4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
12 * This code is derived from software contributed to Berkeley by
13 * The Mach Operating System project at Carnegie-Mellon University.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. Neither the name of the University nor the names of its contributors
24 * may be used to endorse or promote products derived from this software
25 * without specific prior written permission.
27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
42 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
43 * All rights reserved.
45 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
47 * Permission to use, copy, modify and distribute this software and
48 * its documentation is hereby granted, provided that both the copyright
49 * notice and this permission notice appear in all copies of the
50 * software, derivative works or modified versions, and any portions
51 * thereof, and that both notices appear in supporting documentation.
53 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
54 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
55 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
57 * Carnegie Mellon requests users of this software to return to
59 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
60 * School of Computer Science
61 * Carnegie Mellon University
62 * Pittsburgh PA 15213-3890
64 * any improvements or extensions that they make and grant Carnegie the
65 * rights to redistribute these changes.
67 * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
68 * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $
72 * Page fault handling module.
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
79 #include <sys/vnode.h>
80 #include <sys/resourcevar.h>
81 #include <sys/vmmeter.h>
82 #include <sys/vkernel.h>
84 #include <sys/sysctl.h>
86 #include <cpu/lwbuf.h>
89 #include <vm/vm_param.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vnode_pager.h>
98 #include <vm/vm_extern.h>
100 #include <sys/thread2.h>
101 #include <vm/vm_page2.h>
109 vm_object_t first_object;
110 vm_prot_t first_prot;
112 vm_map_entry_t entry;
113 int lookup_still_valid;
123 static int debug_cluster = 0;
124 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
125 int vm_shared_fault = 1;
126 TUNABLE_INT("vm.shared_fault", &vm_shared_fault);
127 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, &vm_shared_fault, 0,
128 "Allow shared token on vm_object");
129 static long vm_shared_hit = 0;
130 SYSCTL_LONG(_vm, OID_AUTO, shared_hit, CTLFLAG_RW, &vm_shared_hit, 0,
131 "Successful shared faults");
132 static long vm_shared_count = 0;
133 SYSCTL_LONG(_vm, OID_AUTO, shared_count, CTLFLAG_RW, &vm_shared_count, 0,
134 "Shared fault attempts");
135 static long vm_shared_miss = 0;
136 SYSCTL_LONG(_vm, OID_AUTO, shared_miss, CTLFLAG_RW, &vm_shared_miss, 0,
137 "Unsuccessful shared faults");
139 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int);
140 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *,
143 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
145 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry);
146 static void vm_prefault(pmap_t pmap, vm_offset_t addra,
147 vm_map_entry_t entry, int prot, int fault_flags);
148 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
149 vm_map_entry_t entry, int prot, int fault_flags);
152 release_page(struct faultstate *fs)
154 vm_page_deactivate(fs->m);
155 vm_page_wakeup(fs->m);
160 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse
161 * requires relocking and then checking the timestamp.
163 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do
164 * not have to update fs->map_generation here.
166 * NOTE: This function can fail due to a deadlock against the caller's
167 * holding of a vm_page BUSY.
170 relock_map(struct faultstate *fs)
174 if (fs->lookup_still_valid == FALSE && fs->map) {
175 error = vm_map_lock_read_to(fs->map);
177 fs->lookup_still_valid = TRUE;
185 unlock_map(struct faultstate *fs)
187 if (fs->lookup_still_valid && fs->map) {
188 vm_map_lookup_done(fs->map, fs->entry, 0);
189 fs->lookup_still_valid = FALSE;
194 * Clean up after a successful call to vm_fault_object() so another call
195 * to vm_fault_object() can be made.
198 _cleanup_successful_fault(struct faultstate *fs, int relock)
201 * We allocated a junk page for a COW operation that did
202 * not occur, the page must be freed.
204 if (fs->object != fs->first_object) {
205 KKASSERT(fs->first_shared == 0);
206 vm_page_free(fs->first_m);
207 vm_object_pip_wakeup(fs->object);
214 fs->object = fs->first_object;
215 if (relock && fs->lookup_still_valid == FALSE) {
217 vm_map_lock_read(fs->map);
218 fs->lookup_still_valid = TRUE;
223 _unlock_things(struct faultstate *fs, int dealloc)
225 _cleanup_successful_fault(fs, 0);
227 /*vm_object_deallocate(fs->first_object);*/
228 /*fs->first_object = NULL; drop used later on */
231 if (fs->vp != NULL) {
237 #define unlock_things(fs) _unlock_things(fs, 0)
238 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
239 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
244 * Determine if the pager for the current object *might* contain the page.
246 * We only need to try the pager if this is not a default object (default
247 * objects are zero-fill and have no real pager), and if we are not taking
248 * a wiring fault or if the FS entry is wired.
250 #define TRYPAGER(fs) \
251 (fs->object->type != OBJT_DEFAULT && \
252 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
257 * Handle a page fault occuring at the given address, requiring the given
258 * permissions, in the map specified. If successful, the page is inserted
259 * into the associated physical map.
261 * NOTE: The given address should be truncated to the proper page address.
263 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
264 * a standard error specifying why the fault is fatal is returned.
266 * The map in question must be referenced, and remains so.
267 * The caller may hold no locks.
268 * No other requirements.
271 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
274 vm_pindex_t first_pindex;
275 struct faultstate fs;
280 vm_page_pcpu_cache();
282 fs.fault_flags = fault_flags;
284 fs.shared = vm_shared_fault;
285 fs.first_shared = vm_shared_fault;
291 * vm_map interactions
293 if ((lp = curthread->td_lwp) != NULL)
294 lp->lwp_flags |= LWP_PAGING;
295 lwkt_gettoken(&map->token);
299 * Find the vm_map_entry representing the backing store and resolve
300 * the top level object and page index. This may have the side
301 * effect of executing a copy-on-write on the map entry and/or
302 * creating a shadow object, but will not COW any actual VM pages.
304 * On success fs.map is left read-locked and various other fields
305 * are initialized but not otherwise referenced or locked.
307 * NOTE! vm_map_lookup will try to upgrade the fault_type to
308 * VM_FAULT_WRITE if the map entry is a virtual page table and also
309 * writable, so we can set the 'A'accessed bit in the virtual page
313 result = vm_map_lookup(&fs.map, vaddr, fault_type,
314 &fs.entry, &fs.first_object,
315 &first_pindex, &fs.first_prot, &fs.wired);
318 * If the lookup failed or the map protections are incompatible,
319 * the fault generally fails. However, if the caller is trying
320 * to do a user wiring we have more work to do.
322 if (result != KERN_SUCCESS) {
323 if (result != KERN_PROTECTION_FAILURE ||
324 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
326 if (result == KERN_INVALID_ADDRESS && growstack &&
327 map != &kernel_map && curproc != NULL) {
328 result = vm_map_growstack(curproc, vaddr);
329 if (result == KERN_SUCCESS) {
334 result = KERN_FAILURE;
340 * If we are user-wiring a r/w segment, and it is COW, then
341 * we need to do the COW operation. Note that we don't
342 * currently COW RO sections now, because it is NOT desirable
343 * to COW .text. We simply keep .text from ever being COW'ed
344 * and take the heat that one cannot debug wired .text sections.
346 result = vm_map_lookup(&fs.map, vaddr,
347 VM_PROT_READ|VM_PROT_WRITE|
348 VM_PROT_OVERRIDE_WRITE,
349 &fs.entry, &fs.first_object,
350 &first_pindex, &fs.first_prot,
352 if (result != KERN_SUCCESS) {
353 result = KERN_FAILURE;
358 * If we don't COW now, on a user wire, the user will never
359 * be able to write to the mapping. If we don't make this
360 * restriction, the bookkeeping would be nearly impossible.
362 * XXX We have a shared lock, this will have a MP race but
363 * I don't see how it can hurt anything.
365 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
366 fs.entry->max_protection &= ~VM_PROT_WRITE;
370 * fs.map is read-locked
372 * Misc checks. Save the map generation number to detect races.
374 fs.map_generation = fs.map->timestamp;
375 fs.lookup_still_valid = TRUE;
377 fs.object = fs.first_object; /* so unlock_and_deallocate works */
379 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) {
380 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
381 panic("vm_fault: fault on nofault entry, addr: %p",
384 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) &&
385 vaddr >= fs.entry->start &&
386 vaddr < fs.entry->start + PAGE_SIZE) {
387 panic("vm_fault: fault on stack guard, addr: %p",
393 * A system map entry may return a NULL object. No object means
394 * no pager means an unrecoverable kernel fault.
396 if (fs.first_object == NULL) {
397 panic("vm_fault: unrecoverable fault at %p in entry %p",
398 (void *)vaddr, fs.entry);
402 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
405 if ((curthread->td_flags & TDF_NOFAULT) &&
407 fs.first_object->type == OBJT_VNODE ||
408 fs.first_object->backing_object)) {
409 result = KERN_FAILURE;
415 * If the entry is wired we cannot change the page protection.
418 fault_type = fs.first_prot;
421 * We generally want to avoid unnecessary exclusive modes on backing
422 * and terminal objects because this can seriously interfere with
423 * heavily fork()'d processes (particularly /bin/sh scripts).
425 * However, we also want to avoid unnecessary retries due to needed
426 * shared->exclusive promotion for common faults. Exclusive mode is
427 * always needed if any page insertion, rename, or free occurs in an
428 * object (and also indirectly if any I/O is done).
430 * The main issue here is going to be fs.first_shared. If the
431 * first_object has a backing object which isn't shadowed and the
432 * process is single-threaded we might as well use an exclusive
433 * lock/chain right off the bat.
435 if (fs.first_shared && fs.first_object->backing_object &&
436 LIST_EMPTY(&fs.first_object->shadow_head) &&
437 curthread->td_proc && curthread->td_proc->p_nthreads == 1) {
443 * Obtain a top-level object lock, shared or exclusive depending
444 * on fs.first_shared. If a shared lock winds up being insufficient
445 * we will retry with an exclusive lock.
447 * The vnode pager lock is always shared.
450 vm_object_hold_shared(fs.first_object);
452 vm_object_hold(fs.first_object);
454 fs.vp = vnode_pager_lock(fs.first_object);
457 * The page we want is at (first_object, first_pindex), but if the
458 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
459 * page table to figure out the actual pindex.
461 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
464 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
465 result = vm_fault_vpagetable(&fs, &first_pindex,
466 fs.entry->aux.master_pde,
468 if (result == KERN_TRY_AGAIN) {
469 vm_object_drop(fs.first_object);
473 if (result != KERN_SUCCESS)
478 * Now we have the actual (object, pindex), fault in the page. If
479 * vm_fault_object() fails it will unlock and deallocate the FS
480 * data. If it succeeds everything remains locked and fs->object
481 * will have an additional PIP count if it is not equal to
484 * vm_fault_object will set fs->prot for the pmap operation. It is
485 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
486 * page can be safely written. However, it will force a read-only
487 * mapping for a read fault if the memory is managed by a virtual
490 * If the fault code uses the shared object lock shortcut
491 * we must not try to burst (we can't allocate VM pages).
493 result = vm_fault_object(&fs, first_pindex, fault_type, 1);
494 if (result == KERN_TRY_AGAIN) {
495 vm_object_drop(fs.first_object);
499 if (result != KERN_SUCCESS)
503 * On success vm_fault_object() does not unlock or deallocate, and fs.m
504 * will contain a busied page.
506 * Enter the page into the pmap and do pmap-related adjustments.
508 KKASSERT(fs.lookup_still_valid == TRUE);
509 vm_page_flag_set(fs.m, PG_REFERENCED);
510 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, fs.entry);
511 mycpu->gd_cnt.v_vm_faults++;
512 if (curthread->td_lwp)
513 ++curthread->td_lwp->lwp_ru.ru_minflt;
515 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
516 KKASSERT(fs.m->flags & PG_BUSY);
519 * If the page is not wired down, then put it where the pageout daemon
522 if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
526 vm_page_unwire(fs.m, 1);
528 vm_page_activate(fs.m);
530 vm_page_wakeup(fs.m);
533 * Burst in a few more pages if possible. The fs.map should still
534 * be locked. To avoid interlocking against a vnode->getblk
535 * operation we had to be sure to unbusy our primary vm_page above
538 * A normal burst can continue down backing store, only execute
539 * if we are holding an exclusive lock, otherwise the exclusive
540 * locks the burst code gets might cause excessive SMP collisions.
542 * A quick burst can be utilized when there is no backing object
543 * (i.e. a shared file mmap).
545 if ((fault_flags & VM_FAULT_BURST) &&
546 (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
548 if (fs.first_shared == 0 && fs.shared == 0) {
549 vm_prefault(fs.map->pmap, vaddr,
550 fs.entry, fs.prot, fault_flags);
552 vm_prefault_quick(fs.map->pmap, vaddr,
553 fs.entry, fs.prot, fault_flags);
558 * Unlock everything, and return
562 if (curthread->td_lwp) {
564 curthread->td_lwp->lwp_ru.ru_majflt++;
566 curthread->td_lwp->lwp_ru.ru_minflt++;
570 /*vm_object_deallocate(fs.first_object);*/
572 /*fs.first_object = NULL; must still drop later */
574 result = KERN_SUCCESS;
577 vm_object_drop(fs.first_object);
579 lwkt_reltoken(&map->token);
581 lp->lwp_flags &= ~LWP_PAGING;
582 if (vm_shared_fault && fs.shared == 0)
588 * Fault in the specified virtual address in the current process map,
589 * returning a held VM page or NULL. See vm_fault_page() for more
595 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
597 struct lwp *lp = curthread->td_lwp;
600 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
601 fault_type, VM_FAULT_NORMAL, errorp);
606 * Fault in the specified virtual address in the specified map, doing all
607 * necessary manipulation of the object store and all necessary I/O. Return
608 * a held VM page or NULL, and set *errorp. The related pmap is not
611 * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
612 * and marked PG_REFERENCED as well.
614 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
615 * error will be returned.
620 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
621 int fault_flags, int *errorp)
623 vm_pindex_t first_pindex;
624 struct faultstate fs;
627 vm_prot_t orig_fault_type = fault_type;
630 fs.fault_flags = fault_flags;
631 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
634 * Dive the pmap (concurrency possible). If we find the
635 * appropriate page we can terminate early and quickly.
637 fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type);
644 * Otherwise take a concurrency hit and do a formal page
647 fs.shared = vm_shared_fault;
648 fs.first_shared = vm_shared_fault;
650 lwkt_gettoken(&map->token);
654 * Find the vm_map_entry representing the backing store and resolve
655 * the top level object and page index. This may have the side
656 * effect of executing a copy-on-write on the map entry and/or
657 * creating a shadow object, but will not COW any actual VM pages.
659 * On success fs.map is left read-locked and various other fields
660 * are initialized but not otherwise referenced or locked.
662 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
663 * if the map entry is a virtual page table and also writable,
664 * so we can set the 'A'accessed bit in the virtual page table entry.
667 result = vm_map_lookup(&fs.map, vaddr, fault_type,
668 &fs.entry, &fs.first_object,
669 &first_pindex, &fs.first_prot, &fs.wired);
671 if (result != KERN_SUCCESS) {
678 * fs.map is read-locked
680 * Misc checks. Save the map generation number to detect races.
682 fs.map_generation = fs.map->timestamp;
683 fs.lookup_still_valid = TRUE;
685 fs.object = fs.first_object; /* so unlock_and_deallocate works */
687 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
688 panic("vm_fault: fault on nofault entry, addr: %lx",
693 * A system map entry may return a NULL object. No object means
694 * no pager means an unrecoverable kernel fault.
696 if (fs.first_object == NULL) {
697 panic("vm_fault: unrecoverable fault at %p in entry %p",
698 (void *)vaddr, fs.entry);
702 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT
705 if ((curthread->td_flags & TDF_NOFAULT) &&
707 fs.first_object->type == OBJT_VNODE ||
708 fs.first_object->backing_object)) {
709 *errorp = KERN_FAILURE;
715 * If the entry is wired we cannot change the page protection.
718 fault_type = fs.first_prot;
721 * Make a reference to this object to prevent its disposal while we
722 * are messing with it. Once we have the reference, the map is free
723 * to be diddled. Since objects reference their shadows (and copies),
724 * they will stay around as well.
726 * The reference should also prevent an unexpected collapse of the
727 * parent that might move pages from the current object into the
728 * parent unexpectedly, resulting in corruption.
730 * Bump the paging-in-progress count to prevent size changes (e.g.
731 * truncation operations) during I/O. This must be done after
732 * obtaining the vnode lock in order to avoid possible deadlocks.
735 vm_object_hold_shared(fs.first_object);
737 vm_object_hold(fs.first_object);
739 fs.vp = vnode_pager_lock(fs.first_object); /* shared */
742 * The page we want is at (first_object, first_pindex), but if the
743 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
744 * page table to figure out the actual pindex.
746 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
749 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
750 result = vm_fault_vpagetable(&fs, &first_pindex,
751 fs.entry->aux.master_pde,
753 if (result == KERN_TRY_AGAIN) {
754 vm_object_drop(fs.first_object);
758 if (result != KERN_SUCCESS) {
766 * Now we have the actual (object, pindex), fault in the page. If
767 * vm_fault_object() fails it will unlock and deallocate the FS
768 * data. If it succeeds everything remains locked and fs->object
769 * will have an additinal PIP count if it is not equal to
773 result = vm_fault_object(&fs, first_pindex, fault_type, 1);
775 if (result == KERN_TRY_AGAIN) {
776 vm_object_drop(fs.first_object);
780 if (result != KERN_SUCCESS) {
786 if ((orig_fault_type & VM_PROT_WRITE) &&
787 (fs.prot & VM_PROT_WRITE) == 0) {
788 *errorp = KERN_PROTECTION_FAILURE;
789 unlock_and_deallocate(&fs);
795 * DO NOT UPDATE THE PMAP!!! This function may be called for
796 * a pmap unrelated to the current process pmap, in which case
797 * the current cpu core will not be listed in the pmap's pm_active
798 * mask. Thus invalidation interlocks will fail to work properly.
800 * (for example, 'ps' uses procfs to read program arguments from
801 * each process's stack).
803 * In addition to the above this function will be called to acquire
804 * a page that might already be faulted in, re-faulting it
805 * continuously is a waste of time.
807 * XXX could this have been the cause of our random seg-fault
808 * issues? procfs accesses user stacks.
810 vm_page_flag_set(fs.m, PG_REFERENCED);
812 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL);
813 mycpu->gd_cnt.v_vm_faults++;
814 if (curthread->td_lwp)
815 ++curthread->td_lwp->lwp_ru.ru_minflt;
819 * On success vm_fault_object() does not unlock or deallocate, and fs.m
820 * will contain a busied page. So we must unlock here after having
821 * messed with the pmap.
826 * Return a held page. We are not doing any pmap manipulation so do
827 * not set PG_MAPPED. However, adjust the page flags according to
828 * the fault type because the caller may not use a managed pmapping
829 * (so we don't want to lose the fact that the page will be dirtied
830 * if a write fault was specified).
833 vm_page_activate(fs.m);
834 if (fault_type & VM_PROT_WRITE)
837 if (curthread->td_lwp) {
839 curthread->td_lwp->lwp_ru.ru_majflt++;
841 curthread->td_lwp->lwp_ru.ru_minflt++;
846 * Unlock everything, and return the held page.
848 vm_page_wakeup(fs.m);
849 /*vm_object_deallocate(fs.first_object);*/
850 /*fs.first_object = NULL; */
855 vm_object_drop(fs.first_object);
857 lwkt_reltoken(&map->token);
862 * Fault in the specified (object,offset), dirty the returned page as
863 * needed. If the requested fault_type cannot be done NULL and an
866 * A held (but not busied) page is returned.
868 * The passed in object must be held as specified by the shared
872 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
873 vm_prot_t fault_type, int fault_flags,
874 int *sharedp, int *errorp)
877 vm_pindex_t first_pindex;
878 struct faultstate fs;
879 struct vm_map_entry entry;
881 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
882 bzero(&entry, sizeof(entry));
883 entry.object.vm_object = object;
884 entry.maptype = VM_MAPTYPE_NORMAL;
885 entry.protection = entry.max_protection = fault_type;
888 fs.fault_flags = fault_flags;
890 fs.shared = vm_shared_fault;
891 fs.first_shared = *sharedp;
893 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
896 *sharedp = fs.first_shared;
897 first_pindex = OFF_TO_IDX(offset);
898 fs.first_object = object;
900 fs.first_prot = fault_type;
902 /*fs.map_generation = 0; unused */
905 * Make a reference to this object to prevent its disposal while we
906 * are messing with it. Once we have the reference, the map is free
907 * to be diddled. Since objects reference their shadows (and copies),
908 * they will stay around as well.
910 * The reference should also prevent an unexpected collapse of the
911 * parent that might move pages from the current object into the
912 * parent unexpectedly, resulting in corruption.
914 * Bump the paging-in-progress count to prevent size changes (e.g.
915 * truncation operations) during I/O. This must be done after
916 * obtaining the vnode lock in order to avoid possible deadlocks.
919 fs.vp = vnode_pager_lock(fs.first_object);
921 fs.lookup_still_valid = TRUE;
923 fs.object = fs.first_object; /* so unlock_and_deallocate works */
926 /* XXX future - ability to operate on VM object using vpagetable */
927 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
928 result = vm_fault_vpagetable(&fs, &first_pindex,
929 fs.entry->aux.master_pde,
931 if (result == KERN_TRY_AGAIN) {
932 if (fs.first_shared == 0 && *sharedp)
933 vm_object_upgrade(object);
936 if (result != KERN_SUCCESS) {
944 * Now we have the actual (object, pindex), fault in the page. If
945 * vm_fault_object() fails it will unlock and deallocate the FS
946 * data. If it succeeds everything remains locked and fs->object
947 * will have an additinal PIP count if it is not equal to
950 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact.
951 * We may have to upgrade its lock to handle the requested fault.
953 result = vm_fault_object(&fs, first_pindex, fault_type, 0);
955 if (result == KERN_TRY_AGAIN) {
956 if (fs.first_shared == 0 && *sharedp)
957 vm_object_upgrade(object);
960 if (result != KERN_SUCCESS) {
965 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
966 *errorp = KERN_PROTECTION_FAILURE;
967 unlock_and_deallocate(&fs);
972 * On success vm_fault_object() does not unlock or deallocate, so we
973 * do it here. Note that the returned fs.m will be busied.
978 * Return a held page. We are not doing any pmap manipulation so do
979 * not set PG_MAPPED. However, adjust the page flags according to
980 * the fault type because the caller may not use a managed pmapping
981 * (so we don't want to lose the fact that the page will be dirtied
982 * if a write fault was specified).
985 vm_page_activate(fs.m);
986 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY))
988 if (fault_flags & VM_FAULT_UNSWAP)
989 swap_pager_unswapped(fs.m);
992 * Indicate that the page was accessed.
994 vm_page_flag_set(fs.m, PG_REFERENCED);
996 if (curthread->td_lwp) {
998 curthread->td_lwp->lwp_ru.ru_majflt++;
1000 curthread->td_lwp->lwp_ru.ru_minflt++;
1005 * Unlock everything, and return the held page.
1007 vm_page_wakeup(fs.m);
1008 /*vm_object_deallocate(fs.first_object);*/
1009 /*fs.first_object = NULL; */
1016 * Translate the virtual page number (first_pindex) that is relative
1017 * to the address space into a logical page number that is relative to the
1018 * backing object. Use the virtual page table pointed to by (vpte).
1020 * This implements an N-level page table. Any level can terminate the
1021 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
1022 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
1026 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
1027 vpte_t vpte, int fault_type, int allow_nofault)
1030 struct lwbuf lwb_cache;
1031 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
1032 int result = KERN_SUCCESS;
1035 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1038 * We cannot proceed if the vpte is not valid, not readable
1039 * for a read fault, or not writable for a write fault.
1041 if ((vpte & VPTE_V) == 0) {
1042 unlock_and_deallocate(fs);
1043 return (KERN_FAILURE);
1045 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) {
1046 unlock_and_deallocate(fs);
1047 return (KERN_FAILURE);
1049 if ((vpte & VPTE_PS) || vshift == 0)
1051 KKASSERT(vshift >= VPTE_PAGE_BITS);
1054 * Get the page table page. Nominally we only read the page
1055 * table, but since we are actively setting VPTE_M and VPTE_A,
1056 * tell vm_fault_object() that we are writing it.
1058 * There is currently no real need to optimize this.
1060 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
1061 VM_PROT_READ|VM_PROT_WRITE,
1063 if (result != KERN_SUCCESS)
1067 * Process the returned fs.m and look up the page table
1068 * entry in the page table page.
1070 vshift -= VPTE_PAGE_BITS;
1071 lwb = lwbuf_alloc(fs->m, &lwb_cache);
1072 ptep = ((vpte_t *)lwbuf_kva(lwb) +
1073 ((*pindex >> vshift) & VPTE_PAGE_MASK));
1077 * Page table write-back. If the vpte is valid for the
1078 * requested operation, do a write-back to the page table.
1080 * XXX VPTE_M is not set properly for page directory pages.
1081 * It doesn't get set in the page directory if the page table
1082 * is modified during a read access.
1084 vm_page_activate(fs->m);
1085 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
1087 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
1088 atomic_set_long(ptep, VPTE_M | VPTE_A);
1089 vm_page_dirty(fs->m);
1092 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V)) {
1093 if ((vpte & VPTE_A) == 0) {
1094 atomic_set_long(ptep, VPTE_A);
1095 vm_page_dirty(fs->m);
1099 vm_page_flag_set(fs->m, PG_REFERENCED);
1100 vm_page_wakeup(fs->m);
1102 cleanup_successful_fault(fs);
1105 * Combine remaining address bits with the vpte.
1107 /* JG how many bits from each? */
1108 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
1109 (*pindex & ((1L << vshift) - 1));
1110 return (KERN_SUCCESS);
1115 * This is the core of the vm_fault code.
1117 * Do all operations required to fault-in (fs.first_object, pindex). Run
1118 * through the shadow chain as necessary and do required COW or virtual
1119 * copy operations. The caller has already fully resolved the vm_map_entry
1120 * and, if appropriate, has created a copy-on-write layer. All we need to
1121 * do is iterate the object chain.
1123 * On failure (fs) is unlocked and deallocated and the caller may return or
1124 * retry depending on the failure code. On success (fs) is NOT unlocked or
1125 * deallocated, fs.m will contained a resolved, busied page, and fs.object
1126 * will have an additional PIP count if it is not equal to fs.first_object.
1128 * If locks based on fs->first_shared or fs->shared are insufficient,
1129 * clear the appropriate field(s) and return RETRY. COWs require that
1130 * first_shared be 0, while page allocations (or frees) require that
1131 * shared be 0. Renames require that both be 0.
1133 * fs->first_object must be held on call.
1137 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex,
1138 vm_prot_t fault_type, int allow_nofault)
1140 vm_object_t next_object;
1144 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1145 fs->prot = fs->first_prot;
1146 fs->object = fs->first_object;
1147 pindex = first_pindex;
1149 vm_object_chain_acquire(fs->first_object, fs->shared);
1150 vm_object_pip_add(fs->first_object, 1);
1153 * If a read fault occurs we try to make the page writable if
1154 * possible. There are three cases where we cannot make the
1155 * page mapping writable:
1157 * (1) The mapping is read-only or the VM object is read-only,
1158 * fs->prot above will simply not have VM_PROT_WRITE set.
1160 * (2) If the mapping is a virtual page table we need to be able
1161 * to detect writes so we can set VPTE_M in the virtual page
1164 * (3) If the VM page is read-only or copy-on-write, upgrading would
1165 * just result in an unnecessary COW fault.
1167 * VM_PROT_VPAGED is set if faulting via a virtual page table and
1168 * causes adjustments to the 'M'odify bit to also turn off write
1169 * access to force a re-fault.
1171 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1172 if ((fault_type & VM_PROT_WRITE) == 0)
1173 fs->prot &= ~VM_PROT_WRITE;
1176 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace &&
1177 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) {
1178 if ((fault_type & VM_PROT_WRITE) == 0)
1179 fs->prot &= ~VM_PROT_WRITE;
1182 /* vm_object_hold(fs->object); implied b/c object == first_object */
1186 * The entire backing chain from first_object to object
1187 * inclusive is chainlocked.
1189 * If the object is dead, we stop here
1191 if (fs->object->flags & OBJ_DEAD) {
1192 vm_object_pip_wakeup(fs->first_object);
1193 vm_object_chain_release_all(fs->first_object,
1195 if (fs->object != fs->first_object)
1196 vm_object_drop(fs->object);
1197 unlock_and_deallocate(fs);
1198 return (KERN_PROTECTION_FAILURE);
1202 * See if the page is resident. Wait/Retry if the page is
1203 * busy (lots of stuff may have changed so we can't continue
1206 * We can theoretically allow the soft-busy case on a read
1207 * fault if the page is marked valid, but since such
1208 * pages are typically already pmap'd, putting that
1209 * special case in might be more effort then it is
1210 * worth. We cannot under any circumstances mess
1211 * around with a vm_page_t->busy page except, perhaps,
1214 fs->m = vm_page_lookup_busy_try(fs->object, pindex,
1217 vm_object_pip_wakeup(fs->first_object);
1218 vm_object_chain_release_all(fs->first_object,
1220 if (fs->object != fs->first_object)
1221 vm_object_drop(fs->object);
1223 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1224 mycpu->gd_cnt.v_intrans++;
1225 /*vm_object_deallocate(fs->first_object);*/
1226 /*fs->first_object = NULL;*/
1228 return (KERN_TRY_AGAIN);
1232 * The page is busied for us.
1234 * If reactivating a page from PQ_CACHE we may have
1237 int queue = fs->m->queue;
1238 vm_page_unqueue_nowakeup(fs->m);
1240 if ((queue - fs->m->pc) == PQ_CACHE &&
1241 vm_page_count_severe()) {
1242 vm_page_activate(fs->m);
1243 vm_page_wakeup(fs->m);
1245 vm_object_pip_wakeup(fs->first_object);
1246 vm_object_chain_release_all(fs->first_object,
1248 if (fs->object != fs->first_object)
1249 vm_object_drop(fs->object);
1250 unlock_and_deallocate(fs);
1251 if (allow_nofault == 0 ||
1252 (curthread->td_flags & TDF_NOFAULT) == 0) {
1255 return (KERN_TRY_AGAIN);
1259 * If it still isn't completely valid (readable),
1260 * or if a read-ahead-mark is set on the VM page,
1261 * jump to readrest, else we found the page and
1264 * We can release the spl once we have marked the
1267 if (fs->m->object != &kernel_object) {
1268 if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1272 if (fs->m->flags & PG_RAM) {
1275 vm_page_flag_clear(fs->m, PG_RAM);
1279 break; /* break to PAGE HAS BEEN FOUND */
1283 * Page is not resident, If this is the search termination
1284 * or the pager might contain the page, allocate a new page.
1286 if (TRYPAGER(fs) || fs->object == fs->first_object) {
1288 * Allocating, must be exclusive.
1290 if (fs->object == fs->first_object &&
1292 fs->first_shared = 0;
1293 vm_object_pip_wakeup(fs->first_object);
1294 vm_object_chain_release_all(fs->first_object,
1296 if (fs->object != fs->first_object)
1297 vm_object_drop(fs->object);
1298 unlock_and_deallocate(fs);
1299 return (KERN_TRY_AGAIN);
1301 if (fs->object != fs->first_object &&
1303 fs->first_shared = 0;
1305 vm_object_pip_wakeup(fs->first_object);
1306 vm_object_chain_release_all(fs->first_object,
1308 if (fs->object != fs->first_object)
1309 vm_object_drop(fs->object);
1310 unlock_and_deallocate(fs);
1311 return (KERN_TRY_AGAIN);
1315 * If the page is beyond the object size we fail
1317 if (pindex >= fs->object->size) {
1318 vm_object_pip_wakeup(fs->first_object);
1319 vm_object_chain_release_all(fs->first_object,
1321 if (fs->object != fs->first_object)
1322 vm_object_drop(fs->object);
1323 unlock_and_deallocate(fs);
1324 return (KERN_PROTECTION_FAILURE);
1328 * Allocate a new page for this object/offset pair.
1330 * It is possible for the allocation to race, so
1334 if (!vm_page_count_severe()) {
1335 fs->m = vm_page_alloc(fs->object, pindex,
1336 ((fs->vp || fs->object->backing_object) ?
1337 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL :
1338 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1339 VM_ALLOC_USE_GD | VM_ALLOC_ZERO));
1341 if (fs->m == NULL) {
1342 vm_object_pip_wakeup(fs->first_object);
1343 vm_object_chain_release_all(fs->first_object,
1345 if (fs->object != fs->first_object)
1346 vm_object_drop(fs->object);
1347 unlock_and_deallocate(fs);
1348 if (allow_nofault == 0 ||
1349 (curthread->td_flags & TDF_NOFAULT) == 0) {
1352 return (KERN_TRY_AGAIN);
1356 * Fall through to readrest. We have a new page which
1357 * will have to be paged (since m->valid will be 0).
1363 * We have found an invalid or partially valid page, a
1364 * page with a read-ahead mark which might be partially or
1365 * fully valid (and maybe dirty too), or we have allocated
1368 * Attempt to fault-in the page if there is a chance that the
1369 * pager has it, and potentially fault in additional pages
1372 * If TRYPAGER is true then fs.m will be non-NULL and busied
1378 u_char behavior = vm_map_entry_behavior(fs->entry);
1380 if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1386 * Doing I/O may synchronously insert additional
1387 * pages so we can't be shared at this point either.
1389 * NOTE: We can't free fs->m here in the allocated
1390 * case (fs->object != fs->first_object) as
1391 * this would require an exclusively locked
1394 if (fs->object == fs->first_object &&
1396 vm_page_deactivate(fs->m);
1397 vm_page_wakeup(fs->m);
1399 fs->first_shared = 0;
1400 vm_object_pip_wakeup(fs->first_object);
1401 vm_object_chain_release_all(fs->first_object,
1403 if (fs->object != fs->first_object)
1404 vm_object_drop(fs->object);
1405 unlock_and_deallocate(fs);
1406 return (KERN_TRY_AGAIN);
1408 if (fs->object != fs->first_object &&
1410 vm_page_deactivate(fs->m);
1411 vm_page_wakeup(fs->m);
1413 fs->first_shared = 0;
1415 vm_object_pip_wakeup(fs->first_object);
1416 vm_object_chain_release_all(fs->first_object,
1418 if (fs->object != fs->first_object)
1419 vm_object_drop(fs->object);
1420 unlock_and_deallocate(fs);
1421 return (KERN_TRY_AGAIN);
1425 * Avoid deadlocking against the map when doing I/O.
1426 * fs.object and the page is PG_BUSY'd.
1428 * NOTE: Once unlocked, fs->entry can become stale
1429 * so this will NULL it out.
1431 * NOTE: fs->entry is invalid until we relock the
1432 * map and verify that the timestamp has not
1438 * Acquire the page data. We still hold a ref on
1439 * fs.object and the page has been PG_BUSY's.
1441 * The pager may replace the page (for example, in
1442 * order to enter a fictitious page into the
1443 * object). If it does so it is responsible for
1444 * cleaning up the passed page and properly setting
1445 * the new page PG_BUSY.
1447 * If we got here through a PG_RAM read-ahead
1448 * mark the page may be partially dirty and thus
1449 * not freeable. Don't bother checking to see
1450 * if the pager has the page because we can't free
1451 * it anyway. We have to depend on the get_page
1452 * operation filling in any gaps whether there is
1453 * backing store or not.
1455 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1457 if (rv == VM_PAGER_OK) {
1459 * Relookup in case pager changed page. Pager
1460 * is responsible for disposition of old page
1463 * XXX other code segments do relookups too.
1464 * It's a bad abstraction that needs to be
1467 fs->m = vm_page_lookup(fs->object, pindex);
1468 if (fs->m == NULL) {
1469 vm_object_pip_wakeup(fs->first_object);
1470 vm_object_chain_release_all(
1471 fs->first_object, fs->object);
1472 if (fs->object != fs->first_object)
1473 vm_object_drop(fs->object);
1474 unlock_and_deallocate(fs);
1475 return (KERN_TRY_AGAIN);
1478 break; /* break to PAGE HAS BEEN FOUND */
1482 * Remove the bogus page (which does not exist at this
1483 * object/offset); before doing so, we must get back
1484 * our object lock to preserve our invariant.
1486 * Also wake up any other process that may want to bring
1489 * If this is the top-level object, we must leave the
1490 * busy page to prevent another process from rushing
1491 * past us, and inserting the page in that object at
1492 * the same time that we are.
1494 if (rv == VM_PAGER_ERROR) {
1496 kprintf("vm_fault: pager read error, "
1501 kprintf("vm_fault: pager read error, "
1509 * Data outside the range of the pager or an I/O error
1511 * The page may have been wired during the pagein,
1512 * e.g. by the buffer cache, and cannot simply be
1513 * freed. Call vnode_pager_freepage() to deal with it.
1515 * Also note that we cannot free the page if we are
1516 * holding the related object shared. XXX not sure
1517 * what to do in that case.
1519 if (fs->object != fs->first_object) {
1520 vnode_pager_freepage(fs->m);
1523 * XXX - we cannot just fall out at this
1524 * point, m has been freed and is invalid!
1528 * XXX - the check for kernel_map is a kludge to work
1529 * around having the machine panic on a kernel space
1530 * fault w/ I/O error.
1532 if (((fs->map != &kernel_map) &&
1533 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1535 if (fs->first_shared) {
1536 vm_page_deactivate(fs->m);
1537 vm_page_wakeup(fs->m);
1539 vnode_pager_freepage(fs->m);
1543 vm_object_pip_wakeup(fs->first_object);
1544 vm_object_chain_release_all(fs->first_object,
1546 if (fs->object != fs->first_object)
1547 vm_object_drop(fs->object);
1548 unlock_and_deallocate(fs);
1549 if (rv == VM_PAGER_ERROR)
1550 return (KERN_FAILURE);
1552 return (KERN_PROTECTION_FAILURE);
1558 * We get here if the object has a default pager (or unwiring)
1559 * or the pager doesn't have the page.
1561 * fs->first_m will be used for the COW unless we find a
1562 * deeper page to be mapped read-only, in which case the
1563 * unlock*(fs) will free first_m.
1565 if (fs->object == fs->first_object)
1566 fs->first_m = fs->m;
1569 * Move on to the next object. The chain lock should prevent
1570 * the backing_object from getting ripped out from under us.
1572 * The object lock for the next object is governed by
1575 if ((next_object = fs->object->backing_object) != NULL) {
1577 vm_object_hold_shared(next_object);
1579 vm_object_hold(next_object);
1580 vm_object_chain_acquire(next_object, fs->shared);
1581 KKASSERT(next_object == fs->object->backing_object);
1582 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1585 if (next_object == NULL) {
1587 * If there's no object left, fill the page in the top
1588 * object with zeros.
1590 if (fs->object != fs->first_object) {
1592 if (fs->first_object->backing_object !=
1594 vm_object_hold(fs->first_object->backing_object);
1597 vm_object_chain_release_all(
1598 fs->first_object->backing_object,
1601 if (fs->first_object->backing_object !=
1603 vm_object_drop(fs->first_object->backing_object);
1606 vm_object_pip_wakeup(fs->object);
1607 vm_object_drop(fs->object);
1608 fs->object = fs->first_object;
1609 pindex = first_pindex;
1610 fs->m = fs->first_m;
1615 * Zero the page if necessary and mark it valid.
1617 if ((fs->m->flags & PG_ZERO) == 0) {
1618 vm_page_zero_fill(fs->m);
1621 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m));
1623 vm_page_flag_clear(fs->m, PG_ZERO);
1624 mycpu->gd_cnt.v_ozfod++;
1626 mycpu->gd_cnt.v_zfod++;
1627 fs->m->valid = VM_PAGE_BITS_ALL;
1628 break; /* break to PAGE HAS BEEN FOUND */
1630 if (fs->object != fs->first_object) {
1631 vm_object_pip_wakeup(fs->object);
1632 vm_object_lock_swap();
1633 vm_object_drop(fs->object);
1635 KASSERT(fs->object != next_object,
1636 ("object loop %p", next_object));
1637 fs->object = next_object;
1638 vm_object_pip_add(fs->object, 1);
1642 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1645 * object still held.
1647 * local shared variable may be different from fs->shared.
1649 * If the page is being written, but isn't already owned by the
1650 * top-level object, we have to copy it into a new page owned by the
1653 KASSERT((fs->m->flags & PG_BUSY) != 0,
1654 ("vm_fault: not busy after main loop"));
1656 if (fs->object != fs->first_object) {
1658 * We only really need to copy if we want to write it.
1660 if (fault_type & VM_PROT_WRITE) {
1662 * This allows pages to be virtually copied from a
1663 * backing_object into the first_object, where the
1664 * backing object has no other refs to it, and cannot
1665 * gain any more refs. Instead of a bcopy, we just
1666 * move the page from the backing object to the
1667 * first object. Note that we must mark the page
1668 * dirty in the first object so that it will go out
1669 * to swap when needed.
1673 * Must be holding exclusive locks
1675 fs->first_shared == 0 &&
1678 * Map, if present, has not changed
1681 fs->map_generation == fs->map->timestamp) &&
1683 * Only one shadow object
1685 (fs->object->shadow_count == 1) &&
1687 * No COW refs, except us
1689 (fs->object->ref_count == 1) &&
1691 * No one else can look this object up
1693 (fs->object->handle == NULL) &&
1695 * No other ways to look the object up
1697 ((fs->object->type == OBJT_DEFAULT) ||
1698 (fs->object->type == OBJT_SWAP)) &&
1700 * We don't chase down the shadow chain
1702 (fs->object == fs->first_object->backing_object) &&
1705 * grab the lock if we need to
1707 (fs->lookup_still_valid ||
1709 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1712 * (first_m) and (m) are both busied. We have
1713 * move (m) into (first_m)'s object/pindex
1714 * in an atomic fashion, then free (first_m).
1716 * first_object is held so second remove
1717 * followed by the rename should wind
1718 * up being atomic. vm_page_free() might
1719 * block so we don't do it until after the
1722 fs->lookup_still_valid = 1;
1723 vm_page_protect(fs->first_m, VM_PROT_NONE);
1724 vm_page_remove(fs->first_m);
1725 vm_page_rename(fs->m, fs->first_object,
1727 vm_page_free(fs->first_m);
1728 fs->first_m = fs->m;
1730 mycpu->gd_cnt.v_cow_optim++;
1733 * Oh, well, lets copy it.
1735 * Why are we unmapping the original page
1736 * here? Well, in short, not all accessors
1737 * of user memory go through the pmap. The
1738 * procfs code doesn't have access user memory
1739 * via a local pmap, so vm_fault_page*()
1740 * can't call pmap_enter(). And the umtx*()
1741 * code may modify the COW'd page via a DMAP
1742 * or kernel mapping and not via the pmap,
1743 * leaving the original page still mapped
1744 * read-only into the pmap.
1746 * So we have to remove the page from at
1747 * least the current pmap if it is in it.
1748 * Just remove it from all pmaps.
1750 KKASSERT(fs->first_shared == 0);
1751 vm_page_copy(fs->m, fs->first_m);
1752 vm_page_protect(fs->m, VM_PROT_NONE);
1753 vm_page_event(fs->m, VMEVENT_COW);
1757 * We no longer need the old page or object.
1763 * We intend to revert to first_object, undo the
1764 * chain lock through to that.
1767 if (fs->first_object->backing_object != fs->object)
1768 vm_object_hold(fs->first_object->backing_object);
1770 vm_object_chain_release_all(
1771 fs->first_object->backing_object,
1774 if (fs->first_object->backing_object != fs->object)
1775 vm_object_drop(fs->first_object->backing_object);
1779 * fs->object != fs->first_object due to above
1782 vm_object_pip_wakeup(fs->object);
1783 vm_object_drop(fs->object);
1786 * Only use the new page below...
1788 mycpu->gd_cnt.v_cow_faults++;
1789 fs->m = fs->first_m;
1790 fs->object = fs->first_object;
1791 pindex = first_pindex;
1794 * If it wasn't a write fault avoid having to copy
1795 * the page by mapping it read-only.
1797 fs->prot &= ~VM_PROT_WRITE;
1802 * Relock the map if necessary, then check the generation count.
1803 * relock_map() will update fs->timestamp to account for the
1804 * relocking if necessary.
1806 * If the count has changed after relocking then all sorts of
1807 * crap may have happened and we have to retry.
1809 * NOTE: The relock_map() can fail due to a deadlock against
1810 * the vm_page we are holding BUSY.
1812 if (fs->lookup_still_valid == FALSE && fs->map) {
1813 if (relock_map(fs) ||
1814 fs->map->timestamp != fs->map_generation) {
1816 vm_object_pip_wakeup(fs->first_object);
1817 vm_object_chain_release_all(fs->first_object,
1819 if (fs->object != fs->first_object)
1820 vm_object_drop(fs->object);
1821 unlock_and_deallocate(fs);
1822 return (KERN_TRY_AGAIN);
1827 * If the fault is a write, we know that this page is being
1828 * written NOW so dirty it explicitly to save on pmap_is_modified()
1831 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1832 * if the page is already dirty to prevent data written with
1833 * the expectation of being synced from not being synced.
1834 * Likewise if this entry does not request NOSYNC then make
1835 * sure the page isn't marked NOSYNC. Applications sharing
1836 * data should use the same flags to avoid ping ponging.
1838 * Also tell the backing pager, if any, that it should remove
1839 * any swap backing since the page is now dirty.
1841 vm_page_activate(fs->m);
1842 if (fs->prot & VM_PROT_WRITE) {
1843 vm_object_set_writeable_dirty(fs->m->object);
1844 vm_set_nosync(fs->m, fs->entry);
1845 if (fs->fault_flags & VM_FAULT_DIRTY) {
1846 vm_page_dirty(fs->m);
1847 swap_pager_unswapped(fs->m);
1851 vm_object_pip_wakeup(fs->first_object);
1852 vm_object_chain_release_all(fs->first_object, fs->object);
1853 if (fs->object != fs->first_object)
1854 vm_object_drop(fs->object);
1857 * Page had better still be busy. We are still locked up and
1858 * fs->object will have another PIP reference if it is not equal
1859 * to fs->first_object.
1861 KASSERT(fs->m->flags & PG_BUSY,
1862 ("vm_fault: page %p not busy!", fs->m));
1865 * Sanity check: page must be completely valid or it is not fit to
1866 * map into user space. vm_pager_get_pages() ensures this.
1868 if (fs->m->valid != VM_PAGE_BITS_ALL) {
1869 vm_page_zero_invalid(fs->m, TRUE);
1870 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1872 vm_page_flag_clear(fs->m, PG_ZERO);
1874 return (KERN_SUCCESS);
1878 * Hold each of the physical pages that are mapped by the specified range of
1879 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1880 * and allow the specified types of access, "prot". If all of the implied
1881 * pages are successfully held, then the number of held pages is returned
1882 * together with pointers to those pages in the array "ma". However, if any
1883 * of the pages cannot be held, -1 is returned.
1886 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1887 vm_prot_t prot, vm_page_t *ma, int max_count)
1889 vm_offset_t start, end;
1890 int i, npages, error;
1892 start = trunc_page(addr);
1893 end = round_page(addr + len);
1895 npages = howmany(end - start, PAGE_SIZE);
1897 if (npages > max_count)
1900 for (i = 0; i < npages; i++) {
1901 // XXX error handling
1902 ma[i] = vm_fault_page_quick(start + (i * PAGE_SIZE),
1911 * Wire down a range of virtual addresses in a map. The entry in question
1912 * should be marked in-transition and the map must be locked. We must
1913 * release the map temporarily while faulting-in the page to avoid a
1914 * deadlock. Note that the entry may be clipped while we are blocked but
1915 * will never be freed.
1920 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1922 boolean_t fictitious;
1931 lwkt_gettoken(&map->token);
1933 pmap = vm_map_pmap(map);
1934 start = entry->start;
1936 fictitious = entry->object.vm_object &&
1937 ((entry->object.vm_object->type == OBJT_DEVICE) ||
1938 (entry->object.vm_object->type == OBJT_MGTDEVICE));
1939 if (entry->eflags & MAP_ENTRY_KSTACK)
1945 * We simulate a fault to get the page and enter it in the physical
1948 for (va = start; va < end; va += PAGE_SIZE) {
1950 rv = vm_fault(map, va, VM_PROT_READ,
1951 VM_FAULT_USER_WIRE);
1953 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1954 VM_FAULT_CHANGE_WIRING);
1957 while (va > start) {
1959 if ((pa = pmap_extract(pmap, va)) == 0)
1961 pmap_change_wiring(pmap, va, FALSE, entry);
1963 m = PHYS_TO_VM_PAGE(pa);
1964 vm_page_busy_wait(m, FALSE, "vmwrpg");
1965 vm_page_unwire(m, 1);
1975 lwkt_reltoken(&map->token);
1980 * Unwire a range of virtual addresses in a map. The map should be
1984 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1986 boolean_t fictitious;
1994 lwkt_gettoken(&map->token);
1996 pmap = vm_map_pmap(map);
1997 start = entry->start;
1999 fictitious = entry->object.vm_object &&
2000 ((entry->object.vm_object->type == OBJT_DEVICE) ||
2001 (entry->object.vm_object->type == OBJT_MGTDEVICE));
2002 if (entry->eflags & MAP_ENTRY_KSTACK)
2006 * Since the pages are wired down, we must be able to get their
2007 * mappings from the physical map system.
2009 for (va = start; va < end; va += PAGE_SIZE) {
2010 pa = pmap_extract(pmap, va);
2012 pmap_change_wiring(pmap, va, FALSE, entry);
2014 m = PHYS_TO_VM_PAGE(pa);
2015 vm_page_busy_wait(m, FALSE, "vmwupg");
2016 vm_page_unwire(m, 1);
2021 lwkt_reltoken(&map->token);
2025 * Copy all of the pages from a wired-down map entry to another.
2027 * The source and destination maps must be locked for write.
2028 * The source and destination maps token must be held
2029 * The source map entry must be wired down (or be a sharing map
2030 * entry corresponding to a main map entry that is wired down).
2032 * No other requirements.
2034 * XXX do segment optimization
2037 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
2038 vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
2040 vm_object_t dst_object;
2041 vm_object_t src_object;
2042 vm_ooffset_t dst_offset;
2043 vm_ooffset_t src_offset;
2049 src_object = src_entry->object.vm_object;
2050 src_offset = src_entry->offset;
2053 * Create the top-level object for the destination entry. (Doesn't
2054 * actually shadow anything - we copy the pages directly.)
2056 vm_map_entry_allocate_object(dst_entry);
2057 dst_object = dst_entry->object.vm_object;
2059 prot = dst_entry->max_protection;
2062 * Loop through all of the pages in the entry's range, copying each
2063 * one from the source object (it should be there) to the destination
2066 vm_object_hold(src_object);
2067 vm_object_hold(dst_object);
2068 for (vaddr = dst_entry->start, dst_offset = 0;
2069 vaddr < dst_entry->end;
2070 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
2073 * Allocate a page in the destination object
2076 dst_m = vm_page_alloc(dst_object,
2077 OFF_TO_IDX(dst_offset),
2079 if (dst_m == NULL) {
2082 } while (dst_m == NULL);
2085 * Find the page in the source object, and copy it in.
2086 * (Because the source is wired down, the page will be in
2089 src_m = vm_page_lookup(src_object,
2090 OFF_TO_IDX(dst_offset + src_offset));
2092 panic("vm_fault_copy_wired: page missing");
2094 vm_page_copy(src_m, dst_m);
2095 vm_page_event(src_m, VMEVENT_COW);
2098 * Enter it in the pmap...
2101 vm_page_flag_clear(dst_m, PG_ZERO);
2102 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry);
2105 * Mark it no longer busy, and put it on the active list.
2107 vm_page_activate(dst_m);
2108 vm_page_wakeup(dst_m);
2110 vm_object_drop(dst_object);
2111 vm_object_drop(src_object);
2117 * This routine checks around the requested page for other pages that
2118 * might be able to be faulted in. This routine brackets the viable
2119 * pages for the pages to be paged in.
2122 * m, rbehind, rahead
2125 * marray (array of vm_page_t), reqpage (index of requested page)
2128 * number of pages in marray
2131 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
2132 vm_page_t *marray, int *reqpage)
2136 vm_pindex_t pindex, startpindex, endpindex, tpindex;
2138 int cbehind, cahead;
2144 * we don't fault-ahead for device pager
2146 if ((object->type == OBJT_DEVICE) ||
2147 (object->type == OBJT_MGTDEVICE)) {
2154 * if the requested page is not available, then give up now
2156 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
2157 *reqpage = 0; /* not used by caller, fix compiler warn */
2161 if ((cbehind == 0) && (cahead == 0)) {
2167 if (rahead > cahead) {
2171 if (rbehind > cbehind) {
2176 * Do not do any readahead if we have insufficient free memory.
2178 * XXX code was broken disabled before and has instability
2179 * with this conditonal fixed, so shortcut for now.
2181 if (burst_fault == 0 || vm_page_count_severe()) {
2188 * scan backward for the read behind pages -- in memory
2190 * Assume that if the page is not found an interrupt will not
2191 * create it. Theoretically interrupts can only remove (busy)
2192 * pages, not create new associations.
2195 if (rbehind > pindex) {
2199 startpindex = pindex - rbehind;
2202 vm_object_hold(object);
2203 for (tpindex = pindex; tpindex > startpindex; --tpindex) {
2204 if (vm_page_lookup(object, tpindex - 1))
2209 while (tpindex < pindex) {
2210 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2213 for (j = 0; j < i; j++) {
2214 vm_page_free(marray[j]);
2216 vm_object_drop(object);
2225 vm_object_drop(object);
2231 * Assign requested page
2238 * Scan forwards for read-ahead pages
2240 tpindex = pindex + 1;
2241 endpindex = tpindex + rahead;
2242 if (endpindex > object->size)
2243 endpindex = object->size;
2245 vm_object_hold(object);
2246 while (tpindex < endpindex) {
2247 if (vm_page_lookup(object, tpindex))
2249 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2257 vm_object_drop(object);
2265 * vm_prefault() provides a quick way of clustering pagefaults into a
2266 * processes address space. It is a "cousin" of pmap_object_init_pt,
2267 * except it runs at page fault time instead of mmap time.
2269 * vm.fast_fault Enables pre-faulting zero-fill pages
2271 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to
2272 * prefault. Scan stops in either direction when
2273 * a page is found to already exist.
2275 * This code used to be per-platform pmap_prefault(). It is now
2276 * machine-independent and enhanced to also pre-fault zero-fill pages
2277 * (see vm.fast_fault) as well as make them writable, which greatly
2278 * reduces the number of page faults programs incur.
2280 * Application performance when pre-faulting zero-fill pages is heavily
2281 * dependent on the application. Very tiny applications like /bin/echo
2282 * lose a little performance while applications of any appreciable size
2283 * gain performance. Prefaulting multiple pages also reduces SMP
2284 * congestion and can improve SMP performance significantly.
2286 * NOTE! prot may allow writing but this only applies to the top level
2287 * object. If we wind up mapping a page extracted from a backing
2288 * object we have to make sure it is read-only.
2290 * NOTE! The caller has already handled any COW operations on the
2291 * vm_map_entry via the normal fault code. Do NOT call this
2292 * shortcut unless the normal fault code has run on this entry.
2294 * The related map must be locked.
2295 * No other requirements.
2297 static int vm_prefault_pages = 8;
2298 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0,
2299 "Maximum number of pages to pre-fault");
2300 static int vm_fast_fault = 1;
2301 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
2302 "Burst fault zero-fill regions");
2305 * Set PG_NOSYNC if the map entry indicates so, but only if the page
2306 * is not already dirty by other means. This will prevent passive
2307 * filesystem syncing as well as 'sync' from writing out the page.
2310 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
2312 if (entry->eflags & MAP_ENTRY_NOSYNC) {
2314 vm_page_flag_set(m, PG_NOSYNC);
2316 vm_page_flag_clear(m, PG_NOSYNC);
2321 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot,
2337 * Get stable max count value, disabled if set to 0
2339 maxpages = vm_prefault_pages;
2345 * We do not currently prefault mappings that use virtual page
2346 * tables. We do not prefault foreign pmaps.
2348 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2350 lp = curthread->td_lwp;
2351 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2355 * Limit pre-fault count to 1024 pages.
2357 if (maxpages > 1024)
2360 object = entry->object.vm_object;
2361 KKASSERT(object != NULL);
2362 KKASSERT(object == entry->object.vm_object);
2363 vm_object_hold(object);
2364 vm_object_chain_acquire(object, 0);
2368 for (i = 0; i < maxpages; ++i) {
2369 vm_object_t lobject;
2370 vm_object_t nobject;
2375 * This can eat a lot of time on a heavily contended
2376 * machine so yield on the tick if needed.
2382 * Calculate the page to pre-fault, stopping the scan in
2383 * each direction separately if the limit is reached.
2388 addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2392 addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2394 if (addr < entry->start) {
2400 if (addr >= entry->end) {
2408 * Skip pages already mapped, and stop scanning in that
2409 * direction. When the scan terminates in both directions
2412 if (pmap_prefault_ok(pmap, addr) == 0) {
2423 * Follow the VM object chain to obtain the page to be mapped
2426 * If we reach the terminal object without finding a page
2427 * and we determine it would be advantageous, then allocate
2428 * a zero-fill page for the base object. The base object
2429 * is guaranteed to be OBJT_DEFAULT for this case.
2431 * In order to not have to check the pager via *haspage*()
2432 * we stop if any non-default object is encountered. e.g.
2433 * a vnode or swap object would stop the loop.
2435 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2440 KKASSERT(lobject == entry->object.vm_object);
2441 /*vm_object_hold(lobject); implied */
2443 while ((m = vm_page_lookup_busy_try(lobject, pindex,
2444 TRUE, &error)) == NULL) {
2445 if (lobject->type != OBJT_DEFAULT)
2447 if (lobject->backing_object == NULL) {
2448 if (vm_fast_fault == 0)
2450 if ((prot & VM_PROT_WRITE) == 0 ||
2451 vm_page_count_min(0)) {
2456 * NOTE: Allocated from base object
2458 m = vm_page_alloc(object, index,
2467 /* lobject = object .. not needed */
2470 if (lobject->backing_object_offset & PAGE_MASK)
2472 nobject = lobject->backing_object;
2473 vm_object_hold(nobject);
2474 KKASSERT(nobject == lobject->backing_object);
2475 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2476 if (lobject != object) {
2477 vm_object_lock_swap();
2478 vm_object_drop(lobject);
2481 pprot &= ~VM_PROT_WRITE;
2482 vm_object_chain_acquire(lobject, 0);
2486 * NOTE: A non-NULL (m) will be associated with lobject if
2487 * it was found there, otherwise it is probably a
2488 * zero-fill page associated with the base object.
2490 * Give-up if no page is available.
2493 if (lobject != object) {
2495 if (object->backing_object != lobject)
2496 vm_object_hold(object->backing_object);
2498 vm_object_chain_release_all(
2499 object->backing_object, lobject);
2501 if (object->backing_object != lobject)
2502 vm_object_drop(object->backing_object);
2504 vm_object_drop(lobject);
2510 * The object must be marked dirty if we are mapping a
2511 * writable page. m->object is either lobject or object,
2512 * both of which are still held. Do this before we
2513 * potentially drop the object.
2515 if (pprot & VM_PROT_WRITE)
2516 vm_object_set_writeable_dirty(m->object);
2519 * Do not conditionalize on PG_RAM. If pages are present in
2520 * the VM system we assume optimal caching. If caching is
2521 * not optimal the I/O gravy train will be restarted when we
2522 * hit an unavailable page. We do not want to try to restart
2523 * the gravy train now because we really don't know how much
2524 * of the object has been cached. The cost for restarting
2525 * the gravy train should be low (since accesses will likely
2526 * be I/O bound anyway).
2528 if (lobject != object) {
2530 if (object->backing_object != lobject)
2531 vm_object_hold(object->backing_object);
2533 vm_object_chain_release_all(object->backing_object,
2536 if (object->backing_object != lobject)
2537 vm_object_drop(object->backing_object);
2539 vm_object_drop(lobject);
2543 * Enter the page into the pmap if appropriate. If we had
2544 * allocated the page we have to place it on a queue. If not
2545 * we just have to make sure it isn't on the cache queue
2546 * (pages on the cache queue are not allowed to be mapped).
2550 * Page must be zerod.
2552 if ((m->flags & PG_ZERO) == 0) {
2553 vm_page_zero_fill(m);
2556 pmap_page_assertzero(
2557 VM_PAGE_TO_PHYS(m));
2559 vm_page_flag_clear(m, PG_ZERO);
2560 mycpu->gd_cnt.v_ozfod++;
2562 mycpu->gd_cnt.v_zfod++;
2563 m->valid = VM_PAGE_BITS_ALL;
2566 * Handle dirty page case
2568 if (pprot & VM_PROT_WRITE)
2569 vm_set_nosync(m, entry);
2570 pmap_enter(pmap, addr, m, pprot, 0, entry);
2571 mycpu->gd_cnt.v_vm_faults++;
2572 if (curthread->td_lwp)
2573 ++curthread->td_lwp->lwp_ru.ru_minflt;
2574 vm_page_deactivate(m);
2575 if (pprot & VM_PROT_WRITE) {
2576 /*vm_object_set_writeable_dirty(m->object);*/
2577 vm_set_nosync(m, entry);
2578 if (fault_flags & VM_FAULT_DIRTY) {
2581 swap_pager_unswapped(m);
2586 /* couldn't busy page, no wakeup */
2588 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2589 (m->flags & PG_FICTITIOUS) == 0) {
2591 * A fully valid page not undergoing soft I/O can
2592 * be immediately entered into the pmap.
2594 if ((m->queue - m->pc) == PQ_CACHE)
2595 vm_page_deactivate(m);
2596 if (pprot & VM_PROT_WRITE) {
2597 /*vm_object_set_writeable_dirty(m->object);*/
2598 vm_set_nosync(m, entry);
2599 if (fault_flags & VM_FAULT_DIRTY) {
2602 swap_pager_unswapped(m);
2605 if (pprot & VM_PROT_WRITE)
2606 vm_set_nosync(m, entry);
2607 pmap_enter(pmap, addr, m, pprot, 0, entry);
2608 mycpu->gd_cnt.v_vm_faults++;
2609 if (curthread->td_lwp)
2610 ++curthread->td_lwp->lwp_ru.ru_minflt;
2616 vm_object_chain_release(object);
2617 vm_object_drop(object);
2621 * Object can be held shared
2624 vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
2625 vm_map_entry_t entry, int prot, int fault_flags)
2638 * Get stable max count value, disabled if set to 0
2640 maxpages = vm_prefault_pages;
2646 * We do not currently prefault mappings that use virtual page
2647 * tables. We do not prefault foreign pmaps.
2649 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2651 lp = curthread->td_lwp;
2652 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2654 object = entry->object.vm_object;
2655 if (object->backing_object != NULL)
2657 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2660 * Limit pre-fault count to 1024 pages.
2662 if (maxpages > 1024)
2667 for (i = 0; i < maxpages; ++i) {
2671 * Calculate the page to pre-fault, stopping the scan in
2672 * each direction separately if the limit is reached.
2677 addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2681 addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2683 if (addr < entry->start) {
2689 if (addr >= entry->end) {
2697 * Skip pages already mapped, and stop scanning in that
2698 * direction. When the scan terminates in both directions
2701 if (pmap_prefault_ok(pmap, addr) == 0) {
2712 * Follow the VM object chain to obtain the page to be mapped
2713 * into the pmap. This version of the prefault code only
2714 * works with terminal objects.
2716 * WARNING! We cannot call swap_pager_unswapped() with a
2719 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2721 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2722 if (m == NULL || error)
2725 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2726 (m->flags & PG_FICTITIOUS) == 0 &&
2727 ((m->flags & PG_SWAPPED) == 0 ||
2728 (prot & VM_PROT_WRITE) == 0 ||
2729 (fault_flags & VM_FAULT_DIRTY) == 0)) {
2731 * A fully valid page not undergoing soft I/O can
2732 * be immediately entered into the pmap.
2734 if ((m->queue - m->pc) == PQ_CACHE)
2735 vm_page_deactivate(m);
2736 if (prot & VM_PROT_WRITE) {
2737 vm_object_set_writeable_dirty(m->object);
2738 vm_set_nosync(m, entry);
2739 if (fault_flags & VM_FAULT_DIRTY) {
2742 swap_pager_unswapped(m);
2745 pmap_enter(pmap, addr, m, prot, 0, entry);
2746 mycpu->gd_cnt.v_vm_faults++;
2747 if (curthread->td_lwp)
2748 ++curthread->td_lwp->lwp_ru.ru_minflt;