2 * Copyright (c) 2003-2014 The DragonFly Project. All rights reserved.
4 * This code is derived from software contributed to The DragonFly Project
5 * by Matthew Dillon <dillon@backplane.com>
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * 3. Neither the name of The DragonFly Project nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific, prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * Copyright (c) 1991, 1993
37 * The Regents of the University of California. All rights reserved.
38 * Copyright (c) 1994 John S. Dyson
39 * All rights reserved.
40 * Copyright (c) 1994 David Greenman
41 * All rights reserved.
44 * This code is derived from software contributed to Berkeley by
45 * The Mach Operating System project at Carnegie-Mellon University.
47 * Redistribution and use in source and binary forms, with or without
48 * modification, are permitted provided that the following conditions
50 * 1. Redistributions of source code must retain the above copyright
51 * notice, this list of conditions and the following disclaimer.
52 * 2. Redistributions in binary form must reproduce the above copyright
53 * notice, this list of conditions and the following disclaimer in the
54 * documentation and/or other materials provided with the distribution.
55 * 3. Neither the name of the University nor the names of its contributors
56 * may be used to endorse or promote products derived from this software
57 * without specific prior written permission.
59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
73 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
74 * All rights reserved.
76 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
78 * Permission to use, copy, modify and distribute this software and
79 * its documentation is hereby granted, provided that both the copyright
80 * notice and this permission notice appear in all copies of the
81 * software, derivative works or modified versions, and any portions
82 * thereof, and that both notices appear in supporting documentation.
84 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
85 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
86 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
88 * Carnegie Mellon requests users of this software to return to
90 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
91 * School of Computer Science
92 * Carnegie Mellon University
93 * Pittsburgh PA 15213-3890
95 * any improvements or extensions that they make and grant Carnegie the
96 * rights to redistribute these changes.
100 * Page fault handling module.
103 #include <sys/param.h>
104 #include <sys/systm.h>
105 #include <sys/kernel.h>
106 #include <sys/proc.h>
107 #include <sys/vnode.h>
108 #include <sys/resourcevar.h>
109 #include <sys/vmmeter.h>
110 #include <sys/vkernel.h>
111 #include <sys/lock.h>
112 #include <sys/sysctl.h>
114 #include <cpu/lwbuf.h>
117 #include <vm/vm_param.h>
119 #include <vm/vm_map.h>
120 #include <vm/vm_object.h>
121 #include <vm/vm_page.h>
122 #include <vm/vm_pageout.h>
123 #include <vm/vm_kern.h>
124 #include <vm/vm_pager.h>
125 #include <vm/vnode_pager.h>
126 #include <vm/vm_extern.h>
128 #include <sys/thread2.h>
129 #include <vm/vm_page2.h>
137 vm_object_t first_object
;
138 vm_prot_t first_prot
;
140 vm_map_entry_t entry
;
141 int lookup_still_valid
;
151 static int debug_fault
= 0;
152 SYSCTL_INT(_vm
, OID_AUTO
, debug_fault
, CTLFLAG_RW
, &debug_fault
, 0, "");
153 static int debug_cluster
= 0;
154 SYSCTL_INT(_vm
, OID_AUTO
, debug_cluster
, CTLFLAG_RW
, &debug_cluster
, 0, "");
155 int vm_shared_fault
= 1;
156 TUNABLE_INT("vm.shared_fault", &vm_shared_fault
);
157 SYSCTL_INT(_vm
, OID_AUTO
, shared_fault
, CTLFLAG_RW
, &vm_shared_fault
, 0,
158 "Allow shared token on vm_object");
160 static int vm_fault_object(struct faultstate
*, vm_pindex_t
, vm_prot_t
, int);
161 static int vm_fault_vpagetable(struct faultstate
*, vm_pindex_t
*,
164 static int vm_fault_additional_pages (vm_page_t
, int, int, vm_page_t
*, int *);
166 static void vm_set_nosync(vm_page_t m
, vm_map_entry_t entry
);
167 static void vm_prefault(pmap_t pmap
, vm_offset_t addra
,
168 vm_map_entry_t entry
, int prot
, int fault_flags
);
169 static void vm_prefault_quick(pmap_t pmap
, vm_offset_t addra
,
170 vm_map_entry_t entry
, int prot
, int fault_flags
);
173 release_page(struct faultstate
*fs
)
175 vm_page_deactivate(fs
->m
);
176 vm_page_wakeup(fs
->m
);
181 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse
182 * requires relocking and then checking the timestamp.
184 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do
185 * not have to update fs->map_generation here.
187 * NOTE: This function can fail due to a deadlock against the caller's
188 * holding of a vm_page BUSY.
191 relock_map(struct faultstate
*fs
)
195 if (fs
->lookup_still_valid
== FALSE
&& fs
->map
) {
196 error
= vm_map_lock_read_to(fs
->map
);
198 fs
->lookup_still_valid
= TRUE
;
206 unlock_map(struct faultstate
*fs
)
208 if (fs
->lookup_still_valid
&& fs
->map
) {
209 vm_map_lookup_done(fs
->map
, fs
->entry
, 0);
210 fs
->lookup_still_valid
= FALSE
;
215 * Clean up after a successful call to vm_fault_object() so another call
216 * to vm_fault_object() can be made.
219 _cleanup_successful_fault(struct faultstate
*fs
, int relock
)
222 * We allocated a junk page for a COW operation that did
223 * not occur, the page must be freed.
225 if (fs
->object
!= fs
->first_object
) {
226 KKASSERT(fs
->first_shared
== 0);
227 vm_page_free(fs
->first_m
);
228 vm_object_pip_wakeup(fs
->object
);
235 fs
->object
= fs
->first_object
;
236 if (relock
&& fs
->lookup_still_valid
== FALSE
) {
238 vm_map_lock_read(fs
->map
);
239 fs
->lookup_still_valid
= TRUE
;
244 _unlock_things(struct faultstate
*fs
, int dealloc
)
246 _cleanup_successful_fault(fs
, 0);
248 /*vm_object_deallocate(fs->first_object);*/
249 /*fs->first_object = NULL; drop used later on */
252 if (fs
->vp
!= NULL
) {
258 #define unlock_things(fs) _unlock_things(fs, 0)
259 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
260 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
265 * Determine if the pager for the current object *might* contain the page.
267 * We only need to try the pager if this is not a default object (default
268 * objects are zero-fill and have no real pager), and if we are not taking
269 * a wiring fault or if the FS entry is wired.
271 #define TRYPAGER(fs) \
272 (fs->object->type != OBJT_DEFAULT && \
273 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
278 * Handle a page fault occuring at the given address, requiring the given
279 * permissions, in the map specified. If successful, the page is inserted
280 * into the associated physical map.
282 * NOTE: The given address should be truncated to the proper page address.
284 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
285 * a standard error specifying why the fault is fatal is returned.
287 * The map in question must be referenced, and remains so.
288 * The caller may hold no locks.
289 * No other requirements.
292 vm_fault(vm_map_t map
, vm_offset_t vaddr
, vm_prot_t fault_type
, int fault_flags
)
295 vm_pindex_t first_pindex
;
296 struct faultstate fs
;
300 struct vm_map_ilock ilock
;
306 inherit_prot
= fault_type
& VM_PROT_NOSYNC
;
308 fs
.fault_flags
= fault_flags
;
310 fs
.shared
= vm_shared_fault
;
311 fs
.first_shared
= vm_shared_fault
;
315 * vm_map interactions
318 if ((lp
= td
->td_lwp
) != NULL
)
319 lp
->lwp_flags
|= LWP_PAGING
;
323 * Find the vm_map_entry representing the backing store and resolve
324 * the top level object and page index. This may have the side
325 * effect of executing a copy-on-write on the map entry and/or
326 * creating a shadow object, but will not COW any actual VM pages.
328 * On success fs.map is left read-locked and various other fields
329 * are initialized but not otherwise referenced or locked.
331 * NOTE! vm_map_lookup will try to upgrade the fault_type to
332 * VM_FAULT_WRITE if the map entry is a virtual page table
333 * and also writable, so we can set the 'A'accessed bit in
334 * the virtual page table entry.
337 result
= vm_map_lookup(&fs
.map
, vaddr
, fault_type
,
338 &fs
.entry
, &fs
.first_object
,
339 &first_pindex
, &fs
.first_prot
, &fs
.wired
);
342 * If the lookup failed or the map protections are incompatible,
343 * the fault generally fails.
345 * The failure could be due to TDF_NOFAULT if vm_map_lookup()
346 * tried to do a COW fault.
348 * If the caller is trying to do a user wiring we have more work
351 if (result
!= KERN_SUCCESS
) {
352 if (result
== KERN_FAILURE_NOFAULT
) {
353 result
= KERN_FAILURE
;
356 if (result
!= KERN_PROTECTION_FAILURE
||
357 (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) != VM_FAULT_USER_WIRE
)
359 if (result
== KERN_INVALID_ADDRESS
&& growstack
&&
360 map
!= &kernel_map
&& curproc
!= NULL
) {
361 result
= vm_map_growstack(map
, vaddr
);
362 if (result
== KERN_SUCCESS
) {
367 result
= KERN_FAILURE
;
373 * If we are user-wiring a r/w segment, and it is COW, then
374 * we need to do the COW operation. Note that we don't
375 * currently COW RO sections now, because it is NOT desirable
376 * to COW .text. We simply keep .text from ever being COW'ed
377 * and take the heat that one cannot debug wired .text sections.
379 result
= vm_map_lookup(&fs
.map
, vaddr
,
380 VM_PROT_READ
|VM_PROT_WRITE
|
381 VM_PROT_OVERRIDE_WRITE
,
382 &fs
.entry
, &fs
.first_object
,
383 &first_pindex
, &fs
.first_prot
,
385 if (result
!= KERN_SUCCESS
) {
386 /* could also be KERN_FAILURE_NOFAULT */
387 result
= KERN_FAILURE
;
392 * If we don't COW now, on a user wire, the user will never
393 * be able to write to the mapping. If we don't make this
394 * restriction, the bookkeeping would be nearly impossible.
396 * XXX We have a shared lock, this will have a MP race but
397 * I don't see how it can hurt anything.
399 if ((fs
.entry
->protection
& VM_PROT_WRITE
) == 0) {
400 atomic_clear_char(&fs
.entry
->max_protection
,
406 * fs.map is read-locked
408 * Misc checks. Save the map generation number to detect races.
410 fs
.map_generation
= fs
.map
->timestamp
;
411 fs
.lookup_still_valid
= TRUE
;
413 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
414 fs
.prot
= fs
.first_prot
; /* default (used by uksmap) */
416 if (fs
.entry
->eflags
& (MAP_ENTRY_NOFAULT
| MAP_ENTRY_KSTACK
)) {
417 if (fs
.entry
->eflags
& MAP_ENTRY_NOFAULT
) {
418 panic("vm_fault: fault on nofault entry, addr: %p",
421 if ((fs
.entry
->eflags
& MAP_ENTRY_KSTACK
) &&
422 vaddr
>= fs
.entry
->start
&&
423 vaddr
< fs
.entry
->start
+ PAGE_SIZE
) {
424 panic("vm_fault: fault on stack guard, addr: %p",
430 * A user-kernel shared map has no VM object and bypasses
431 * everything. We execute the uksmap function with a temporary
432 * fictitious vm_page. The address is directly mapped with no
435 if (fs
.entry
->maptype
== VM_MAPTYPE_UKSMAP
) {
436 struct vm_page fakem
;
438 bzero(&fakem
, sizeof(fakem
));
439 fakem
.pindex
= first_pindex
;
440 fakem
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_UNMANAGED
;
441 fakem
.valid
= VM_PAGE_BITS_ALL
;
442 fakem
.pat_mode
= VM_MEMATTR_DEFAULT
;
443 if (fs
.entry
->object
.uksmap(fs
.entry
->aux
.dev
, &fakem
)) {
444 result
= KERN_FAILURE
;
448 pmap_enter(fs
.map
->pmap
, vaddr
, &fakem
, fs
.prot
| inherit_prot
,
454 * A system map entry may return a NULL object. No object means
455 * no pager means an unrecoverable kernel fault.
457 if (fs
.first_object
== NULL
) {
458 panic("vm_fault: unrecoverable fault at %p in entry %p",
459 (void *)vaddr
, fs
.entry
);
463 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
466 * Unfortunately a deadlock can occur if we are forced to page-in
467 * from swap, but diving all the way into the vm_pager_get_page()
468 * function to find out is too much. Just check the object type.
470 * The deadlock is a CAM deadlock on a busy VM page when trying
471 * to finish an I/O if another process gets stuck in
472 * vop_helper_read_shortcut() due to a swap fault.
474 if ((td
->td_flags
& TDF_NOFAULT
) &&
476 fs
.first_object
->type
== OBJT_VNODE
||
477 fs
.first_object
->type
== OBJT_SWAP
||
478 fs
.first_object
->backing_object
)) {
479 result
= KERN_FAILURE
;
485 * If the entry is wired we cannot change the page protection.
488 fault_type
= fs
.first_prot
;
491 * We generally want to avoid unnecessary exclusive modes on backing
492 * and terminal objects because this can seriously interfere with
493 * heavily fork()'d processes (particularly /bin/sh scripts).
495 * However, we also want to avoid unnecessary retries due to needed
496 * shared->exclusive promotion for common faults. Exclusive mode is
497 * always needed if any page insertion, rename, or free occurs in an
498 * object (and also indirectly if any I/O is done).
500 * The main issue here is going to be fs.first_shared. If the
501 * first_object has a backing object which isn't shadowed and the
502 * process is single-threaded we might as well use an exclusive
503 * lock/chain right off the bat.
505 if (fs
.first_shared
&& fs
.first_object
->backing_object
&&
506 LIST_EMPTY(&fs
.first_object
->shadow_head
) &&
507 td
->td_proc
&& td
->td_proc
->p_nthreads
== 1) {
512 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
513 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but
514 * we can try shared first.
516 if (fault_flags
& VM_FAULT_UNSWAP
) {
521 * Obtain a top-level object lock, shared or exclusive depending
522 * on fs.first_shared. If a shared lock winds up being insufficient
523 * we will retry with an exclusive lock.
525 * The vnode pager lock is always shared.
528 vm_object_hold_shared(fs
.first_object
);
530 vm_object_hold(fs
.first_object
);
532 fs
.vp
= vnode_pager_lock(fs
.first_object
);
535 * The page we want is at (first_object, first_pindex), but if the
536 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
537 * page table to figure out the actual pindex.
539 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
543 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
544 vm_map_interlock(fs
.map
, &ilock
, vaddr
, vaddr
+ PAGE_SIZE
);
546 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
547 fs
.entry
->aux
.master_pde
,
549 if (result
== KERN_TRY_AGAIN
) {
550 vm_map_deinterlock(fs
.map
, &ilock
);
551 vm_object_drop(fs
.first_object
);
555 if (result
!= KERN_SUCCESS
) {
556 vm_map_deinterlock(fs
.map
, &ilock
);
562 * Now we have the actual (object, pindex), fault in the page. If
563 * vm_fault_object() fails it will unlock and deallocate the FS
564 * data. If it succeeds everything remains locked and fs->object
565 * will have an additional PIP count if it is not equal to
568 * vm_fault_object will set fs->prot for the pmap operation. It is
569 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
570 * page can be safely written. However, it will force a read-only
571 * mapping for a read fault if the memory is managed by a virtual
574 * If the fault code uses the shared object lock shortcut
575 * we must not try to burst (we can't allocate VM pages).
577 result
= vm_fault_object(&fs
, first_pindex
, fault_type
, 1);
579 if (debug_fault
> 0) {
581 kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x "
582 "fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n",
583 result
, (intmax_t)vaddr
, fault_type
, fault_flags
,
584 fs
.m
, fs
.prot
, fs
.wired
, fs
.entry
);
587 if (result
== KERN_TRY_AGAIN
) {
589 vm_map_deinterlock(fs
.map
, &ilock
);
590 vm_object_drop(fs
.first_object
);
594 if (result
!= KERN_SUCCESS
) {
596 vm_map_deinterlock(fs
.map
, &ilock
);
601 * On success vm_fault_object() does not unlock or deallocate, and fs.m
602 * will contain a busied page.
604 * Enter the page into the pmap and do pmap-related adjustments.
606 KKASSERT(fs
.lookup_still_valid
== TRUE
);
607 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
608 pmap_enter(fs
.map
->pmap
, vaddr
, fs
.m
, fs
.prot
| inherit_prot
,
612 vm_map_deinterlock(fs
.map
, &ilock
);
614 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
615 KKASSERT(fs
.m
->flags
& PG_BUSY
);
618 * If the page is not wired down, then put it where the pageout daemon
621 if (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) {
625 vm_page_unwire(fs
.m
, 1);
627 vm_page_activate(fs
.m
);
629 vm_page_wakeup(fs
.m
);
632 * Burst in a few more pages if possible. The fs.map should still
633 * be locked. To avoid interlocking against a vnode->getblk
634 * operation we had to be sure to unbusy our primary vm_page above
637 * A normal burst can continue down backing store, only execute
638 * if we are holding an exclusive lock, otherwise the exclusive
639 * locks the burst code gets might cause excessive SMP collisions.
641 * A quick burst can be utilized when there is no backing object
642 * (i.e. a shared file mmap).
644 if ((fault_flags
& VM_FAULT_BURST
) &&
645 (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) == 0 &&
647 if (fs
.first_shared
== 0 && fs
.shared
== 0) {
648 vm_prefault(fs
.map
->pmap
, vaddr
,
649 fs
.entry
, fs
.prot
, fault_flags
);
651 vm_prefault_quick(fs
.map
->pmap
, vaddr
,
652 fs
.entry
, fs
.prot
, fault_flags
);
657 mycpu
->gd_cnt
.v_vm_faults
++;
659 ++td
->td_lwp
->lwp_ru
.ru_minflt
;
662 * Unlock everything, and return
668 td
->td_lwp
->lwp_ru
.ru_majflt
++;
670 td
->td_lwp
->lwp_ru
.ru_minflt
++;
674 /*vm_object_deallocate(fs.first_object);*/
676 /*fs.first_object = NULL; must still drop later */
678 result
= KERN_SUCCESS
;
681 vm_object_drop(fs
.first_object
);
684 lp
->lwp_flags
&= ~LWP_PAGING
;
686 #if !defined(NO_SWAPPING)
688 * Check the process RSS limit and force deactivation and
689 * (asynchronous) paging if necessary. This is a complex operation,
690 * only do it for direct user-mode faults, for now.
692 * To reduce overhead implement approximately a ~16MB hysteresis.
695 if ((fault_flags
& VM_FAULT_USERMODE
) && lp
&&
696 p
->p_limit
&& map
->pmap
&& vm_pageout_memuse_mode
>= 1 &&
697 map
!= &kernel_map
) {
701 limit
= OFF_TO_IDX(qmin(p
->p_rlimit
[RLIMIT_RSS
].rlim_cur
,
702 p
->p_rlimit
[RLIMIT_RSS
].rlim_max
));
703 size
= pmap_resident_tlnw_count(map
->pmap
);
704 if (limit
>= 0 && size
> 4096 && size
- 4096 >= limit
) {
705 vm_pageout_map_deactivate_pages(map
, limit
);
714 * Fault in the specified virtual address in the current process map,
715 * returning a held VM page or NULL. See vm_fault_page() for more
721 vm_fault_page_quick(vm_offset_t va
, vm_prot_t fault_type
,
722 int *errorp
, int *busyp
)
724 struct lwp
*lp
= curthread
->td_lwp
;
727 m
= vm_fault_page(&lp
->lwp_vmspace
->vm_map
, va
,
728 fault_type
, VM_FAULT_NORMAL
,
734 * Fault in the specified virtual address in the specified map, doing all
735 * necessary manipulation of the object store and all necessary I/O. Return
736 * a held VM page or NULL, and set *errorp. The related pmap is not
739 * If busyp is not NULL then *busyp will be set to TRUE if this routine
740 * decides to return a busied page (aka VM_PROT_WRITE), or FALSE if it
741 * does not (VM_PROT_WRITE not specified or busyp is NULL). If busyp is
742 * NULL the returned page is only held.
744 * If the caller has no intention of writing to the page's contents, busyp
745 * can be passed as NULL along with VM_PROT_WRITE to force a COW operation
746 * without busying the page.
748 * The returned page will also be marked PG_REFERENCED.
750 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
751 * error will be returned.
756 vm_fault_page(vm_map_t map
, vm_offset_t vaddr
, vm_prot_t fault_type
,
757 int fault_flags
, int *errorp
, int *busyp
)
759 vm_pindex_t first_pindex
;
760 struct faultstate fs
;
764 vm_prot_t orig_fault_type
= fault_type
;
768 fs
.fault_flags
= fault_flags
;
769 KKASSERT((fault_flags
& VM_FAULT_WIRE_MASK
) == 0);
772 * Dive the pmap (concurrency possible). If we find the
773 * appropriate page we can terminate early and quickly.
775 * This works great for normal programs but will always return
776 * NULL for host lookups of vkernel maps in VMM mode.
778 fs
.m
= pmap_fault_page_quick(map
->pmap
, vaddr
, fault_type
, busyp
);
780 if (fault_type
& (VM_PROT_WRITE
|VM_PROT_OVERRIDE_WRITE
))
787 * Otherwise take a concurrency hit and do a formal page
791 fs
.shared
= vm_shared_fault
;
792 fs
.first_shared
= vm_shared_fault
;
796 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
797 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but
798 * we can try shared first.
800 if (fault_flags
& VM_FAULT_UNSWAP
) {
806 * Find the vm_map_entry representing the backing store and resolve
807 * the top level object and page index. This may have the side
808 * effect of executing a copy-on-write on the map entry and/or
809 * creating a shadow object, but will not COW any actual VM pages.
811 * On success fs.map is left read-locked and various other fields
812 * are initialized but not otherwise referenced or locked.
814 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
815 * if the map entry is a virtual page table and also writable,
816 * so we can set the 'A'accessed bit in the virtual page table
820 result
= vm_map_lookup(&fs
.map
, vaddr
, fault_type
,
821 &fs
.entry
, &fs
.first_object
,
822 &first_pindex
, &fs
.first_prot
, &fs
.wired
);
824 if (result
!= KERN_SUCCESS
) {
825 if (result
== KERN_FAILURE_NOFAULT
) {
826 *errorp
= KERN_FAILURE
;
830 if (result
!= KERN_PROTECTION_FAILURE
||
831 (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) != VM_FAULT_USER_WIRE
)
833 if (result
== KERN_INVALID_ADDRESS
&& growstack
&&
834 map
!= &kernel_map
&& curproc
!= NULL
) {
835 result
= vm_map_growstack(map
, vaddr
);
836 if (result
== KERN_SUCCESS
) {
841 result
= KERN_FAILURE
;
849 * If we are user-wiring a r/w segment, and it is COW, then
850 * we need to do the COW operation. Note that we don't
851 * currently COW RO sections now, because it is NOT desirable
852 * to COW .text. We simply keep .text from ever being COW'ed
853 * and take the heat that one cannot debug wired .text sections.
855 result
= vm_map_lookup(&fs
.map
, vaddr
,
856 VM_PROT_READ
|VM_PROT_WRITE
|
857 VM_PROT_OVERRIDE_WRITE
,
858 &fs
.entry
, &fs
.first_object
,
859 &first_pindex
, &fs
.first_prot
,
861 if (result
!= KERN_SUCCESS
) {
862 /* could also be KERN_FAILURE_NOFAULT */
863 *errorp
= KERN_FAILURE
;
869 * If we don't COW now, on a user wire, the user will never
870 * be able to write to the mapping. If we don't make this
871 * restriction, the bookkeeping would be nearly impossible.
873 * XXX We have a shared lock, this will have a MP race but
874 * I don't see how it can hurt anything.
876 if ((fs
.entry
->protection
& VM_PROT_WRITE
) == 0) {
877 atomic_clear_char(&fs
.entry
->max_protection
,
883 * fs.map is read-locked
885 * Misc checks. Save the map generation number to detect races.
887 fs
.map_generation
= fs
.map
->timestamp
;
888 fs
.lookup_still_valid
= TRUE
;
890 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
892 if (fs
.entry
->eflags
& MAP_ENTRY_NOFAULT
) {
893 panic("vm_fault: fault on nofault entry, addr: %lx",
898 * A user-kernel shared map has no VM object and bypasses
899 * everything. We execute the uksmap function with a temporary
900 * fictitious vm_page. The address is directly mapped with no
903 if (fs
.entry
->maptype
== VM_MAPTYPE_UKSMAP
) {
904 struct vm_page fakem
;
906 bzero(&fakem
, sizeof(fakem
));
907 fakem
.pindex
= first_pindex
;
908 fakem
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_UNMANAGED
;
909 fakem
.valid
= VM_PAGE_BITS_ALL
;
910 fakem
.pat_mode
= VM_MEMATTR_DEFAULT
;
911 if (fs
.entry
->object
.uksmap(fs
.entry
->aux
.dev
, &fakem
)) {
912 *errorp
= KERN_FAILURE
;
917 fs
.m
= PHYS_TO_VM_PAGE(fakem
.phys_addr
);
920 *busyp
= 0; /* don't need to busy R or W */
928 * A system map entry may return a NULL object. No object means
929 * no pager means an unrecoverable kernel fault.
931 if (fs
.first_object
== NULL
) {
932 panic("vm_fault: unrecoverable fault at %p in entry %p",
933 (void *)vaddr
, fs
.entry
);
937 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
940 * Unfortunately a deadlock can occur if we are forced to page-in
941 * from swap, but diving all the way into the vm_pager_get_page()
942 * function to find out is too much. Just check the object type.
944 if ((curthread
->td_flags
& TDF_NOFAULT
) &&
946 fs
.first_object
->type
== OBJT_VNODE
||
947 fs
.first_object
->type
== OBJT_SWAP
||
948 fs
.first_object
->backing_object
)) {
949 *errorp
= KERN_FAILURE
;
956 * If the entry is wired we cannot change the page protection.
959 fault_type
= fs
.first_prot
;
962 * Make a reference to this object to prevent its disposal while we
963 * are messing with it. Once we have the reference, the map is free
964 * to be diddled. Since objects reference their shadows (and copies),
965 * they will stay around as well.
967 * The reference should also prevent an unexpected collapse of the
968 * parent that might move pages from the current object into the
969 * parent unexpectedly, resulting in corruption.
971 * Bump the paging-in-progress count to prevent size changes (e.g.
972 * truncation operations) during I/O. This must be done after
973 * obtaining the vnode lock in order to avoid possible deadlocks.
976 vm_object_hold_shared(fs
.first_object
);
978 vm_object_hold(fs
.first_object
);
980 fs
.vp
= vnode_pager_lock(fs
.first_object
); /* shared */
983 * The page we want is at (first_object, first_pindex), but if the
984 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
985 * page table to figure out the actual pindex.
987 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
990 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
991 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
992 fs
.entry
->aux
.master_pde
,
994 if (result
== KERN_TRY_AGAIN
) {
995 vm_object_drop(fs
.first_object
);
999 if (result
!= KERN_SUCCESS
) {
1007 * Now we have the actual (object, pindex), fault in the page. If
1008 * vm_fault_object() fails it will unlock and deallocate the FS
1009 * data. If it succeeds everything remains locked and fs->object
1010 * will have an additinal PIP count if it is not equal to
1014 result
= vm_fault_object(&fs
, first_pindex
, fault_type
, 1);
1016 if (result
== KERN_TRY_AGAIN
) {
1017 vm_object_drop(fs
.first_object
);
1021 if (result
!= KERN_SUCCESS
) {
1027 if ((orig_fault_type
& VM_PROT_WRITE
) &&
1028 (fs
.prot
& VM_PROT_WRITE
) == 0) {
1029 *errorp
= KERN_PROTECTION_FAILURE
;
1030 unlock_and_deallocate(&fs
);
1036 * DO NOT UPDATE THE PMAP!!! This function may be called for
1037 * a pmap unrelated to the current process pmap, in which case
1038 * the current cpu core will not be listed in the pmap's pm_active
1039 * mask. Thus invalidation interlocks will fail to work properly.
1041 * (for example, 'ps' uses procfs to read program arguments from
1042 * each process's stack).
1044 * In addition to the above this function will be called to acquire
1045 * a page that might already be faulted in, re-faulting it
1046 * continuously is a waste of time.
1048 * XXX could this have been the cause of our random seg-fault
1049 * issues? procfs accesses user stacks.
1051 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
1053 pmap_enter(fs
.map
->pmap
, vaddr
, fs
.m
, fs
.prot
, fs
.wired
, NULL
);
1054 mycpu
->gd_cnt
.v_vm_faults
++;
1055 if (curthread
->td_lwp
)
1056 ++curthread
->td_lwp
->lwp_ru
.ru_minflt
;
1060 * On success vm_fault_object() does not unlock or deallocate, and fs.m
1061 * will contain a busied page. So we must unlock here after having
1062 * messed with the pmap.
1067 * Return a held page. We are not doing any pmap manipulation so do
1068 * not set PG_MAPPED. However, adjust the page flags according to
1069 * the fault type because the caller may not use a managed pmapping
1070 * (so we don't want to lose the fact that the page will be dirtied
1071 * if a write fault was specified).
1073 if (fault_type
& VM_PROT_WRITE
)
1074 vm_page_dirty(fs
.m
);
1075 vm_page_activate(fs
.m
);
1077 if (curthread
->td_lwp
) {
1079 curthread
->td_lwp
->lwp_ru
.ru_majflt
++;
1081 curthread
->td_lwp
->lwp_ru
.ru_minflt
++;
1086 * Unlock everything, and return the held or busied page.
1089 if (fault_type
& (VM_PROT_WRITE
|VM_PROT_OVERRIDE_WRITE
)) {
1090 vm_page_dirty(fs
.m
);
1095 vm_page_wakeup(fs
.m
);
1099 vm_page_wakeup(fs
.m
);
1101 /*vm_object_deallocate(fs.first_object);*/
1102 /*fs.first_object = NULL; */
1106 if (fs
.first_object
)
1107 vm_object_drop(fs
.first_object
);
1113 * Fault in the specified (object,offset), dirty the returned page as
1114 * needed. If the requested fault_type cannot be done NULL and an
1115 * error is returned.
1117 * A held (but not busied) page is returned.
1119 * The passed in object must be held as specified by the shared
1123 vm_fault_object_page(vm_object_t object
, vm_ooffset_t offset
,
1124 vm_prot_t fault_type
, int fault_flags
,
1125 int *sharedp
, int *errorp
)
1128 vm_pindex_t first_pindex
;
1129 struct faultstate fs
;
1130 struct vm_map_entry entry
;
1132 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1133 bzero(&entry
, sizeof(entry
));
1134 entry
.object
.vm_object
= object
;
1135 entry
.maptype
= VM_MAPTYPE_NORMAL
;
1136 entry
.protection
= entry
.max_protection
= fault_type
;
1139 fs
.fault_flags
= fault_flags
;
1141 fs
.shared
= vm_shared_fault
;
1142 fs
.first_shared
= *sharedp
;
1144 KKASSERT((fault_flags
& VM_FAULT_WIRE_MASK
) == 0);
1147 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
1148 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but
1149 * we can try shared first.
1151 if (fs
.first_shared
&& (fault_flags
& VM_FAULT_UNSWAP
)) {
1152 fs
.first_shared
= 0;
1153 vm_object_upgrade(object
);
1157 * Retry loop as needed (typically for shared->exclusive transitions)
1160 *sharedp
= fs
.first_shared
;
1161 first_pindex
= OFF_TO_IDX(offset
);
1162 fs
.first_object
= object
;
1164 fs
.first_prot
= fault_type
;
1166 /*fs.map_generation = 0; unused */
1169 * Make a reference to this object to prevent its disposal while we
1170 * are messing with it. Once we have the reference, the map is free
1171 * to be diddled. Since objects reference their shadows (and copies),
1172 * they will stay around as well.
1174 * The reference should also prevent an unexpected collapse of the
1175 * parent that might move pages from the current object into the
1176 * parent unexpectedly, resulting in corruption.
1178 * Bump the paging-in-progress count to prevent size changes (e.g.
1179 * truncation operations) during I/O. This must be done after
1180 * obtaining the vnode lock in order to avoid possible deadlocks.
1183 fs
.vp
= vnode_pager_lock(fs
.first_object
);
1185 fs
.lookup_still_valid
= TRUE
;
1187 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
1190 /* XXX future - ability to operate on VM object using vpagetable */
1191 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
1192 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
1193 fs
.entry
->aux
.master_pde
,
1195 if (result
== KERN_TRY_AGAIN
) {
1196 if (fs
.first_shared
== 0 && *sharedp
)
1197 vm_object_upgrade(object
);
1200 if (result
!= KERN_SUCCESS
) {
1208 * Now we have the actual (object, pindex), fault in the page. If
1209 * vm_fault_object() fails it will unlock and deallocate the FS
1210 * data. If it succeeds everything remains locked and fs->object
1211 * will have an additinal PIP count if it is not equal to
1214 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact.
1215 * We may have to upgrade its lock to handle the requested fault.
1217 result
= vm_fault_object(&fs
, first_pindex
, fault_type
, 0);
1219 if (result
== KERN_TRY_AGAIN
) {
1220 if (fs
.first_shared
== 0 && *sharedp
)
1221 vm_object_upgrade(object
);
1224 if (result
!= KERN_SUCCESS
) {
1229 if ((fault_type
& VM_PROT_WRITE
) && (fs
.prot
& VM_PROT_WRITE
) == 0) {
1230 *errorp
= KERN_PROTECTION_FAILURE
;
1231 unlock_and_deallocate(&fs
);
1236 * On success vm_fault_object() does not unlock or deallocate, so we
1237 * do it here. Note that the returned fs.m will be busied.
1242 * Return a held page. We are not doing any pmap manipulation so do
1243 * not set PG_MAPPED. However, adjust the page flags according to
1244 * the fault type because the caller may not use a managed pmapping
1245 * (so we don't want to lose the fact that the page will be dirtied
1246 * if a write fault was specified).
1249 vm_page_activate(fs
.m
);
1250 if ((fault_type
& VM_PROT_WRITE
) || (fault_flags
& VM_FAULT_DIRTY
))
1251 vm_page_dirty(fs
.m
);
1252 if (fault_flags
& VM_FAULT_UNSWAP
)
1253 swap_pager_unswapped(fs
.m
);
1256 * Indicate that the page was accessed.
1258 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
1260 if (curthread
->td_lwp
) {
1262 curthread
->td_lwp
->lwp_ru
.ru_majflt
++;
1264 curthread
->td_lwp
->lwp_ru
.ru_minflt
++;
1269 * Unlock everything, and return the held page.
1271 vm_page_wakeup(fs
.m
);
1272 /*vm_object_deallocate(fs.first_object);*/
1273 /*fs.first_object = NULL; */
1280 * Translate the virtual page number (first_pindex) that is relative
1281 * to the address space into a logical page number that is relative to the
1282 * backing object. Use the virtual page table pointed to by (vpte).
1284 * Possibly downgrade the protection based on the vpte bits.
1286 * This implements an N-level page table. Any level can terminate the
1287 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
1288 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
1292 vm_fault_vpagetable(struct faultstate
*fs
, vm_pindex_t
*pindex
,
1293 vpte_t vpte
, int fault_type
, int allow_nofault
)
1296 struct lwbuf lwb_cache
;
1297 int vshift
= VPTE_FRAME_END
- PAGE_SHIFT
; /* index bits remaining */
1301 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs
->first_object
));
1304 * We cannot proceed if the vpte is not valid, not readable
1305 * for a read fault, not writable for a write fault, or
1306 * not executable for an instruction execution fault.
1308 if ((vpte
& VPTE_V
) == 0) {
1309 unlock_and_deallocate(fs
);
1310 return (KERN_FAILURE
);
1312 if ((fault_type
& VM_PROT_WRITE
) && (vpte
& VPTE_RW
) == 0) {
1313 unlock_and_deallocate(fs
);
1314 return (KERN_FAILURE
);
1316 if ((fault_type
& VM_PROT_EXECUTE
) && (vpte
& VPTE_NX
)) {
1317 unlock_and_deallocate(fs
);
1318 return (KERN_FAILURE
);
1320 if ((vpte
& VPTE_PS
) || vshift
== 0)
1324 * Get the page table page. Nominally we only read the page
1325 * table, but since we are actively setting VPTE_M and VPTE_A,
1326 * tell vm_fault_object() that we are writing it.
1328 * There is currently no real need to optimize this.
1330 result
= vm_fault_object(fs
, (vpte
& VPTE_FRAME
) >> PAGE_SHIFT
,
1331 VM_PROT_READ
|VM_PROT_WRITE
,
1333 if (result
!= KERN_SUCCESS
)
1337 * Process the returned fs.m and look up the page table
1338 * entry in the page table page.
1340 vshift
-= VPTE_PAGE_BITS
;
1341 lwb
= lwbuf_alloc(fs
->m
, &lwb_cache
);
1342 ptep
= ((vpte_t
*)lwbuf_kva(lwb
) +
1343 ((*pindex
>> vshift
) & VPTE_PAGE_MASK
));
1344 vm_page_activate(fs
->m
);
1347 * Page table write-back - entire operation including
1348 * validation of the pte must be atomic to avoid races
1349 * against the vkernel changing the pte.
1351 * If the vpte is valid for the* requested operation, do
1352 * a write-back to the page table.
1354 * XXX VPTE_M is not set properly for page directory pages.
1355 * It doesn't get set in the page directory if the page table
1356 * is modified during a read access.
1362 * Reload for the cmpset, but make sure the pte is
1369 if ((vpte
& VPTE_V
) == 0)
1372 if ((fault_type
& VM_PROT_WRITE
) && (vpte
& VPTE_RW
))
1373 nvpte
|= VPTE_M
| VPTE_A
;
1374 if (fault_type
& (VM_PROT_READ
| VM_PROT_EXECUTE
))
1378 if (atomic_cmpset_long(ptep
, vpte
, nvpte
)) {
1379 vm_page_dirty(fs
->m
);
1384 vm_page_flag_set(fs
->m
, PG_REFERENCED
);
1385 vm_page_wakeup(fs
->m
);
1387 cleanup_successful_fault(fs
);
1391 * When the vkernel sets VPTE_RW it expects the real kernel to
1392 * reflect VPTE_M back when the page is modified via the mapping.
1393 * In order to accomplish this the real kernel must map the page
1394 * read-only for read faults and use write faults to reflect VPTE_M
1397 * Once VPTE_M has been set, the real kernel's pte allows writing.
1398 * If the vkernel clears VPTE_M the vkernel must be sure to
1399 * MADV_INVAL the real kernel's mappings to force the real kernel
1400 * to re-fault on the next write so oit can set VPTE_M again.
1402 if ((fault_type
& VM_PROT_WRITE
) == 0 &&
1403 (vpte
& (VPTE_RW
| VPTE_M
)) != (VPTE_RW
| VPTE_M
)) {
1404 fs
->first_prot
&= ~VM_PROT_WRITE
;
1408 * Disable EXECUTE perms if NX bit is set.
1411 fs
->first_prot
&= ~VM_PROT_EXECUTE
;
1414 * Combine remaining address bits with the vpte.
1416 *pindex
= ((vpte
& VPTE_FRAME
) >> PAGE_SHIFT
) +
1417 (*pindex
& ((1L << vshift
) - 1));
1418 return (KERN_SUCCESS
);
1423 * This is the core of the vm_fault code.
1425 * Do all operations required to fault-in (fs.first_object, pindex). Run
1426 * through the shadow chain as necessary and do required COW or virtual
1427 * copy operations. The caller has already fully resolved the vm_map_entry
1428 * and, if appropriate, has created a copy-on-write layer. All we need to
1429 * do is iterate the object chain.
1431 * On failure (fs) is unlocked and deallocated and the caller may return or
1432 * retry depending on the failure code. On success (fs) is NOT unlocked or
1433 * deallocated, fs.m will contained a resolved, busied page, and fs.object
1434 * will have an additional PIP count if it is not equal to fs.first_object.
1436 * If locks based on fs->first_shared or fs->shared are insufficient,
1437 * clear the appropriate field(s) and return RETRY. COWs require that
1438 * first_shared be 0, while page allocations (or frees) require that
1439 * shared be 0. Renames require that both be 0.
1441 * NOTE! fs->[first_]shared might be set with VM_FAULT_DIRTY also set.
1442 * we will have to retry with it exclusive if the vm_page is
1445 * fs->first_object must be held on call.
1449 vm_fault_object(struct faultstate
*fs
, vm_pindex_t first_pindex
,
1450 vm_prot_t fault_type
, int allow_nofault
)
1452 vm_object_t next_object
;
1456 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs
->first_object
));
1457 fs
->prot
= fs
->first_prot
;
1458 fs
->object
= fs
->first_object
;
1459 pindex
= first_pindex
;
1461 vm_object_chain_acquire(fs
->first_object
, fs
->shared
);
1462 vm_object_pip_add(fs
->first_object
, 1);
1465 * If a read fault occurs we try to upgrade the page protection
1466 * and make it also writable if possible. There are three cases
1467 * where we cannot make the page mapping writable:
1469 * (1) The mapping is read-only or the VM object is read-only,
1470 * fs->prot above will simply not have VM_PROT_WRITE set.
1472 * (2) If the mapping is a virtual page table fs->first_prot will
1473 * have already been properly adjusted by vm_fault_vpagetable().
1474 * to detect writes so we can set VPTE_M in the virtual page
1475 * table. Used by vkernels.
1477 * (3) If the VM page is read-only or copy-on-write, upgrading would
1478 * just result in an unnecessary COW fault.
1480 * (4) If the pmap specifically requests A/M bit emulation, downgrade
1484 /* see vpagetable code */
1485 if (fs
->entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
1486 if ((fault_type
& VM_PROT_WRITE
) == 0)
1487 fs
->prot
&= ~VM_PROT_WRITE
;
1491 if (curthread
->td_lwp
&& curthread
->td_lwp
->lwp_vmspace
&&
1492 pmap_emulate_ad_bits(&curthread
->td_lwp
->lwp_vmspace
->vm_pmap
)) {
1493 if ((fault_type
& VM_PROT_WRITE
) == 0)
1494 fs
->prot
&= ~VM_PROT_WRITE
;
1497 /* vm_object_hold(fs->object); implied b/c object == first_object */
1501 * The entire backing chain from first_object to object
1502 * inclusive is chainlocked.
1504 * If the object is dead, we stop here
1506 if (fs
->object
->flags
& OBJ_DEAD
) {
1507 vm_object_pip_wakeup(fs
->first_object
);
1508 vm_object_chain_release_all(fs
->first_object
,
1510 if (fs
->object
!= fs
->first_object
)
1511 vm_object_drop(fs
->object
);
1512 unlock_and_deallocate(fs
);
1513 return (KERN_PROTECTION_FAILURE
);
1517 * See if the page is resident. Wait/Retry if the page is
1518 * busy (lots of stuff may have changed so we can't continue
1521 * We can theoretically allow the soft-busy case on a read
1522 * fault if the page is marked valid, but since such
1523 * pages are typically already pmap'd, putting that
1524 * special case in might be more effort then it is
1525 * worth. We cannot under any circumstances mess
1526 * around with a vm_page_t->busy page except, perhaps,
1529 fs
->m
= vm_page_lookup_busy_try(fs
->object
, pindex
,
1532 vm_object_pip_wakeup(fs
->first_object
);
1533 vm_object_chain_release_all(fs
->first_object
,
1535 if (fs
->object
!= fs
->first_object
)
1536 vm_object_drop(fs
->object
);
1538 vm_page_sleep_busy(fs
->m
, TRUE
, "vmpfw");
1539 mycpu
->gd_cnt
.v_intrans
++;
1540 /*vm_object_deallocate(fs->first_object);*/
1541 /*fs->first_object = NULL;*/
1543 return (KERN_TRY_AGAIN
);
1547 * The page is busied for us.
1549 * If reactivating a page from PQ_CACHE we may have
1552 int queue
= fs
->m
->queue
;
1553 vm_page_unqueue_nowakeup(fs
->m
);
1555 if ((queue
- fs
->m
->pc
) == PQ_CACHE
&&
1556 vm_page_count_severe()) {
1557 vm_page_activate(fs
->m
);
1558 vm_page_wakeup(fs
->m
);
1560 vm_object_pip_wakeup(fs
->first_object
);
1561 vm_object_chain_release_all(fs
->first_object
,
1563 if (fs
->object
!= fs
->first_object
)
1564 vm_object_drop(fs
->object
);
1565 unlock_and_deallocate(fs
);
1566 if (allow_nofault
== 0 ||
1567 (curthread
->td_flags
& TDF_NOFAULT
) == 0) {
1572 if (td
->td_proc
&& (td
->td_proc
->p_flags
& P_LOWMEMKILL
))
1573 return (KERN_PROTECTION_FAILURE
);
1575 return (KERN_TRY_AGAIN
);
1579 * If it still isn't completely valid (readable),
1580 * or if a read-ahead-mark is set on the VM page,
1581 * jump to readrest, else we found the page and
1584 * We can release the spl once we have marked the
1587 if (fs
->m
->object
!= &kernel_object
) {
1588 if ((fs
->m
->valid
& VM_PAGE_BITS_ALL
) !=
1592 if (fs
->m
->flags
& PG_RAM
) {
1595 vm_page_flag_clear(fs
->m
, PG_RAM
);
1599 break; /* break to PAGE HAS BEEN FOUND */
1603 * Page is not resident, If this is the search termination
1604 * or the pager might contain the page, allocate a new page.
1606 if (TRYPAGER(fs
) || fs
->object
== fs
->first_object
) {
1608 * Allocating, must be exclusive.
1610 if (fs
->object
== fs
->first_object
&&
1612 fs
->first_shared
= 0;
1613 vm_object_pip_wakeup(fs
->first_object
);
1614 vm_object_chain_release_all(fs
->first_object
,
1616 if (fs
->object
!= fs
->first_object
)
1617 vm_object_drop(fs
->object
);
1618 unlock_and_deallocate(fs
);
1619 return (KERN_TRY_AGAIN
);
1621 if (fs
->object
!= fs
->first_object
&&
1623 fs
->first_shared
= 0;
1625 vm_object_pip_wakeup(fs
->first_object
);
1626 vm_object_chain_release_all(fs
->first_object
,
1628 if (fs
->object
!= fs
->first_object
)
1629 vm_object_drop(fs
->object
);
1630 unlock_and_deallocate(fs
);
1631 return (KERN_TRY_AGAIN
);
1635 * If the page is beyond the object size we fail
1637 if (pindex
>= fs
->object
->size
) {
1638 vm_object_pip_wakeup(fs
->first_object
);
1639 vm_object_chain_release_all(fs
->first_object
,
1641 if (fs
->object
!= fs
->first_object
)
1642 vm_object_drop(fs
->object
);
1643 unlock_and_deallocate(fs
);
1644 return (KERN_PROTECTION_FAILURE
);
1648 * Allocate a new page for this object/offset pair.
1650 * It is possible for the allocation to race, so
1654 if (!vm_page_count_severe()) {
1655 fs
->m
= vm_page_alloc(fs
->object
, pindex
,
1656 ((fs
->vp
|| fs
->object
->backing_object
) ?
1657 VM_ALLOC_NULL_OK
| VM_ALLOC_NORMAL
:
1658 VM_ALLOC_NULL_OK
| VM_ALLOC_NORMAL
|
1659 VM_ALLOC_USE_GD
| VM_ALLOC_ZERO
));
1661 if (fs
->m
== NULL
) {
1662 vm_object_pip_wakeup(fs
->first_object
);
1663 vm_object_chain_release_all(fs
->first_object
,
1665 if (fs
->object
!= fs
->first_object
)
1666 vm_object_drop(fs
->object
);
1667 unlock_and_deallocate(fs
);
1668 if (allow_nofault
== 0 ||
1669 (curthread
->td_flags
& TDF_NOFAULT
) == 0) {
1674 if (td
->td_proc
&& (td
->td_proc
->p_flags
& P_LOWMEMKILL
))
1675 return (KERN_PROTECTION_FAILURE
);
1677 return (KERN_TRY_AGAIN
);
1681 * Fall through to readrest. We have a new page which
1682 * will have to be paged (since m->valid will be 0).
1688 * We have found an invalid or partially valid page, a
1689 * page with a read-ahead mark which might be partially or
1690 * fully valid (and maybe dirty too), or we have allocated
1693 * Attempt to fault-in the page if there is a chance that the
1694 * pager has it, and potentially fault in additional pages
1697 * If TRYPAGER is true then fs.m will be non-NULL and busied
1703 u_char behavior
= vm_map_entry_behavior(fs
->entry
);
1705 if (behavior
== MAP_ENTRY_BEHAV_RANDOM
)
1711 * Doing I/O may synchronously insert additional
1712 * pages so we can't be shared at this point either.
1714 * NOTE: We can't free fs->m here in the allocated
1715 * case (fs->object != fs->first_object) as
1716 * this would require an exclusively locked
1719 if (fs
->object
== fs
->first_object
&&
1721 vm_page_deactivate(fs
->m
);
1722 vm_page_wakeup(fs
->m
);
1724 fs
->first_shared
= 0;
1725 vm_object_pip_wakeup(fs
->first_object
);
1726 vm_object_chain_release_all(fs
->first_object
,
1728 if (fs
->object
!= fs
->first_object
)
1729 vm_object_drop(fs
->object
);
1730 unlock_and_deallocate(fs
);
1731 return (KERN_TRY_AGAIN
);
1733 if (fs
->object
!= fs
->first_object
&&
1735 vm_page_deactivate(fs
->m
);
1736 vm_page_wakeup(fs
->m
);
1738 fs
->first_shared
= 0;
1740 vm_object_pip_wakeup(fs
->first_object
);
1741 vm_object_chain_release_all(fs
->first_object
,
1743 if (fs
->object
!= fs
->first_object
)
1744 vm_object_drop(fs
->object
);
1745 unlock_and_deallocate(fs
);
1746 return (KERN_TRY_AGAIN
);
1750 * Avoid deadlocking against the map when doing I/O.
1751 * fs.object and the page is PG_BUSY'd.
1753 * NOTE: Once unlocked, fs->entry can become stale
1754 * so this will NULL it out.
1756 * NOTE: fs->entry is invalid until we relock the
1757 * map and verify that the timestamp has not
1763 * Acquire the page data. We still hold a ref on
1764 * fs.object and the page has been PG_BUSY's.
1766 * The pager may replace the page (for example, in
1767 * order to enter a fictitious page into the
1768 * object). If it does so it is responsible for
1769 * cleaning up the passed page and properly setting
1770 * the new page PG_BUSY.
1772 * If we got here through a PG_RAM read-ahead
1773 * mark the page may be partially dirty and thus
1774 * not freeable. Don't bother checking to see
1775 * if the pager has the page because we can't free
1776 * it anyway. We have to depend on the get_page
1777 * operation filling in any gaps whether there is
1778 * backing store or not.
1780 rv
= vm_pager_get_page(fs
->object
, &fs
->m
, seqaccess
);
1782 if (rv
== VM_PAGER_OK
) {
1784 * Relookup in case pager changed page. Pager
1785 * is responsible for disposition of old page
1788 * XXX other code segments do relookups too.
1789 * It's a bad abstraction that needs to be
1792 fs
->m
= vm_page_lookup(fs
->object
, pindex
);
1793 if (fs
->m
== NULL
) {
1794 vm_object_pip_wakeup(fs
->first_object
);
1795 vm_object_chain_release_all(
1796 fs
->first_object
, fs
->object
);
1797 if (fs
->object
!= fs
->first_object
)
1798 vm_object_drop(fs
->object
);
1799 unlock_and_deallocate(fs
);
1800 return (KERN_TRY_AGAIN
);
1803 break; /* break to PAGE HAS BEEN FOUND */
1807 * Remove the bogus page (which does not exist at this
1808 * object/offset); before doing so, we must get back
1809 * our object lock to preserve our invariant.
1811 * Also wake up any other process that may want to bring
1814 * If this is the top-level object, we must leave the
1815 * busy page to prevent another process from rushing
1816 * past us, and inserting the page in that object at
1817 * the same time that we are.
1819 if (rv
== VM_PAGER_ERROR
) {
1821 kprintf("vm_fault: pager read error, "
1826 kprintf("vm_fault: pager read error, "
1834 * Data outside the range of the pager or an I/O error
1836 * The page may have been wired during the pagein,
1837 * e.g. by the buffer cache, and cannot simply be
1838 * freed. Call vnode_pager_freepage() to deal with it.
1840 * Also note that we cannot free the page if we are
1841 * holding the related object shared. XXX not sure
1842 * what to do in that case.
1844 if (fs
->object
!= fs
->first_object
) {
1845 vnode_pager_freepage(fs
->m
);
1848 * XXX - we cannot just fall out at this
1849 * point, m has been freed and is invalid!
1853 * XXX - the check for kernel_map is a kludge to work
1854 * around having the machine panic on a kernel space
1855 * fault w/ I/O error.
1857 if (((fs
->map
!= &kernel_map
) &&
1858 (rv
== VM_PAGER_ERROR
)) || (rv
== VM_PAGER_BAD
)) {
1860 if (fs
->first_shared
) {
1861 vm_page_deactivate(fs
->m
);
1862 vm_page_wakeup(fs
->m
);
1864 vnode_pager_freepage(fs
->m
);
1868 vm_object_pip_wakeup(fs
->first_object
);
1869 vm_object_chain_release_all(fs
->first_object
,
1871 if (fs
->object
!= fs
->first_object
)
1872 vm_object_drop(fs
->object
);
1873 unlock_and_deallocate(fs
);
1874 if (rv
== VM_PAGER_ERROR
)
1875 return (KERN_FAILURE
);
1877 return (KERN_PROTECTION_FAILURE
);
1883 * We get here if the object has a default pager (or unwiring)
1884 * or the pager doesn't have the page.
1886 * fs->first_m will be used for the COW unless we find a
1887 * deeper page to be mapped read-only, in which case the
1888 * unlock*(fs) will free first_m.
1890 if (fs
->object
== fs
->first_object
)
1891 fs
->first_m
= fs
->m
;
1894 * Move on to the next object. The chain lock should prevent
1895 * the backing_object from getting ripped out from under us.
1897 * The object lock for the next object is governed by
1900 if ((next_object
= fs
->object
->backing_object
) != NULL
) {
1902 vm_object_hold_shared(next_object
);
1904 vm_object_hold(next_object
);
1905 vm_object_chain_acquire(next_object
, fs
->shared
);
1906 KKASSERT(next_object
== fs
->object
->backing_object
);
1907 pindex
+= OFF_TO_IDX(fs
->object
->backing_object_offset
);
1910 if (next_object
== NULL
) {
1912 * If there's no object left, fill the page in the top
1913 * object with zeros.
1915 if (fs
->object
!= fs
->first_object
) {
1917 if (fs
->first_object
->backing_object
!=
1919 vm_object_hold(fs
->first_object
->backing_object
);
1922 vm_object_chain_release_all(
1923 fs
->first_object
->backing_object
,
1926 if (fs
->first_object
->backing_object
!=
1928 vm_object_drop(fs
->first_object
->backing_object
);
1931 vm_object_pip_wakeup(fs
->object
);
1932 vm_object_drop(fs
->object
);
1933 fs
->object
= fs
->first_object
;
1934 pindex
= first_pindex
;
1935 fs
->m
= fs
->first_m
;
1940 * Zero the page and mark it valid.
1942 vm_page_zero_fill(fs
->m
);
1943 mycpu
->gd_cnt
.v_zfod
++;
1944 fs
->m
->valid
= VM_PAGE_BITS_ALL
;
1945 break; /* break to PAGE HAS BEEN FOUND */
1947 if (fs
->object
!= fs
->first_object
) {
1948 vm_object_pip_wakeup(fs
->object
);
1949 vm_object_lock_swap();
1950 vm_object_drop(fs
->object
);
1952 KASSERT(fs
->object
!= next_object
,
1953 ("object loop %p", next_object
));
1954 fs
->object
= next_object
;
1955 vm_object_pip_add(fs
->object
, 1);
1959 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1962 * object still held.
1964 * local shared variable may be different from fs->shared.
1966 * If the page is being written, but isn't already owned by the
1967 * top-level object, we have to copy it into a new page owned by the
1970 KASSERT((fs
->m
->flags
& PG_BUSY
) != 0,
1971 ("vm_fault: not busy after main loop"));
1973 if (fs
->object
!= fs
->first_object
) {
1975 * We only really need to copy if we want to write it.
1977 if (fault_type
& VM_PROT_WRITE
) {
1979 * This allows pages to be virtually copied from a
1980 * backing_object into the first_object, where the
1981 * backing object has no other refs to it, and cannot
1982 * gain any more refs. Instead of a bcopy, we just
1983 * move the page from the backing object to the
1984 * first object. Note that we must mark the page
1985 * dirty in the first object so that it will go out
1986 * to swap when needed.
1990 * Must be holding exclusive locks
1992 fs
->first_shared
== 0 &&
1995 * Map, if present, has not changed
1998 fs
->map_generation
== fs
->map
->timestamp
) &&
2000 * Only one shadow object
2002 (fs
->object
->shadow_count
== 1) &&
2004 * No COW refs, except us
2006 (fs
->object
->ref_count
== 1) &&
2008 * No one else can look this object up
2010 (fs
->object
->handle
== NULL
) &&
2012 * No other ways to look the object up
2014 ((fs
->object
->type
== OBJT_DEFAULT
) ||
2015 (fs
->object
->type
== OBJT_SWAP
)) &&
2017 * We don't chase down the shadow chain
2019 (fs
->object
== fs
->first_object
->backing_object
) &&
2022 * grab the lock if we need to
2024 (fs
->lookup_still_valid
||
2026 lockmgr(&fs
->map
->lock
, LK_EXCLUSIVE
|LK_NOWAIT
) == 0)
2029 * (first_m) and (m) are both busied. We have
2030 * move (m) into (first_m)'s object/pindex
2031 * in an atomic fashion, then free (first_m).
2033 * first_object is held so second remove
2034 * followed by the rename should wind
2035 * up being atomic. vm_page_free() might
2036 * block so we don't do it until after the
2039 fs
->lookup_still_valid
= 1;
2040 vm_page_protect(fs
->first_m
, VM_PROT_NONE
);
2041 vm_page_remove(fs
->first_m
);
2042 vm_page_rename(fs
->m
, fs
->first_object
,
2044 vm_page_free(fs
->first_m
);
2045 fs
->first_m
= fs
->m
;
2047 mycpu
->gd_cnt
.v_cow_optim
++;
2050 * Oh, well, lets copy it.
2052 * Why are we unmapping the original page
2053 * here? Well, in short, not all accessors
2054 * of user memory go through the pmap. The
2055 * procfs code doesn't have access user memory
2056 * via a local pmap, so vm_fault_page*()
2057 * can't call pmap_enter(). And the umtx*()
2058 * code may modify the COW'd page via a DMAP
2059 * or kernel mapping and not via the pmap,
2060 * leaving the original page still mapped
2061 * read-only into the pmap.
2063 * So we have to remove the page from at
2064 * least the current pmap if it is in it.
2065 * Just remove it from all pmaps.
2067 KKASSERT(fs
->first_shared
== 0);
2068 vm_page_copy(fs
->m
, fs
->first_m
);
2069 vm_page_protect(fs
->m
, VM_PROT_NONE
);
2070 vm_page_event(fs
->m
, VMEVENT_COW
);
2074 * We no longer need the old page or object.
2080 * We intend to revert to first_object, undo the
2081 * chain lock through to that.
2084 if (fs
->first_object
->backing_object
!= fs
->object
)
2085 vm_object_hold(fs
->first_object
->backing_object
);
2087 vm_object_chain_release_all(
2088 fs
->first_object
->backing_object
,
2091 if (fs
->first_object
->backing_object
!= fs
->object
)
2092 vm_object_drop(fs
->first_object
->backing_object
);
2096 * fs->object != fs->first_object due to above
2099 vm_object_pip_wakeup(fs
->object
);
2100 vm_object_drop(fs
->object
);
2103 * Only use the new page below...
2105 mycpu
->gd_cnt
.v_cow_faults
++;
2106 fs
->m
= fs
->first_m
;
2107 fs
->object
= fs
->first_object
;
2108 pindex
= first_pindex
;
2111 * If it wasn't a write fault avoid having to copy
2112 * the page by mapping it read-only.
2114 fs
->prot
&= ~VM_PROT_WRITE
;
2119 * Relock the map if necessary, then check the generation count.
2120 * relock_map() will update fs->timestamp to account for the
2121 * relocking if necessary.
2123 * If the count has changed after relocking then all sorts of
2124 * crap may have happened and we have to retry.
2126 * NOTE: The relock_map() can fail due to a deadlock against
2127 * the vm_page we are holding BUSY.
2129 if (fs
->lookup_still_valid
== FALSE
&& fs
->map
) {
2130 if (relock_map(fs
) ||
2131 fs
->map
->timestamp
!= fs
->map_generation
) {
2133 vm_object_pip_wakeup(fs
->first_object
);
2134 vm_object_chain_release_all(fs
->first_object
,
2136 if (fs
->object
!= fs
->first_object
)
2137 vm_object_drop(fs
->object
);
2138 unlock_and_deallocate(fs
);
2139 return (KERN_TRY_AGAIN
);
2144 * If the fault is a write, we know that this page is being
2145 * written NOW so dirty it explicitly to save on pmap_is_modified()
2148 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
2149 * if the page is already dirty to prevent data written with
2150 * the expectation of being synced from not being synced.
2151 * Likewise if this entry does not request NOSYNC then make
2152 * sure the page isn't marked NOSYNC. Applications sharing
2153 * data should use the same flags to avoid ping ponging.
2155 * Also tell the backing pager, if any, that it should remove
2156 * any swap backing since the page is now dirty.
2158 vm_page_activate(fs
->m
);
2159 if (fs
->prot
& VM_PROT_WRITE
) {
2160 vm_object_set_writeable_dirty(fs
->m
->object
);
2161 vm_set_nosync(fs
->m
, fs
->entry
);
2162 if (fs
->fault_flags
& VM_FAULT_DIRTY
) {
2163 vm_page_dirty(fs
->m
);
2164 if (fs
->m
->flags
& PG_SWAPPED
) {
2166 * If the page is swapped out we have to call
2167 * swap_pager_unswapped() which requires an
2168 * exclusive object lock. If we are shared,
2169 * we must clear the shared flag and retry.
2171 if ((fs
->object
== fs
->first_object
&&
2172 fs
->first_shared
) ||
2173 (fs
->object
!= fs
->first_object
&&
2175 vm_page_wakeup(fs
->m
);
2177 if (fs
->object
== fs
->first_object
)
2178 fs
->first_shared
= 0;
2181 vm_object_pip_wakeup(fs
->first_object
);
2182 vm_object_chain_release_all(
2183 fs
->first_object
, fs
->object
);
2184 if (fs
->object
!= fs
->first_object
)
2185 vm_object_drop(fs
->object
);
2186 unlock_and_deallocate(fs
);
2187 return (KERN_TRY_AGAIN
);
2189 swap_pager_unswapped(fs
->m
);
2194 vm_object_pip_wakeup(fs
->first_object
);
2195 vm_object_chain_release_all(fs
->first_object
, fs
->object
);
2196 if (fs
->object
!= fs
->first_object
)
2197 vm_object_drop(fs
->object
);
2200 * Page had better still be busy. We are still locked up and
2201 * fs->object will have another PIP reference if it is not equal
2202 * to fs->first_object.
2204 KASSERT(fs
->m
->flags
& PG_BUSY
,
2205 ("vm_fault: page %p not busy!", fs
->m
));
2208 * Sanity check: page must be completely valid or it is not fit to
2209 * map into user space. vm_pager_get_pages() ensures this.
2211 if (fs
->m
->valid
!= VM_PAGE_BITS_ALL
) {
2212 vm_page_zero_invalid(fs
->m
, TRUE
);
2213 kprintf("Warning: page %p partially invalid on fault\n", fs
->m
);
2216 return (KERN_SUCCESS
);
2220 * Wire down a range of virtual addresses in a map. The entry in question
2221 * should be marked in-transition and the map must be locked. We must
2222 * release the map temporarily while faulting-in the page to avoid a
2223 * deadlock. Note that the entry may be clipped while we are blocked but
2224 * will never be freed.
2229 vm_fault_wire(vm_map_t map
, vm_map_entry_t entry
,
2230 boolean_t user_wire
, int kmflags
)
2232 boolean_t fictitious
;
2243 wire_prot
= VM_PROT_READ
;
2244 fault_flags
= VM_FAULT_USER_WIRE
;
2246 wire_prot
= VM_PROT_READ
| VM_PROT_WRITE
;
2247 fault_flags
= VM_FAULT_CHANGE_WIRING
;
2249 if (kmflags
& KM_NOTLBSYNC
)
2250 wire_prot
|= VM_PROT_NOSYNC
;
2252 pmap
= vm_map_pmap(map
);
2253 start
= entry
->start
;
2256 switch(entry
->maptype
) {
2257 case VM_MAPTYPE_NORMAL
:
2258 case VM_MAPTYPE_VPAGETABLE
:
2259 fictitious
= entry
->object
.vm_object
&&
2260 ((entry
->object
.vm_object
->type
== OBJT_DEVICE
) ||
2261 (entry
->object
.vm_object
->type
== OBJT_MGTDEVICE
));
2263 case VM_MAPTYPE_UKSMAP
:
2271 if (entry
->eflags
& MAP_ENTRY_KSTACK
)
2277 * We simulate a fault to get the page and enter it in the physical
2280 for (va
= start
; va
< end
; va
+= PAGE_SIZE
) {
2281 rv
= vm_fault(map
, va
, wire_prot
, fault_flags
);
2283 while (va
> start
) {
2285 m
= pmap_unwire(pmap
, va
);
2286 if (m
&& !fictitious
) {
2287 vm_page_busy_wait(m
, FALSE
, "vmwrpg");
2288 vm_page_unwire(m
, 1);
2303 * Unwire a range of virtual addresses in a map. The map should be
2307 vm_fault_unwire(vm_map_t map
, vm_map_entry_t entry
)
2309 boolean_t fictitious
;
2316 pmap
= vm_map_pmap(map
);
2317 start
= entry
->start
;
2319 fictitious
= entry
->object
.vm_object
&&
2320 ((entry
->object
.vm_object
->type
== OBJT_DEVICE
) ||
2321 (entry
->object
.vm_object
->type
== OBJT_MGTDEVICE
));
2322 if (entry
->eflags
& MAP_ENTRY_KSTACK
)
2326 * Since the pages are wired down, we must be able to get their
2327 * mappings from the physical map system.
2329 for (va
= start
; va
< end
; va
+= PAGE_SIZE
) {
2330 m
= pmap_unwire(pmap
, va
);
2331 if (m
&& !fictitious
) {
2332 vm_page_busy_wait(m
, FALSE
, "vmwrpg");
2333 vm_page_unwire(m
, 1);
2340 * Copy all of the pages from a wired-down map entry to another.
2342 * The source and destination maps must be locked for write.
2343 * The source and destination maps token must be held
2344 * The source map entry must be wired down (or be a sharing map
2345 * entry corresponding to a main map entry that is wired down).
2347 * No other requirements.
2349 * XXX do segment optimization
2352 vm_fault_copy_entry(vm_map_t dst_map
, vm_map_t src_map
,
2353 vm_map_entry_t dst_entry
, vm_map_entry_t src_entry
)
2355 vm_object_t dst_object
;
2356 vm_object_t src_object
;
2357 vm_ooffset_t dst_offset
;
2358 vm_ooffset_t src_offset
;
2364 src_object
= src_entry
->object
.vm_object
;
2365 src_offset
= src_entry
->offset
;
2368 * Create the top-level object for the destination entry. (Doesn't
2369 * actually shadow anything - we copy the pages directly.)
2371 vm_map_entry_allocate_object(dst_entry
);
2372 dst_object
= dst_entry
->object
.vm_object
;
2374 prot
= dst_entry
->max_protection
;
2377 * Loop through all of the pages in the entry's range, copying each
2378 * one from the source object (it should be there) to the destination
2381 vm_object_hold(src_object
);
2382 vm_object_hold(dst_object
);
2383 for (vaddr
= dst_entry
->start
, dst_offset
= 0;
2384 vaddr
< dst_entry
->end
;
2385 vaddr
+= PAGE_SIZE
, dst_offset
+= PAGE_SIZE
) {
2388 * Allocate a page in the destination object
2391 dst_m
= vm_page_alloc(dst_object
,
2392 OFF_TO_IDX(dst_offset
),
2394 if (dst_m
== NULL
) {
2397 } while (dst_m
== NULL
);
2400 * Find the page in the source object, and copy it in.
2401 * (Because the source is wired down, the page will be in
2404 src_m
= vm_page_lookup(src_object
,
2405 OFF_TO_IDX(dst_offset
+ src_offset
));
2407 panic("vm_fault_copy_wired: page missing");
2409 vm_page_copy(src_m
, dst_m
);
2410 vm_page_event(src_m
, VMEVENT_COW
);
2413 * Enter it in the pmap...
2415 pmap_enter(dst_map
->pmap
, vaddr
, dst_m
, prot
, FALSE
, dst_entry
);
2418 * Mark it no longer busy, and put it on the active list.
2420 vm_page_activate(dst_m
);
2421 vm_page_wakeup(dst_m
);
2423 vm_object_drop(dst_object
);
2424 vm_object_drop(src_object
);
2430 * This routine checks around the requested page for other pages that
2431 * might be able to be faulted in. This routine brackets the viable
2432 * pages for the pages to be paged in.
2435 * m, rbehind, rahead
2438 * marray (array of vm_page_t), reqpage (index of requested page)
2441 * number of pages in marray
2444 vm_fault_additional_pages(vm_page_t m
, int rbehind
, int rahead
,
2445 vm_page_t
*marray
, int *reqpage
)
2449 vm_pindex_t pindex
, startpindex
, endpindex
, tpindex
;
2451 int cbehind
, cahead
;
2457 * we don't fault-ahead for device pager
2459 if ((object
->type
== OBJT_DEVICE
) ||
2460 (object
->type
== OBJT_MGTDEVICE
)) {
2467 * if the requested page is not available, then give up now
2469 if (!vm_pager_has_page(object
, pindex
, &cbehind
, &cahead
)) {
2470 *reqpage
= 0; /* not used by caller, fix compiler warn */
2474 if ((cbehind
== 0) && (cahead
== 0)) {
2480 if (rahead
> cahead
) {
2484 if (rbehind
> cbehind
) {
2489 * Do not do any readahead if we have insufficient free memory.
2491 * XXX code was broken disabled before and has instability
2492 * with this conditonal fixed, so shortcut for now.
2494 if (burst_fault
== 0 || vm_page_count_severe()) {
2501 * scan backward for the read behind pages -- in memory
2503 * Assume that if the page is not found an interrupt will not
2504 * create it. Theoretically interrupts can only remove (busy)
2505 * pages, not create new associations.
2508 if (rbehind
> pindex
) {
2512 startpindex
= pindex
- rbehind
;
2515 vm_object_hold(object
);
2516 for (tpindex
= pindex
; tpindex
> startpindex
; --tpindex
) {
2517 if (vm_page_lookup(object
, tpindex
- 1))
2522 while (tpindex
< pindex
) {
2523 rtm
= vm_page_alloc(object
, tpindex
, VM_ALLOC_SYSTEM
|
2526 for (j
= 0; j
< i
; j
++) {
2527 vm_page_free(marray
[j
]);
2529 vm_object_drop(object
);
2538 vm_object_drop(object
);
2544 * Assign requested page
2551 * Scan forwards for read-ahead pages
2553 tpindex
= pindex
+ 1;
2554 endpindex
= tpindex
+ rahead
;
2555 if (endpindex
> object
->size
)
2556 endpindex
= object
->size
;
2558 vm_object_hold(object
);
2559 while (tpindex
< endpindex
) {
2560 if (vm_page_lookup(object
, tpindex
))
2562 rtm
= vm_page_alloc(object
, tpindex
, VM_ALLOC_SYSTEM
|
2570 vm_object_drop(object
);
2578 * vm_prefault() provides a quick way of clustering pagefaults into a
2579 * processes address space. It is a "cousin" of pmap_object_init_pt,
2580 * except it runs at page fault time instead of mmap time.
2582 * vm.fast_fault Enables pre-faulting zero-fill pages
2584 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to
2585 * prefault. Scan stops in either direction when
2586 * a page is found to already exist.
2588 * This code used to be per-platform pmap_prefault(). It is now
2589 * machine-independent and enhanced to also pre-fault zero-fill pages
2590 * (see vm.fast_fault) as well as make them writable, which greatly
2591 * reduces the number of page faults programs incur.
2593 * Application performance when pre-faulting zero-fill pages is heavily
2594 * dependent on the application. Very tiny applications like /bin/echo
2595 * lose a little performance while applications of any appreciable size
2596 * gain performance. Prefaulting multiple pages also reduces SMP
2597 * congestion and can improve SMP performance significantly.
2599 * NOTE! prot may allow writing but this only applies to the top level
2600 * object. If we wind up mapping a page extracted from a backing
2601 * object we have to make sure it is read-only.
2603 * NOTE! The caller has already handled any COW operations on the
2604 * vm_map_entry via the normal fault code. Do NOT call this
2605 * shortcut unless the normal fault code has run on this entry.
2607 * The related map must be locked.
2608 * No other requirements.
2610 static int vm_prefault_pages
= 8;
2611 SYSCTL_INT(_vm
, OID_AUTO
, prefault_pages
, CTLFLAG_RW
, &vm_prefault_pages
, 0,
2612 "Maximum number of pages to pre-fault");
2613 static int vm_fast_fault
= 1;
2614 SYSCTL_INT(_vm
, OID_AUTO
, fast_fault
, CTLFLAG_RW
, &vm_fast_fault
, 0,
2615 "Burst fault zero-fill regions");
2618 * Set PG_NOSYNC if the map entry indicates so, but only if the page
2619 * is not already dirty by other means. This will prevent passive
2620 * filesystem syncing as well as 'sync' from writing out the page.
2623 vm_set_nosync(vm_page_t m
, vm_map_entry_t entry
)
2625 if (entry
->eflags
& MAP_ENTRY_NOSYNC
) {
2627 vm_page_flag_set(m
, PG_NOSYNC
);
2629 vm_page_flag_clear(m
, PG_NOSYNC
);
2634 vm_prefault(pmap_t pmap
, vm_offset_t addra
, vm_map_entry_t entry
, int prot
,
2650 * Get stable max count value, disabled if set to 0
2652 maxpages
= vm_prefault_pages
;
2658 * We do not currently prefault mappings that use virtual page
2659 * tables. We do not prefault foreign pmaps.
2661 if (entry
->maptype
!= VM_MAPTYPE_NORMAL
)
2663 lp
= curthread
->td_lwp
;
2664 if (lp
== NULL
|| (pmap
!= vmspace_pmap(lp
->lwp_vmspace
)))
2668 * Limit pre-fault count to 1024 pages.
2670 if (maxpages
> 1024)
2673 object
= entry
->object
.vm_object
;
2674 KKASSERT(object
!= NULL
);
2675 KKASSERT(object
== entry
->object
.vm_object
);
2678 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively
2679 * now (or do something more complex XXX).
2681 vm_object_hold(object
);
2682 vm_object_chain_acquire(object
, 0);
2686 for (i
= 0; i
< maxpages
; ++i
) {
2687 vm_object_t lobject
;
2688 vm_object_t nobject
;
2693 * This can eat a lot of time on a heavily contended
2694 * machine so yield on the tick if needed.
2700 * Calculate the page to pre-fault, stopping the scan in
2701 * each direction separately if the limit is reached.
2706 addr
= addra
- ((i
+ 1) >> 1) * PAGE_SIZE
;
2710 addr
= addra
+ ((i
+ 2) >> 1) * PAGE_SIZE
;
2712 if (addr
< entry
->start
) {
2718 if (addr
>= entry
->end
) {
2726 * Skip pages already mapped, and stop scanning in that
2727 * direction. When the scan terminates in both directions
2730 if (pmap_prefault_ok(pmap
, addr
) == 0) {
2741 * Follow the VM object chain to obtain the page to be mapped
2744 * If we reach the terminal object without finding a page
2745 * and we determine it would be advantageous, then allocate
2746 * a zero-fill page for the base object. The base object
2747 * is guaranteed to be OBJT_DEFAULT for this case.
2749 * In order to not have to check the pager via *haspage*()
2750 * we stop if any non-default object is encountered. e.g.
2751 * a vnode or swap object would stop the loop.
2753 index
= ((addr
- entry
->start
) + entry
->offset
) >> PAGE_SHIFT
;
2758 KKASSERT(lobject
== entry
->object
.vm_object
);
2759 /*vm_object_hold(lobject); implied */
2761 while ((m
= vm_page_lookup_busy_try(lobject
, pindex
,
2762 TRUE
, &error
)) == NULL
) {
2763 if (lobject
->type
!= OBJT_DEFAULT
)
2765 if (lobject
->backing_object
== NULL
) {
2766 if (vm_fast_fault
== 0)
2768 if ((prot
& VM_PROT_WRITE
) == 0 ||
2769 vm_page_count_min(0)) {
2774 * NOTE: Allocated from base object
2776 m
= vm_page_alloc(object
, index
,
2785 /* lobject = object .. not needed */
2788 if (lobject
->backing_object_offset
& PAGE_MASK
)
2790 nobject
= lobject
->backing_object
;
2791 vm_object_hold(nobject
);
2792 KKASSERT(nobject
== lobject
->backing_object
);
2793 pindex
+= lobject
->backing_object_offset
>> PAGE_SHIFT
;
2794 if (lobject
!= object
) {
2795 vm_object_lock_swap();
2796 vm_object_drop(lobject
);
2799 pprot
&= ~VM_PROT_WRITE
;
2800 vm_object_chain_acquire(lobject
, 0);
2804 * NOTE: A non-NULL (m) will be associated with lobject if
2805 * it was found there, otherwise it is probably a
2806 * zero-fill page associated with the base object.
2808 * Give-up if no page is available.
2811 if (lobject
!= object
) {
2813 if (object
->backing_object
!= lobject
)
2814 vm_object_hold(object
->backing_object
);
2816 vm_object_chain_release_all(
2817 object
->backing_object
, lobject
);
2819 if (object
->backing_object
!= lobject
)
2820 vm_object_drop(object
->backing_object
);
2822 vm_object_drop(lobject
);
2828 * The object must be marked dirty if we are mapping a
2829 * writable page. m->object is either lobject or object,
2830 * both of which are still held. Do this before we
2831 * potentially drop the object.
2833 if (pprot
& VM_PROT_WRITE
)
2834 vm_object_set_writeable_dirty(m
->object
);
2837 * Do not conditionalize on PG_RAM. If pages are present in
2838 * the VM system we assume optimal caching. If caching is
2839 * not optimal the I/O gravy train will be restarted when we
2840 * hit an unavailable page. We do not want to try to restart
2841 * the gravy train now because we really don't know how much
2842 * of the object has been cached. The cost for restarting
2843 * the gravy train should be low (since accesses will likely
2844 * be I/O bound anyway).
2846 if (lobject
!= object
) {
2848 if (object
->backing_object
!= lobject
)
2849 vm_object_hold(object
->backing_object
);
2851 vm_object_chain_release_all(object
->backing_object
,
2854 if (object
->backing_object
!= lobject
)
2855 vm_object_drop(object
->backing_object
);
2857 vm_object_drop(lobject
);
2861 * Enter the page into the pmap if appropriate. If we had
2862 * allocated the page we have to place it on a queue. If not
2863 * we just have to make sure it isn't on the cache queue
2864 * (pages on the cache queue are not allowed to be mapped).
2868 * Page must be zerod.
2870 vm_page_zero_fill(m
);
2871 mycpu
->gd_cnt
.v_zfod
++;
2872 m
->valid
= VM_PAGE_BITS_ALL
;
2875 * Handle dirty page case
2877 if (pprot
& VM_PROT_WRITE
)
2878 vm_set_nosync(m
, entry
);
2879 pmap_enter(pmap
, addr
, m
, pprot
, 0, entry
);
2880 mycpu
->gd_cnt
.v_vm_faults
++;
2881 if (curthread
->td_lwp
)
2882 ++curthread
->td_lwp
->lwp_ru
.ru_minflt
;
2883 vm_page_deactivate(m
);
2884 if (pprot
& VM_PROT_WRITE
) {
2885 /*vm_object_set_writeable_dirty(m->object);*/
2886 vm_set_nosync(m
, entry
);
2887 if (fault_flags
& VM_FAULT_DIRTY
) {
2890 swap_pager_unswapped(m
);
2895 /* couldn't busy page, no wakeup */
2897 ((m
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
2898 (m
->flags
& PG_FICTITIOUS
) == 0) {
2900 * A fully valid page not undergoing soft I/O can
2901 * be immediately entered into the pmap.
2903 if ((m
->queue
- m
->pc
) == PQ_CACHE
)
2904 vm_page_deactivate(m
);
2905 if (pprot
& VM_PROT_WRITE
) {
2906 /*vm_object_set_writeable_dirty(m->object);*/
2907 vm_set_nosync(m
, entry
);
2908 if (fault_flags
& VM_FAULT_DIRTY
) {
2911 swap_pager_unswapped(m
);
2914 if (pprot
& VM_PROT_WRITE
)
2915 vm_set_nosync(m
, entry
);
2916 pmap_enter(pmap
, addr
, m
, pprot
, 0, entry
);
2917 mycpu
->gd_cnt
.v_vm_faults
++;
2918 if (curthread
->td_lwp
)
2919 ++curthread
->td_lwp
->lwp_ru
.ru_minflt
;
2925 vm_object_chain_release(object
);
2926 vm_object_drop(object
);
2930 * Object can be held shared
2933 vm_prefault_quick(pmap_t pmap
, vm_offset_t addra
,
2934 vm_map_entry_t entry
, int prot
, int fault_flags
)
2947 * Get stable max count value, disabled if set to 0
2949 maxpages
= vm_prefault_pages
;
2955 * We do not currently prefault mappings that use virtual page
2956 * tables. We do not prefault foreign pmaps.
2958 if (entry
->maptype
!= VM_MAPTYPE_NORMAL
)
2960 lp
= curthread
->td_lwp
;
2961 if (lp
== NULL
|| (pmap
!= vmspace_pmap(lp
->lwp_vmspace
)))
2963 object
= entry
->object
.vm_object
;
2964 if (object
->backing_object
!= NULL
)
2966 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2969 * Limit pre-fault count to 1024 pages.
2971 if (maxpages
> 1024)
2976 for (i
= 0; i
< maxpages
; ++i
) {
2980 * Calculate the page to pre-fault, stopping the scan in
2981 * each direction separately if the limit is reached.
2986 addr
= addra
- ((i
+ 1) >> 1) * PAGE_SIZE
;
2990 addr
= addra
+ ((i
+ 2) >> 1) * PAGE_SIZE
;
2992 if (addr
< entry
->start
) {
2998 if (addr
>= entry
->end
) {
3006 * Follow the VM object chain to obtain the page to be mapped
3007 * into the pmap. This version of the prefault code only
3008 * works with terminal objects.
3010 * The page must already exist. If we encounter a problem
3013 * WARNING! We cannot call swap_pager_unswapped() or insert
3014 * a new vm_page with a shared token.
3016 pindex
= ((addr
- entry
->start
) + entry
->offset
) >> PAGE_SHIFT
;
3018 m
= vm_page_lookup_busy_try(object
, pindex
, TRUE
, &error
);
3019 if (m
== NULL
|| error
)
3023 * Skip pages already mapped, and stop scanning in that
3024 * direction. When the scan terminates in both directions
3027 if (pmap_prefault_ok(pmap
, addr
) == 0) {
3039 * Stop if the page cannot be trivially entered into the
3042 if (((m
->valid
& VM_PAGE_BITS_ALL
) != VM_PAGE_BITS_ALL
) ||
3043 (m
->flags
& PG_FICTITIOUS
) ||
3044 ((m
->flags
& PG_SWAPPED
) &&
3045 (prot
& VM_PROT_WRITE
) &&
3046 (fault_flags
& VM_FAULT_DIRTY
))) {
3052 * Enter the page into the pmap. The object might be held
3053 * shared so we can't do any (serious) modifying operation
3056 if ((m
->queue
- m
->pc
) == PQ_CACHE
)
3057 vm_page_deactivate(m
);
3058 if (prot
& VM_PROT_WRITE
) {
3059 vm_object_set_writeable_dirty(m
->object
);
3060 vm_set_nosync(m
, entry
);
3061 if (fault_flags
& VM_FAULT_DIRTY
) {
3063 /* can't happeen due to conditional above */
3064 /* swap_pager_unswapped(m); */
3067 pmap_enter(pmap
, addr
, m
, prot
, 0, entry
);
3068 mycpu
->gd_cnt
.v_vm_faults
++;
3069 if (curthread
->td_lwp
)
3070 ++curthread
->td_lwp
->lwp_ru
.ru_minflt
;