shadows (and copies),
720          * they will stay around as well.
721          *
722          * Bump the paging-in-progress count to prevent size changes (e.g.
723          * truncation operations) during I/O.  This must be done after
724          * obtaining the vnode lock in order to avoid possible deadlocks.
725          */
726         lwkt_gettoken(&vm_token);
727         vm_object_reference(fs.first_object);
728         fs.vp = vnode_pager_lock(fs.first_object);
729         vm_object_pip_add(fs.first_object, 1);
730         lwkt_reltoken(&vm_token);
731
732         fs.lookup_still_valid = TRUE;
733         fs.first_m = NULL;
734         fs.object = fs.first_object;    /* so unlock_and_deallocate works */
735
736 #if 0
737         /* XXX future - ability to operate on VM object using vpagetable */
738         if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
739                 result = vm_fault_vpagetable(&fs, &first_pindex,
740                                              fs.entry->aux.master_pde,
741                                              fault_type);
742                 if (result == KERN_TRY_AGAIN)
743                         goto RetryFault;
744                 if (result != KERN_SUCCESS) {
745                         *errorp = result;
746                         return (NULL);
747                 }
748         }
749 #endif
750
751         /*
752          * Now we have the actual (object, pindex), fault in the page.  If
753          * vm_fault_object() fails it will unlock and deallocate the FS
754          * data.   If it succeeds everything remains locked and fs->object
755          * will have an additinal PIP count if it is not equal to
756          * fs->first_object
757          */
758         result = vm_fault_object(&fs, first_pindex, fault_type);
759
760         if (result == KERN_TRY_AGAIN)
761                 goto RetryFault;
762         if (result != KERN_SUCCESS) {
763                 *errorp = result;
764                 return(NULL);
765         }
766
767         if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
768                 *errorp = KERN_PROTECTION_FAILURE;
769                 unlock_and_deallocate(&fs);
770                 return(NULL);
771         }
772
773         /*
774          * On success vm_fault_object() does not unlock or deallocate, and fs.m
775          * will contain a busied page.
776          */
777         unlock_things(&fs);
778
779         /*
780          * Return a held page.  We are not doing any pmap manipulation so do
781          * not set PG_MAPPED.  However, adjust the page flags according to
782          * the fault type because the caller may not use a managed pmapping
783          * (so we don't want to lose the fact that the page will be dirtied
784          * if a write fault was specified).
785          */
786         lwkt_gettoken(&vm_token);
787         vm_page_hold(fs.m);
788         vm_page_flag_clear(fs.m, PG_ZERO);
789         if (fault_type & VM_PROT_WRITE)
790                 vm_page_dirty(fs.m);
791
792         /*
793          * Indicate that the page was accessed.
794          */
795         vm_page_flag_set(fs.m, PG_REFERENCED);
796
797         /*
798          * Unbusy the page by activating it.  It remains held and will not
799          * be reclaimed.
800          */
801         vm_page_activate(fs.m);
802
803         if (curthread->td_lwp) {
804                 if (fs.hardfault) {
805                         mycpu->gd_cnt.v_vm_faults++;
806                         curthread->td_lwp->lwp_ru.ru_majflt++;
807                 } else {
808                         curthread->td_lwp->lwp_ru.ru_minflt++;
809                 }
810         }
811
812         /*
813          * Unlock everything, and return the held page.
814          */
815         vm_page_wakeup(fs.m);
816         vm_object_deallocate(fs.first_object);
817         lwkt_reltoken(&vm_token);
818
819         *errorp = 0;
820         return(fs.m);
821 }
822
823 /*
824  * Translate the virtual page number (first_pindex) that is relative
825  * to the address space into a logical page number that is relative to the
826  * backing object.  Use the virtual page table pointed to by (vpte).
827  *
828  * This implements an N-level page table.  Any level can terminate the
829  * scan by setting VPTE_PS.   A linear mapping is accomplished by setting
830  * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
831  *
832  * No requirements (vm_token need not be held).
833  */
834 static
835 int
836 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
837                     vpte_t vpte, int fault_type)
838 {
839         struct lwbuf *lwb;
840         int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
841         int result = KERN_SUCCESS;
842         vpte_t *ptep;
843
844         for (;;) {
845                 /*
846                  * We cannot proceed if the vpte is not valid, not readable
847                  * for a read fault, or not writable for a write fault.
848                  */
849                 if ((vpte & VPTE_V) == 0) {
850                         unlock_and_deallocate(fs);
851                         return (KERN_FAILURE);
852                 }
853                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
854                         unlock_and_deallocate(fs);
855                         return (KERN_FAILURE);
856                 }
857                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
858                         unlock_and_deallocate(fs);
859                         return (KERN_FAILURE);
860                 }
861                 if ((vpte & VPTE_PS) || vshift == 0)
862                         break;
863                 KKASSERT(vshift >= VPTE_PAGE_BITS);
864
865                 /*
866                  * Get the page table page.  Nominally we only read the page
867                  * table, but since we are actively setting VPTE_M and VPTE_A,
868                  * tell vm_fault_object() that we are writing it. 
869                  *
870                  * There is currently no real need to optimize this.
871                  */
872                 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
873                                          VM_PROT_READ|VM_PROT_WRITE);
874                 if (result != KERN_SUCCESS)
875                         return (result);
876
877                 /*
878                  * Process the returned fs.m and look up the page table
879                  * entry in the page table page.
880                  */
881                 vshift -= VPTE_PAGE_BITS;
882                 lwb = lwbuf_alloc(fs->m);
883                 ptep = ((vpte_t *)lwbuf_kva(lwb) +
884                         ((*pindex >> vshift) & VPTE_PAGE_MASK));
885                 vpte = *ptep;
886
887                 /*
888                  * Page table write-back.  If the vpte is valid for the
889                  * requested operation, do a write-back to the page table.
890                  *
891                  * XXX VPTE_M is not set properly for page directory pages.
892                  * It doesn't get set in the page directory if the page table
893                  * is modified during a read access.
894                  */
895                 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
896                     (vpte & VPTE_W)) {
897                         if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
898                                 atomic_set_long(ptep, VPTE_M | VPTE_A);
899                                 vm_page_dirty(fs->m);
900                         }
901                 }
902                 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
903                     (vpte & VPTE_R)) {
904                         if ((vpte & VPTE_A) == 0) {
905                                 atomic_set_long(ptep, VPTE_A);
906                                 vm_page_dirty(fs->m);
907                         }
908                 }
909                 lwbuf_free(lwb);
910                 vm_page_flag_set(fs->m, PG_REFERENCED);
911                 vm_page_activate(fs->m);
912                 vm_page_wakeup(fs->m);
913                 cleanup_successful_fault(fs);
914         }
915         /*
916          * Combine remaining address bits with the vpte.
917          */
918         /* JG how many bits from each? */
919         *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
920                   (*pindex & ((1L << vshift) - 1));
921         return (KERN_SUCCESS);
922 }
923
924
925 /*
926  * This is the core of the vm_fault code.
927  *
928  * Do all operations required to fault-in (fs.first_object, pindex).  Run
929  * through the shadow chain as necessary and do required COW or virtual
930  * copy operations.  The caller has already fully resolved the vm_map_entry
931  * and, if appropriate, has created a copy-on-write layer.  All we need to
932  * do is iterate the object chain.
933  *
934  * On failure (fs) is unlocked and deallocated and the caller may return or
935  * retry depending on the failure code.  On success (fs) is NOT unlocked or
936  * deallocated, fs.m will contained a resolved, busied page, and fs.object
937  * will have an additional PIP count if it is not equal to fs.first_object.
938  *
939  * No requirements.
940  */
941 static
942 int
943 vm_fault_object(struct faultstate *fs,
944                 vm_pindex_t first_pindex, vm_prot_t fault_type)
945 {
946         vm_object_t next_object;
947         vm_pindex_t pindex;
948
949         fs->prot = fs->first_prot;
950         fs->object = fs->first_object;
951         pindex = first_pindex;
952
953         /* 
954          * If a read fault occurs we try to make the page writable if
955          * possible.  There are three cases where we cannot make the
956          * page mapping writable:
957          *
958          * (1) The mapping is read-only or the VM object is read-only,
959          *     fs->prot above will simply not have VM_PROT_WRITE set.
960          *
961          * (2) If the mapping is a virtual page table we need to be able
962          *     to detect writes so we can set VPTE_M in the virtual page
963          *     table.
964          *
965          * (3) If the VM page is read-only or copy-on-write, upgrading would
966          *     just result in an unnecessary COW fault.
967          *
968          * VM_PROT_VPAGED is set if faulting via a virtual page table and
969          * causes adjustments to the 'M'odify bit to also turn off write
970          * access to force a re-fault.
971          */
972         if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
973                 if ((fault_type & VM_PROT_WRITE) == 0)
974                         fs->prot &= ~VM_PROT_WRITE;
975         }
976
977         lwkt_gettoken(&vm_token);
978
979         for (;;) {
980                 /*
981                  * If the object is dead, we stop here
982                  */
983                 if (fs->object->flags & OBJ_DEAD) {
984                         unlock_and_deallocate(fs);
985                         lwkt_reltoken(&vm_token);
986                         return (KERN_PROTECTION_FAILURE);
987                 }
988
989                 /*
990                  * See if page is resident.  spl protection is required
991                  * to avoid an interrupt unbusy/free race against our
992                  * lookup.  We must hold the protection through a page
993                  * allocation or busy.
994                  */
995                 crit_enter();
996                 fs->m = vm_page_lookup(fs->object, pindex);
997                 if (fs->m != NULL) {
998                         int queue;
999                         /*
1000                          * Wait/Retry if the page is busy.  We have to do this
1001                          * if the page is busy via either PG_BUSY or 
1002                          * vm_page_t->busy because the vm_pager may be using
1003                          * vm_page_t->busy for pageouts ( and even pageins if
1004                          * it is the vnode pager ), and we could end up trying
1005                          * to pagein and pageout the same page simultaneously.
1006                          *
1007                          * We can theoretically allow the busy case on a read
1008                          * fault if the page is marked valid, but since such
1009                          * pages are typically already pmap'd, putting that
1010                          * special case in might be more effort then it is 
1011                          * worth.  We cannot under any circumstances mess
1012                          * around with a vm_page_t->busy page except, perhaps,
1013                          * to pmap it.
1014                          */
1015                         if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
1016                                 unlock_things(fs);
1017                                 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1018                                 mycpu->gd_cnt.v_intrans++;
1019                                 vm_object_deallocate(fs->first_object);
1020                                 fs->first_object = NULL;
1021                                 lwkt_reltoken(&vm_token);
1022                                 crit_exit();
1023                                 return (KERN_TRY_AGAIN);
1024                         }
1025
1026                         /*
1027                          * If reactivating a page from PQ_CACHE we may have
1028                          * to rate-limit.
1029                          */
1030                         queue = fs->m->queue;
1031                         vm_page_unqueue_nowakeup(fs->m);
1032
1033                         if ((queue - fs->m->pc) == PQ_CACHE && 
1034                             vm_page_count_severe()) {
1035                                 vm_page_activate(fs->m);
1036                                 unlock_and_deallocate(fs);
1037                                 vm_waitpfault();
1038                                 lwkt_reltoken(&vm_token);
1039                                 crit_exit();
1040                                 return (KERN_TRY_AGAIN);
1041                         }
1042
1043                         /*
1044                          * Mark page busy for other processes, and the 
1045                          * pagedaemon.  If it still isn't completely valid
1046                          * (readable), or if a read-ahead-mark is set on
1047                          * the VM page, jump to readrest, else we found the
1048                          * page and can return.
1049                          *
1050                          * We can release the spl once we have marked the
1051                          * page busy.
1052                          */
1053                         vm_page_busy(fs->m);
1054                         crit_exit();
1055
1056                         if (fs->m->object != &kernel_object) {
1057                                 if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1058                                     VM_PAGE_BITS_ALL) {
1059                                         goto readrest;
1060                                 }
1061                                 if (fs->m->flags & PG_RAM) {
1062                                         if (debug_cluster)
1063                                                 kprintf("R");
1064                                         vm_page_flag_clear(fs->m, PG_RAM);
1065                                         goto readrest;
1066                                 }
1067                         }
1068                         break; /* break to PAGE HAS BEEN FOUND */
1069                 }
1070
1071                 /*
1072                  * Page is not resident, If this is the search termination
1073                  * or the pager might contain the page, allocate a new page.
1074                  *
1075                  * NOTE: We are still in a critical section.
1076                  */
1077                 if (TRYPAGER(fs) || fs->object == fs->first_object) {
1078                         /*
1079                          * If the page is beyond the object size we fail
1080                          */
1081                         if (pindex >= fs->object->size) {
1082                                 lwkt_reltoken(&vm_token);
1083                                 crit_exit();
1084                                 unlock_and_deallocate(fs);
1085                                 return (KERN_PROTECTION_FAILURE);
1086                         }
1087
1088                         /*
1089                          * Ratelimit.
1090                          */
1091                         if (fs->didlimit == 0 && curproc != NULL) {
1092                                 int limticks;
1093
1094                                 limticks = vm_fault_ratelimit(curproc->p_vmspace);
1095                                 if (limticks) {
1096                                         lwkt_reltoken(&vm_token);
1097                                         crit_exit();
1098                                         unlock_and_deallocate(fs);
1099                                         tsleep(curproc, 0, "vmrate", limticks);
1100                                         fs->didlimit = 1;
1101                                         return (KERN_TRY_AGAIN);
1102                                 }
1103                         }
1104
1105                         /*
1106                          * Allocate a new page for this object/offset pair.
1107                          */
1108                         fs->m = NULL;
1109                         if (!vm_page_count_severe()) {
1110                                 fs->m = vm_page_alloc(fs->object, pindex,
1111                                     (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
1112                         }
1113                         if (fs->m == NULL) {
1114                                 lwkt_reltoken(&vm_token);
1115                                 crit_exit();
1116                                 unlock_and_deallocate(fs);
1117                                 vm_waitpfault();
1118                                 return (KERN_TRY_AGAIN);
1119                         }
1120                 }
1121                 crit_exit();
1122
1123 readrest:
1124                 /*
1125                  * We have found an invalid or partially valid page, a
1126                  * page with a read-ahead mark which might be partially or
1127                  * fully valid (and maybe dirty too), or we have allocated
1128                  * a new page.
1129                  *
1130                  * Attempt to fault-in the page if there is a chance that the
1131                  * pager has it, and potentially fault in additional pages
1132                  * at the same time.
1133                  *
1134                  * We are NOT in splvm here and if TRYPAGER is true then
1135                  * fs.m will be non-NULL and will be PG_BUSY for us.
1136                  */
1137                 if (TRYPAGER(fs)) {
1138                         int rv;
1139                         int seqaccess;
1140                         u_char behavior = vm_map_entry_behavior(fs->entry);
1141
1142                         if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1143                                 seqaccess = 0;
1144                         else
1145                                 seqaccess = -1;
1146
1147                         /*
1148                          * If sequential access is detected then attempt
1149                          * to deactivate/cache pages behind the scan to
1150                          * prevent resource hogging.
1151                          *
1152                          * Use of PG_RAM to detect sequential access
1153                          * also simulates multi-zone sequential access
1154                          * detection for free.
1155                          *
1156                          * NOTE: Partially valid dirty pages cannot be
1157                          *       deactivated without causing NFS picemeal
1158                          *       writes to barf.
1159                          */
1160                         if ((fs->first_object->type != OBJT_DEVICE) &&
1161                             (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1162                                 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1163                                  (fs->m->flags & PG_RAM)))
1164                         ) {
1165                                 vm_pindex_t scan_pindex;
1166                                 int scan_count = 16;
1167
1168                                 if (first_pindex < 16) {
1169                                         scan_pindex = 0;
1170                                         scan_count = 0;
1171                                 } else {
1172                                         scan_pindex = first_pindex - 16;
1173                                         if (scan_pindex < 16)
1174                                                 scan_count = scan_pindex;
1175                                         else
1176                                                 scan_count = 16;
1177                                 }
1178
1179                                 crit_enter();
1180                                 while (scan_count) {
1181                                         vm_page_t mt;
1182
1183                                         mt = vm_page_lookup(fs->first_object,
1184                                                             scan_pindex);
1185                                         if (mt == NULL ||
1186                                             (mt->valid != VM_PAGE_BITS_ALL)) {
1187                                                 break;
1188                                         }
1189                                         if (mt->busy ||
1190                                             (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
1191                                             mt->hold_count ||
1192                                             mt->wire_count)  {
1193                                                 goto skip;
1194                                         }
1195                                         if (mt->dirty == 0)
1196                                                 vm_page_test_dirty(mt);
1197                                         if (mt->dirty) {
1198                                                 vm_page_busy(mt);
1199                                                 vm_page_protect(mt,
1200                                                                 VM_PROT_NONE);
1201                                                 vm_page_deactivate(mt);
1202                                                 vm_page_wakeup(mt);
1203                                         } else {
1204                                                 vm_page_cache(mt);
1205                                         }
1206 skip:
1207                                         --scan_count;
1208                                         --scan_pindex;
1209                                 }
1210                                 crit_exit();
1211
1212                                 seqaccess = 1;
1213                         }
1214
1215                         /*
1216                          * Avoid deadlocking against the map when doing I/O.
1217                          * fs.object and the page is PG_BUSY'd.
1218                          */
1219                         unlock_map(fs);
1220
1221                         /*
1222                          * Acquire the page data.  We still hold a ref on
1223                          * fs.object and the page has been PG_BUSY's.
1224                          *
1225                          * The pager may replace the page (for example, in
1226                          * order to enter a fictitious page into the
1227                          * object).  If it does so it is responsible for
1228                          * cleaning up the passed page and properly setting
1229                          * the new page PG_BUSY.
1230                          *
1231                          * If we got here through a PG_RAM read-ahead
1232                          * mark the page may be partially dirty and thus
1233                          * not freeable.  Don't bother checking to see
1234                          * if the pager has the page because we can't free
1235                          * it anyway.  We have to depend on the get_page
1236                          * operation filling in any gaps whether there is
1237                          * backing store or not.
1238                          */
1239                         rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1240
1241                         if (rv == VM_PAGER_OK) {
1242                                 /*
1243                                  * Relookup in case pager changed page. Pager
1244                                  * is responsible for disposition of old page
1245                                  * if moved.
1246                                  *
1247                                  * XXX other code segments do relookups too.
1248                                  * It's a bad abstraction that needs to be
1249                                  * fixed/removed.
1250                                  */
1251                                 fs->m = vm_page_lookup(fs->object, pindex);
1252                                 if (fs->m == NULL) {
1253                                         lwkt_reltoken(&vm_token);
1254                                         unlock_and_deallocate(fs);
1255                                         return (KERN_TRY_AGAIN);
1256                                 }
1257
1258                                 ++fs->hardfault;
1259                                 break; /* break to PAGE HAS BEEN FOUND */
1260                         }
1261
1262                         /*
1263                          * Remove the bogus page (which does not exist at this
1264                          * object/offset); before doing so, we must get back
1265                          * our object lock to preserve our invariant.
1266                          *
1267                          * Also wake up any other process that may want to bring
1268                          * in this page.
1269                          *
1270                          * If this is the top-level object, we must leave the
1271                          * busy page to prevent another process from rushing
1272                          * past us, and inserting the page in that object at
1273                          * the same time that we are.
1274                          */
1275                         if (rv == VM_PAGER_ERROR) {
1276                                 if (curproc)
1277                                         kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
1278                                 else
1279                                         kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
1280                         }
1281
1282                         /*
1283                          * Data outside the range of the pager or an I/O error
1284                          *
1285                          * The page may have been wired during the pagein,
1286                          * e.g. by the buffer cache, and cannot simply be
1287                          * freed.  Call vnode_pager_freepage() to deal with it.
1288                          */
1289                         /*
1290                          * XXX - the check for kernel_map is a kludge to work
1291                          * around having the machine panic on a kernel space
1292                          * fault w/ I/O error.
1293                          */
1294                         if (((fs->map != &kernel_map) &&
1295                             (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1296                                 vnode_pager_freepage(fs->m);
1297                                 lwkt_reltoken(&vm_token);
1298                                 fs->m = NULL;
1299                                 unlock_and_deallocate(fs);
1300                                 if (rv == VM_PAGER_ERROR)
1301                                         return (KERN_FAILURE);
1302                                 else
1303                                         return (KERN_PROTECTION_FAILURE);
1304                                 /* NOT REACHED */
1305                         }
1306                         if (fs->object != fs->first_object) {
1307                                 vnode_pager_freepage(fs->m);
1308                                 fs->m = NULL;
1309                                 /*
1310                                  * XXX - we cannot just fall out at this
1311                                  * point, m has been freed and is invalid!
1312                                  */
1313                         }
1314                 }
1315
1316                 /*
1317                  * We get here if the object has a default pager (or unwiring) 
1318                  * or the pager doesn't have the page.
1319                  */
1320                 if (fs->object == fs->first_object)
1321                         fs->first_m = fs->m;
1322
1323                 /*
1324                  * Move on to the next object.  Lock the next object before
1325                  * unlocking the current one.
1326                  */
1327                 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1328                 next_object = fs->object->backing_object;
1329                 if (next_object == NULL) {
1330                         /*
1331                          * If there's no object left, fill the page in the top
1332                          * object with zeros.
1333                          */
1334                         if (fs->object != fs->first_object) {
1335                                 vm_object_pip_wakeup(fs->object);
1336
1337                                 fs->object = fs->first_object;
1338                                 pindex = first_pindex;
1339                                 fs->m = fs->first_m;
1340                         }
1341                         fs->first_m = NULL;
1342
1343                         /*
1344                          * Zero the page if necessary and mark it valid.
1345                          */
1346                         if ((fs->m->flags & PG_ZERO) == 0) {
1347                                 vm_page_zero_fill(fs->m);
1348                         } else {
1349                                 mycpu->gd_cnt.v_ozfod++;
1350                         }
1351                         mycpu->gd_cnt.v_zfod++;
1352                         fs->m->valid = VM_PAGE_BITS_ALL;
1353                         break;  /* break to PAGE HAS BEEN FOUND */
1354                 }
1355                 if (fs->object != fs->first_object) {
1356                         vm_object_pip_wakeup(fs->object);
1357                 }
1358                 KASSERT(fs->object != next_object,
1359                         ("object loop %p", next_object));
1360                 fs->object = next_object;
1361                 vm_object_pip_add(fs->object, 1);
1362         }
1363
1364         /*
1365          * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1366          * is held.]
1367          *
1368          * vm_token is still held
1369          *
1370          * If the page is being written, but isn't already owned by the
1371          * top-level object, we have to copy it into a new page owned by the
1372          * top-level object.
1373          */
1374         KASSERT((fs->m->flags & PG_BUSY) != 0,
1375                 ("vm_fault: not busy after main loop"));
1376
1377         if (fs->object != fs->first_object) {
1378                 /*
1379                  * We only really need to copy if we want to write it.
1380                  */
1381                 if (fault_type & VM_PROT_WRITE) {
1382                         /*
1383                          * This allows pages to be virtually copied from a 
1384                          * backing_object into the first_object, where the 
1385                          * backing object has no other refs to it, and cannot
1386                          * gain any more refs.  Instead of a bcopy, we just 
1387                          * move the page from the backing object to the 
1388                          * first object.  Note that we must mark the page 
1389                          * dirty in the first object so that it will go out 
1390                          * to swap when needed.
1391                          */
1392                         if (
1393                                 /*
1394                                  * Map, if present, has not changed
1395                                  */
1396                                 (fs->map == NULL ||
1397                                 fs->map_generation == fs->map->timestamp) &&
1398                                 /*
1399                                  * Only one shadow object
1400                                  */
1401                                 (fs->object->shadow_count == 1) &&
1402                                 /*
1403                                  * No COW refs, except us
1404                                  */
1405                                 (fs->object->ref_count == 1) &&
1406                                 /*
1407                                  * No one else can look this object up
1408                                  */
1409                                 (fs->object->handle == NULL) &&
1410                                 /*
1411                                  * No other ways to look the object up
1412                                  */
1413                                 ((fs->object->type == OBJT_DEFAULT) ||
1414                                  (fs->object->type == OBJT_SWAP)) &&
1415                                 /*
1416                                  * We don't chase down the shadow chain
1417                                  */
1418                                 (fs->object == fs->first_object->backing_object) &&
1419
1420                                 /*
1421                                  * grab the lock if we need to
1422                                  */
1423                                 (fs->lookup_still_valid ||
1424                                  fs->map == NULL ||
1425                                  lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1426                             ) {
1427                                 
1428                                 fs->lookup_still_valid = 1;
1429                                 /*
1430                                  * get rid of the unnecessary page
1431                                  */
1432                                 vm_page_protect(fs->first_m, VM_PROT_NONE);
1433                                 vm_page_free(fs->first_m);
1434                                 fs->first_m = NULL;
1435
1436                                 /*
1437                                  * grab the page and put it into the 
1438                                  * process'es object.  The page is 
1439                                  * automatically made dirty.
1440                                  */
1441                                 vm_page_rename(fs->m, fs->first_object, first_pindex);
1442                                 fs->first_m = fs->m;
1443                                 vm_page_busy(fs->first_m);
1444                                 fs->m = NULL;
1445                                 mycpu->gd_cnt.v_cow_optim++;
1446                         } else {
1447                                 /*
1448                                  * Oh, well, lets copy it.
1449                                  */
1450                                 vm_page_copy(fs->m, fs->first_m);
1451                                 vm_page_event(fs->m, VMEVENT_COW);
1452                         }
1453
1454                         if (fs->m) {
1455                                 /*
1456                                  * We no longer need the old page or object.
1457                                  */
1458                                 release_page(fs);
1459                         }
1460
1461                         /*
1462                          * fs->object != fs->first_object due to above 
1463                          * conditional
1464                          */
1465                         vm_object_pip_wakeup(fs->object);
1466
1467                         /*
1468                          * Only use the new page below...
1469                          */
1470
1471                         mycpu->gd_cnt.v_cow_faults++;
1472                         fs->m = fs->first_m;
1473                         fs->object = fs->first_object;
1474                         pindex = first_pindex;
1475                 } else {
1476                         /*
1477                          * If it wasn't a write fault avoid having to copy
1478                          * the page by mapping it read-only.
1479                          */
1480                         fs->prot &= ~VM_PROT_WRITE;
1481                 }
1482         }
1483
1484         /*
1485          * We may have had to unlock a map to do I/O.  If we did then
1486          * lookup_still_valid will be FALSE.  If the map generation count
1487          * also changed then all sorts of things could have happened while
1488          * we were doing the I/O and we need to retry.
1489          */
1490
1491         if (!fs->lookup_still_valid &&
1492             fs->map != NULL &&
1493             (fs->map->timestamp != fs->map_generation)) {
1494                 release_page(fs);
1495                 lwkt_reltoken(&vm_token);
1496                 unlock_and_deallocate(fs);
1497                 return (KERN_TRY_AGAIN);
1498         }
1499
1500         /*
1501          * If the fault is a write, we know that this page is being
1502          * written NOW so dirty it explicitly to save on pmap_is_modified()
1503          * calls later.
1504          *
1505          * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1506          * if the page is already dirty to prevent data written with
1507          * the expectation of being synced from not being synced.
1508          * Likewise if this entry does not request NOSYNC then make
1509          * sure the page isn't marked NOSYNC.  Applications sharing
1510          * data should use the same flags to avoid ping ponging.
1511          *
1512          * Also tell the backing pager, if any, that it should remove
1513          * any swap backing since the page is now dirty.
1514          */
1515         if (fs->prot & VM_PROT_WRITE) {
1516                 vm_object_set_writeable_dirty(fs->m->object);
1517                 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) {
1518                         if (fs->m->dirty == 0)
1519                                 vm_page_flag_set(fs->m, PG_NOSYNC);
1520                 } else {
1521                         vm_page_flag_clear(fs->m, PG_NOSYNC);
1522                 }
1523                 if (fs->fault_flags & VM_FAULT_DIRTY) {
1524                         crit_enter();
1525                         vm_page_dirty(fs->m);
1526                         swap_pager_unswapped(fs->m);
1527                         crit_exit();
1528                 }
1529         }
1530
1531         lwkt_reltoken(&vm_token);
1532
1533         /*
1534          * Page had better still be busy.  We are still locked up and 
1535          * fs->object will have another PIP reference if it is not equal
1536          * to fs->first_object.
1537          */
1538         KASSERT(fs->m->flags & PG_BUSY,
1539                 ("vm_fault: page %p not busy!", fs->m));
1540
1541         /*
1542          * Sanity check: page must be completely valid or it is not fit to
1543          * map into user space.  vm_pager_get_pages() ensures this.
1544          */
1545         if (fs->m->valid != VM_PAGE_BITS_ALL) {
1546                 vm_page_zero_invalid(fs->m, TRUE);
1547                 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1548         }
1549
1550         return (KERN_SUCCESS);
1551 }
1552
1553 /*
1554  * Wire down a range of virtual addresses in a map.  The entry in question
1555  * should be marked in-transition and the map must be locked.  We must
1556  * release the map temporarily while faulting-in the page to avoid a
1557  * deadlock.  Note that the entry may be clipped while we are blocked but
1558  * will never be freed.
1559  *
1560  * No requirements.
1561  */
1562 int
1563 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1564 {
1565         boolean_t fictitious;
1566         vm_offset_t start;
1567         vm_offset_t end;
1568         vm_offset_t va;
1569         vm_paddr_t pa;
1570         pmap_t pmap;
1571         int rv;
1572
1573         pmap = vm_map_pmap(map);
1574         start = entry->start;
1575         end = entry->end;
1576         fictitious = entry->object.vm_object &&
1577                         (entry->object.vm_object->type == OBJT_DEVICE);
1578
1579         lwkt_gettoken(&vm_token);
1580         vm_map_unlock(map);
1581         map->timestamp++;
1582
1583         /*
1584          * We simulate a fault to get the page and enter it in the physical
1585          * map.
1586          */
1587         for (va = start; va < end; va += PAGE_SIZE) {
1588                 if (user_wire) {
1589                         rv = vm_fault(map, va, VM_PROT_READ, 
1590                                         VM_FAULT_USER_WIRE);
1591                 } else {
1592                         rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1593                                         VM_FAULT_CHANGE_WIRING);
1594                 }
1595                 if (rv) {
1596                         while (va > start) {
1597                                 va -= PAGE_SIZE;
1598                                 if ((pa = pmap_extract(pmap, va)) == 0)
1599                                         continue;
1600                                 pmap_change_wiring(pmap, va, FALSE);
1601                                 if (!fictitious)
1602                                         vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1603                         }
1604                         vm_map_lock(map);
1605                         lwkt_reltoken(&vm_token);
1606                         return (rv);
1607                 }
1608         }
1609         vm_map_lock(map);
1610         lwkt_reltoken(&vm_token);
1611         return (KERN_SUCCESS);
1612 }
1613
1614 /*
1615  * Unwire a range of virtual addresses in a map.  The map should be
1616  * locked.
1617  */
1618 void
1619 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1620 {
1621         boolean_t fictitious;
1622         vm_offset_t start;
1623         vm_offset_t end;
1624         vm_offset_t va;
1625         vm_paddr_t pa;
1626         pmap_t pmap;
1627
1628         pmap = vm_map_pmap(map);
1629         start = entry->start;
1630         end = entry->end;
1631         fictitious = entry->object.vm_object &&
1632                         (entry->object.vm_object->type == OBJT_DEVICE);
1633
1634         /*
1635          * Since the pages are wired down, we must be able to get their
1636          * mappings from the physical map system.
1637          */
1638         lwkt_gettoken(&vm_token);
1639         for (va = start; va < end; va += PAGE_SIZE) {
1640                 pa = pmap_extract(pmap, va);
1641                 if (pa != 0) {
1642                         pmap_change_wiring(pmap, va, FALSE);
1643                         if (!fictitious)
1644                                 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1645                 }
1646         }
1647         lwkt_reltoken(&vm_token);
1648 }
1649
1650 /*
1651  * Reduce the rate at which memory is allocated to a process based
1652  * on the perceived load on the VM system. As the load increases
1653  * the allocation burst rate goes down and the delay increases. 
1654  *
1655  * Rate limiting does not apply when faulting active or inactive
1656  * pages.  When faulting 'cache' pages, rate limiting only applies
1657  * if the system currently has a severe page deficit.
1658  *
1659  * XXX vm_pagesupply should be increased when a page is freed.
1660  *
1661  * We sleep up to 1/10 of a second.
1662  */
1663 static int
1664 vm_fault_ratelimit(struct vmspace *vmspace)
1665 {
1666         if (vm_load_enable == 0)
1667                 return(0);
1668         if (vmspace->vm_pagesupply > 0) {
1669                 --vmspace->vm_pagesupply;       /* SMP race ok */
1670                 return(0);
1671         }
1672 #ifdef INVARIANTS
1673         if (vm_load_debug) {
1674                 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1675                         vm_load, 
1676                         (1000 - vm_load ) / 10, vm_load * hz / 10000,
1677                         curproc->p_pid, curproc->p_comm);
1678         }
1679 #endif
1680         vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1681         return(vm_load * hz / 10000);
1682 }
1683
1684 /*
1685  * Copy all of the pages from a wired-down map entry to another.
1686  *
1687  * The source and destination maps must be locked for write.
1688  * The source map entry must be wired down (or be a sharing map
1689  * entry corresponding to a main map entry that is wired down).
1690  *
1691  * No other requirements.
1692  */
1693 void
1694 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1695                     vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1696 {
1697         vm_object_t dst_object;
1698         vm_object_t src_object;
1699         vm_ooffset_t dst_offset;
1700         vm_ooffset_t src_offset;
1701         vm_prot_t prot;
1702         vm_offset_t vaddr;
1703         vm_page_t dst_m;
1704         vm_page_t src_m;
1705
1706 #ifdef  lint
1707         src_map++;
1708 #endif  /* lint */
1709
1710         src_object = src_entry->object.vm_object;
1711         src_offset = src_entry->offset;
1712
1713         /*
1714          * Create the top-level object for the destination entry. (Doesn't
1715          * actually shadow anything - we copy the pages directly.)
1716          */
1717         vm_map_entry_allocate_object(dst_entry);
1718         dst_object = dst_entry->object.vm_object;
1719
1720         prot = dst_entry->max_protection;
1721
1722         /*
1723          * Loop through all of the pages in the entry's range, copying each
1724          * one from the source object (it should be there) to the destination
1725          * object.
1726          */
1727         for (vaddr = dst_entry->start, dst_offset = 0;
1728             vaddr < dst_entry->end;
1729             vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1730
1731                 /*
1732                  * Allocate a page in the destination object
1733                  */
1734                 do {
1735                         dst_m = vm_page_alloc(dst_object,
1736                                 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1737                         if (dst_m == NULL) {
1738                                 vm_wait(0);
1739                         }
1740                 } while (dst_m == NULL);
1741
1742                 /*
1743                  * Find the page in the source object, and copy it in.
1744                  * (Because the source is wired down, the page will be in
1745                  * memory.)
1746                  */
1747                 src_m = vm_page_lookup(src_object,
1748                         OFF_TO_IDX(dst_offset + src_offset));
1749                 if (src_m == NULL)
1750                         panic("vm_fault_copy_wired: page missing");
1751
1752                 vm_page_copy(src_m, dst_m);
1753                 vm_page_event(src_m, VMEVENT_COW);
1754
1755                 /*
1756                  * Enter it in the pmap...
1757                  */
1758
1759                 vm_page_flag_clear(dst_m, PG_ZERO);
1760                 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1761
1762                 /*
1763                  * Mark it no longer busy, and put it on the active list.
1764                  */
1765                 vm_page_activate(dst_m);
1766                 vm_page_wakeup(dst_m);
1767         }
1768 }
1769
1770 #if 0
1771
1772 /*
1773  * This routine checks around the requested page for other pages that
1774  * might be able to be faulted in.  This routine brackets the viable
1775  * pages for the pages to be paged in.
1776  *
1777  * Inputs:
1778  *      m, rbehind, rahead
1779  *
1780  * Outputs:
1781  *  marray (array of vm_page_t), reqpage (index of requested page)
1782  *
1783  * Return value:
1784  *  number of pages in marray
1785  */
1786 static int
1787 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1788                           vm_page_t *marray, int *reqpage)
1789 {
1790         int i,j;
1791         vm_object_t object;
1792         vm_pindex_t pindex, startpindex, endpindex, tpindex;
1793         vm_page_t rtm;
1794         int cbehind, cahead;
1795
1796         object = m->object;
1797         pindex = m->pindex;
1798
1799         /*
1800          * we don't fault-ahead for device pager
1801          */
1802         if (object->type == OBJT_DEVICE) {
1803                 *reqpage = 0;
1804                 marray[0] = m;
1805                 return 1;
1806         }
1807
1808         /*
1809          * if the requested page is not available, then give up now
1810          */
1811         if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1812                 *reqpage = 0;   /* not used by caller, fix compiler warn */
1813                 return 0;
1814         }
1815
1816         if ((cbehind == 0) && (cahead == 0)) {
1817                 *reqpage = 0;
1818                 marray[0] = m;
1819                 return 1;
1820         }
1821
1822         if (rahead > cahead) {
1823                 rahead = cahead;
1824         }
1825
1826         if (rbehind > cbehind) {
1827                 rbehind = cbehind;
1828         }
1829
1830         /*
1831          * Do not do any readahead if we have insufficient free memory.
1832          *
1833          * XXX code was broken disabled before and has instability
1834          * with this conditonal fixed, so shortcut for now.
1835          */
1836         if (burst_fault == 0 || vm_page_count_severe()) {
1837                 marray[0] = m;
1838                 *reqpage = 0;
1839                 return 1;
1840         }
1841
1842         /*
1843          * scan backward for the read behind pages -- in memory 
1844          *
1845          * Assume that if the page is not found an interrupt will not
1846          * create it.  Theoretically interrupts can only remove (busy)
1847          * pages, not create new associations.
1848          */
1849         if (pindex > 0) {
1850                 if (rbehind > pindex) {
1851                         rbehind = pindex;
1852                         startpindex = 0;
1853                 } else {
1854                         startpindex = pindex - rbehind;
1855                 }
1856
1857                 crit_enter();
1858                 lwkt_gettoken(&vm_token);
1859                 for (tpindex = pindex; tpindex > startpindex; --tpindex) {
1860                         if (vm_page_lookup(object, tpindex - 1))
1861                                 break;
1862                 }
1863
1864                 i = 0;
1865                 while (tpindex < pindex) {
1866                         rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1867                         if (rtm == NULL) {
1868                                 lwkt_reltoken(&vm_token);
1869                                 crit_exit();
1870                                 for (j = 0; j < i; j++) {
1871                                         vm_page_free(marray[j]);
1872                                 }
1873                                 marray[0] = m;
1874                                 *reqpage = 0;
1875                                 return 1;
1876                         }
1877                         marray[i] = rtm;
1878                         ++i;
1879                         ++tpindex;
1880                 }
1881                 lwkt_reltoken(&vm_token);
1882                 crit_exit();
1883         } else {
1884                 i = 0;
1885         }
1886
1887         /*
1888          * Assign requested page
1889          */
1890         marray[i] = m;
1891         *reqpage = i;
1892         ++i;
1893
1894         /*
1895          * Scan forwards for read-ahead pages
1896          */
1897         tpindex = pindex + 1;
1898         endpindex = tpindex + rahead;
1899         if (endpindex > object->size)
1900                 endpindex = object->size;
1901
1902         crit_enter();
1903         lwkt_gettoken(&vm_token);
1904         while (tpindex < endpindex) {
1905                 if (vm_page_lookup(object, tpindex))
1906                         break;
1907                 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1908                 if (rtm == NULL)
1909                         break;
1910                 marray[i] = rtm;
1911                 ++i;
1912                 ++tpindex;
1913         }
1914         lwkt_reltoken(&vm_token);
1915         crit_exit();
1916
1917         return (i);
1918 }
1919
1920 #endif
1921
1922 /*
1923  * vm_prefault() provides a quick way of clustering pagefaults into a
1924  * processes address space.  It is a "cousin" of pmap_object_init_pt,
1925  * except it runs at page fault time instead of mmap time.
1926  *
1927  * This code used to be per-platform pmap_prefault().  It is now
1928  * machine-independent and enhanced to also pre-fault zero-fill pages
1929  * (see vm.fast_fault) as well as make them writable, which greatly
1930  * reduces the number of page faults programs incur.
1931  *
1932  * Application performance when pre-faulting zero-fill pages is heavily
1933  * dependent on the application.  Very tiny applications like /bin/echo
1934  * lose a little performance while applications of any appreciable size
1935  * gain performance.  Prefaulting multiple pages also reduces SMP
1936  * congestion and can improve SMP performance significantly.
1937  *
1938  * NOTE!  prot may allow writing but this only applies to the top level
1939  *        object.  If we wind up mapping a page extracted from a backing
1940  *        object we have to make sure it is read-only.
1941  *
1942  * NOTE!  The caller has already handled any COW operations on the
1943  *        vm_map_entry via the normal fault code.  Do NOT call this
1944  *        shortcut unless the normal fault code has run on this entry.
1945  *
1946  * No other requirements.
1947  */
1948 #define PFBAK 4
1949 #define PFFOR 4
1950 #define PAGEORDER_SIZE (PFBAK+PFFOR)
1951
1952 static int vm_prefault_pageorder[] = {
1953         -PAGE_SIZE, PAGE_SIZE,
1954         -2 * PAGE_SIZE, 2 * PAGE_SIZE,
1955         -3 * PAGE_SIZE, 3 * PAGE_SIZE,
1956         -4 * PAGE_SIZE, 4 * PAGE_SIZE
1957 };
1958
1959 static void
1960 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot)
1961 {
1962         struct lwp *lp;
1963         vm_page_t m;
1964         vm_offset_t starta;
1965         vm_offset_t addr;
1966         vm_pindex_t index;
1967         vm_pindex_t pindex;
1968         vm_object_t object;
1969         int pprot;
1970         int i;
1971
1972         /*
1973          * We do not currently prefault mappings that use virtual page
1974          * tables.  We do not prefault foreign pmaps.
1975          */
1976         if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
1977                 return;
1978         lp = curthread->td_lwp;
1979         if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
1980                 return;
1981
1982         object = entry->object.vm_object;
1983
1984         starta = addra - PFBAK * PAGE_SIZE;
1985         if (starta < entry->start)
1986                 starta = entry->start;
1987         else if (starta > addra)
1988                 starta = 0;
1989
1990         /*
1991          * critical section protection is required to maintain the
1992          * page/object association, interrupts can free pages and remove
1993          * them from their objects.
1994          */
1995         crit_enter();
1996         lwkt_gettoken(&vm_token);
1997         for (i = 0; i < PAGEORDER_SIZE; i++) {
1998                 vm_object_t lobject;
1999                 int allocated = 0;
2000
2001                 addr = addra + vm_prefault_pageorder[i];
2002                 if (addr > addra + (PFFOR * PAGE_SIZE))
2003                         addr = 0;
2004
2005                 if (addr < starta || addr >= entry->end)
2006                         continue;
2007
2008                 if (pmap_prefault_ok(pmap, addr) == 0)
2009                         continue;
2010
2011                 /*
2012                  * Follow the VM object chain to obtain the page to be mapped
2013                  * into the pmap.
2014                  *
2015                  * If we reach the terminal object without finding a page
2016                  * and we determine it would be advantageous, then allocate
2017                  * a zero-fill page for the base object.  The base object
2018                  * is guaranteed to be OBJT_DEFAULT for this case.
2019                  *
2020                  * In order to not have to check the pager via *haspage*()
2021                  * we stop if any non-default object is encountered.  e.g.
2022                  * a vnode or swap object would stop the loop.
2023                  */
2024                 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2025                 lobject = object;
2026                 pindex = index;
2027                 pprot = prot;
2028
2029                 while ((m = vm_page_lookup(lobject, pindex)) == NULL) {
2030                         if (lobject->type != OBJT_DEFAULT)
2031                                 break;
2032                         if (lobject->backing_object == NULL) {
2033                                 if (vm_fast_fault == 0)
2034                                         break;
2035                                 if (vm_prefault_pageorder[i] < 0 ||
2036                                     (prot & VM_PROT_WRITE) == 0 ||
2037                                     vm_page_count_min(0)) {
2038                                         break;
2039                                 }
2040                                 /* note: allocate from base object */
2041                                 m = vm_page_alloc(object, index,
2042                                               VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
2043
2044                                 if ((m->flags & PG_ZERO) == 0) {
2045                                         vm_page_zero_fill(m);
2046                                 } else {
2047                                         vm_page_flag_clear(m, PG_ZERO);
2048                                         mycpu->gd_cnt.v_ozfod++;
2049                                 }
2050                                 mycpu->gd_cnt.v_zfod++;
2051                                 m->valid = VM_PAGE_BITS_ALL;
2052                                 allocated = 1;
2053                                 pprot = prot;
2054                                 /* lobject = object .. not needed */
2055                                 break;
2056                         }
2057                         if (lobject->backing_object_offset & PAGE_MASK)
2058                                 break;
2059                         pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2060                         lobject = lobject->backing_object;
2061                         pprot &= ~VM_PROT_WRITE;
2062                 }
2063                 /*
2064                  * NOTE: lobject now invalid (if we did a zero-fill we didn't
2065                  *       bother assigning lobject = object).
2066                  *
2067                  * Give-up if the page is not available.
2068                  */
2069                 if (m == NULL)
2070                         break;
2071
2072                 /*
2073                  * Do not conditionalize on PG_RAM.  If pages are present in
2074                  * the VM system we assume optimal caching.  If caching is
2075                  * not optimal the I/O gravy train will be restarted when we
2076                  * hit an unavailable page.  We do not want to try to restart
2077                  * the gravy train now because we really don't know how much
2078                  * of the object has been cached.  The cost for restarting
2079                  * the gravy train should be low (since accesses will likely
2080                  * be I/O bound anyway).
2081                  *
2082                  * The object must be marked dirty if we are mapping a
2083                  * writable page.
2084                  */
2085                 if (pprot & VM_PROT_WRITE)
2086                         vm_object_set_writeable_dirty(m->object);
2087
2088                 /*
2089                  * Enter the page into the pmap if appropriate.  If we had
2090                  * allocated the page we have to place it on a queue.  If not
2091                  * we just have to make sure it isn't on the cache queue
2092                  * (pages on the cache queue are not allowed to be mapped).
2093                  */
2094                 if (allocated) {
2095                         pmap_enter(pmap, addr, m, pprot, 0);
2096                         vm_page_deactivate(m);
2097                         vm_page_wakeup(m);
2098                 } else if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2099                     (m->busy == 0) &&
2100                     (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2101
2102                         if ((m->queue - m->pc) == PQ_CACHE) {
2103                                 vm_page_deactivate(m);
2104                         }
2105                         vm_page_busy(m);
2106                         pmap_enter(pmap, addr, m, pprot, 0);
2107                         vm_page_wakeup(m);
2108                 }
2109         }
2110         lwkt_reltoken(&vm_token);
2111         crit_exit();
2112 }