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
;
304 inherit_prot
= fault_type
& VM_PROT_NOSYNC
;
306 fs
.fault_flags
= fault_flags
;
308 fs
.shared
= vm_shared_fault
;
309 fs
.first_shared
= vm_shared_fault
;
313 * vm_map interactions
316 if ((lp
= td
->td_lwp
) != NULL
)
317 lp
->lwp_flags
|= LWP_PAGING
;
318 lwkt_gettoken(&map
->token
);
322 * Find the vm_map_entry representing the backing store and resolve
323 * the top level object and page index. This may have the side
324 * effect of executing a copy-on-write on the map entry and/or
325 * creating a shadow object, but will not COW any actual VM pages.
327 * On success fs.map is left read-locked and various other fields
328 * are initialized but not otherwise referenced or locked.
330 * NOTE! vm_map_lookup will try to upgrade the fault_type to
331 * VM_FAULT_WRITE if the map entry is a virtual page table
332 * and also writable, so we can set the 'A'accessed bit in
333 * the virtual page table entry.
336 result
= vm_map_lookup(&fs
.map
, vaddr
, fault_type
,
337 &fs
.entry
, &fs
.first_object
,
338 &first_pindex
, &fs
.first_prot
, &fs
.wired
);
341 * If the lookup failed or the map protections are incompatible,
342 * the fault generally fails.
344 * The failure could be due to TDF_NOFAULT if vm_map_lookup()
345 * tried to do a COW fault.
347 * If the caller is trying to do a user wiring we have more work
350 if (result
!= KERN_SUCCESS
) {
351 if (result
== KERN_FAILURE_NOFAULT
) {
352 result
= KERN_FAILURE
;
355 if (result
!= KERN_PROTECTION_FAILURE
||
356 (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) != VM_FAULT_USER_WIRE
)
358 if (result
== KERN_INVALID_ADDRESS
&& growstack
&&
359 map
!= &kernel_map
&& curproc
!= NULL
) {
360 result
= vm_map_growstack(map
, vaddr
);
361 if (result
== KERN_SUCCESS
) {
366 result
= KERN_FAILURE
;
372 * If we are user-wiring a r/w segment, and it is COW, then
373 * we need to do the COW operation. Note that we don't
374 * currently COW RO sections now, because it is NOT desirable
375 * to COW .text. We simply keep .text from ever being COW'ed
376 * and take the heat that one cannot debug wired .text sections.
378 result
= vm_map_lookup(&fs
.map
, vaddr
,
379 VM_PROT_READ
|VM_PROT_WRITE
|
380 VM_PROT_OVERRIDE_WRITE
,
381 &fs
.entry
, &fs
.first_object
,
382 &first_pindex
, &fs
.first_prot
,
384 if (result
!= KERN_SUCCESS
) {
385 /* could also be KERN_FAILURE_NOFAULT */
386 result
= KERN_FAILURE
;
391 * If we don't COW now, on a user wire, the user will never
392 * be able to write to the mapping. If we don't make this
393 * restriction, the bookkeeping would be nearly impossible.
395 * XXX We have a shared lock, this will have a MP race but
396 * I don't see how it can hurt anything.
398 if ((fs
.entry
->protection
& VM_PROT_WRITE
) == 0) {
399 atomic_clear_char(&fs
.entry
->max_protection
,
405 * fs.map is read-locked
407 * Misc checks. Save the map generation number to detect races.
409 fs
.map_generation
= fs
.map
->timestamp
;
410 fs
.lookup_still_valid
= TRUE
;
412 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
413 fs
.prot
= fs
.first_prot
; /* default (used by uksmap) */
415 if (fs
.entry
->eflags
& (MAP_ENTRY_NOFAULT
| MAP_ENTRY_KSTACK
)) {
416 if (fs
.entry
->eflags
& MAP_ENTRY_NOFAULT
) {
417 panic("vm_fault: fault on nofault entry, addr: %p",
420 if ((fs
.entry
->eflags
& MAP_ENTRY_KSTACK
) &&
421 vaddr
>= fs
.entry
->start
&&
422 vaddr
< fs
.entry
->start
+ PAGE_SIZE
) {
423 panic("vm_fault: fault on stack guard, addr: %p",
429 * A user-kernel shared map has no VM object and bypasses
430 * everything. We execute the uksmap function with a temporary
431 * fictitious vm_page. The address is directly mapped with no
434 if (fs
.entry
->maptype
== VM_MAPTYPE_UKSMAP
) {
435 struct vm_page fakem
;
437 bzero(&fakem
, sizeof(fakem
));
438 fakem
.pindex
= first_pindex
;
439 fakem
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_UNMANAGED
;
440 fakem
.valid
= VM_PAGE_BITS_ALL
;
441 fakem
.pat_mode
= VM_MEMATTR_DEFAULT
;
442 if (fs
.entry
->object
.uksmap(fs
.entry
->aux
.dev
, &fakem
)) {
443 result
= KERN_FAILURE
;
447 pmap_enter(fs
.map
->pmap
, vaddr
, &fakem
, fs
.prot
| inherit_prot
,
453 * A system map entry may return a NULL object. No object means
454 * no pager means an unrecoverable kernel fault.
456 if (fs
.first_object
== NULL
) {
457 panic("vm_fault: unrecoverable fault at %p in entry %p",
458 (void *)vaddr
, fs
.entry
);
462 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
465 * Unfortunately a deadlock can occur if we are forced to page-in
466 * from swap, but diving all the way into the vm_pager_get_page()
467 * function to find out is too much. Just check the object type.
469 * The deadlock is a CAM deadlock on a busy VM page when trying
470 * to finish an I/O if another process gets stuck in
471 * vop_helper_read_shortcut() due to a swap fault.
473 if ((td
->td_flags
& TDF_NOFAULT
) &&
475 fs
.first_object
->type
== OBJT_VNODE
||
476 fs
.first_object
->type
== OBJT_SWAP
||
477 fs
.first_object
->backing_object
)) {
478 result
= KERN_FAILURE
;
484 * If the entry is wired we cannot change the page protection.
487 fault_type
= fs
.first_prot
;
490 * We generally want to avoid unnecessary exclusive modes on backing
491 * and terminal objects because this can seriously interfere with
492 * heavily fork()'d processes (particularly /bin/sh scripts).
494 * However, we also want to avoid unnecessary retries due to needed
495 * shared->exclusive promotion for common faults. Exclusive mode is
496 * always needed if any page insertion, rename, or free occurs in an
497 * object (and also indirectly if any I/O is done).
499 * The main issue here is going to be fs.first_shared. If the
500 * first_object has a backing object which isn't shadowed and the
501 * process is single-threaded we might as well use an exclusive
502 * lock/chain right off the bat.
504 if (fs
.first_shared
&& fs
.first_object
->backing_object
&&
505 LIST_EMPTY(&fs
.first_object
->shadow_head
) &&
506 td
->td_proc
&& td
->td_proc
->p_nthreads
== 1) {
511 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
512 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but
513 * we can try shared first.
515 if (fault_flags
& VM_FAULT_UNSWAP
) {
520 * Obtain a top-level object lock, shared or exclusive depending
521 * on fs.first_shared. If a shared lock winds up being insufficient
522 * we will retry with an exclusive lock.
524 * The vnode pager lock is always shared.
527 vm_object_hold_shared(fs
.first_object
);
529 vm_object_hold(fs
.first_object
);
531 fs
.vp
= vnode_pager_lock(fs
.first_object
);
534 * The page we want is at (first_object, first_pindex), but if the
535 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
536 * page table to figure out the actual pindex.
538 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
541 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
542 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
543 fs
.entry
->aux
.master_pde
,
545 if (result
== KERN_TRY_AGAIN
) {
546 vm_object_drop(fs
.first_object
);
550 if (result
!= KERN_SUCCESS
)
555 * Now we have the actual (object, pindex), fault in the page. If
556 * vm_fault_object() fails it will unlock and deallocate the FS
557 * data. If it succeeds everything remains locked and fs->object
558 * will have an additional PIP count if it is not equal to
561 * vm_fault_object will set fs->prot for the pmap operation. It is
562 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
563 * page can be safely written. However, it will force a read-only
564 * mapping for a read fault if the memory is managed by a virtual
567 * If the fault code uses the shared object lock shortcut
568 * we must not try to burst (we can't allocate VM pages).
570 result
= vm_fault_object(&fs
, first_pindex
, fault_type
, 1);
572 if (debug_fault
> 0) {
574 kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x "
575 "fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n",
576 result
, (intmax_t)vaddr
, fault_type
, fault_flags
,
577 fs
.m
, fs
.prot
, fs
.wired
, fs
.entry
);
580 if (result
== KERN_TRY_AGAIN
) {
581 vm_object_drop(fs
.first_object
);
585 if (result
!= KERN_SUCCESS
)
589 * On success vm_fault_object() does not unlock or deallocate, and fs.m
590 * will contain a busied page.
592 * Enter the page into the pmap and do pmap-related adjustments.
594 KKASSERT(fs
.lookup_still_valid
== TRUE
);
595 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
596 pmap_enter(fs
.map
->pmap
, vaddr
, fs
.m
, fs
.prot
| inherit_prot
,
599 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
600 KKASSERT(fs
.m
->flags
& PG_BUSY
);
603 * If the page is not wired down, then put it where the pageout daemon
606 if (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) {
610 vm_page_unwire(fs
.m
, 1);
612 vm_page_activate(fs
.m
);
614 vm_page_wakeup(fs
.m
);
617 * Burst in a few more pages if possible. The fs.map should still
618 * be locked. To avoid interlocking against a vnode->getblk
619 * operation we had to be sure to unbusy our primary vm_page above
622 * A normal burst can continue down backing store, only execute
623 * if we are holding an exclusive lock, otherwise the exclusive
624 * locks the burst code gets might cause excessive SMP collisions.
626 * A quick burst can be utilized when there is no backing object
627 * (i.e. a shared file mmap).
629 if ((fault_flags
& VM_FAULT_BURST
) &&
630 (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) == 0 &&
632 if (fs
.first_shared
== 0 && fs
.shared
== 0) {
633 vm_prefault(fs
.map
->pmap
, vaddr
,
634 fs
.entry
, fs
.prot
, fault_flags
);
636 vm_prefault_quick(fs
.map
->pmap
, vaddr
,
637 fs
.entry
, fs
.prot
, fault_flags
);
642 mycpu
->gd_cnt
.v_vm_faults
++;
644 ++td
->td_lwp
->lwp_ru
.ru_minflt
;
647 * Unlock everything, and return
653 td
->td_lwp
->lwp_ru
.ru_majflt
++;
655 td
->td_lwp
->lwp_ru
.ru_minflt
++;
659 /*vm_object_deallocate(fs.first_object);*/
661 /*fs.first_object = NULL; must still drop later */
663 result
= KERN_SUCCESS
;
666 vm_object_drop(fs
.first_object
);
668 lwkt_reltoken(&map
->token
);
670 lp
->lwp_flags
&= ~LWP_PAGING
;
672 #if !defined(NO_SWAPPING)
674 * Check the process RSS limit and force deactivation and
675 * (asynchronous) paging if necessary. This is a complex operation,
676 * only do it for direct user-mode faults, for now.
678 * To reduce overhead implement approximately a ~16MB hysteresis.
681 if ((fault_flags
& VM_FAULT_USERMODE
) && lp
&&
682 p
->p_limit
&& map
->pmap
&& vm_pageout_memuse_mode
>= 1 &&
683 map
!= &kernel_map
) {
687 limit
= OFF_TO_IDX(qmin(p
->p_rlimit
[RLIMIT_RSS
].rlim_cur
,
688 p
->p_rlimit
[RLIMIT_RSS
].rlim_max
));
689 size
= pmap_resident_tlnw_count(map
->pmap
);
690 if (limit
>= 0 && size
> 4096 && size
- 4096 >= limit
) {
691 vm_pageout_map_deactivate_pages(map
, limit
);
700 * Fault in the specified virtual address in the current process map,
701 * returning a held VM page or NULL. See vm_fault_page() for more
707 vm_fault_page_quick(vm_offset_t va
, vm_prot_t fault_type
,
708 int *errorp
, int *busyp
)
710 struct lwp
*lp
= curthread
->td_lwp
;
713 m
= vm_fault_page(&lp
->lwp_vmspace
->vm_map
, va
,
714 fault_type
, VM_FAULT_NORMAL
,
720 * Fault in the specified virtual address in the specified map, doing all
721 * necessary manipulation of the object store and all necessary I/O. Return
722 * a held VM page or NULL, and set *errorp. The related pmap is not
725 * If busyp is not NULL then *busyp will be set to TRUE if this routine
726 * decides to return a busied page (aka VM_PROT_WRITE), or FALSE if it
727 * does not (VM_PROT_WRITE not specified or busyp is NULL). If busyp is
728 * NULL the returned page is only held.
730 * If the caller has no intention of writing to the page's contents, busyp
731 * can be passed as NULL along with VM_PROT_WRITE to force a COW operation
732 * without busying the page.
734 * The returned page will also be marked PG_REFERENCED.
736 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
737 * error will be returned.
742 vm_fault_page(vm_map_t map
, vm_offset_t vaddr
, vm_prot_t fault_type
,
743 int fault_flags
, int *errorp
, int *busyp
)
745 vm_pindex_t first_pindex
;
746 struct faultstate fs
;
750 vm_prot_t orig_fault_type
= fault_type
;
754 fs
.fault_flags
= fault_flags
;
755 KKASSERT((fault_flags
& VM_FAULT_WIRE_MASK
) == 0);
758 * Dive the pmap (concurrency possible). If we find the
759 * appropriate page we can terminate early and quickly.
761 * This works great for normal programs but will always return
762 * NULL for host lookups of vkernel maps in VMM mode.
764 fs
.m
= pmap_fault_page_quick(map
->pmap
, vaddr
, fault_type
, busyp
);
766 if (fault_type
& (VM_PROT_WRITE
|VM_PROT_OVERRIDE_WRITE
))
773 * Otherwise take a concurrency hit and do a formal page
777 fs
.shared
= vm_shared_fault
;
778 fs
.first_shared
= vm_shared_fault
;
780 lwkt_gettoken(&map
->token
);
783 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
784 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but
785 * we can try shared first.
787 if (fault_flags
& VM_FAULT_UNSWAP
) {
793 * Find the vm_map_entry representing the backing store and resolve
794 * the top level object and page index. This may have the side
795 * effect of executing a copy-on-write on the map entry and/or
796 * creating a shadow object, but will not COW any actual VM pages.
798 * On success fs.map is left read-locked and various other fields
799 * are initialized but not otherwise referenced or locked.
801 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
802 * if the map entry is a virtual page table and also writable,
803 * so we can set the 'A'accessed bit in the virtual page table
807 result
= vm_map_lookup(&fs
.map
, vaddr
, fault_type
,
808 &fs
.entry
, &fs
.first_object
,
809 &first_pindex
, &fs
.first_prot
, &fs
.wired
);
811 if (result
!= KERN_SUCCESS
) {
812 if (result
== KERN_FAILURE_NOFAULT
) {
813 *errorp
= KERN_FAILURE
;
817 if (result
!= KERN_PROTECTION_FAILURE
||
818 (fs
.fault_flags
& VM_FAULT_WIRE_MASK
) != VM_FAULT_USER_WIRE
)
820 if (result
== KERN_INVALID_ADDRESS
&& growstack
&&
821 map
!= &kernel_map
&& curproc
!= NULL
) {
822 result
= vm_map_growstack(map
, vaddr
);
823 if (result
== KERN_SUCCESS
) {
828 result
= KERN_FAILURE
;
836 * If we are user-wiring a r/w segment, and it is COW, then
837 * we need to do the COW operation. Note that we don't
838 * currently COW RO sections now, because it is NOT desirable
839 * to COW .text. We simply keep .text from ever being COW'ed
840 * and take the heat that one cannot debug wired .text sections.
842 result
= vm_map_lookup(&fs
.map
, vaddr
,
843 VM_PROT_READ
|VM_PROT_WRITE
|
844 VM_PROT_OVERRIDE_WRITE
,
845 &fs
.entry
, &fs
.first_object
,
846 &first_pindex
, &fs
.first_prot
,
848 if (result
!= KERN_SUCCESS
) {
849 /* could also be KERN_FAILURE_NOFAULT */
850 *errorp
= KERN_FAILURE
;
856 * If we don't COW now, on a user wire, the user will never
857 * be able to write to the mapping. If we don't make this
858 * restriction, the bookkeeping would be nearly impossible.
860 * XXX We have a shared lock, this will have a MP race but
861 * I don't see how it can hurt anything.
863 if ((fs
.entry
->protection
& VM_PROT_WRITE
) == 0) {
864 atomic_clear_char(&fs
.entry
->max_protection
,
870 * fs.map is read-locked
872 * Misc checks. Save the map generation number to detect races.
874 fs
.map_generation
= fs
.map
->timestamp
;
875 fs
.lookup_still_valid
= TRUE
;
877 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
879 if (fs
.entry
->eflags
& MAP_ENTRY_NOFAULT
) {
880 panic("vm_fault: fault on nofault entry, addr: %lx",
885 * A user-kernel shared map has no VM object and bypasses
886 * everything. We execute the uksmap function with a temporary
887 * fictitious vm_page. The address is directly mapped with no
890 if (fs
.entry
->maptype
== VM_MAPTYPE_UKSMAP
) {
891 struct vm_page fakem
;
893 bzero(&fakem
, sizeof(fakem
));
894 fakem
.pindex
= first_pindex
;
895 fakem
.flags
= PG_BUSY
| PG_FICTITIOUS
| PG_UNMANAGED
;
896 fakem
.valid
= VM_PAGE_BITS_ALL
;
897 fakem
.pat_mode
= VM_MEMATTR_DEFAULT
;
898 if (fs
.entry
->object
.uksmap(fs
.entry
->aux
.dev
, &fakem
)) {
899 *errorp
= KERN_FAILURE
;
904 fs
.m
= PHYS_TO_VM_PAGE(fakem
.phys_addr
);
907 *busyp
= 0; /* don't need to busy R or W */
915 * A system map entry may return a NULL object. No object means
916 * no pager means an unrecoverable kernel fault.
918 if (fs
.first_object
== NULL
) {
919 panic("vm_fault: unrecoverable fault at %p in entry %p",
920 (void *)vaddr
, fs
.entry
);
924 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
927 * Unfortunately a deadlock can occur if we are forced to page-in
928 * from swap, but diving all the way into the vm_pager_get_page()
929 * function to find out is too much. Just check the object type.
931 if ((curthread
->td_flags
& TDF_NOFAULT
) &&
933 fs
.first_object
->type
== OBJT_VNODE
||
934 fs
.first_object
->type
== OBJT_SWAP
||
935 fs
.first_object
->backing_object
)) {
936 *errorp
= KERN_FAILURE
;
943 * If the entry is wired we cannot change the page protection.
946 fault_type
= fs
.first_prot
;
949 * Make a reference to this object to prevent its disposal while we
950 * are messing with it. Once we have the reference, the map is free
951 * to be diddled. Since objects reference their shadows (and copies),
952 * they will stay around as well.
954 * The reference should also prevent an unexpected collapse of the
955 * parent that might move pages from the current object into the
956 * parent unexpectedly, resulting in corruption.
958 * Bump the paging-in-progress count to prevent size changes (e.g.
959 * truncation operations) during I/O. This must be done after
960 * obtaining the vnode lock in order to avoid possible deadlocks.
963 vm_object_hold_shared(fs
.first_object
);
965 vm_object_hold(fs
.first_object
);
967 fs
.vp
= vnode_pager_lock(fs
.first_object
); /* shared */
970 * The page we want is at (first_object, first_pindex), but if the
971 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
972 * page table to figure out the actual pindex.
974 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
977 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
978 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
979 fs
.entry
->aux
.master_pde
,
981 if (result
== KERN_TRY_AGAIN
) {
982 vm_object_drop(fs
.first_object
);
986 if (result
!= KERN_SUCCESS
) {
994 * Now we have the actual (object, pindex), fault in the page. If
995 * vm_fault_object() fails it will unlock and deallocate the FS
996 * data. If it succeeds everything remains locked and fs->object
997 * will have an additinal PIP count if it is not equal to
1001 result
= vm_fault_object(&fs
, first_pindex
, fault_type
, 1);
1003 if (result
== KERN_TRY_AGAIN
) {
1004 vm_object_drop(fs
.first_object
);
1008 if (result
!= KERN_SUCCESS
) {
1014 if ((orig_fault_type
& VM_PROT_WRITE
) &&
1015 (fs
.prot
& VM_PROT_WRITE
) == 0) {
1016 *errorp
= KERN_PROTECTION_FAILURE
;
1017 unlock_and_deallocate(&fs
);
1023 * DO NOT UPDATE THE PMAP!!! This function may be called for
1024 * a pmap unrelated to the current process pmap, in which case
1025 * the current cpu core will not be listed in the pmap's pm_active
1026 * mask. Thus invalidation interlocks will fail to work properly.
1028 * (for example, 'ps' uses procfs to read program arguments from
1029 * each process's stack).
1031 * In addition to the above this function will be called to acquire
1032 * a page that might already be faulted in, re-faulting it
1033 * continuously is a waste of time.
1035 * XXX could this have been the cause of our random seg-fault
1036 * issues? procfs accesses user stacks.
1038 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
1040 pmap_enter(fs
.map
->pmap
, vaddr
, fs
.m
, fs
.prot
, fs
.wired
, NULL
);
1041 mycpu
->gd_cnt
.v_vm_faults
++;
1042 if (curthread
->td_lwp
)
1043 ++curthread
->td_lwp
->lwp_ru
.ru_minflt
;
1047 * On success vm_fault_object() does not unlock or deallocate, and fs.m
1048 * will contain a busied page. So we must unlock here after having
1049 * messed with the pmap.
1054 * Return a held page. We are not doing any pmap manipulation so do
1055 * not set PG_MAPPED. However, adjust the page flags according to
1056 * the fault type because the caller may not use a managed pmapping
1057 * (so we don't want to lose the fact that the page will be dirtied
1058 * if a write fault was specified).
1060 if (fault_type
& VM_PROT_WRITE
)
1061 vm_page_dirty(fs
.m
);
1062 vm_page_activate(fs
.m
);
1064 if (curthread
->td_lwp
) {
1066 curthread
->td_lwp
->lwp_ru
.ru_majflt
++;
1068 curthread
->td_lwp
->lwp_ru
.ru_minflt
++;
1073 * Unlock everything, and return the held or busied page.
1076 if (fault_type
& (VM_PROT_WRITE
|VM_PROT_OVERRIDE_WRITE
)) {
1077 vm_page_dirty(fs
.m
);
1082 vm_page_wakeup(fs
.m
);
1086 vm_page_wakeup(fs
.m
);
1088 /*vm_object_deallocate(fs.first_object);*/
1089 /*fs.first_object = NULL; */
1093 if (fs
.first_object
)
1094 vm_object_drop(fs
.first_object
);
1096 lwkt_reltoken(&map
->token
);
1101 * Fault in the specified (object,offset), dirty the returned page as
1102 * needed. If the requested fault_type cannot be done NULL and an
1103 * error is returned.
1105 * A held (but not busied) page is returned.
1107 * The passed in object must be held as specified by the shared
1111 vm_fault_object_page(vm_object_t object
, vm_ooffset_t offset
,
1112 vm_prot_t fault_type
, int fault_flags
,
1113 int *sharedp
, int *errorp
)
1116 vm_pindex_t first_pindex
;
1117 struct faultstate fs
;
1118 struct vm_map_entry entry
;
1120 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
1121 bzero(&entry
, sizeof(entry
));
1122 entry
.object
.vm_object
= object
;
1123 entry
.maptype
= VM_MAPTYPE_NORMAL
;
1124 entry
.protection
= entry
.max_protection
= fault_type
;
1127 fs
.fault_flags
= fault_flags
;
1129 fs
.shared
= vm_shared_fault
;
1130 fs
.first_shared
= *sharedp
;
1132 KKASSERT((fault_flags
& VM_FAULT_WIRE_MASK
) == 0);
1135 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object
1136 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but
1137 * we can try shared first.
1139 if (fs
.first_shared
&& (fault_flags
& VM_FAULT_UNSWAP
)) {
1140 fs
.first_shared
= 0;
1141 vm_object_upgrade(object
);
1145 * Retry loop as needed (typically for shared->exclusive transitions)
1148 *sharedp
= fs
.first_shared
;
1149 first_pindex
= OFF_TO_IDX(offset
);
1150 fs
.first_object
= object
;
1152 fs
.first_prot
= fault_type
;
1154 /*fs.map_generation = 0; unused */
1157 * Make a reference to this object to prevent its disposal while we
1158 * are messing with it. Once we have the reference, the map is free
1159 * to be diddled. Since objects reference their shadows (and copies),
1160 * they will stay around as well.
1162 * The reference should also prevent an unexpected collapse of the
1163 * parent that might move pages from the current object into the
1164 * parent unexpectedly, resulting in corruption.
1166 * Bump the paging-in-progress count to prevent size changes (e.g.
1167 * truncation operations) during I/O. This must be done after
1168 * obtaining the vnode lock in order to avoid possible deadlocks.
1171 fs
.vp
= vnode_pager_lock(fs
.first_object
);
1173 fs
.lookup_still_valid
= TRUE
;
1175 fs
.object
= fs
.first_object
; /* so unlock_and_deallocate works */
1178 /* XXX future - ability to operate on VM object using vpagetable */
1179 if (fs
.entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
1180 result
= vm_fault_vpagetable(&fs
, &first_pindex
,
1181 fs
.entry
->aux
.master_pde
,
1183 if (result
== KERN_TRY_AGAIN
) {
1184 if (fs
.first_shared
== 0 && *sharedp
)
1185 vm_object_upgrade(object
);
1188 if (result
!= KERN_SUCCESS
) {
1196 * Now we have the actual (object, pindex), fault in the page. If
1197 * vm_fault_object() fails it will unlock and deallocate the FS
1198 * data. If it succeeds everything remains locked and fs->object
1199 * will have an additinal PIP count if it is not equal to
1202 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact.
1203 * We may have to upgrade its lock to handle the requested fault.
1205 result
= vm_fault_object(&fs
, first_pindex
, fault_type
, 0);
1207 if (result
== KERN_TRY_AGAIN
) {
1208 if (fs
.first_shared
== 0 && *sharedp
)
1209 vm_object_upgrade(object
);
1212 if (result
!= KERN_SUCCESS
) {
1217 if ((fault_type
& VM_PROT_WRITE
) && (fs
.prot
& VM_PROT_WRITE
) == 0) {
1218 *errorp
= KERN_PROTECTION_FAILURE
;
1219 unlock_and_deallocate(&fs
);
1224 * On success vm_fault_object() does not unlock or deallocate, so we
1225 * do it here. Note that the returned fs.m will be busied.
1230 * Return a held page. We are not doing any pmap manipulation so do
1231 * not set PG_MAPPED. However, adjust the page flags according to
1232 * the fault type because the caller may not use a managed pmapping
1233 * (so we don't want to lose the fact that the page will be dirtied
1234 * if a write fault was specified).
1237 vm_page_activate(fs
.m
);
1238 if ((fault_type
& VM_PROT_WRITE
) || (fault_flags
& VM_FAULT_DIRTY
))
1239 vm_page_dirty(fs
.m
);
1240 if (fault_flags
& VM_FAULT_UNSWAP
)
1241 swap_pager_unswapped(fs
.m
);
1244 * Indicate that the page was accessed.
1246 vm_page_flag_set(fs
.m
, PG_REFERENCED
);
1248 if (curthread
->td_lwp
) {
1250 curthread
->td_lwp
->lwp_ru
.ru_majflt
++;
1252 curthread
->td_lwp
->lwp_ru
.ru_minflt
++;
1257 * Unlock everything, and return the held page.
1259 vm_page_wakeup(fs
.m
);
1260 /*vm_object_deallocate(fs.first_object);*/
1261 /*fs.first_object = NULL; */
1268 * Translate the virtual page number (first_pindex) that is relative
1269 * to the address space into a logical page number that is relative to the
1270 * backing object. Use the virtual page table pointed to by (vpte).
1272 * Possibly downgrade the protection based on the vpte bits.
1274 * This implements an N-level page table. Any level can terminate the
1275 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
1276 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
1280 vm_fault_vpagetable(struct faultstate
*fs
, vm_pindex_t
*pindex
,
1281 vpte_t vpte
, int fault_type
, int allow_nofault
)
1284 struct lwbuf lwb_cache
;
1285 int vshift
= VPTE_FRAME_END
- PAGE_SHIFT
; /* index bits remaining */
1289 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs
->first_object
));
1292 * We cannot proceed if the vpte is not valid, not readable
1293 * for a read fault, or not writable for a write fault.
1295 if ((vpte
& VPTE_V
) == 0) {
1296 unlock_and_deallocate(fs
);
1297 return (KERN_FAILURE
);
1299 if ((fault_type
& VM_PROT_WRITE
) && (vpte
& VPTE_RW
) == 0) {
1300 unlock_and_deallocate(fs
);
1301 return (KERN_FAILURE
);
1303 if ((vpte
& VPTE_PS
) || vshift
== 0)
1307 * Get the page table page. Nominally we only read the page
1308 * table, but since we are actively setting VPTE_M and VPTE_A,
1309 * tell vm_fault_object() that we are writing it.
1311 * There is currently no real need to optimize this.
1313 result
= vm_fault_object(fs
, (vpte
& VPTE_FRAME
) >> PAGE_SHIFT
,
1314 VM_PROT_READ
|VM_PROT_WRITE
,
1316 if (result
!= KERN_SUCCESS
)
1320 * Process the returned fs.m and look up the page table
1321 * entry in the page table page.
1323 vshift
-= VPTE_PAGE_BITS
;
1324 lwb
= lwbuf_alloc(fs
->m
, &lwb_cache
);
1325 ptep
= ((vpte_t
*)lwbuf_kva(lwb
) +
1326 ((*pindex
>> vshift
) & VPTE_PAGE_MASK
));
1327 vm_page_activate(fs
->m
);
1330 * Page table write-back - entire operation including
1331 * validation of the pte must be atomic to avoid races
1332 * against the vkernel changing the pte.
1334 * If the vpte is valid for the* requested operation, do
1335 * a write-back to the page table.
1337 * XXX VPTE_M is not set properly for page directory pages.
1338 * It doesn't get set in the page directory if the page table
1339 * is modified during a read access.
1345 * Reload for the cmpset, but make sure the pte is
1352 if ((vpte
& VPTE_V
) == 0)
1355 if ((fault_type
& VM_PROT_WRITE
) && (vpte
& VPTE_RW
))
1356 nvpte
|= VPTE_M
| VPTE_A
;
1357 if (fault_type
& VM_PROT_READ
)
1361 if (atomic_cmpset_long(ptep
, vpte
, nvpte
)) {
1362 vm_page_dirty(fs
->m
);
1367 vm_page_flag_set(fs
->m
, PG_REFERENCED
);
1368 vm_page_wakeup(fs
->m
);
1370 cleanup_successful_fault(fs
);
1374 * When the vkernel sets VPTE_RW it expects the real kernel to
1375 * reflect VPTE_M back when the page is modified via the mapping.
1376 * In order to accomplish this the real kernel must map the page
1377 * read-only for read faults and use write faults to reflect VPTE_M
1380 * Once VPTE_M has been set, the real kernel's pte allows writing.
1381 * If the vkernel clears VPTE_M the vkernel must be sure to
1382 * MADV_INVAL the real kernel's mappings to force the real kernel
1383 * to re-fault on the next write so oit can set VPTE_M again.
1385 if ((fault_type
& VM_PROT_WRITE
) == 0 &&
1386 (vpte
& (VPTE_RW
| VPTE_M
)) != (VPTE_RW
| VPTE_M
)) {
1387 fs
->first_prot
&= ~VM_PROT_WRITE
;
1391 * Combine remaining address bits with the vpte.
1393 *pindex
= ((vpte
& VPTE_FRAME
) >> PAGE_SHIFT
) +
1394 (*pindex
& ((1L << vshift
) - 1));
1395 return (KERN_SUCCESS
);
1400 * This is the core of the vm_fault code.
1402 * Do all operations required to fault-in (fs.first_object, pindex). Run
1403 * through the shadow chain as necessary and do required COW or virtual
1404 * copy operations. The caller has already fully resolved the vm_map_entry
1405 * and, if appropriate, has created a copy-on-write layer. All we need to
1406 * do is iterate the object chain.
1408 * On failure (fs) is unlocked and deallocated and the caller may return or
1409 * retry depending on the failure code. On success (fs) is NOT unlocked or
1410 * deallocated, fs.m will contained a resolved, busied page, and fs.object
1411 * will have an additional PIP count if it is not equal to fs.first_object.
1413 * If locks based on fs->first_shared or fs->shared are insufficient,
1414 * clear the appropriate field(s) and return RETRY. COWs require that
1415 * first_shared be 0, while page allocations (or frees) require that
1416 * shared be 0. Renames require that both be 0.
1418 * NOTE! fs->[first_]shared might be set with VM_FAULT_DIRTY also set.
1419 * we will have to retry with it exclusive if the vm_page is
1422 * fs->first_object must be held on call.
1426 vm_fault_object(struct faultstate
*fs
, vm_pindex_t first_pindex
,
1427 vm_prot_t fault_type
, int allow_nofault
)
1429 vm_object_t next_object
;
1433 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs
->first_object
));
1434 fs
->prot
= fs
->first_prot
;
1435 fs
->object
= fs
->first_object
;
1436 pindex
= first_pindex
;
1438 vm_object_chain_acquire(fs
->first_object
, fs
->shared
);
1439 vm_object_pip_add(fs
->first_object
, 1);
1442 * If a read fault occurs we try to upgrade the page protection
1443 * and make it also writable if possible. There are three cases
1444 * where we cannot make the page mapping writable:
1446 * (1) The mapping is read-only or the VM object is read-only,
1447 * fs->prot above will simply not have VM_PROT_WRITE set.
1449 * (2) If the mapping is a virtual page table fs->first_prot will
1450 * have already been properly adjusted by vm_fault_vpagetable().
1451 * to detect writes so we can set VPTE_M in the virtual page
1452 * table. Used by vkernels.
1454 * (3) If the VM page is read-only or copy-on-write, upgrading would
1455 * just result in an unnecessary COW fault.
1457 * (4) If the pmap specifically requests A/M bit emulation, downgrade
1461 /* see vpagetable code */
1462 if (fs
->entry
->maptype
== VM_MAPTYPE_VPAGETABLE
) {
1463 if ((fault_type
& VM_PROT_WRITE
) == 0)
1464 fs
->prot
&= ~VM_PROT_WRITE
;
1468 if (curthread
->td_lwp
&& curthread
->td_lwp
->lwp_vmspace
&&
1469 pmap_emulate_ad_bits(&curthread
->td_lwp
->lwp_vmspace
->vm_pmap
)) {
1470 if ((fault_type
& VM_PROT_WRITE
) == 0)
1471 fs
->prot
&= ~VM_PROT_WRITE
;
1474 /* vm_object_hold(fs->object); implied b/c object == first_object */
1478 * The entire backing chain from first_object to object
1479 * inclusive is chainlocked.
1481 * If the object is dead, we stop here
1483 if (fs
->object
->flags
& OBJ_DEAD
) {
1484 vm_object_pip_wakeup(fs
->first_object
);
1485 vm_object_chain_release_all(fs
->first_object
,
1487 if (fs
->object
!= fs
->first_object
)
1488 vm_object_drop(fs
->object
);
1489 unlock_and_deallocate(fs
);
1490 return (KERN_PROTECTION_FAILURE
);
1494 * See if the page is resident. Wait/Retry if the page is
1495 * busy (lots of stuff may have changed so we can't continue
1498 * We can theoretically allow the soft-busy case on a read
1499 * fault if the page is marked valid, but since such
1500 * pages are typically already pmap'd, putting that
1501 * special case in might be more effort then it is
1502 * worth. We cannot under any circumstances mess
1503 * around with a vm_page_t->busy page except, perhaps,
1506 fs
->m
= vm_page_lookup_busy_try(fs
->object
, pindex
,
1509 vm_object_pip_wakeup(fs
->first_object
);
1510 vm_object_chain_release_all(fs
->first_object
,
1512 if (fs
->object
!= fs
->first_object
)
1513 vm_object_drop(fs
->object
);
1515 vm_page_sleep_busy(fs
->m
, TRUE
, "vmpfw");
1516 mycpu
->gd_cnt
.v_intrans
++;
1517 /*vm_object_deallocate(fs->first_object);*/
1518 /*fs->first_object = NULL;*/
1520 return (KERN_TRY_AGAIN
);
1524 * The page is busied for us.
1526 * If reactivating a page from PQ_CACHE we may have
1529 int queue
= fs
->m
->queue
;
1530 vm_page_unqueue_nowakeup(fs
->m
);
1532 if ((queue
- fs
->m
->pc
) == PQ_CACHE
&&
1533 vm_page_count_severe()) {
1534 vm_page_activate(fs
->m
);
1535 vm_page_wakeup(fs
->m
);
1537 vm_object_pip_wakeup(fs
->first_object
);
1538 vm_object_chain_release_all(fs
->first_object
,
1540 if (fs
->object
!= fs
->first_object
)
1541 vm_object_drop(fs
->object
);
1542 unlock_and_deallocate(fs
);
1543 if (allow_nofault
== 0 ||
1544 (curthread
->td_flags
& TDF_NOFAULT
) == 0) {
1549 if (td
->td_proc
&& (td
->td_proc
->p_flags
& P_LOWMEMKILL
))
1550 return (KERN_PROTECTION_FAILURE
);
1552 return (KERN_TRY_AGAIN
);
1556 * If it still isn't completely valid (readable),
1557 * or if a read-ahead-mark is set on the VM page,
1558 * jump to readrest, else we found the page and
1561 * We can release the spl once we have marked the
1564 if (fs
->m
->object
!= &kernel_object
) {
1565 if ((fs
->m
->valid
& VM_PAGE_BITS_ALL
) !=
1569 if (fs
->m
->flags
& PG_RAM
) {
1572 vm_page_flag_clear(fs
->m
, PG_RAM
);
1576 break; /* break to PAGE HAS BEEN FOUND */
1580 * Page is not resident, If this is the search termination
1581 * or the pager might contain the page, allocate a new page.
1583 if (TRYPAGER(fs
) || fs
->object
== fs
->first_object
) {
1585 * Allocating, must be exclusive.
1587 if (fs
->object
== fs
->first_object
&&
1589 fs
->first_shared
= 0;
1590 vm_object_pip_wakeup(fs
->first_object
);
1591 vm_object_chain_release_all(fs
->first_object
,
1593 if (fs
->object
!= fs
->first_object
)
1594 vm_object_drop(fs
->object
);
1595 unlock_and_deallocate(fs
);
1596 return (KERN_TRY_AGAIN
);
1598 if (fs
->object
!= fs
->first_object
&&
1600 fs
->first_shared
= 0;
1602 vm_object_pip_wakeup(fs
->first_object
);
1603 vm_object_chain_release_all(fs
->first_object
,
1605 if (fs
->object
!= fs
->first_object
)
1606 vm_object_drop(fs
->object
);
1607 unlock_and_deallocate(fs
);
1608 return (KERN_TRY_AGAIN
);
1612 * If the page is beyond the object size we fail
1614 if (pindex
>= fs
->object
->size
) {
1615 vm_object_pip_wakeup(fs
->first_object
);
1616 vm_object_chain_release_all(fs
->first_object
,
1618 if (fs
->object
!= fs
->first_object
)
1619 vm_object_drop(fs
->object
);
1620 unlock_and_deallocate(fs
);
1621 return (KERN_PROTECTION_FAILURE
);
1625 * Allocate a new page for this object/offset pair.
1627 * It is possible for the allocation to race, so
1631 if (!vm_page_count_severe()) {
1632 fs
->m
= vm_page_alloc(fs
->object
, pindex
,
1633 ((fs
->vp
|| fs
->object
->backing_object
) ?
1634 VM_ALLOC_NULL_OK
| VM_ALLOC_NORMAL
:
1635 VM_ALLOC_NULL_OK
| VM_ALLOC_NORMAL
|
1636 VM_ALLOC_USE_GD
| VM_ALLOC_ZERO
));
1638 if (fs
->m
== NULL
) {
1639 vm_object_pip_wakeup(fs
->first_object
);
1640 vm_object_chain_release_all(fs
->first_object
,
1642 if (fs
->object
!= fs
->first_object
)
1643 vm_object_drop(fs
->object
);
1644 unlock_and_deallocate(fs
);
1645 if (allow_nofault
== 0 ||
1646 (curthread
->td_flags
& TDF_NOFAULT
) == 0) {
1651 if (td
->td_proc
&& (td
->td_proc
->p_flags
& P_LOWMEMKILL
))
1652 return (KERN_PROTECTION_FAILURE
);
1654 return (KERN_TRY_AGAIN
);
1658 * Fall through to readrest. We have a new page which
1659 * will have to be paged (since m->valid will be 0).
1665 * We have found an invalid or partially valid page, a
1666 * page with a read-ahead mark which might be partially or
1667 * fully valid (and maybe dirty too), or we have allocated
1670 * Attempt to fault-in the page if there is a chance that the
1671 * pager has it, and potentially fault in additional pages
1674 * If TRYPAGER is true then fs.m will be non-NULL and busied
1680 u_char behavior
= vm_map_entry_behavior(fs
->entry
);
1682 if (behavior
== MAP_ENTRY_BEHAV_RANDOM
)
1688 * Doing I/O may synchronously insert additional
1689 * pages so we can't be shared at this point either.
1691 * NOTE: We can't free fs->m here in the allocated
1692 * case (fs->object != fs->first_object) as
1693 * this would require an exclusively locked
1696 if (fs
->object
== fs
->first_object
&&
1698 vm_page_deactivate(fs
->m
);
1699 vm_page_wakeup(fs
->m
);
1701 fs
->first_shared
= 0;
1702 vm_object_pip_wakeup(fs
->first_object
);
1703 vm_object_chain_release_all(fs
->first_object
,
1705 if (fs
->object
!= fs
->first_object
)
1706 vm_object_drop(fs
->object
);
1707 unlock_and_deallocate(fs
);
1708 return (KERN_TRY_AGAIN
);
1710 if (fs
->object
!= fs
->first_object
&&
1712 vm_page_deactivate(fs
->m
);
1713 vm_page_wakeup(fs
->m
);
1715 fs
->first_shared
= 0;
1717 vm_object_pip_wakeup(fs
->first_object
);
1718 vm_object_chain_release_all(fs
->first_object
,
1720 if (fs
->object
!= fs
->first_object
)
1721 vm_object_drop(fs
->object
);
1722 unlock_and_deallocate(fs
);
1723 return (KERN_TRY_AGAIN
);
1727 * Avoid deadlocking against the map when doing I/O.
1728 * fs.object and the page is PG_BUSY'd.
1730 * NOTE: Once unlocked, fs->entry can become stale
1731 * so this will NULL it out.
1733 * NOTE: fs->entry is invalid until we relock the
1734 * map and verify that the timestamp has not
1740 * Acquire the page data. We still hold a ref on
1741 * fs.object and the page has been PG_BUSY's.
1743 * The pager may replace the page (for example, in
1744 * order to enter a fictitious page into the
1745 * object). If it does so it is responsible for
1746 * cleaning up the passed page and properly setting
1747 * the new page PG_BUSY.
1749 * If we got here through a PG_RAM read-ahead
1750 * mark the page may be partially dirty and thus
1751 * not freeable. Don't bother checking to see
1752 * if the pager has the page because we can't free
1753 * it anyway. We have to depend on the get_page
1754 * operation filling in any gaps whether there is
1755 * backing store or not.
1757 rv
= vm_pager_get_page(fs
->object
, &fs
->m
, seqaccess
);
1759 if (rv
== VM_PAGER_OK
) {
1761 * Relookup in case pager changed page. Pager
1762 * is responsible for disposition of old page
1765 * XXX other code segments do relookups too.
1766 * It's a bad abstraction that needs to be
1769 fs
->m
= vm_page_lookup(fs
->object
, pindex
);
1770 if (fs
->m
== NULL
) {
1771 vm_object_pip_wakeup(fs
->first_object
);
1772 vm_object_chain_release_all(
1773 fs
->first_object
, fs
->object
);
1774 if (fs
->object
!= fs
->first_object
)
1775 vm_object_drop(fs
->object
);
1776 unlock_and_deallocate(fs
);
1777 return (KERN_TRY_AGAIN
);
1780 break; /* break to PAGE HAS BEEN FOUND */
1784 * Remove the bogus page (which does not exist at this
1785 * object/offset); before doing so, we must get back
1786 * our object lock to preserve our invariant.
1788 * Also wake up any other process that may want to bring
1791 * If this is the top-level object, we must leave the
1792 * busy page to prevent another process from rushing
1793 * past us, and inserting the page in that object at
1794 * the same time that we are.
1796 if (rv
== VM_PAGER_ERROR
) {
1798 kprintf("vm_fault: pager read error, "
1803 kprintf("vm_fault: pager read error, "
1811 * Data outside the range of the pager or an I/O error
1813 * The page may have been wired during the pagein,
1814 * e.g. by the buffer cache, and cannot simply be
1815 * freed. Call vnode_pager_freepage() to deal with it.
1817 * Also note that we cannot free the page if we are
1818 * holding the related object shared. XXX not sure
1819 * what to do in that case.
1821 if (fs
->object
!= fs
->first_object
) {
1822 vnode_pager_freepage(fs
->m
);
1825 * XXX - we cannot just fall out at this
1826 * point, m has been freed and is invalid!
1830 * XXX - the check for kernel_map is a kludge to work
1831 * around having the machine panic on a kernel space
1832 * fault w/ I/O error.
1834 if (((fs
->map
!= &kernel_map
) &&
1835 (rv
== VM_PAGER_ERROR
)) || (rv
== VM_PAGER_BAD
)) {
1837 if (fs
->first_shared
) {
1838 vm_page_deactivate(fs
->m
);
1839 vm_page_wakeup(fs
->m
);
1841 vnode_pager_freepage(fs
->m
);
1845 vm_object_pip_wakeup(fs
->first_object
);
1846 vm_object_chain_release_all(fs
->first_object
,
1848 if (fs
->object
!= fs
->first_object
)
1849 vm_object_drop(fs
->object
);
1850 unlock_and_deallocate(fs
);
1851 if (rv
== VM_PAGER_ERROR
)
1852 return (KERN_FAILURE
);
1854 return (KERN_PROTECTION_FAILURE
);
1860 * We get here if the object has a default pager (or unwiring)
1861 * or the pager doesn't have the page.
1863 * fs->first_m will be used for the COW unless we find a
1864 * deeper page to be mapped read-only, in which case the
1865 * unlock*(fs) will free first_m.
1867 if (fs
->object
== fs
->first_object
)
1868 fs
->first_m
= fs
->m
;
1871 * Move on to the next object. The chain lock should prevent
1872 * the backing_object from getting ripped out from under us.
1874 * The object lock for the next object is governed by
1877 if ((next_object
= fs
->object
->backing_object
) != NULL
) {
1879 vm_object_hold_shared(next_object
);
1881 vm_object_hold(next_object
);
1882 vm_object_chain_acquire(next_object
, fs
->shared
);
1883 KKASSERT(next_object
== fs
->object
->backing_object
);
1884 pindex
+= OFF_TO_IDX(fs
->object
->backing_object_offset
);
1887 if (next_object
== NULL
) {
1889 * If there's no object left, fill the page in the top
1890 * object with zeros.
1892 if (fs
->object
!= fs
->first_object
) {
1894 if (fs
->first_object
->backing_object
!=
1896 vm_object_hold(fs
->first_object
->backing_object
);
1899 vm_object_chain_release_all(
1900 fs
->first_object
->backing_object
,
1903 if (fs
->first_object
->backing_object
!=
1905 vm_object_drop(fs
->first_object
->backing_object
);
1908 vm_object_pip_wakeup(fs
->object
);
1909 vm_object_drop(fs
->object
);
1910 fs
->object
= fs
->first_object
;
1911 pindex
= first_pindex
;
1912 fs
->m
= fs
->first_m
;
1917 * Zero the page and mark it valid.
1919 vm_page_zero_fill(fs
->m
);
1920 mycpu
->gd_cnt
.v_zfod
++;
1921 fs
->m
->valid
= VM_PAGE_BITS_ALL
;
1922 break; /* break to PAGE HAS BEEN FOUND */
1924 if (fs
->object
!= fs
->first_object
) {
1925 vm_object_pip_wakeup(fs
->object
);
1926 vm_object_lock_swap();
1927 vm_object_drop(fs
->object
);
1929 KASSERT(fs
->object
!= next_object
,
1930 ("object loop %p", next_object
));
1931 fs
->object
= next_object
;
1932 vm_object_pip_add(fs
->object
, 1);
1936 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1939 * object still held.
1941 * local shared variable may be different from fs->shared.
1943 * If the page is being written, but isn't already owned by the
1944 * top-level object, we have to copy it into a new page owned by the
1947 KASSERT((fs
->m
->flags
& PG_BUSY
) != 0,
1948 ("vm_fault: not busy after main loop"));
1950 if (fs
->object
!= fs
->first_object
) {
1952 * We only really need to copy if we want to write it.
1954 if (fault_type
& VM_PROT_WRITE
) {
1956 * This allows pages to be virtually copied from a
1957 * backing_object into the first_object, where the
1958 * backing object has no other refs to it, and cannot
1959 * gain any more refs. Instead of a bcopy, we just
1960 * move the page from the backing object to the
1961 * first object. Note that we must mark the page
1962 * dirty in the first object so that it will go out
1963 * to swap when needed.
1967 * Must be holding exclusive locks
1969 fs
->first_shared
== 0 &&
1972 * Map, if present, has not changed
1975 fs
->map_generation
== fs
->map
->timestamp
) &&
1977 * Only one shadow object
1979 (fs
->object
->shadow_count
== 1) &&
1981 * No COW refs, except us
1983 (fs
->object
->ref_count
== 1) &&
1985 * No one else can look this object up
1987 (fs
->object
->handle
== NULL
) &&
1989 * No other ways to look the object up
1991 ((fs
->object
->type
== OBJT_DEFAULT
) ||
1992 (fs
->object
->type
== OBJT_SWAP
)) &&
1994 * We don't chase down the shadow chain
1996 (fs
->object
== fs
->first_object
->backing_object
) &&
1999 * grab the lock if we need to
2001 (fs
->lookup_still_valid
||
2003 lockmgr(&fs
->map
->lock
, LK_EXCLUSIVE
|LK_NOWAIT
) == 0)
2006 * (first_m) and (m) are both busied. We have
2007 * move (m) into (first_m)'s object/pindex
2008 * in an atomic fashion, then free (first_m).
2010 * first_object is held so second remove
2011 * followed by the rename should wind
2012 * up being atomic. vm_page_free() might
2013 * block so we don't do it until after the
2016 fs
->lookup_still_valid
= 1;
2017 vm_page_protect(fs
->first_m
, VM_PROT_NONE
);
2018 vm_page_remove(fs
->first_m
);
2019 vm_page_rename(fs
->m
, fs
->first_object
,
2021 vm_page_free(fs
->first_m
);
2022 fs
->first_m
= fs
->m
;
2024 mycpu
->gd_cnt
.v_cow_optim
++;
2027 * Oh, well, lets copy it.
2029 * Why are we unmapping the original page
2030 * here? Well, in short, not all accessors
2031 * of user memory go through the pmap. The
2032 * procfs code doesn't have access user memory
2033 * via a local pmap, so vm_fault_page*()
2034 * can't call pmap_enter(). And the umtx*()
2035 * code may modify the COW'd page via a DMAP
2036 * or kernel mapping and not via the pmap,
2037 * leaving the original page still mapped
2038 * read-only into the pmap.
2040 * So we have to remove the page from at
2041 * least the current pmap if it is in it.
2042 * Just remove it from all pmaps.
2044 KKASSERT(fs
->first_shared
== 0);
2045 vm_page_copy(fs
->m
, fs
->first_m
);
2046 vm_page_protect(fs
->m
, VM_PROT_NONE
);
2047 vm_page_event(fs
->m
, VMEVENT_COW
);
2051 * We no longer need the old page or object.
2057 * We intend to revert to first_object, undo the
2058 * chain lock through to that.
2061 if (fs
->first_object
->backing_object
!= fs
->object
)
2062 vm_object_hold(fs
->first_object
->backing_object
);
2064 vm_object_chain_release_all(
2065 fs
->first_object
->backing_object
,
2068 if (fs
->first_object
->backing_object
!= fs
->object
)
2069 vm_object_drop(fs
->first_object
->backing_object
);
2073 * fs->object != fs->first_object due to above
2076 vm_object_pip_wakeup(fs
->object
);
2077 vm_object_drop(fs
->object
);
2080 * Only use the new page below...
2082 mycpu
->gd_cnt
.v_cow_faults
++;
2083 fs
->m
= fs
->first_m
;
2084 fs
->object
= fs
->first_object
;
2085 pindex
= first_pindex
;
2088 * If it wasn't a write fault avoid having to copy
2089 * the page by mapping it read-only.
2091 fs
->prot
&= ~VM_PROT_WRITE
;
2096 * Relock the map if necessary, then check the generation count.
2097 * relock_map() will update fs->timestamp to account for the
2098 * relocking if necessary.
2100 * If the count has changed after relocking then all sorts of
2101 * crap may have happened and we have to retry.
2103 * NOTE: The relock_map() can fail due to a deadlock against
2104 * the vm_page we are holding BUSY.
2106 if (fs
->lookup_still_valid
== FALSE
&& fs
->map
) {
2107 if (relock_map(fs
) ||
2108 fs
->map
->timestamp
!= fs
->map_generation
) {
2110 vm_object_pip_wakeup(fs
->first_object
);
2111 vm_object_chain_release_all(fs
->first_object
,
2113 if (fs
->object
!= fs
->first_object
)
2114 vm_object_drop(fs
->object
);
2115 unlock_and_deallocate(fs
);
2116 return (KERN_TRY_AGAIN
);
2121 * If the fault is a write, we know that this page is being
2122 * written NOW so dirty it explicitly to save on pmap_is_modified()
2125 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
2126 * if the page is already dirty to prevent data written with
2127 * the expectation of being synced from not being synced.
2128 * Likewise if this entry does not request NOSYNC then make
2129 * sure the page isn't marked NOSYNC. Applications sharing
2130 * data should use the same flags to avoid ping ponging.
2132 * Also tell the backing pager, if any, that it should remove
2133 * any swap backing since the page is now dirty.
2135 vm_page_activate(fs
->m
);
2136 if (fs
->prot
& VM_PROT_WRITE
) {
2137 vm_object_set_writeable_dirty(fs
->m
->object
);
2138 vm_set_nosync(fs
->m
, fs
->entry
);
2139 if (fs
->fault_flags
& VM_FAULT_DIRTY
) {
2140 vm_page_dirty(fs
->m
);
2141 if (fs
->m
->flags
& PG_SWAPPED
) {
2143 * If the page is swapped out we have to call
2144 * swap_pager_unswapped() which requires an
2145 * exclusive object lock. If we are shared,
2146 * we must clear the shared flag and retry.
2148 if ((fs
->object
== fs
->first_object
&&
2149 fs
->first_shared
) ||
2150 (fs
->object
!= fs
->first_object
&&
2152 vm_page_wakeup(fs
->m
);
2154 if (fs
->object
== fs
->first_object
)
2155 fs
->first_shared
= 0;
2158 vm_object_pip_wakeup(fs
->first_object
);
2159 vm_object_chain_release_all(
2160 fs
->first_object
, fs
->object
);
2161 if (fs
->object
!= fs
->first_object
)
2162 vm_object_drop(fs
->object
);
2163 unlock_and_deallocate(fs
);
2164 return (KERN_TRY_AGAIN
);
2166 swap_pager_unswapped(fs
->m
);
2171 vm_object_pip_wakeup(fs
->first_object
);
2172 vm_object_chain_release_all(fs
->first_object
, fs
->object
);
2173 if (fs
->object
!= fs
->first_object
)
2174 vm_object_drop(fs
->object
);
2177 * Page had better still be busy. We are still locked up and
2178 * fs->object will have another PIP reference if it is not equal
2179 * to fs->first_object.
2181 KASSERT(fs
->m
->flags
& PG_BUSY
,
2182 ("vm_fault: page %p not busy!", fs
->m
));
2185 * Sanity check: page must be completely valid or it is not fit to
2186 * map into user space. vm_pager_get_pages() ensures this.
2188 if (fs
->m
->valid
!= VM_PAGE_BITS_ALL
) {
2189 vm_page_zero_invalid(fs
->m
, TRUE
);
2190 kprintf("Warning: page %p partially invalid on fault\n", fs
->m
);
2193 return (KERN_SUCCESS
);
2197 * Wire down a range of virtual addresses in a map. The entry in question
2198 * should be marked in-transition and the map must be locked. We must
2199 * release the map temporarily while faulting-in the page to avoid a
2200 * deadlock. Note that the entry may be clipped while we are blocked but
2201 * will never be freed.
2206 vm_fault_wire(vm_map_t map
, vm_map_entry_t entry
,
2207 boolean_t user_wire
, int kmflags
)
2209 boolean_t fictitious
;
2219 lwkt_gettoken(&map
->token
);
2222 wire_prot
= VM_PROT_READ
;
2223 fault_flags
= VM_FAULT_USER_WIRE
;
2225 wire_prot
= VM_PROT_READ
| VM_PROT_WRITE
;
2226 fault_flags
= VM_FAULT_CHANGE_WIRING
;
2228 if (kmflags
& KM_NOTLBSYNC
)
2229 wire_prot
|= VM_PROT_NOSYNC
;
2231 pmap
= vm_map_pmap(map
);
2232 start
= entry
->start
;
2234 switch(entry
->maptype
) {
2235 case VM_MAPTYPE_NORMAL
:
2236 case VM_MAPTYPE_VPAGETABLE
:
2237 fictitious
= entry
->object
.vm_object
&&
2238 ((entry
->object
.vm_object
->type
== OBJT_DEVICE
) ||
2239 (entry
->object
.vm_object
->type
== OBJT_MGTDEVICE
));
2241 case VM_MAPTYPE_UKSMAP
:
2249 if (entry
->eflags
& MAP_ENTRY_KSTACK
)
2255 * We simulate a fault to get the page and enter it in the physical
2258 for (va
= start
; va
< end
; va
+= PAGE_SIZE
) {
2259 rv
= vm_fault(map
, va
, wire_prot
, fault_flags
);
2261 while (va
> start
) {
2263 m
= pmap_unwire(pmap
, va
);
2264 if (m
&& !fictitious
) {
2265 vm_page_busy_wait(m
, FALSE
, "vmwrpg");
2266 vm_page_unwire(m
, 1);
2276 lwkt_reltoken(&map
->token
);
2281 * Unwire a range of virtual addresses in a map. The map should be
2285 vm_fault_unwire(vm_map_t map
, vm_map_entry_t entry
)
2287 boolean_t fictitious
;
2294 lwkt_gettoken(&map
->token
);
2296 pmap
= vm_map_pmap(map
);
2297 start
= entry
->start
;
2299 fictitious
= entry
->object
.vm_object
&&
2300 ((entry
->object
.vm_object
->type
== OBJT_DEVICE
) ||
2301 (entry
->object
.vm_object
->type
== OBJT_MGTDEVICE
));
2302 if (entry
->eflags
& MAP_ENTRY_KSTACK
)
2306 * Since the pages are wired down, we must be able to get their
2307 * mappings from the physical map system.
2309 for (va
= start
; va
< end
; va
+= PAGE_SIZE
) {
2310 m
= pmap_unwire(pmap
, va
);
2311 if (m
&& !fictitious
) {
2312 vm_page_busy_wait(m
, FALSE
, "vmwrpg");
2313 vm_page_unwire(m
, 1);
2317 lwkt_reltoken(&map
->token
);
2321 * Copy all of the pages from a wired-down map entry to another.
2323 * The source and destination maps must be locked for write.
2324 * The source and destination maps token must be held
2325 * The source map entry must be wired down (or be a sharing map
2326 * entry corresponding to a main map entry that is wired down).
2328 * No other requirements.
2330 * XXX do segment optimization
2333 vm_fault_copy_entry(vm_map_t dst_map
, vm_map_t src_map
,
2334 vm_map_entry_t dst_entry
, vm_map_entry_t src_entry
)
2336 vm_object_t dst_object
;
2337 vm_object_t src_object
;
2338 vm_ooffset_t dst_offset
;
2339 vm_ooffset_t src_offset
;
2345 src_object
= src_entry
->object
.vm_object
;
2346 src_offset
= src_entry
->offset
;
2349 * Create the top-level object for the destination entry. (Doesn't
2350 * actually shadow anything - we copy the pages directly.)
2352 vm_map_entry_allocate_object(dst_entry
);
2353 dst_object
= dst_entry
->object
.vm_object
;
2355 prot
= dst_entry
->max_protection
;
2358 * Loop through all of the pages in the entry's range, copying each
2359 * one from the source object (it should be there) to the destination
2362 vm_object_hold(src_object
);
2363 vm_object_hold(dst_object
);
2364 for (vaddr
= dst_entry
->start
, dst_offset
= 0;
2365 vaddr
< dst_entry
->end
;
2366 vaddr
+= PAGE_SIZE
, dst_offset
+= PAGE_SIZE
) {
2369 * Allocate a page in the destination object
2372 dst_m
= vm_page_alloc(dst_object
,
2373 OFF_TO_IDX(dst_offset
),
2375 if (dst_m
== NULL
) {
2378 } while (dst_m
== NULL
);
2381 * Find the page in the source object, and copy it in.
2382 * (Because the source is wired down, the page will be in
2385 src_m
= vm_page_lookup(src_object
,
2386 OFF_TO_IDX(dst_offset
+ src_offset
));
2388 panic("vm_fault_copy_wired: page missing");
2390 vm_page_copy(src_m
, dst_m
);
2391 vm_page_event(src_m
, VMEVENT_COW
);
2394 * Enter it in the pmap...
2396 pmap_enter(dst_map
->pmap
, vaddr
, dst_m
, prot
, FALSE
, dst_entry
);
2399 * Mark it no longer busy, and put it on the active list.
2401 vm_page_activate(dst_m
);
2402 vm_page_wakeup(dst_m
);
2404 vm_object_drop(dst_object
);
2405 vm_object_drop(src_object
);
2411 * This routine checks around the requested page for other pages that
2412 * might be able to be faulted in. This routine brackets the viable
2413 * pages for the pages to be paged in.
2416 * m, rbehind, rahead
2419 * marray (array of vm_page_t), reqpage (index of requested page)
2422 * number of pages in marray
2425 vm_fault_additional_pages(vm_page_t m
, int rbehind
, int rahead
,
2426 vm_page_t
*marray
, int *reqpage
)
2430 vm_pindex_t pindex
, startpindex
, endpindex
, tpindex
;
2432 int cbehind
, cahead
;
2438 * we don't fault-ahead for device pager
2440 if ((object
->type
== OBJT_DEVICE
) ||
2441 (object
->type
== OBJT_MGTDEVICE
)) {
2448 * if the requested page is not available, then give up now
2450 if (!vm_pager_has_page(object
, pindex
, &cbehind
, &cahead
)) {
2451 *reqpage
= 0; /* not used by caller, fix compiler warn */
2455 if ((cbehind
== 0) && (cahead
== 0)) {
2461 if (rahead
> cahead
) {
2465 if (rbehind
> cbehind
) {
2470 * Do not do any readahead if we have insufficient free memory.
2472 * XXX code was broken disabled before and has instability
2473 * with this conditonal fixed, so shortcut for now.
2475 if (burst_fault
== 0 || vm_page_count_severe()) {
2482 * scan backward for the read behind pages -- in memory
2484 * Assume that if the page is not found an interrupt will not
2485 * create it. Theoretically interrupts can only remove (busy)
2486 * pages, not create new associations.
2489 if (rbehind
> pindex
) {
2493 startpindex
= pindex
- rbehind
;
2496 vm_object_hold(object
);
2497 for (tpindex
= pindex
; tpindex
> startpindex
; --tpindex
) {
2498 if (vm_page_lookup(object
, tpindex
- 1))
2503 while (tpindex
< pindex
) {
2504 rtm
= vm_page_alloc(object
, tpindex
, VM_ALLOC_SYSTEM
|
2507 for (j
= 0; j
< i
; j
++) {
2508 vm_page_free(marray
[j
]);
2510 vm_object_drop(object
);
2519 vm_object_drop(object
);
2525 * Assign requested page
2532 * Scan forwards for read-ahead pages
2534 tpindex
= pindex
+ 1;
2535 endpindex
= tpindex
+ rahead
;
2536 if (endpindex
> object
->size
)
2537 endpindex
= object
->size
;
2539 vm_object_hold(object
);
2540 while (tpindex
< endpindex
) {
2541 if (vm_page_lookup(object
, tpindex
))
2543 rtm
= vm_page_alloc(object
, tpindex
, VM_ALLOC_SYSTEM
|
2551 vm_object_drop(object
);
2559 * vm_prefault() provides a quick way of clustering pagefaults into a
2560 * processes address space. It is a "cousin" of pmap_object_init_pt,
2561 * except it runs at page fault time instead of mmap time.
2563 * vm.fast_fault Enables pre-faulting zero-fill pages
2565 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to
2566 * prefault. Scan stops in either direction when
2567 * a page is found to already exist.
2569 * This code used to be per-platform pmap_prefault(). It is now
2570 * machine-independent and enhanced to also pre-fault zero-fill pages
2571 * (see vm.fast_fault) as well as make them writable, which greatly
2572 * reduces the number of page faults programs incur.
2574 * Application performance when pre-faulting zero-fill pages is heavily
2575 * dependent on the application. Very tiny applications like /bin/echo
2576 * lose a little performance while applications of any appreciable size
2577 * gain performance. Prefaulting multiple pages also reduces SMP
2578 * congestion and can improve SMP performance significantly.
2580 * NOTE! prot may allow writing but this only applies to the top level
2581 * object. If we wind up mapping a page extracted from a backing
2582 * object we have to make sure it is read-only.
2584 * NOTE! The caller has already handled any COW operations on the
2585 * vm_map_entry via the normal fault code. Do NOT call this
2586 * shortcut unless the normal fault code has run on this entry.
2588 * The related map must be locked.
2589 * No other requirements.
2591 static int vm_prefault_pages
= 8;
2592 SYSCTL_INT(_vm
, OID_AUTO
, prefault_pages
, CTLFLAG_RW
, &vm_prefault_pages
, 0,
2593 "Maximum number of pages to pre-fault");
2594 static int vm_fast_fault
= 1;
2595 SYSCTL_INT(_vm
, OID_AUTO
, fast_fault
, CTLFLAG_RW
, &vm_fast_fault
, 0,
2596 "Burst fault zero-fill regions");
2599 * Set PG_NOSYNC if the map entry indicates so, but only if the page
2600 * is not already dirty by other means. This will prevent passive
2601 * filesystem syncing as well as 'sync' from writing out the page.
2604 vm_set_nosync(vm_page_t m
, vm_map_entry_t entry
)
2606 if (entry
->eflags
& MAP_ENTRY_NOSYNC
) {
2608 vm_page_flag_set(m
, PG_NOSYNC
);
2610 vm_page_flag_clear(m
, PG_NOSYNC
);
2615 vm_prefault(pmap_t pmap
, vm_offset_t addra
, vm_map_entry_t entry
, int prot
,
2631 * Get stable max count value, disabled if set to 0
2633 maxpages
= vm_prefault_pages
;
2639 * We do not currently prefault mappings that use virtual page
2640 * tables. We do not prefault foreign pmaps.
2642 if (entry
->maptype
!= VM_MAPTYPE_NORMAL
)
2644 lp
= curthread
->td_lwp
;
2645 if (lp
== NULL
|| (pmap
!= vmspace_pmap(lp
->lwp_vmspace
)))
2649 * Limit pre-fault count to 1024 pages.
2651 if (maxpages
> 1024)
2654 object
= entry
->object
.vm_object
;
2655 KKASSERT(object
!= NULL
);
2656 KKASSERT(object
== entry
->object
.vm_object
);
2659 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively
2660 * now (or do something more complex XXX).
2662 vm_object_hold(object
);
2663 vm_object_chain_acquire(object
, 0);
2667 for (i
= 0; i
< maxpages
; ++i
) {
2668 vm_object_t lobject
;
2669 vm_object_t nobject
;
2674 * This can eat a lot of time on a heavily contended
2675 * machine so yield on the tick if needed.
2681 * Calculate the page to pre-fault, stopping the scan in
2682 * each direction separately if the limit is reached.
2687 addr
= addra
- ((i
+ 1) >> 1) * PAGE_SIZE
;
2691 addr
= addra
+ ((i
+ 2) >> 1) * PAGE_SIZE
;
2693 if (addr
< entry
->start
) {
2699 if (addr
>= entry
->end
) {
2707 * Skip pages already mapped, and stop scanning in that
2708 * direction. When the scan terminates in both directions
2711 if (pmap_prefault_ok(pmap
, addr
) == 0) {
2722 * Follow the VM object chain to obtain the page to be mapped
2725 * If we reach the terminal object without finding a page
2726 * and we determine it would be advantageous, then allocate
2727 * a zero-fill page for the base object. The base object
2728 * is guaranteed to be OBJT_DEFAULT for this case.
2730 * In order to not have to check the pager via *haspage*()
2731 * we stop if any non-default object is encountered. e.g.
2732 * a vnode or swap object would stop the loop.
2734 index
= ((addr
- entry
->start
) + entry
->offset
) >> PAGE_SHIFT
;
2739 KKASSERT(lobject
== entry
->object
.vm_object
);
2740 /*vm_object_hold(lobject); implied */
2742 while ((m
= vm_page_lookup_busy_try(lobject
, pindex
,
2743 TRUE
, &error
)) == NULL
) {
2744 if (lobject
->type
!= OBJT_DEFAULT
)
2746 if (lobject
->backing_object
== NULL
) {
2747 if (vm_fast_fault
== 0)
2749 if ((prot
& VM_PROT_WRITE
) == 0 ||
2750 vm_page_count_min(0)) {
2755 * NOTE: Allocated from base object
2757 m
= vm_page_alloc(object
, index
,
2766 /* lobject = object .. not needed */
2769 if (lobject
->backing_object_offset
& PAGE_MASK
)
2771 nobject
= lobject
->backing_object
;
2772 vm_object_hold(nobject
);
2773 KKASSERT(nobject
== lobject
->backing_object
);
2774 pindex
+= lobject
->backing_object_offset
>> PAGE_SHIFT
;
2775 if (lobject
!= object
) {
2776 vm_object_lock_swap();
2777 vm_object_drop(lobject
);
2780 pprot
&= ~VM_PROT_WRITE
;
2781 vm_object_chain_acquire(lobject
, 0);
2785 * NOTE: A non-NULL (m) will be associated with lobject if
2786 * it was found there, otherwise it is probably a
2787 * zero-fill page associated with the base object.
2789 * Give-up if no page is available.
2792 if (lobject
!= object
) {
2794 if (object
->backing_object
!= lobject
)
2795 vm_object_hold(object
->backing_object
);
2797 vm_object_chain_release_all(
2798 object
->backing_object
, lobject
);
2800 if (object
->backing_object
!= lobject
)
2801 vm_object_drop(object
->backing_object
);
2803 vm_object_drop(lobject
);
2809 * The object must be marked dirty if we are mapping a
2810 * writable page. m->object is either lobject or object,
2811 * both of which are still held. Do this before we
2812 * potentially drop the object.
2814 if (pprot
& VM_PROT_WRITE
)
2815 vm_object_set_writeable_dirty(m
->object
);
2818 * Do not conditionalize on PG_RAM. If pages are present in
2819 * the VM system we assume optimal caching. If caching is
2820 * not optimal the I/O gravy train will be restarted when we
2821 * hit an unavailable page. We do not want to try to restart
2822 * the gravy train now because we really don't know how much
2823 * of the object has been cached. The cost for restarting
2824 * the gravy train should be low (since accesses will likely
2825 * be I/O bound anyway).
2827 if (lobject
!= object
) {
2829 if (object
->backing_object
!= lobject
)
2830 vm_object_hold(object
->backing_object
);
2832 vm_object_chain_release_all(object
->backing_object
,
2835 if (object
->backing_object
!= lobject
)
2836 vm_object_drop(object
->backing_object
);
2838 vm_object_drop(lobject
);
2842 * Enter the page into the pmap if appropriate. If we had
2843 * allocated the page we have to place it on a queue. If not
2844 * we just have to make sure it isn't on the cache queue
2845 * (pages on the cache queue are not allowed to be mapped).
2849 * Page must be zerod.
2851 vm_page_zero_fill(m
);
2852 mycpu
->gd_cnt
.v_zfod
++;
2853 m
->valid
= VM_PAGE_BITS_ALL
;
2856 * Handle dirty page case
2858 if (pprot
& VM_PROT_WRITE
)
2859 vm_set_nosync(m
, entry
);
2860 pmap_enter(pmap
, addr
, m
, pprot
, 0, entry
);
2861 mycpu
->gd_cnt
.v_vm_faults
++;
2862 if (curthread
->td_lwp
)
2863 ++curthread
->td_lwp
->lwp_ru
.ru_minflt
;
2864 vm_page_deactivate(m
);
2865 if (pprot
& VM_PROT_WRITE
) {
2866 /*vm_object_set_writeable_dirty(m->object);*/
2867 vm_set_nosync(m
, entry
);
2868 if (fault_flags
& VM_FAULT_DIRTY
) {
2871 swap_pager_unswapped(m
);
2876 /* couldn't busy page, no wakeup */
2878 ((m
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
2879 (m
->flags
& PG_FICTITIOUS
) == 0) {
2881 * A fully valid page not undergoing soft I/O can
2882 * be immediately entered into the pmap.
2884 if ((m
->queue
- m
->pc
) == PQ_CACHE
)
2885 vm_page_deactivate(m
);
2886 if (pprot
& VM_PROT_WRITE
) {
2887 /*vm_object_set_writeable_dirty(m->object);*/
2888 vm_set_nosync(m
, entry
);
2889 if (fault_flags
& VM_FAULT_DIRTY
) {
2892 swap_pager_unswapped(m
);
2895 if (pprot
& VM_PROT_WRITE
)
2896 vm_set_nosync(m
, entry
);
2897 pmap_enter(pmap
, addr
, m
, pprot
, 0, entry
);
2898 mycpu
->gd_cnt
.v_vm_faults
++;
2899 if (curthread
->td_lwp
)
2900 ++curthread
->td_lwp
->lwp_ru
.ru_minflt
;
2906 vm_object_chain_release(object
);
2907 vm_object_drop(object
);
2911 * Object can be held shared
2914 vm_prefault_quick(pmap_t pmap
, vm_offset_t addra
,
2915 vm_map_entry_t entry
, int prot
, int fault_flags
)
2928 * Get stable max count value, disabled if set to 0
2930 maxpages
= vm_prefault_pages
;
2936 * We do not currently prefault mappings that use virtual page
2937 * tables. We do not prefault foreign pmaps.
2939 if (entry
->maptype
!= VM_MAPTYPE_NORMAL
)
2941 lp
= curthread
->td_lwp
;
2942 if (lp
== NULL
|| (pmap
!= vmspace_pmap(lp
->lwp_vmspace
)))
2944 object
= entry
->object
.vm_object
;
2945 if (object
->backing_object
!= NULL
)
2947 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object
));
2950 * Limit pre-fault count to 1024 pages.
2952 if (maxpages
> 1024)
2957 for (i
= 0; i
< maxpages
; ++i
) {
2961 * Calculate the page to pre-fault, stopping the scan in
2962 * each direction separately if the limit is reached.
2967 addr
= addra
- ((i
+ 1) >> 1) * PAGE_SIZE
;
2971 addr
= addra
+ ((i
+ 2) >> 1) * PAGE_SIZE
;
2973 if (addr
< entry
->start
) {
2979 if (addr
>= entry
->end
) {
2987 * Follow the VM object chain to obtain the page to be mapped
2988 * into the pmap. This version of the prefault code only
2989 * works with terminal objects.
2991 * The page must already exist. If we encounter a problem
2994 * WARNING! We cannot call swap_pager_unswapped() or insert
2995 * a new vm_page with a shared token.
2997 pindex
= ((addr
- entry
->start
) + entry
->offset
) >> PAGE_SHIFT
;
2999 m
= vm_page_lookup_busy_try(object
, pindex
, TRUE
, &error
);
3000 if (m
== NULL
|| error
)
3004 * Skip pages already mapped, and stop scanning in that
3005 * direction. When the scan terminates in both directions
3008 if (pmap_prefault_ok(pmap
, addr
) == 0) {
3020 * Stop if the page cannot be trivially entered into the
3023 if (((m
->valid
& VM_PAGE_BITS_ALL
) != VM_PAGE_BITS_ALL
) ||
3024 (m
->flags
& PG_FICTITIOUS
) ||
3025 ((m
->flags
& PG_SWAPPED
) &&
3026 (prot
& VM_PROT_WRITE
) &&
3027 (fault_flags
& VM_FAULT_DIRTY
))) {
3033 * Enter the page into the pmap. The object might be held
3034 * shared so we can't do any (serious) modifying operation
3037 if ((m
->queue
- m
->pc
) == PQ_CACHE
)
3038 vm_page_deactivate(m
);
3039 if (prot
& VM_PROT_WRITE
) {
3040 vm_object_set_writeable_dirty(m
->object
);
3041 vm_set_nosync(m
, entry
);
3042 if (fault_flags
& VM_FAULT_DIRTY
) {
3044 /* can't happeen due to conditional above */
3045 /* swap_pager_unswapped(m); */
3048 pmap_enter(pmap
, addr
, m
, prot
, 0, entry
);
3049 mycpu
->gd_cnt
.v_vm_faults
++;
3050 if (curthread
->td_lwp
)
3051 ++curthread
->td_lwp
->lwp_ru
.ru_minflt
;