2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
10 * This code is derived from software contributed to Berkeley by
11 * The Mach Operating System project at Carnegie-Mellon University.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
44 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45 * All rights reserved.
47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49 * Permission to use, copy, modify and distribute this software and
50 * its documentation is hereby granted, provided that both the copyright
51 * notice and this permission notice appear in all copies of the
52 * software, derivative works or modified versions, and any portions
53 * thereof, and that both notices appear in supporting documentation.
55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59 * Carnegie Mellon requests users of this software to return to
61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
62 * School of Computer Science
63 * Carnegie Mellon University
64 * Pittsburgh PA 15213-3890
66 * any improvements or extensions that they make and grant Carnegie the
67 * rights to redistribute these changes.
69 * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
70 * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $
74 * Page fault handling module.
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
81 #include <sys/vnode.h>
82 #include <sys/resourcevar.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vkernel.h>
85 #include <sys/sfbuf.h>
87 #include <sys/sysctl.h>
90 #include <vm/vm_param.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_object.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_pageout.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_pager.h>
98 #include <vm/vnode_pager.h>
99 #include <vm/vm_extern.h>
101 #include <sys/thread2.h>
102 #include <vm/vm_page2.h>
110 vm_object_t first_object
;
111 vm_prot_t first_prot
;
113 vm_map_entry_t entry
;
114 int lookup_still_valid
;
123 static int vm_fast_fault
= 1;
124 SYSCTL_INT(_vm
, OID_AUTO
, fast_fault
, CTLFLAG_RW
, &vm_fast_fault
, 0, "");
125 static int debug_cluster
= 0;
126 SYSCTL_INT(_vm
, OID_AUTO
, debug_cluster
, CTLFLAG_RW
, &debug_cluster
, 0, "");
128 static int vm_fault_object(struct faultstate
*, vm_pindex_t
, vm_prot_t
);
129 static int vm_fault_vpagetable(struct faultstate
*, vm_pindex_t
*, vpte_t
, int);
131 static int vm_fault_additional_pages (vm_page_t
, int, int, vm_page_t
*, int *);
133 static int vm_fault_ratelimit(struct vmspace
*);
134 static void vm_prefault(pmap_t pmap
, vm_offset_t addra
, vm_map_entry_t entry
,
138 release_page(struct faultstate
*fs
)
140 vm_page_deactivate(fs
->m
);
141 vm_page_wakeup(fs
->m
);
146 unlock_map(struct faultstate
*fs
)
148 if (fs
->lookup_still_valid
&& fs
->map
) {
149 vm_map_lookup_done(fs
->map
, fs
->entry
, 0);
150 fs
->lookup_still_valid
= FALSE
;
155 * Clean up after a successful call to vm_fault_object() so another call
156 * to vm_fault_object() can be made.
159 _cleanup_successful_fault(struct faultstate
*fs
, int relock
)
161 if (fs
->object
!= fs
->first_object
) {
162 vm_page_free(fs
->first_m
);
163 vm_object_pip_wakeup(fs
->object
);
166 fs
->object
= fs
->first_object
;
167 if (relock
&& fs
->lookup_still_valid
== FALSE
) {
169 vm_map_lock_read(fs
->map
);
170 fs
->lookup_still_valid
= TRUE
;
175 _unlock_things(struct faultstate
*fs
, int dealloc
)
177 vm_object_pip_wakeup(fs
->first_object
);
178 _cleanup_successful_fault(fs
, 0);
180 vm_object_deallocate(fs
->first_object
);
181 fs
->first_object
= NULL
;
184 if (fs
->vp
!= NULL
) {
190 #define unlock_things(fs) _unlock_things(fs, 0)
191 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
192 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
197 * Determine if the pager for the current object *might* contain the page.
199 * We only need to try the pager if this is not a default object (default
200 * objects are zero-fill and have no real pager), and if we are not taking
201 * a wiring fault or if the FS entry is wired.
203 #define TRYPAGER(fs) \
204 (fs->object->type != OBJT_DEFAULT && \
205 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
210 * Handle a page fault occuring at the given address, requiring the given
211 * permissions, in the map specified. If successful, the page is inserted
212 * into the associated physical map.
214 * NOTE: The given address should be truncated to the proper page address.
216 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
217 * a standard error specifying why the fault is fatal is returned.
219 * The map in question must be referenced, and remains so.
220 * The caller may hold no locks.
223 vm_fault(vm_map_t map
, vm_offset_t vaddr
, vm_prot_t fault_type
, int fault_flags
)
226 vm_pindex_t first_pindex
;
227 struct faultstate fs
;
229 mycpu
->gd_cnt
.v_vm_faults
++;
233 fs
.fault_flags
= fault_flags
;
237 * Find the vm_map_entry representing the backing store and resolve
238 * the top level object and page index. This may have the side
239 * effect of executing a copy-on-write on the map entry and/or
240 * creating a shadow object, but will not COW any actual VM pages.
242 * On success fs.map is left read-locked and various other fields
243 * are initialized but not otherwise referenced or locked.
245 * NOTE! vm_map_lookup will try to upgrade the fault_type to
246 * VM_FAULT_WRITE if the map entry is a virtual page table and also
247 * writable, so we can set the 'A'accessed bit in the virtual page
251 result
= vm_map_lookup(&fs
.map
, vaddr
, fault_type
,
252 &fs
.entry
, &fs
.first_object
,
253 &first_pindex
, &fs
.first_prot
, &fs
.wired
);
256 * If the lookup failed or the map protections are incompatible,
257 * the fault generally fails. However, if the caller is trying
258 * to do a user wiring we have more work to do.
260 if (result
!= KERN_SUCCESS
) {
261 if (result
!= KERN_PROTECTION_FAILURE
)
263 if ((fs
.fault_flags
& VM_FAULT_WIRE_MASK
) != VM_FAULT_USER_WIRE
)
267 * If we are user-wiring a r/w segment, and it is COW, then
268 * we need to do the COW operation. Note that we don't
269 * currently COW RO sections now, because it is NOT desirable
270 * to COW .text. We simply keep .text from ever being COW'ed
271 * and take the heat that one cannot debug wired .text sections.
273 result
= vm_map_lookup(&fs
.map
, vaddr
,
274 VM_PROT_READ
|VM_PROT_WRITE
|
275 VM_PROT_OVERRIDE_WRITE
,
276 &fs
.entry
, &fs
.first_object
,
277 &first_pindex
, &fs
.first_prot
,
279 if (result
!= KERN_SUCCESS
)
283 * If we don't COW now, on a user wire, the user will never
284 * be able to write to the mapping. If we don't make this
285 * restriction, the bookkeeping would be nearly impossible.
287 if ((fs
.entry
->protection
& VM_PROT_WRITE
) == 0)
288 fs
.entry
->max_protection
&= ~VM_PROT_WRITE
;
292 * fs.map is read-locked
294 * Misc checks. Save the map generation number to detect races.
296 fs
.map_generation
= fs
.map
->timestamp
;
298 if (fs
.entry
->eflags
& MAP_ENTRY_NOFAULT
) {
299 panic("vm_fault: fault on nofault entry, addr: %lx",
304 * A system map entry may return a NULL object. No object means
305 * no pager means an unrecoverable kernel fault.
307 if (fs
.first_object
== NULL
) {
308 panic("vm_fault: unrecoverable fault at %p in entry %p",
309 (void *)vaddr
, fs
.entry
);
313 * Make a reference to this object to prevent its disposal while we
314 * are messing with it. Once we have the reference, the map is free
315 * to be diddled. Since objects reference their shadows (and copies),
316 * they will stay around as well.
318 * Bump the paging-in-progress count to prevent size changes (e.g.
319 * truncation operations) during I/O. This must be done after
320 * obtaining the vnode lock in order to avoid possible deadlocks.
322 vm_object_reference(fs
.first_object
);
323 fs
.vp
= vnode_pager_lock(fs
.first_object
);
324 vm_object_pip_add(fs
.first_object
, 1);
326 fs
.lookup_still_valid
= TRUE
;
328 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
331 * If the entry is wired we cannot change the page protection.
334 fault_type
= fs
.first_prot
;
337 * The page we want is at (first_object, first_pindex), but if the
338 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
339 * page table to figure out the actual pindex.
341 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
344 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
345 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
346 fs
.entry
->aux
.master_pde
,
348 if (result
== KERN_TRY_AGAIN
)
350 if (result
!= KERN_SUCCESS
)
355 * Now we have the actual (object, pindex), fault in the page. If
356 * vm_fault_object() fails it will unlock and deallocate the FS
357 * data. If it succeeds everything remains locked and fs->object
358 * will have an additinal PIP count if it is not equal to
361 * vm_fault_object will set fs->prot for the pmap operation. It is
362 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
363 * page can be safely written. However, it will force a read-only
364 * mapping for a read fault if the memory is managed by a virtual
367 result
= vm_fault_object(&fs
, first_pindex
, fault_type
);
369 if (result
== KERN_TRY_AGAIN
)
371 if (result
!= KERN_SUCCESS
)
375 * On success vm_fault_object() does not unlock or deallocate, and fs.m
376 * will contain a busied page.
378 * Enter the page into the pmap and do pmap-related adjustments.
380 pmap_enter(fs
.map
->pmap
, vaddr
, fs
.m
, fs
.prot
, fs
.wired
);
383 * Burst in a few more pages if possible. The fs.map should still
386 if (fault_flags
& VM_FAULT_BURST
) {
387 if ((fs
.fault_flags
& VM_FAULT_WIRE_MASK
) == 0 &&
389 vm_prefault(fs
.map
->pmap
, vaddr
, fs
.entry
, fs
.prot
);
394 vm_page_flag_clear(fs
.m
, PG_ZERO
);
395 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
398 * If the page is not wired down, then put it where the pageout daemon
401 if (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) {
405 vm_page_unwire(fs
.m
, 1);
407 vm_page_activate(fs
.m
);
410 if (curthread
->td_lwp
) {
412 curthread
->td_lwp
->lwp_ru
.ru_majflt
++;
414 curthread
->td_lwp
->lwp_ru
.ru_minflt
++;
419 * Unlock everything, and return
421 vm_page_wakeup(fs
.m
);
422 vm_object_deallocate(fs
.first_object
);
424 return (KERN_SUCCESS
);
428 * Fault in the specified virtual address in the current process map,
429 * returning a held VM page or NULL. See vm_fault_page() for more
433 vm_fault_page_quick(vm_offset_t va
, vm_prot_t fault_type
, int *errorp
)
435 struct lwp
*lp
= curthread
->td_lwp
;
438 m
= vm_fault_page(&lp
->lwp_vmspace
->vm_map
, va
,
439 fault_type
, VM_FAULT_NORMAL
, errorp
);
444 * Fault in the specified virtual address in the specified map, doing all
445 * necessary manipulation of the object store and all necessary I/O. Return
446 * a held VM page or NULL, and set *errorp. The related pmap is not
449 * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
450 * and marked PG_REFERENCED as well.
452 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
453 * error will be returned.
456 vm_fault_page(vm_map_t map
, vm_offset_t vaddr
, vm_prot_t fault_type
,
457 int fault_flags
, int *errorp
)
459 vm_pindex_t first_pindex
;
460 struct faultstate fs
;
462 vm_prot_t orig_fault_type
= fault_type
;
464 mycpu
->gd_cnt
.v_vm_faults
++;
468 fs
.fault_flags
= fault_flags
;
469 KKASSERT((fault_flags
& VM_FAULT_WIRE_MASK
) == 0);
473 * Find the vm_map_entry representing the backing store and resolve
474 * the top level object and page index. This may have the side
475 * effect of executing a copy-on-write on the map entry and/or
476 * creating a shadow object, but will not COW any actual VM pages.
478 * On success fs.map is left read-locked and various other fields
479 * are initialized but not otherwise referenced or locked.
481 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
482 * if the map entry is a virtual page table and also writable,
483 * so we can set the 'A'accessed bit in the virtual page table entry.
486 result
= vm_map_lookup(&fs
.map
, vaddr
, fault_type
,
487 &fs
.entry
, &fs
.first_object
,
488 &first_pindex
, &fs
.first_prot
, &fs
.wired
);
490 if (result
!= KERN_SUCCESS
) {
496 * fs.map is read-locked
498 * Misc checks. Save the map generation number to detect races.
500 fs
.map_generation
= fs
.map
->timestamp
;
502 if (fs
.entry
->eflags
& MAP_ENTRY_NOFAULT
) {
503 panic("vm_fault: fault on nofault entry, addr: %lx",
508 * A system map entry may return a NULL object. No object means
509 * no pager means an unrecoverable kernel fault.
511 if (fs
.first_object
== NULL
) {
512 panic("vm_fault: unrecoverable fault at %p in entry %p",
513 (void *)vaddr
, fs
.entry
);
517 * Make a reference to this object to prevent its disposal while we
518 * are messing with it. Once we have the reference, the map is free
519 * to be diddled. Since objects reference their shadows (and copies),
520 * they will stay around as well.
522 * Bump the paging-in-progress count to prevent size changes (e.g.
523 * truncation operations) during I/O. This must be done after
524 * obtaining the vnode lock in order to avoid possible deadlocks.
526 vm_object_reference(fs
.first_object
);
527 fs
.vp
= vnode_pager_lock(fs
.first_object
);
528 vm_object_pip_add(fs
.first_object
, 1);
530 fs
.lookup_still_valid
= TRUE
;
532 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
535 * If the entry is wired we cannot change the page protection.
538 fault_type
= fs
.first_prot
;
541 * The page we want is at (first_object, first_pindex), but if the
542 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
543 * page table to figure out the actual pindex.
545 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
548 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
549 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
550 fs
.entry
->aux
.master_pde
,
552 if (result
== KERN_TRY_AGAIN
)
554 if (result
!= KERN_SUCCESS
) {
561 * Now we have the actual (object, pindex), fault in the page. If
562 * vm_fault_object() fails it will unlock and deallocate the FS
563 * data. If it succeeds everything remains locked and fs->object
564 * will have an additinal PIP count if it is not equal to
567 result
= vm_fault_object(&fs
, first_pindex
, fault_type
);
569 if (result
== KERN_TRY_AGAIN
)
571 if (result
!= KERN_SUCCESS
) {
576 if ((orig_fault_type
& VM_PROT_WRITE
) &&
577 (fs
.prot
& VM_PROT_WRITE
) == 0) {
578 *errorp
= KERN_PROTECTION_FAILURE
;
579 unlock_and_deallocate(&fs
);
584 * On success vm_fault_object() does not unlock or deallocate, and fs.m
585 * will contain a busied page.
590 * Return a held page. We are not doing any pmap manipulation so do
591 * not set PG_MAPPED. However, adjust the page flags according to
592 * the fault type because the caller may not use a managed pmapping
593 * (so we don't want to lose the fact that the page will be dirtied
594 * if a write fault was specified).
597 vm_page_flag_clear(fs
.m
, PG_ZERO
);
598 if (fault_type
& VM_PROT_WRITE
)
602 * Update the pmap. We really only have to do this if a COW
603 * occured to replace the read-only page with the new page. For
604 * now just do it unconditionally. XXX
606 pmap_enter(fs
.map
->pmap
, vaddr
, fs
.m
, fs
.prot
, fs
.wired
);
607 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
610 * Unbusy the page by activating it. It remains held and will not
613 vm_page_activate(fs
.m
);
615 if (curthread
->td_lwp
) {
617 curthread
->td_lwp
->lwp_ru
.ru_majflt
++;
619 curthread
->td_lwp
->lwp_ru
.ru_minflt
++;
624 * Unlock everything, and return the held page.
626 vm_page_wakeup(fs
.m
);
627 vm_object_deallocate(fs
.first_object
);
634 * Fault in the specified (object,offset), dirty the returned page as
635 * needed. If the requested fault_type cannot be done NULL and an
639 vm_fault_object_page(vm_object_t object
, vm_ooffset_t offset
,
640 vm_prot_t fault_type
, int fault_flags
, int *errorp
)
643 vm_pindex_t first_pindex
;
644 struct faultstate fs
;
645 struct vm_map_entry entry
;
647 bzero(&entry
, sizeof(entry
));
648 entry
.object
.vm_object
= object
;
649 entry
.maptype
= VM_MAPTYPE_NORMAL
;
650 entry
.protection
= entry
.max_protection
= fault_type
;
654 fs
.fault_flags
= fault_flags
;
656 KKASSERT((fault_flags
& VM_FAULT_WIRE_MASK
) == 0);
660 fs
.first_object
= object
;
661 first_pindex
= OFF_TO_IDX(offset
);
663 fs
.first_prot
= fault_type
;
665 /*fs.map_generation = 0; unused */
668 * Make a reference to this object to prevent its disposal while we
669 * are messing with it. Once we have the reference, the map is free
670 * to be diddled. Since objects reference their shadows (and copies),
671 * they will stay around as well.
673 * Bump the paging-in-progress count to prevent size changes (e.g.
674 * truncation operations) during I/O. This must be done after
675 * obtaining the vnode lock in order to avoid possible deadlocks.
677 vm_object_reference(fs
.first_object
);
678 fs
.vp
= vnode_pager_lock(fs
.first_object
);
679 vm_object_pip_add(fs
.first_object
, 1);
681 fs
.lookup_still_valid
= TRUE
;
683 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
686 /* XXX future - ability to operate on VM object using vpagetable */
687 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
688 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
689 fs
.entry
->aux
.master_pde
,
691 if (result
== KERN_TRY_AGAIN
)
693 if (result
!= KERN_SUCCESS
) {
701 * Now we have the actual (object, pindex), fault in the page. If
702 * vm_fault_object() fails it will unlock and deallocate the FS
703 * data. If it succeeds everything remains locked and fs->object
704 * will have an additinal PIP count if it is not equal to
707 result
= vm_fault_object(&fs
, first_pindex
, fault_type
);
709 if (result
== KERN_TRY_AGAIN
)
711 if (result
!= KERN_SUCCESS
) {
716 if ((fault_type
& VM_PROT_WRITE
) && (fs
.prot
& VM_PROT_WRITE
) == 0) {
717 *errorp
= KERN_PROTECTION_FAILURE
;
718 unlock_and_deallocate(&fs
);
723 * On success vm_fault_object() does not unlock or deallocate, and fs.m
724 * will contain a busied page.
729 * Return a held page. We are not doing any pmap manipulation so do
730 * not set PG_MAPPED. However, adjust the page flags according to
731 * the fault type because the caller may not use a managed pmapping
732 * (so we don't want to lose the fact that the page will be dirtied
733 * if a write fault was specified).
736 vm_page_flag_clear(fs
.m
, PG_ZERO
);
737 if (fault_type
& VM_PROT_WRITE
)
741 * Indicate that the page was accessed.
743 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
746 * Unbusy the page by activating it. It remains held and will not
749 vm_page_activate(fs
.m
);
751 if (curthread
->td_lwp
) {
753 mycpu
->gd_cnt
.v_vm_faults
++;
754 curthread
->td_lwp
->lwp_ru
.ru_majflt
++;
756 curthread
->td_lwp
->lwp_ru
.ru_minflt
++;
761 * Unlock everything, and return the held page.
763 vm_page_wakeup(fs
.m
);
764 vm_object_deallocate(fs
.first_object
);
771 * Translate the virtual page number (first_pindex) that is relative
772 * to the address space into a logical page number that is relative to the
773 * backing object. Use the virtual page table pointed to by (vpte).
775 * This implements an N-level page table. Any level can terminate the
776 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
777 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
781 vm_fault_vpagetable(struct faultstate
*fs
, vm_pindex_t
*pindex
,
782 vpte_t vpte
, int fault_type
)
785 int vshift
= 32 - PAGE_SHIFT
; /* page index bits remaining */
786 int result
= KERN_SUCCESS
;
791 * We cannot proceed if the vpte is not valid, not readable
792 * for a read fault, or not writable for a write fault.
794 if ((vpte
& VPTE_V
) == 0) {
795 unlock_and_deallocate(fs
);
796 return (KERN_FAILURE
);
798 if ((fault_type
& VM_PROT_READ
) && (vpte
& VPTE_R
) == 0) {
799 unlock_and_deallocate(fs
);
800 return (KERN_FAILURE
);
802 if ((fault_type
& VM_PROT_WRITE
) && (vpte
& VPTE_W
) == 0) {
803 unlock_and_deallocate(fs
);
804 return (KERN_FAILURE
);
806 if ((vpte
& VPTE_PS
) || vshift
== 0)
808 KKASSERT(vshift
>= VPTE_PAGE_BITS
);
811 * Get the page table page. Nominally we only read the page
812 * table, but since we are actively setting VPTE_M and VPTE_A,
813 * tell vm_fault_object() that we are writing it.
815 * There is currently no real need to optimize this.
817 result
= vm_fault_object(fs
, vpte
>> PAGE_SHIFT
,
818 VM_PROT_READ
|VM_PROT_WRITE
);
819 if (result
!= KERN_SUCCESS
)
823 * Process the returned fs.m and look up the page table
824 * entry in the page table page.
826 vshift
-= VPTE_PAGE_BITS
;
827 sf
= sf_buf_alloc(fs
->m
, SFB_CPUPRIVATE
);
828 ptep
= ((vpte_t
*)sf_buf_kva(sf
) +
829 ((*pindex
>> vshift
) & VPTE_PAGE_MASK
));
833 * Page table write-back. If the vpte is valid for the
834 * requested operation, do a write-back to the page table.
836 * XXX VPTE_M is not set properly for page directory pages.
837 * It doesn't get set in the page directory if the page table
838 * is modified during a read access.
840 if ((fault_type
& VM_PROT_WRITE
) && (vpte
& VPTE_V
) &&
842 if ((vpte
& (VPTE_M
|VPTE_A
)) != (VPTE_M
|VPTE_A
)) {
843 atomic_set_int(ptep
, VPTE_M
|VPTE_A
);
844 vm_page_dirty(fs
->m
);
847 if ((fault_type
& VM_PROT_READ
) && (vpte
& VPTE_V
) &&
849 if ((vpte
& VPTE_A
) == 0) {
850 atomic_set_int(ptep
, VPTE_A
);
851 vm_page_dirty(fs
->m
);
855 vm_page_flag_set(fs
->m
, PG_REFERENCED
);
856 vm_page_activate(fs
->m
);
857 vm_page_wakeup(fs
->m
);
858 cleanup_successful_fault(fs
);
861 * Combine remaining address bits with the vpte.
863 *pindex
= (vpte
>> PAGE_SHIFT
) +
864 (*pindex
& ((1 << vshift
) - 1));
865 return (KERN_SUCCESS
);
870 * Do all operations required to fault-in (fs.first_object, pindex). Run
871 * through the shadow chain as necessary and do required COW or virtual
872 * copy operations. The caller has already fully resolved the vm_map_entry
873 * and, if appropriate, has created a copy-on-write layer. All we need to
874 * do is iterate the object chain.
876 * On failure (fs) is unlocked and deallocated and the caller may return or
877 * retry depending on the failure code. On success (fs) is NOT unlocked or
878 * deallocated, fs.m will contained a resolved, busied page, and fs.object
879 * will have an additional PIP count if it is not equal to fs.first_object.
883 vm_fault_object(struct faultstate
*fs
,
884 vm_pindex_t first_pindex
, vm_prot_t fault_type
)
886 vm_object_t next_object
;
889 fs
->prot
= fs
->first_prot
;
890 fs
->object
= fs
->first_object
;
891 pindex
= first_pindex
;
894 * If a read fault occurs we try to make the page writable if
895 * possible. There are three cases where we cannot make the
896 * page mapping writable:
898 * (1) The mapping is read-only or the VM object is read-only,
899 * fs->prot above will simply not have VM_PROT_WRITE set.
901 * (2) If the mapping is a virtual page table we need to be able
902 * to detect writes so we can set VPTE_M in the virtual page
905 * (3) If the VM page is read-only or copy-on-write, upgrading would
906 * just result in an unnecessary COW fault.
908 * VM_PROT_VPAGED is set if faulting via a virtual page table and
909 * causes adjustments to the 'M'odify bit to also turn off write
910 * access to force a re-fault.
912 if (fs
->entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
913 if ((fault_type
& VM_PROT_WRITE
) == 0)
914 fs
->prot
&= ~VM_PROT_WRITE
;
919 * If the object is dead, we stop here
921 if (fs
->object
->flags
& OBJ_DEAD
) {
922 unlock_and_deallocate(fs
);
923 return (KERN_PROTECTION_FAILURE
);
927 * See if page is resident. spl protection is required
928 * to avoid an interrupt unbusy/free race against our
929 * lookup. We must hold the protection through a page
930 * allocation or busy.
933 fs
->m
= vm_page_lookup(fs
->object
, pindex
);
937 * Wait/Retry if the page is busy. We have to do this
938 * if the page is busy via either PG_BUSY or
939 * vm_page_t->busy because the vm_pager may be using
940 * vm_page_t->busy for pageouts ( and even pageins if
941 * it is the vnode pager ), and we could end up trying
942 * to pagein and pageout the same page simultaneously.
944 * We can theoretically allow the busy case on a read
945 * fault if the page is marked valid, but since such
946 * pages are typically already pmap'd, putting that
947 * special case in might be more effort then it is
948 * worth. We cannot under any circumstances mess
949 * around with a vm_page_t->busy page except, perhaps,
952 if ((fs
->m
->flags
& PG_BUSY
) || fs
->m
->busy
) {
954 vm_page_sleep_busy(fs
->m
, TRUE
, "vmpfw");
955 mycpu
->gd_cnt
.v_intrans
++;
956 vm_object_deallocate(fs
->first_object
);
957 fs
->first_object
= NULL
;
959 return (KERN_TRY_AGAIN
);
963 * If reactivating a page from PQ_CACHE we may have
966 queue
= fs
->m
->queue
;
967 vm_page_unqueue_nowakeup(fs
->m
);
969 if ((queue
- fs
->m
->pc
) == PQ_CACHE
&&
970 vm_page_count_severe()) {
971 vm_page_activate(fs
->m
);
972 unlock_and_deallocate(fs
);
975 return (KERN_TRY_AGAIN
);
979 * Mark page busy for other processes, and the
980 * pagedaemon. If it still isn't completely valid
981 * (readable), or if a read-ahead-mark is set on
982 * the VM page, jump to readrest, else we found the
983 * page and can return.
985 * We can release the spl once we have marked the
991 if (fs
->m
->object
!= &kernel_object
) {
992 if ((fs
->m
->valid
& VM_PAGE_BITS_ALL
) !=
996 if (fs
->m
->flags
& PG_RAM
) {
999 vm_page_flag_clear(fs
->m
, PG_RAM
);
1003 break; /* break to PAGE HAS BEEN FOUND */
1007 * Page is not resident, If this is the search termination
1008 * or the pager might contain the page, allocate a new page.
1010 * NOTE: We are still in a critical section.
1012 if (TRYPAGER(fs
) || fs
->object
== fs
->first_object
) {
1014 * If the page is beyond the object size we fail
1016 if (pindex
>= fs
->object
->size
) {
1018 unlock_and_deallocate(fs
);
1019 return (KERN_PROTECTION_FAILURE
);
1025 if (fs
->didlimit
== 0 && curproc
!= NULL
) {
1028 limticks
= vm_fault_ratelimit(curproc
->p_vmspace
);
1031 unlock_and_deallocate(fs
);
1032 tsleep(curproc
, 0, "vmrate", limticks
);
1034 return (KERN_TRY_AGAIN
);
1039 * Allocate a new page for this object/offset pair.
1042 if (!vm_page_count_severe()) {
1043 fs
->m
= vm_page_alloc(fs
->object
, pindex
,
1044 (fs
->vp
|| fs
->object
->backing_object
) ? VM_ALLOC_NORMAL
: VM_ALLOC_NORMAL
| VM_ALLOC_ZERO
);
1046 if (fs
->m
== NULL
) {
1048 unlock_and_deallocate(fs
);
1050 return (KERN_TRY_AGAIN
);
1057 * We have found an invalid or partially valid page, a
1058 * potentially fully valid page with a read-ahead mark,
1059 * or we have allocated a new page.
1061 * Attempt to fault-in the page if there is a chance that the
1062 * pager has it, and potentially fault in additional pages
1065 * We are NOT in splvm here and if TRYPAGER is true then
1066 * fs.m will be non-NULL and will be PG_BUSY for us.
1072 u_char behavior
= vm_map_entry_behavior(fs
->entry
);
1074 if (behavior
== MAP_ENTRY_BEHAV_RANDOM
)
1080 * If sequential access is detected then attempt
1081 * to deactivate/cache pages behind the scan to
1082 * prevent resource hogging.
1084 * Use of PG_RAM to detect sequential access
1085 * also simulates multi-zone sequential access
1086 * detection for free.
1088 * NOTE: Partially valid dirty pages cannot be
1089 * deactivated without causing NFS picemeal
1092 if ((fs
->first_object
->type
!= OBJT_DEVICE
) &&
1093 (behavior
== MAP_ENTRY_BEHAV_SEQUENTIAL
||
1094 (behavior
!= MAP_ENTRY_BEHAV_RANDOM
&&
1095 (fs
->m
->flags
& PG_RAM
)))
1097 vm_pindex_t scan_pindex
;
1098 int scan_count
= 16;
1100 if (first_pindex
< 16) {
1104 scan_pindex
= first_pindex
- 16;
1105 if (scan_pindex
< 16)
1106 scan_count
= scan_pindex
;
1112 while (scan_count
) {
1115 mt
= vm_page_lookup(fs
->first_object
,
1118 (mt
->valid
!= VM_PAGE_BITS_ALL
)) {
1122 (mt
->flags
& (PG_BUSY
| PG_FICTITIOUS
| PG_UNMANAGED
)) ||
1128 vm_page_test_dirty(mt
);
1133 vm_page_deactivate(mt
);
1148 * Avoid deadlocking against the map when doing I/O.
1149 * fs.object and the page is PG_BUSY'd.
1154 * Acquire the page data. We still hold a ref on
1155 * fs.object and the page has been PG_BUSY's.
1157 * The pager may replace the page (for example, in
1158 * order to enter a fictitious page into the
1159 * object). If it does so it is responsible for
1160 * cleaning up the passed page and properly setting
1161 * the new page PG_BUSY.
1163 if (vm_pager_has_page(fs
->object
, pindex
)) {
1164 rv
= vm_pager_get_page(fs
->object
, &fs
->m
,
1170 if (rv
== VM_PAGER_OK
) {
1172 * Relookup in case pager changed page. Pager
1173 * is responsible for disposition of old page
1176 * XXX other code segments do relookups too.
1177 * It's a bad abstraction that needs to be
1180 fs
->m
= vm_page_lookup(fs
->object
, pindex
);
1181 if (fs
->m
== NULL
) {
1182 unlock_and_deallocate(fs
);
1183 return (KERN_TRY_AGAIN
);
1187 break; /* break to PAGE HAS BEEN FOUND */
1191 * Remove the bogus page (which does not exist at this
1192 * object/offset); before doing so, we must get back
1193 * our object lock to preserve our invariant.
1195 * Also wake up any other process that may want to bring
1198 * If this is the top-level object, we must leave the
1199 * busy page to prevent another process from rushing
1200 * past us, and inserting the page in that object at
1201 * the same time that we are.
1203 if (rv
== VM_PAGER_ERROR
) {
1205 kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc
->p_pid
, curproc
->p_comm
);
1207 kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread
, curproc
->p_comm
);
1211 * Data outside the range of the pager or an I/O error
1213 * The page may have been wired during the pagein,
1214 * e.g. by the buffer cache, and cannot simply be
1215 * freed. Call vnode_pager_freepage() to deal with it.
1218 * XXX - the check for kernel_map is a kludge to work
1219 * around having the machine panic on a kernel space
1220 * fault w/ I/O error.
1222 if (((fs
->map
!= &kernel_map
) &&
1223 (rv
== VM_PAGER_ERROR
)) || (rv
== VM_PAGER_BAD
)) {
1224 vnode_pager_freepage(fs
->m
);
1226 unlock_and_deallocate(fs
);
1227 if (rv
== VM_PAGER_ERROR
)
1228 return (KERN_FAILURE
);
1230 return (KERN_PROTECTION_FAILURE
);
1233 if (fs
->object
!= fs
->first_object
) {
1234 vnode_pager_freepage(fs
->m
);
1237 * XXX - we cannot just fall out at this
1238 * point, m has been freed and is invalid!
1244 * We get here if the object has a default pager (or unwiring)
1245 * or the pager doesn't have the page.
1247 if (fs
->object
== fs
->first_object
)
1248 fs
->first_m
= fs
->m
;
1251 * Move on to the next object. Lock the next object before
1252 * unlocking the current one.
1254 pindex
+= OFF_TO_IDX(fs
->object
->backing_object_offset
);
1255 next_object
= fs
->object
->backing_object
;
1256 if (next_object
== NULL
) {
1258 * If there's no object left, fill the page in the top
1259 * object with zeros.
1261 if (fs
->object
!= fs
->first_object
) {
1262 vm_object_pip_wakeup(fs
->object
);
1264 fs
->object
= fs
->first_object
;
1265 pindex
= first_pindex
;
1266 fs
->m
= fs
->first_m
;
1271 * Zero the page if necessary and mark it valid.
1273 if ((fs
->m
->flags
& PG_ZERO
) == 0) {
1274 vm_page_zero_fill(fs
->m
);
1276 mycpu
->gd_cnt
.v_ozfod
++;
1278 mycpu
->gd_cnt
.v_zfod
++;
1279 fs
->m
->valid
= VM_PAGE_BITS_ALL
;
1280 break; /* break to PAGE HAS BEEN FOUND */
1282 if (fs
->object
!= fs
->first_object
) {
1283 vm_object_pip_wakeup(fs
->object
);
1285 KASSERT(fs
->object
!= next_object
, ("object loop %p", next_object
));
1286 fs
->object
= next_object
;
1287 vm_object_pip_add(fs
->object
, 1);
1292 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1295 * If the page is being written, but isn't already owned by the
1296 * top-level object, we have to copy it into a new page owned by the
1299 KASSERT((fs
->m
->flags
& PG_BUSY
) != 0,
1300 ("vm_fault: not busy after main loop"));
1302 if (fs
->object
!= fs
->first_object
) {
1304 * We only really need to copy if we want to write it.
1306 if (fault_type
& VM_PROT_WRITE
) {
1308 * This allows pages to be virtually copied from a
1309 * backing_object into the first_object, where the
1310 * backing object has no other refs to it, and cannot
1311 * gain any more refs. Instead of a bcopy, we just
1312 * move the page from the backing object to the
1313 * first object. Note that we must mark the page
1314 * dirty in the first object so that it will go out
1315 * to swap when needed.
1319 * Map, if present, has not changed
1322 fs
->map_generation
== fs
->map
->timestamp
) &&
1324 * Only one shadow object
1326 (fs
->object
->shadow_count
== 1) &&
1328 * No COW refs, except us
1330 (fs
->object
->ref_count
== 1) &&
1332 * No one else can look this object up
1334 (fs
->object
->handle
== NULL
) &&
1336 * No other ways to look the object up
1338 ((fs
->object
->type
== OBJT_DEFAULT
) ||
1339 (fs
->object
->type
== OBJT_SWAP
)) &&
1341 * We don't chase down the shadow chain
1343 (fs
->object
== fs
->first_object
->backing_object
) &&
1346 * grab the lock if we need to
1348 (fs
->lookup_still_valid
||
1350 lockmgr(&fs
->map
->lock
, LK_EXCLUSIVE
|LK_NOWAIT
) == 0)
1353 fs
->lookup_still_valid
= 1;
1355 * get rid of the unnecessary page
1357 vm_page_protect(fs
->first_m
, VM_PROT_NONE
);
1358 vm_page_free(fs
->first_m
);
1362 * grab the page and put it into the
1363 * process'es object. The page is
1364 * automatically made dirty.
1366 vm_page_rename(fs
->m
, fs
->first_object
, first_pindex
);
1367 fs
->first_m
= fs
->m
;
1368 vm_page_busy(fs
->first_m
);
1370 mycpu
->gd_cnt
.v_cow_optim
++;
1373 * Oh, well, lets copy it.
1375 vm_page_copy(fs
->m
, fs
->first_m
);
1376 vm_page_event(fs
->m
, VMEVENT_COW
);
1381 * We no longer need the old page or object.
1387 * fs->object != fs->first_object due to above
1390 vm_object_pip_wakeup(fs
->object
);
1393 * Only use the new page below...
1396 mycpu
->gd_cnt
.v_cow_faults
++;
1397 fs
->m
= fs
->first_m
;
1398 fs
->object
= fs
->first_object
;
1399 pindex
= first_pindex
;
1402 * If it wasn't a write fault avoid having to copy
1403 * the page by mapping it read-only.
1405 fs
->prot
&= ~VM_PROT_WRITE
;
1410 * We may have had to unlock a map to do I/O. If we did then
1411 * lookup_still_valid will be FALSE. If the map generation count
1412 * also changed then all sorts of things could have happened while
1413 * we were doing the I/O and we need to retry.
1416 if (!fs
->lookup_still_valid
&&
1418 (fs
->map
->timestamp
!= fs
->map_generation
)) {
1420 unlock_and_deallocate(fs
);
1421 return (KERN_TRY_AGAIN
);
1425 * If the fault is a write, we know that this page is being
1426 * written NOW so dirty it explicitly to save on pmap_is_modified()
1429 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1430 * if the page is already dirty to prevent data written with
1431 * the expectation of being synced from not being synced.
1432 * Likewise if this entry does not request NOSYNC then make
1433 * sure the page isn't marked NOSYNC. Applications sharing
1434 * data should use the same flags to avoid ping ponging.
1436 * Also tell the backing pager, if any, that it should remove
1437 * any swap backing since the page is now dirty.
1439 if (fs
->prot
& VM_PROT_WRITE
) {
1440 vm_object_set_writeable_dirty(fs
->m
->object
);
1441 if (fs
->entry
->eflags
& MAP_ENTRY_NOSYNC
) {
1442 if (fs
->m
->dirty
== 0)
1443 vm_page_flag_set(fs
->m
, PG_NOSYNC
);
1445 vm_page_flag_clear(fs
->m
, PG_NOSYNC
);
1447 if (fs
->fault_flags
& VM_FAULT_DIRTY
) {
1449 vm_page_dirty(fs
->m
);
1450 vm_pager_page_unswapped(fs
->m
);
1456 * Page had better still be busy. We are still locked up and
1457 * fs->object will have another PIP reference if it is not equal
1458 * to fs->first_object.
1460 KASSERT(fs
->m
->flags
& PG_BUSY
,
1461 ("vm_fault: page %p not busy!", fs
->m
));
1464 * Sanity check: page must be completely valid or it is not fit to
1465 * map into user space. vm_pager_get_pages() ensures this.
1467 if (fs
->m
->valid
!= VM_PAGE_BITS_ALL
) {
1468 vm_page_zero_invalid(fs
->m
, TRUE
);
1469 kprintf("Warning: page %p partially invalid on fault\n", fs
->m
);
1472 return (KERN_SUCCESS
);
1476 * Wire down a range of virtual addresses in a map. The entry in question
1477 * should be marked in-transition and the map must be locked. We must
1478 * release the map temporarily while faulting-in the page to avoid a
1479 * deadlock. Note that the entry may be clipped while we are blocked but
1480 * will never be freed.
1483 vm_fault_wire(vm_map_t map
, vm_map_entry_t entry
, boolean_t user_wire
)
1485 boolean_t fictitious
;
1493 pmap
= vm_map_pmap(map
);
1494 start
= entry
->start
;
1496 fictitious
= entry
->object
.vm_object
&&
1497 (entry
->object
.vm_object
->type
== OBJT_DEVICE
);
1503 * We simulate a fault to get the page and enter it in the physical
1506 for (va
= start
; va
< end
; va
+= PAGE_SIZE
) {
1508 rv
= vm_fault(map
, va
, VM_PROT_READ
,
1509 VM_FAULT_USER_WIRE
);
1511 rv
= vm_fault(map
, va
, VM_PROT_READ
|VM_PROT_WRITE
,
1512 VM_FAULT_CHANGE_WIRING
);
1515 while (va
> start
) {
1517 if ((pa
= pmap_extract(pmap
, va
)) == 0)
1519 pmap_change_wiring(pmap
, va
, FALSE
);
1521 vm_page_unwire(PHYS_TO_VM_PAGE(pa
), 1);
1528 return (KERN_SUCCESS
);
1532 * Unwire a range of virtual addresses in a map. The map should be
1536 vm_fault_unwire(vm_map_t map
, vm_map_entry_t entry
)
1538 boolean_t fictitious
;
1545 pmap
= vm_map_pmap(map
);
1546 start
= entry
->start
;
1548 fictitious
= entry
->object
.vm_object
&&
1549 (entry
->object
.vm_object
->type
== OBJT_DEVICE
);
1552 * Since the pages are wired down, we must be able to get their
1553 * mappings from the physical map system.
1555 for (va
= start
; va
< end
; va
+= PAGE_SIZE
) {
1556 pa
= pmap_extract(pmap
, va
);
1558 pmap_change_wiring(pmap
, va
, FALSE
);
1560 vm_page_unwire(PHYS_TO_VM_PAGE(pa
), 1);
1566 * Reduce the rate at which memory is allocated to a process based
1567 * on the perceived load on the VM system. As the load increases
1568 * the allocation burst rate goes down and the delay increases.
1570 * Rate limiting does not apply when faulting active or inactive
1571 * pages. When faulting 'cache' pages, rate limiting only applies
1572 * if the system currently has a severe page deficit.
1574 * XXX vm_pagesupply should be increased when a page is freed.
1576 * We sleep up to 1/10 of a second.
1579 vm_fault_ratelimit(struct vmspace
*vmspace
)
1581 if (vm_load_enable
== 0)
1583 if (vmspace
->vm_pagesupply
> 0) {
1584 --vmspace
->vm_pagesupply
;
1588 if (vm_load_debug
) {
1589 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1591 (1000 - vm_load
) / 10, vm_load
* hz
/ 10000,
1592 curproc
->p_pid
, curproc
->p_comm
);
1595 vmspace
->vm_pagesupply
= (1000 - vm_load
) / 10;
1596 return(vm_load
* hz
/ 10000);
1601 * vm_fault_copy_entry
1603 * Copy all of the pages from a wired-down map entry to another.
1605 * In/out conditions:
1606 * The source and destination maps must be locked for write.
1607 * The source map entry must be wired down (or be a sharing map
1608 * entry corresponding to a main map entry that is wired down).
1612 vm_fault_copy_entry(vm_map_t dst_map
, vm_map_t src_map
,
1613 vm_map_entry_t dst_entry
, vm_map_entry_t src_entry
)
1615 vm_object_t dst_object
;
1616 vm_object_t src_object
;
1617 vm_ooffset_t dst_offset
;
1618 vm_ooffset_t src_offset
;
1628 src_object
= src_entry
->object
.vm_object
;
1629 src_offset
= src_entry
->offset
;
1632 * Create the top-level object for the destination entry. (Doesn't
1633 * actually shadow anything - we copy the pages directly.)
1635 vm_map_entry_allocate_object(dst_entry
);
1636 dst_object
= dst_entry
->object
.vm_object
;
1638 prot
= dst_entry
->max_protection
;
1641 * Loop through all of the pages in the entry's range, copying each
1642 * one from the source object (it should be there) to the destination
1645 for (vaddr
= dst_entry
->start
, dst_offset
= 0;
1646 vaddr
< dst_entry
->end
;
1647 vaddr
+= PAGE_SIZE
, dst_offset
+= PAGE_SIZE
) {
1650 * Allocate a page in the destination object
1653 dst_m
= vm_page_alloc(dst_object
,
1654 OFF_TO_IDX(dst_offset
), VM_ALLOC_NORMAL
);
1655 if (dst_m
== NULL
) {
1658 } while (dst_m
== NULL
);
1661 * Find the page in the source object, and copy it in.
1662 * (Because the source is wired down, the page will be in
1665 src_m
= vm_page_lookup(src_object
,
1666 OFF_TO_IDX(dst_offset
+ src_offset
));
1668 panic("vm_fault_copy_wired: page missing");
1670 vm_page_copy(src_m
, dst_m
);
1671 vm_page_event(src_m
, VMEVENT_COW
);
1674 * Enter it in the pmap...
1677 vm_page_flag_clear(dst_m
, PG_ZERO
);
1678 pmap_enter(dst_map
->pmap
, vaddr
, dst_m
, prot
, FALSE
);
1681 * Mark it no longer busy, and put it on the active list.
1683 vm_page_activate(dst_m
);
1684 vm_page_wakeup(dst_m
);
1691 * This routine checks around the requested page for other pages that
1692 * might be able to be faulted in. This routine brackets the viable
1693 * pages for the pages to be paged in.
1696 * m, rbehind, rahead
1699 * marray (array of vm_page_t), reqpage (index of requested page)
1702 * number of pages in marray
1705 vm_fault_additional_pages(vm_page_t m
, int rbehind
, int rahead
,
1706 vm_page_t
*marray
, int *reqpage
)
1710 vm_pindex_t pindex
, startpindex
, endpindex
, tpindex
;
1712 int cbehind
, cahead
;
1718 * we don't fault-ahead for device pager
1720 if (object
->type
== OBJT_DEVICE
) {
1727 * if the requested page is not available, then give up now
1729 if (!vm_pager_has_page(object
, pindex
, &cbehind
, &cahead
)) {
1730 *reqpage
= 0; /* not used by caller, fix compiler warn */
1734 if ((cbehind
== 0) && (cahead
== 0)) {
1740 if (rahead
> cahead
) {
1744 if (rbehind
> cbehind
) {
1749 * Do not do any readahead if we have insufficient free memory.
1751 * XXX code was broken disabled before and has instability
1752 * with this conditonal fixed, so shortcut for now.
1754 if (burst_fault
== 0 || vm_page_count_severe()) {
1761 * scan backward for the read behind pages -- in memory
1763 * Assume that if the page is not found an interrupt will not
1764 * create it. Theoretically interrupts can only remove (busy)
1765 * pages, not create new associations.
1768 if (rbehind
> pindex
) {
1772 startpindex
= pindex
- rbehind
;
1776 for (tpindex
= pindex
; tpindex
> startpindex
; --tpindex
) {
1777 if (vm_page_lookup(object
, tpindex
- 1))
1782 while (tpindex
< pindex
) {
1783 rtm
= vm_page_alloc(object
, tpindex
, VM_ALLOC_SYSTEM
);
1786 for (j
= 0; j
< i
; j
++) {
1787 vm_page_free(marray
[j
]);
1803 * Assign requested page
1810 * Scan forwards for read-ahead pages
1812 tpindex
= pindex
+ 1;
1813 endpindex
= tpindex
+ rahead
;
1814 if (endpindex
> object
->size
)
1815 endpindex
= object
->size
;
1818 while (tpindex
< endpindex
) {
1819 if (vm_page_lookup(object
, tpindex
))
1821 rtm
= vm_page_alloc(object
, tpindex
, VM_ALLOC_SYSTEM
);
1836 * vm_prefault() provides a quick way of clustering pagefaults into a
1837 * processes address space. It is a "cousin" of pmap_object_init_pt,
1838 * except it runs at page fault time instead of mmap time.
1840 * This code used to be per-platform pmap_prefault(). It is now
1841 * machine-independent and enhanced to also pre-fault zero-fill pages
1842 * (see vm.fast_fault) as well as make them writable, which greatly
1843 * reduces the number of page faults programs incur.
1845 * Application performance when pre-faulting zero-fill pages is heavily
1846 * dependent on the application. Very tiny applications like /bin/echo
1847 * lose a little performance while applications of any appreciable size
1848 * gain performance. Prefaulting multiple pages also reduces SMP
1849 * congestion and can improve SMP performance significantly.
1851 * NOTE! prot may allow writing but this only applies to the top level
1852 * object. If we wind up mapping a page extracted from a backing
1853 * object we have to make sure it is read-only.
1855 * NOTE! The caller has already handled any COW operations on the
1856 * vm_map_entry via the normal fault code. Do NOT call this
1857 * shortcut unless the normal fault code has run on this entry.
1861 #define PAGEORDER_SIZE (PFBAK+PFFOR)
1863 static int vm_prefault_pageorder
[] = {
1864 -PAGE_SIZE
, PAGE_SIZE
,
1865 -2 * PAGE_SIZE
, 2 * PAGE_SIZE
,
1866 -3 * PAGE_SIZE
, 3 * PAGE_SIZE
,
1867 -4 * PAGE_SIZE
, 4 * PAGE_SIZE
1871 vm_prefault(pmap_t pmap
, vm_offset_t addra
, vm_map_entry_t entry
, int prot
)
1884 * We do not currently prefault mappings that use virtual page
1885 * tables. We do not prefault foreign pmaps.
1887 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
)
1889 lp
= curthread
->td_lwp
;
1890 if (lp
== NULL
|| (pmap
!= vmspace_pmap(lp
->lwp_vmspace
)))
1893 object
= entry
->object
.vm_object
;
1895 starta
= addra
- PFBAK
* PAGE_SIZE
;
1896 if (starta
< entry
->start
)
1897 starta
= entry
->start
;
1898 else if (starta
> addra
)
1902 * critical section protection is required to maintain the
1903 * page/object association, interrupts can free pages and remove
1904 * them from their objects.
1907 for (i
= 0; i
< PAGEORDER_SIZE
; i
++) {
1908 vm_object_t lobject
;
1910 addr
= addra
+ vm_prefault_pageorder
[i
];
1911 if (addr
> addra
+ (PFFOR
* PAGE_SIZE
))
1914 if (addr
< starta
|| addr
>= entry
->end
)
1917 if (pmap_prefault_ok(pmap
, addr
) == 0)
1921 * Follow the VM object chain to obtain the page to be mapped
1924 * If we reach the terminal object without finding a page
1925 * and we determine it would be advantageous, then allocate
1926 * a zero-fill page for the base object. The base object
1927 * is guaranteed to be OBJT_DEFAULT for this case.
1929 index
= ((addr
- entry
->start
) + entry
->offset
) >> PAGE_SHIFT
;
1934 while ((m
= vm_page_lookup(lobject
, pindex
)) == NULL
) {
1935 if (lobject
->type
!= OBJT_DEFAULT
)
1937 if (lobject
->backing_object
== NULL
) {
1938 if (vm_fast_fault
== 0)
1940 if (vm_prefault_pageorder
[i
] < 0 ||
1941 (prot
& VM_PROT_WRITE
) == 0 ||
1942 vm_page_count_min(0)) {
1945 m
= vm_page_alloc(object
, index
,
1946 VM_ALLOC_NORMAL
| VM_ALLOC_ZERO
);
1948 if ((m
->flags
& PG_ZERO
) == 0) {
1949 vm_page_zero_fill(m
);
1951 vm_page_flag_clear(m
, PG_ZERO
);
1952 mycpu
->gd_cnt
.v_ozfod
++;
1954 mycpu
->gd_cnt
.v_zfod
++;
1955 m
->valid
= VM_PAGE_BITS_ALL
;
1958 /* lobject = object .. not needed */
1961 if (lobject
->backing_object_offset
& PAGE_MASK
)
1963 pindex
+= lobject
->backing_object_offset
>> PAGE_SHIFT
;
1964 lobject
= lobject
->backing_object
;
1965 pprot
&= ~VM_PROT_WRITE
;
1968 * NOTE: lobject now invalid (if we did a zero-fill we didn't
1969 * bother assigning lobject = object).
1971 * Give-up if the page is not available.
1977 * Do not conditionalize on PG_RAM. If pages are present in
1978 * the VM system we assume optimal caching. If caching is
1979 * not optimal the I/O gravy train will be restarted when we
1980 * hit an unavailable page. We do not want to try to restart
1981 * the gravy train now because we really don't know how much
1982 * of the object has been cached. The cost for restarting
1983 * the gravy train should be low (since accesses will likely
1984 * be I/O bound anyway).
1986 * The object must be marked dirty if we are mapping a
1989 if (pprot
& VM_PROT_WRITE
)
1990 vm_object_set_writeable_dirty(m
->object
);
1993 * Enter the page into the pmap if appropriate.
1995 if (((m
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
1997 (m
->flags
& (PG_BUSY
| PG_FICTITIOUS
)) == 0) {
1999 if ((m
->queue
- m
->pc
) == PQ_CACHE
) {
2000 vm_page_deactivate(m
);
2003 pmap_enter(pmap
, addr
, m
, pprot
, 0);