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. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
46 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47 * All rights reserved.
49 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
51 * Permission to use, copy, modify and distribute this software and
52 * its documentation is hereby granted, provided that both the copyright
53 * notice and this permission notice appear in all copies of the
54 * software, derivative works or modified versions, and any portions
55 * thereof, and that both notices appear in supporting documentation.
57 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
58 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
59 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
61 * Carnegie Mellon requests users of this software to return to
63 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
64 * School of Computer Science
65 * Carnegie Mellon University
66 * Pittsburgh PA 15213-3890
68 * any improvements or extensions that they make and grant Carnegie the
69 * rights to redistribute these changes.
71 * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
72 * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $
76 * Page fault handling module.
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
83 #include <sys/vnode.h>
84 #include <sys/resourcevar.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vkernel.h>
88 #include <sys/sysctl.h>
90 #include <cpu/lwbuf.h>
93 #include <vm/vm_param.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_kern.h>
100 #include <vm/vm_pager.h>
101 #include <vm/vnode_pager.h>
102 #include <vm/vm_extern.h>
104 #include <sys/thread2.h>
105 #include <vm/vm_page2.h>
113 vm_object_t first_object;
114 vm_prot_t first_prot;
116 vm_map_entry_t entry;
117 int lookup_still_valid;
126 static int debug_cluster = 0;
127 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
129 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
130 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
132 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
134 static int vm_fault_ratelimit(struct vmspace *);
135 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry);
136 static void vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry,
140 release_page(struct faultstate *fs)
142 vm_page_deactivate(fs->m);
143 vm_page_wakeup(fs->m);
148 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse
149 * requires relocking and then checking the timestamp.
151 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do
152 * not have to update fs->map_generation here.
154 * NOTE: This function can fail due to a deadlock against the caller's
155 * holding of a vm_page BUSY.
158 relock_map(struct faultstate *fs)
162 if (fs->lookup_still_valid == FALSE && fs->map) {
163 error = vm_map_lock_read_to(fs->map);
165 fs->lookup_still_valid = TRUE;
173 unlock_map(struct faultstate *fs)
175 if (fs->lookup_still_valid && fs->map) {
176 vm_map_lookup_done(fs->map, fs->entry, 0);
177 fs->lookup_still_valid = FALSE;
182 * Clean up after a successful call to vm_fault_object() so another call
183 * to vm_fault_object() can be made.
186 _cleanup_successful_fault(struct faultstate *fs, int relock)
188 if (fs->object != fs->first_object) {
189 vm_page_free(fs->first_m);
190 vm_object_pip_wakeup(fs->object);
193 fs->object = fs->first_object;
194 if (relock && fs->lookup_still_valid == FALSE) {
196 vm_map_lock_read(fs->map);
197 fs->lookup_still_valid = TRUE;
202 _unlock_things(struct faultstate *fs, int dealloc)
204 _cleanup_successful_fault(fs, 0);
206 /*vm_object_deallocate(fs->first_object);*/
207 /*fs->first_object = NULL; drop used later on */
210 if (fs->vp != NULL) {
216 #define unlock_things(fs) _unlock_things(fs, 0)
217 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
218 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
223 * Determine if the pager for the current object *might* contain the page.
225 * We only need to try the pager if this is not a default object (default
226 * objects are zero-fill and have no real pager), and if we are not taking
227 * a wiring fault or if the FS entry is wired.
229 #define TRYPAGER(fs) \
230 (fs->object->type != OBJT_DEFAULT && \
231 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
236 * Handle a page fault occuring at the given address, requiring the given
237 * permissions, in the map specified. If successful, the page is inserted
238 * into the associated physical map.
240 * NOTE: The given address should be truncated to the proper page address.
242 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
243 * a standard error specifying why the fault is fatal is returned.
245 * The map in question must be referenced, and remains so.
246 * The caller may hold no locks.
247 * No other requirements.
250 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
253 vm_pindex_t first_pindex;
254 struct faultstate fs;
257 mycpu->gd_cnt.v_vm_faults++;
261 fs.fault_flags = fault_flags;
264 lwkt_gettoken(&map->token);
268 * Find the vm_map_entry representing the backing store and resolve
269 * the top level object and page index. This may have the side
270 * effect of executing a copy-on-write on the map entry and/or
271 * creating a shadow object, but will not COW any actual VM pages.
273 * On success fs.map is left read-locked and various other fields
274 * are initialized but not otherwise referenced or locked.
276 * NOTE! vm_map_lookup will try to upgrade the fault_type to
277 * VM_FAULT_WRITE if the map entry is a virtual page table and also
278 * writable, so we can set the 'A'accessed bit in the virtual page
282 result = vm_map_lookup(&fs.map, vaddr, fault_type,
283 &fs.entry, &fs.first_object,
284 &first_pindex, &fs.first_prot, &fs.wired);
287 * If the lookup failed or the map protections are incompatible,
288 * the fault generally fails. However, if the caller is trying
289 * to do a user wiring we have more work to do.
291 if (result != KERN_SUCCESS) {
292 if (result != KERN_PROTECTION_FAILURE ||
293 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
295 if (result == KERN_INVALID_ADDRESS && growstack &&
296 map != &kernel_map && curproc != NULL) {
297 result = vm_map_growstack(curproc, vaddr);
298 if (result == KERN_SUCCESS) {
302 result = KERN_FAILURE;
308 * If we are user-wiring a r/w segment, and it is COW, then
309 * we need to do the COW operation. Note that we don't
310 * currently COW RO sections now, because it is NOT desirable
311 * to COW .text. We simply keep .text from ever being COW'ed
312 * and take the heat that one cannot debug wired .text sections.
314 result = vm_map_lookup(&fs.map, vaddr,
315 VM_PROT_READ|VM_PROT_WRITE|
316 VM_PROT_OVERRIDE_WRITE,
317 &fs.entry, &fs.first_object,
318 &first_pindex, &fs.first_prot,
320 if (result != KERN_SUCCESS) {
321 result = KERN_FAILURE;
326 * If we don't COW now, on a user wire, the user will never
327 * be able to write to the mapping. If we don't make this
328 * restriction, the bookkeeping would be nearly impossible.
330 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
331 fs.entry->max_protection &= ~VM_PROT_WRITE;
335 * fs.map is read-locked
337 * Misc checks. Save the map generation number to detect races.
339 fs.map_generation = fs.map->timestamp;
341 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) {
342 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
343 panic("vm_fault: fault on nofault entry, addr: %p",
346 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) &&
347 vaddr >= fs.entry->start &&
348 vaddr < fs.entry->start + PAGE_SIZE) {
349 panic("vm_fault: fault on stack guard, addr: %p",
355 * A system map entry may return a NULL object. No object means
356 * no pager means an unrecoverable kernel fault.
358 if (fs.first_object == NULL) {
359 panic("vm_fault: unrecoverable fault at %p in entry %p",
360 (void *)vaddr, fs.entry);
364 * Bump the paging-in-progress count to prevent size changes (e.g.
365 * truncation operations) during I/O. This must be done after
366 * obtaining the vnode lock in order to avoid possible deadlocks.
368 vm_object_hold(fs.first_object);
369 fs.vp = vnode_pager_lock(fs.first_object);
371 fs.lookup_still_valid = TRUE;
373 fs.object = fs.first_object; /* so unlock_and_deallocate works */
376 * If the entry is wired we cannot change the page protection.
379 fault_type = fs.first_prot;
382 * The page we want is at (first_object, first_pindex), but if the
383 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
384 * page table to figure out the actual pindex.
386 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
389 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
390 result = vm_fault_vpagetable(&fs, &first_pindex,
391 fs.entry->aux.master_pde,
393 if (result == KERN_TRY_AGAIN) {
394 vm_object_drop(fs.first_object);
397 if (result != KERN_SUCCESS)
402 * Now we have the actual (object, pindex), fault in the page. If
403 * vm_fault_object() fails it will unlock and deallocate the FS
404 * data. If it succeeds everything remains locked and fs->object
405 * will have an additional PIP count if it is not equal to
408 * vm_fault_object will set fs->prot for the pmap operation. It is
409 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
410 * page can be safely written. However, it will force a read-only
411 * mapping for a read fault if the memory is managed by a virtual
415 result = vm_fault_object(&fs, first_pindex, fault_type);
417 if (result == KERN_TRY_AGAIN) {
418 vm_object_drop(fs.first_object);
421 if (result != KERN_SUCCESS)
425 * On success vm_fault_object() does not unlock or deallocate, and fs.m
426 * will contain a busied page.
428 * Enter the page into the pmap and do pmap-related adjustments.
430 vm_page_flag_set(fs.m, PG_REFERENCED);
431 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
434 * Burst in a few more pages if possible. The fs.map should still
437 if (fault_flags & VM_FAULT_BURST) {
438 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
440 vm_prefault(fs.map->pmap, vaddr, fs.entry, fs.prot);
445 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
446 KKASSERT(fs.m->flags & PG_BUSY);
449 * If the page is not wired down, then put it where the pageout daemon
453 if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
457 vm_page_unwire(fs.m, 1);
459 vm_page_activate(fs.m);
462 if (curthread->td_lwp) {
464 curthread->td_lwp->lwp_ru.ru_majflt++;
466 curthread->td_lwp->lwp_ru.ru_minflt++;
471 * Unlock everything, and return
473 vm_page_wakeup(fs.m);
474 /*vm_object_deallocate(fs.first_object);*/
476 /*fs.first_object = NULL; must still drop later */
478 result = KERN_SUCCESS;
481 vm_object_drop(fs.first_object);
482 lwkt_reltoken(&map->token);
487 * Fault in the specified virtual address in the current process map,
488 * returning a held VM page or NULL. See vm_fault_page() for more
494 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
496 struct lwp *lp = curthread->td_lwp;
499 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
500 fault_type, VM_FAULT_NORMAL, errorp);
505 * Fault in the specified virtual address in the specified map, doing all
506 * necessary manipulation of the object store and all necessary I/O. Return
507 * a held VM page or NULL, and set *errorp. The related pmap is not
510 * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
511 * and marked PG_REFERENCED as well.
513 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
514 * error will be returned.
519 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
520 int fault_flags, int *errorp)
522 vm_pindex_t first_pindex;
523 struct faultstate fs;
525 vm_prot_t orig_fault_type = fault_type;
527 mycpu->gd_cnt.v_vm_faults++;
531 fs.fault_flags = fault_flags;
532 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
534 lwkt_gettoken(&map->token);
538 * Find the vm_map_entry representing the backing store and resolve
539 * the top level object and page index. This may have the side
540 * effect of executing a copy-on-write on the map entry and/or
541 * creating a shadow object, but will not COW any actual VM pages.
543 * On success fs.map is left read-locked and various other fields
544 * are initialized but not otherwise referenced or locked.
546 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
547 * if the map entry is a virtual page table and also writable,
548 * so we can set the 'A'accessed bit in the virtual page table entry.
551 result = vm_map_lookup(&fs.map, vaddr, fault_type,
552 &fs.entry, &fs.first_object,
553 &first_pindex, &fs.first_prot, &fs.wired);
555 if (result != KERN_SUCCESS) {
562 * fs.map is read-locked
564 * Misc checks. Save the map generation number to detect races.
566 fs.map_generation = fs.map->timestamp;
568 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
569 panic("vm_fault: fault on nofault entry, addr: %lx",
574 * A system map entry may return a NULL object. No object means
575 * no pager means an unrecoverable kernel fault.
577 if (fs.first_object == NULL) {
578 panic("vm_fault: unrecoverable fault at %p in entry %p",
579 (void *)vaddr, fs.entry);
583 * Make a reference to this object to prevent its disposal while we
584 * are messing with it. Once we have the reference, the map is free
585 * to be diddled. Since objects reference their shadows (and copies),
586 * they will stay around as well.
588 * The reference should also prevent an unexpected collapse of the
589 * parent that might move pages from the current object into the
590 * parent unexpectedly, resulting in corruption.
592 * Bump the paging-in-progress count to prevent size changes (e.g.
593 * truncation operations) during I/O. This must be done after
594 * obtaining the vnode lock in order to avoid possible deadlocks.
596 vm_object_hold(fs.first_object);
597 fs.vp = vnode_pager_lock(fs.first_object);
599 fs.lookup_still_valid = TRUE;
601 fs.object = fs.first_object; /* so unlock_and_deallocate works */
604 * If the entry is wired we cannot change the page protection.
607 fault_type = fs.first_prot;
610 * The page we want is at (first_object, first_pindex), but if the
611 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
612 * page table to figure out the actual pindex.
614 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
617 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
618 result = vm_fault_vpagetable(&fs, &first_pindex,
619 fs.entry->aux.master_pde,
621 if (result == KERN_TRY_AGAIN) {
622 vm_object_drop(fs.first_object);
625 if (result != KERN_SUCCESS) {
633 * Now we have the actual (object, pindex), fault in the page. If
634 * vm_fault_object() fails it will unlock and deallocate the FS
635 * data. If it succeeds everything remains locked and fs->object
636 * will have an additinal PIP count if it is not equal to
639 result = vm_fault_object(&fs, first_pindex, fault_type);
641 if (result == KERN_TRY_AGAIN) {
642 vm_object_drop(fs.first_object);
645 if (result != KERN_SUCCESS) {
651 if ((orig_fault_type & VM_PROT_WRITE) &&
652 (fs.prot & VM_PROT_WRITE) == 0) {
653 *errorp = KERN_PROTECTION_FAILURE;
654 unlock_and_deallocate(&fs);
660 * Update the pmap. We really only have to do this if a COW
661 * occured to replace the read-only page with the new page. For
662 * now just do it unconditionally. XXX
664 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
665 vm_page_flag_set(fs.m, PG_REFERENCED);
668 * On success vm_fault_object() does not unlock or deallocate, and fs.m
669 * will contain a busied page. So we must unlock here after having
670 * messed with the pmap.
675 * Return a held page. We are not doing any pmap manipulation so do
676 * not set PG_MAPPED. However, adjust the page flags according to
677 * the fault type because the caller may not use a managed pmapping
678 * (so we don't want to lose the fact that the page will be dirtied
679 * if a write fault was specified).
682 if (fault_type & VM_PROT_WRITE)
686 * Unbusy the page by activating it. It remains held and will not
689 vm_page_activate(fs.m);
691 if (curthread->td_lwp) {
693 curthread->td_lwp->lwp_ru.ru_majflt++;
695 curthread->td_lwp->lwp_ru.ru_minflt++;
700 * Unlock everything, and return the held page.
702 vm_page_wakeup(fs.m);
703 /*vm_object_deallocate(fs.first_object);*/
704 /*fs.first_object = NULL; */
709 vm_object_drop(fs.first_object);
710 lwkt_reltoken(&map->token);
715 * Fault in the specified (object,offset), dirty the returned page as
716 * needed. If the requested fault_type cannot be done NULL and an
719 * A held (but not busied) page is returned.
724 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
725 vm_prot_t fault_type, int fault_flags, int *errorp)
728 vm_pindex_t first_pindex;
729 struct faultstate fs;
730 struct vm_map_entry entry;
732 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
733 bzero(&entry, sizeof(entry));
734 entry.object.vm_object = object;
735 entry.maptype = VM_MAPTYPE_NORMAL;
736 entry.protection = entry.max_protection = fault_type;
740 fs.fault_flags = fault_flags;
742 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
746 fs.first_object = object;
747 first_pindex = OFF_TO_IDX(offset);
749 fs.first_prot = fault_type;
751 /*fs.map_generation = 0; unused */
754 * Make a reference to this object to prevent its disposal while we
755 * are messing with it. Once we have the reference, the map is free
756 * to be diddled. Since objects reference their shadows (and copies),
757 * they will stay around as well.
759 * The reference should also prevent an unexpected collapse of the
760 * parent that might move pages from the current object into the
761 * parent unexpectedly, resulting in corruption.
763 * Bump the paging-in-progress count to prevent size changes (e.g.
764 * truncation operations) during I/O. This must be done after
765 * obtaining the vnode lock in order to avoid possible deadlocks.
767 fs.vp = vnode_pager_lock(fs.first_object);
769 fs.lookup_still_valid = TRUE;
771 fs.object = fs.first_object; /* so unlock_and_deallocate works */
774 /* XXX future - ability to operate on VM object using vpagetable */
775 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
776 result = vm_fault_vpagetable(&fs, &first_pindex,
777 fs.entry->aux.master_pde,
779 if (result == KERN_TRY_AGAIN)
781 if (result != KERN_SUCCESS) {
789 * Now we have the actual (object, pindex), fault in the page. If
790 * vm_fault_object() fails it will unlock and deallocate the FS
791 * data. If it succeeds everything remains locked and fs->object
792 * will have an additinal PIP count if it is not equal to
795 result = vm_fault_object(&fs, first_pindex, fault_type);
797 if (result == KERN_TRY_AGAIN)
799 if (result != KERN_SUCCESS) {
804 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
805 *errorp = KERN_PROTECTION_FAILURE;
806 unlock_and_deallocate(&fs);
811 * On success vm_fault_object() does not unlock or deallocate, so we
812 * do it here. Note that the returned fs.m will be busied.
817 * Return a held page. We are not doing any pmap manipulation so do
818 * not set PG_MAPPED. However, adjust the page flags according to
819 * the fault type because the caller may not use a managed pmapping
820 * (so we don't want to lose the fact that the page will be dirtied
821 * if a write fault was specified).
824 if (fault_type & VM_PROT_WRITE)
827 if (fault_flags & VM_FAULT_DIRTY)
829 if (fault_flags & VM_FAULT_UNSWAP)
830 swap_pager_unswapped(fs.m);
833 * Indicate that the page was accessed.
835 vm_page_flag_set(fs.m, PG_REFERENCED);
838 * Unbusy the page by activating it. It remains held and will not
841 vm_page_activate(fs.m);
843 if (curthread->td_lwp) {
845 mycpu->gd_cnt.v_vm_faults++;
846 curthread->td_lwp->lwp_ru.ru_majflt++;
848 curthread->td_lwp->lwp_ru.ru_minflt++;
853 * Unlock everything, and return the held page.
855 vm_page_wakeup(fs.m);
856 /*vm_object_deallocate(fs.first_object);*/
857 /*fs.first_object = NULL; */
864 * Translate the virtual page number (first_pindex) that is relative
865 * to the address space into a logical page number that is relative to the
866 * backing object. Use the virtual page table pointed to by (vpte).
868 * This implements an N-level page table. Any level can terminate the
869 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
870 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
874 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
875 vpte_t vpte, int fault_type)
878 struct lwbuf lwb_cache;
879 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
880 int result = KERN_SUCCESS;
883 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
886 * We cannot proceed if the vpte is not valid, not readable
887 * for a read fault, or not writable for a write fault.
889 if ((vpte & VPTE_V) == 0) {
890 unlock_and_deallocate(fs);
891 return (KERN_FAILURE);
893 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
894 unlock_and_deallocate(fs);
895 return (KERN_FAILURE);
897 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
898 unlock_and_deallocate(fs);
899 return (KERN_FAILURE);
901 if ((vpte & VPTE_PS) || vshift == 0)
903 KKASSERT(vshift >= VPTE_PAGE_BITS);
906 * Get the page table page. Nominally we only read the page
907 * table, but since we are actively setting VPTE_M and VPTE_A,
908 * tell vm_fault_object() that we are writing it.
910 * There is currently no real need to optimize this.
912 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
913 VM_PROT_READ|VM_PROT_WRITE);
914 if (result != KERN_SUCCESS)
918 * Process the returned fs.m and look up the page table
919 * entry in the page table page.
921 vshift -= VPTE_PAGE_BITS;
922 lwb = lwbuf_alloc(fs->m, &lwb_cache);
923 ptep = ((vpte_t *)lwbuf_kva(lwb) +
924 ((*pindex >> vshift) & VPTE_PAGE_MASK));
928 * Page table write-back. If the vpte is valid for the
929 * requested operation, do a write-back to the page table.
931 * XXX VPTE_M is not set properly for page directory pages.
932 * It doesn't get set in the page directory if the page table
933 * is modified during a read access.
935 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
937 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
938 atomic_set_long(ptep, VPTE_M | VPTE_A);
939 vm_page_dirty(fs->m);
942 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
944 if ((vpte & VPTE_A) == 0) {
945 atomic_set_long(ptep, VPTE_A);
946 vm_page_dirty(fs->m);
950 vm_page_flag_set(fs->m, PG_REFERENCED);
951 vm_page_activate(fs->m);
952 vm_page_wakeup(fs->m);
954 cleanup_successful_fault(fs);
957 * Combine remaining address bits with the vpte.
959 /* JG how many bits from each? */
960 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
961 (*pindex & ((1L << vshift) - 1));
962 return (KERN_SUCCESS);
967 * This is the core of the vm_fault code.
969 * Do all operations required to fault-in (fs.first_object, pindex). Run
970 * through the shadow chain as necessary and do required COW or virtual
971 * copy operations. The caller has already fully resolved the vm_map_entry
972 * and, if appropriate, has created a copy-on-write layer. All we need to
973 * do is iterate the object chain.
975 * On failure (fs) is unlocked and deallocated and the caller may return or
976 * retry depending on the failure code. On success (fs) is NOT unlocked or
977 * deallocated, fs.m will contained a resolved, busied page, and fs.object
978 * will have an additional PIP count if it is not equal to fs.first_object.
980 * fs->first_object must be held on call.
984 vm_fault_object(struct faultstate *fs,
985 vm_pindex_t first_pindex, vm_prot_t fault_type)
987 vm_object_t next_object;
991 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
992 fs->prot = fs->first_prot;
993 fs->object = fs->first_object;
994 pindex = first_pindex;
996 vm_object_chain_acquire(fs->first_object);
997 vm_object_pip_add(fs->first_object, 1);
1000 * If a read fault occurs we try to make the page writable if
1001 * possible. There are three cases where we cannot make the
1002 * page mapping writable:
1004 * (1) The mapping is read-only or the VM object is read-only,
1005 * fs->prot above will simply not have VM_PROT_WRITE set.
1007 * (2) If the mapping is a virtual page table we need to be able
1008 * to detect writes so we can set VPTE_M in the virtual page
1011 * (3) If the VM page is read-only or copy-on-write, upgrading would
1012 * just result in an unnecessary COW fault.
1014 * VM_PROT_VPAGED is set if faulting via a virtual page table and
1015 * causes adjustments to the 'M'odify bit to also turn off write
1016 * access to force a re-fault.
1018 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1019 if ((fault_type & VM_PROT_WRITE) == 0)
1020 fs->prot &= ~VM_PROT_WRITE;
1023 /* vm_object_hold(fs->object); implied b/c object == first_object */
1027 * The entire backing chain from first_object to object
1028 * inclusive is chainlocked.
1030 * If the object is dead, we stop here
1032 if (fs->object->flags & OBJ_DEAD) {
1033 vm_object_pip_wakeup(fs->first_object);
1034 vm_object_chain_release_all(fs->first_object,
1036 if (fs->object != fs->first_object)
1037 vm_object_drop(fs->object);
1038 unlock_and_deallocate(fs);
1039 return (KERN_PROTECTION_FAILURE);
1043 * See if the page is resident. Wait/Retry if the page is
1044 * busy (lots of stuff may have changed so we can't continue
1047 * We can theoretically allow the soft-busy case on a read
1048 * fault if the page is marked valid, but since such
1049 * pages are typically already pmap'd, putting that
1050 * special case in might be more effort then it is
1051 * worth. We cannot under any circumstances mess
1052 * around with a vm_page_t->busy page except, perhaps,
1055 fs->m = vm_page_lookup_busy_try(fs->object, pindex,
1058 vm_object_pip_wakeup(fs->first_object);
1059 vm_object_chain_release_all(fs->first_object,
1061 if (fs->object != fs->first_object)
1062 vm_object_drop(fs->object);
1064 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1065 mycpu->gd_cnt.v_intrans++;
1066 /*vm_object_deallocate(fs->first_object);*/
1067 /*fs->first_object = NULL;*/
1069 return (KERN_TRY_AGAIN);
1073 * The page is busied for us.
1075 * If reactivating a page from PQ_CACHE we may have
1078 int queue = fs->m->queue;
1079 vm_page_unqueue_nowakeup(fs->m);
1081 if ((queue - fs->m->pc) == PQ_CACHE &&
1082 vm_page_count_severe()) {
1083 vm_page_activate(fs->m);
1084 vm_page_wakeup(fs->m);
1086 vm_object_pip_wakeup(fs->first_object);
1087 vm_object_chain_release_all(fs->first_object,
1089 if (fs->object != fs->first_object)
1090 vm_object_drop(fs->object);
1091 unlock_and_deallocate(fs);
1093 return (KERN_TRY_AGAIN);
1097 * If it still isn't completely valid (readable),
1098 * or if a read-ahead-mark is set on the VM page,
1099 * jump to readrest, else we found the page and
1102 * We can release the spl once we have marked the
1105 if (fs->m->object != &kernel_object) {
1106 if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1110 if (fs->m->flags & PG_RAM) {
1113 vm_page_flag_clear(fs->m, PG_RAM);
1117 break; /* break to PAGE HAS BEEN FOUND */
1121 * Page is not resident, If this is the search termination
1122 * or the pager might contain the page, allocate a new page.
1124 if (TRYPAGER(fs) || fs->object == fs->first_object) {
1126 * If the page is beyond the object size we fail
1128 if (pindex >= fs->object->size) {
1129 vm_object_pip_wakeup(fs->first_object);
1130 vm_object_chain_release_all(fs->first_object,
1132 if (fs->object != fs->first_object)
1133 vm_object_drop(fs->object);
1134 unlock_and_deallocate(fs);
1135 return (KERN_PROTECTION_FAILURE);
1141 if (fs->didlimit == 0 && curproc != NULL) {
1144 limticks = vm_fault_ratelimit(curproc->p_vmspace);
1146 vm_object_pip_wakeup(fs->first_object);
1147 vm_object_chain_release_all(
1148 fs->first_object, fs->object);
1149 if (fs->object != fs->first_object)
1150 vm_object_drop(fs->object);
1151 unlock_and_deallocate(fs);
1152 tsleep(curproc, 0, "vmrate", limticks);
1154 return (KERN_TRY_AGAIN);
1159 * Allocate a new page for this object/offset pair.
1161 * It is possible for the allocation to race, so
1165 if (!vm_page_count_severe()) {
1166 fs->m = vm_page_alloc(fs->object, pindex,
1167 ((fs->vp || fs->object->backing_object) ?
1168 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL :
1169 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1172 if (fs->m == NULL) {
1173 vm_object_pip_wakeup(fs->first_object);
1174 vm_object_chain_release_all(fs->first_object,
1176 if (fs->object != fs->first_object)
1177 vm_object_drop(fs->object);
1178 unlock_and_deallocate(fs);
1180 return (KERN_TRY_AGAIN);
1184 * Fall through to readrest. We have a new page which
1185 * will have to be paged (since m->valid will be 0).
1191 * We have found an invalid or partially valid page, a
1192 * page with a read-ahead mark which might be partially or
1193 * fully valid (and maybe dirty too), or we have allocated
1196 * Attempt to fault-in the page if there is a chance that the
1197 * pager has it, and potentially fault in additional pages
1200 * If TRYPAGER is true then fs.m will be non-NULL and busied
1206 u_char behavior = vm_map_entry_behavior(fs->entry);
1208 if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1214 * If sequential access is detected then attempt
1215 * to deactivate/cache pages behind the scan to
1216 * prevent resource hogging.
1218 * Use of PG_RAM to detect sequential access
1219 * also simulates multi-zone sequential access
1220 * detection for free.
1222 * NOTE: Partially valid dirty pages cannot be
1223 * deactivated without causing NFS picemeal
1226 if ((fs->first_object->type != OBJT_DEVICE) &&
1227 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1228 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1229 (fs->m->flags & PG_RAM)))
1231 vm_pindex_t scan_pindex;
1232 int scan_count = 16;
1234 if (first_pindex < 16) {
1238 scan_pindex = first_pindex - 16;
1239 if (scan_pindex < 16)
1240 scan_count = scan_pindex;
1245 while (scan_count) {
1248 mt = vm_page_lookup(fs->first_object,
1252 if (vm_page_busy_try(mt, TRUE))
1255 if (mt->valid != VM_PAGE_BITS_ALL) {
1260 (PG_FICTITIOUS | PG_UNMANAGED)) ||
1267 vm_page_test_dirty(mt);
1271 vm_page_deactivate(mt);
1285 * Avoid deadlocking against the map when doing I/O.
1286 * fs.object and the page is PG_BUSY'd.
1288 * NOTE: Once unlocked, fs->entry can become stale
1289 * so this will NULL it out.
1291 * NOTE: fs->entry is invalid until we relock the
1292 * map and verify that the timestamp has not
1298 * Acquire the page data. We still hold a ref on
1299 * fs.object and the page has been PG_BUSY's.
1301 * The pager may replace the page (for example, in
1302 * order to enter a fictitious page into the
1303 * object). If it does so it is responsible for
1304 * cleaning up the passed page and properly setting
1305 * the new page PG_BUSY.
1307 * If we got here through a PG_RAM read-ahead
1308 * mark the page may be partially dirty and thus
1309 * not freeable. Don't bother checking to see
1310 * if the pager has the page because we can't free
1311 * it anyway. We have to depend on the get_page
1312 * operation filling in any gaps whether there is
1313 * backing store or not.
1315 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1317 if (rv == VM_PAGER_OK) {
1319 * Relookup in case pager changed page. Pager
1320 * is responsible for disposition of old page
1323 * XXX other code segments do relookups too.
1324 * It's a bad abstraction that needs to be
1327 fs->m = vm_page_lookup(fs->object, pindex);
1328 if (fs->m == NULL) {
1329 vm_object_pip_wakeup(fs->first_object);
1330 vm_object_chain_release_all(
1331 fs->first_object, fs->object);
1332 if (fs->object != fs->first_object)
1333 vm_object_drop(fs->object);
1334 unlock_and_deallocate(fs);
1335 return (KERN_TRY_AGAIN);
1339 break; /* break to PAGE HAS BEEN FOUND */
1343 * Remove the bogus page (which does not exist at this
1344 * object/offset); before doing so, we must get back
1345 * our object lock to preserve our invariant.
1347 * Also wake up any other process that may want to bring
1350 * If this is the top-level object, we must leave the
1351 * busy page to prevent another process from rushing
1352 * past us, and inserting the page in that object at
1353 * the same time that we are.
1355 if (rv == VM_PAGER_ERROR) {
1357 kprintf("vm_fault: pager read error, "
1362 kprintf("vm_fault: pager read error, "
1370 * Data outside the range of the pager or an I/O error
1372 * The page may have been wired during the pagein,
1373 * e.g. by the buffer cache, and cannot simply be
1374 * freed. Call vnode_pager_freepage() to deal with it.
1377 * XXX - the check for kernel_map is a kludge to work
1378 * around having the machine panic on a kernel space
1379 * fault w/ I/O error.
1381 if (((fs->map != &kernel_map) &&
1382 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1383 vnode_pager_freepage(fs->m);
1385 vm_object_pip_wakeup(fs->first_object);
1386 vm_object_chain_release_all(fs->first_object,
1388 if (fs->object != fs->first_object)
1389 vm_object_drop(fs->object);
1390 unlock_and_deallocate(fs);
1391 if (rv == VM_PAGER_ERROR)
1392 return (KERN_FAILURE);
1394 return (KERN_PROTECTION_FAILURE);
1397 if (fs->object != fs->first_object) {
1398 vnode_pager_freepage(fs->m);
1401 * XXX - we cannot just fall out at this
1402 * point, m has been freed and is invalid!
1408 * We get here if the object has a default pager (or unwiring)
1409 * or the pager doesn't have the page.
1411 if (fs->object == fs->first_object)
1412 fs->first_m = fs->m;
1415 * Move on to the next object. The chain lock should prevent
1416 * the backing_object from getting ripped out from under us.
1418 if ((next_object = fs->object->backing_object) != NULL) {
1419 vm_object_hold(next_object);
1420 vm_object_chain_acquire(next_object);
1421 KKASSERT(next_object == fs->object->backing_object);
1422 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1425 if (next_object == NULL) {
1427 * If there's no object left, fill the page in the top
1428 * object with zeros.
1430 if (fs->object != fs->first_object) {
1431 if (fs->first_object->backing_object !=
1433 vm_object_hold(fs->first_object->backing_object);
1435 vm_object_chain_release_all(
1436 fs->first_object->backing_object,
1438 if (fs->first_object->backing_object !=
1440 vm_object_drop(fs->first_object->backing_object);
1442 vm_object_pip_wakeup(fs->object);
1443 vm_object_drop(fs->object);
1444 fs->object = fs->first_object;
1445 pindex = first_pindex;
1446 fs->m = fs->first_m;
1451 * Zero the page if necessary and mark it valid.
1453 if ((fs->m->flags & PG_ZERO) == 0) {
1454 vm_page_zero_fill(fs->m);
1457 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m));
1459 vm_page_flag_clear(fs->m, PG_ZERO);
1460 mycpu->gd_cnt.v_ozfod++;
1462 mycpu->gd_cnt.v_zfod++;
1463 fs->m->valid = VM_PAGE_BITS_ALL;
1464 break; /* break to PAGE HAS BEEN FOUND */
1466 if (fs->object != fs->first_object) {
1467 vm_object_pip_wakeup(fs->object);
1468 vm_object_lock_swap();
1469 vm_object_drop(fs->object);
1471 KASSERT(fs->object != next_object,
1472 ("object loop %p", next_object));
1473 fs->object = next_object;
1474 vm_object_pip_add(fs->object, 1);
1478 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1481 * object still held.
1483 * If the page is being written, but isn't already owned by the
1484 * top-level object, we have to copy it into a new page owned by the
1487 KASSERT((fs->m->flags & PG_BUSY) != 0,
1488 ("vm_fault: not busy after main loop"));
1490 if (fs->object != fs->first_object) {
1492 * We only really need to copy if we want to write it.
1494 if (fault_type & VM_PROT_WRITE) {
1496 * This allows pages to be virtually copied from a
1497 * backing_object into the first_object, where the
1498 * backing object has no other refs to it, and cannot
1499 * gain any more refs. Instead of a bcopy, we just
1500 * move the page from the backing object to the
1501 * first object. Note that we must mark the page
1502 * dirty in the first object so that it will go out
1503 * to swap when needed.
1507 * Map, if present, has not changed
1510 fs->map_generation == fs->map->timestamp) &&
1512 * Only one shadow object
1514 (fs->object->shadow_count == 1) &&
1516 * No COW refs, except us
1518 (fs->object->ref_count == 1) &&
1520 * No one else can look this object up
1522 (fs->object->handle == NULL) &&
1524 * No other ways to look the object up
1526 ((fs->object->type == OBJT_DEFAULT) ||
1527 (fs->object->type == OBJT_SWAP)) &&
1529 * We don't chase down the shadow chain
1531 (fs->object == fs->first_object->backing_object) &&
1534 * grab the lock if we need to
1536 (fs->lookup_still_valid ||
1538 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1541 * (first_m) and (m) are both busied. We have
1542 * move (m) into (first_m)'s object/pindex
1543 * in an atomic fashion, then free (first_m).
1545 * first_object is held so second remove
1546 * followed by the rename should wind
1547 * up being atomic. vm_page_free() might
1548 * block so we don't do it until after the
1551 fs->lookup_still_valid = 1;
1552 vm_page_protect(fs->first_m, VM_PROT_NONE);
1553 vm_page_remove(fs->first_m);
1554 vm_page_rename(fs->m, fs->first_object,
1556 vm_page_free(fs->first_m);
1557 fs->first_m = fs->m;
1559 mycpu->gd_cnt.v_cow_optim++;
1562 * Oh, well, lets copy it.
1564 vm_page_copy(fs->m, fs->first_m);
1565 vm_page_event(fs->m, VMEVENT_COW);
1570 * We no longer need the old page or object.
1576 * We intend to revert to first_object, undo the
1577 * chain lock through to that.
1579 if (fs->first_object->backing_object != fs->object)
1580 vm_object_hold(fs->first_object->backing_object);
1581 vm_object_chain_release_all(
1582 fs->first_object->backing_object,
1584 if (fs->first_object->backing_object != fs->object)
1585 vm_object_drop(fs->first_object->backing_object);
1588 * fs->object != fs->first_object due to above
1591 vm_object_pip_wakeup(fs->object);
1592 vm_object_drop(fs->object);
1595 * Only use the new page below...
1598 mycpu->gd_cnt.v_cow_faults++;
1599 fs->m = fs->first_m;
1600 fs->object = fs->first_object;
1601 pindex = first_pindex;
1604 * If it wasn't a write fault avoid having to copy
1605 * the page by mapping it read-only.
1607 fs->prot &= ~VM_PROT_WRITE;
1612 * Relock the map if necessary, then check the generation count.
1613 * relock_map() will update fs->timestamp to account for the
1614 * relocking if necessary.
1616 * If the count has changed after relocking then all sorts of
1617 * crap may have happened and we have to retry.
1619 * NOTE: The relock_map() can fail due to a deadlock against
1620 * the vm_page we are holding BUSY.
1622 if (fs->lookup_still_valid == FALSE && fs->map) {
1623 if (relock_map(fs) ||
1624 fs->map->timestamp != fs->map_generation) {
1626 vm_object_pip_wakeup(fs->first_object);
1627 vm_object_chain_release_all(fs->first_object,
1629 if (fs->object != fs->first_object)
1630 vm_object_drop(fs->object);
1631 unlock_and_deallocate(fs);
1632 return (KERN_TRY_AGAIN);
1637 * If the fault is a write, we know that this page is being
1638 * written NOW so dirty it explicitly to save on pmap_is_modified()
1641 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1642 * if the page is already dirty to prevent data written with
1643 * the expectation of being synced from not being synced.
1644 * Likewise if this entry does not request NOSYNC then make
1645 * sure the page isn't marked NOSYNC. Applications sharing
1646 * data should use the same flags to avoid ping ponging.
1648 * Also tell the backing pager, if any, that it should remove
1649 * any swap backing since the page is now dirty.
1651 if (fs->prot & VM_PROT_WRITE) {
1652 vm_object_set_writeable_dirty(fs->m->object);
1653 vm_set_nosync(fs->m, fs->entry);
1654 if (fs->fault_flags & VM_FAULT_DIRTY) {
1655 vm_page_dirty(fs->m);
1656 swap_pager_unswapped(fs->m);
1660 vm_object_pip_wakeup(fs->first_object);
1661 vm_object_chain_release_all(fs->first_object, fs->object);
1662 if (fs->object != fs->first_object)
1663 vm_object_drop(fs->object);
1666 * Page had better still be busy. We are still locked up and
1667 * fs->object will have another PIP reference if it is not equal
1668 * to fs->first_object.
1670 KASSERT(fs->m->flags & PG_BUSY,
1671 ("vm_fault: page %p not busy!", fs->m));
1674 * Sanity check: page must be completely valid or it is not fit to
1675 * map into user space. vm_pager_get_pages() ensures this.
1677 if (fs->m->valid != VM_PAGE_BITS_ALL) {
1678 vm_page_zero_invalid(fs->m, TRUE);
1679 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1681 vm_page_flag_clear(fs->m, PG_ZERO);
1683 return (KERN_SUCCESS);
1687 * Wire down a range of virtual addresses in a map. The entry in question
1688 * should be marked in-transition and the map must be locked. We must
1689 * release the map temporarily while faulting-in the page to avoid a
1690 * deadlock. Note that the entry may be clipped while we are blocked but
1691 * will never be freed.
1696 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1698 boolean_t fictitious;
1707 lwkt_gettoken(&map->token);
1709 pmap = vm_map_pmap(map);
1710 start = entry->start;
1712 fictitious = entry->object.vm_object &&
1713 (entry->object.vm_object->type == OBJT_DEVICE);
1714 if (entry->eflags & MAP_ENTRY_KSTACK)
1720 * We simulate a fault to get the page and enter it in the physical
1723 for (va = start; va < end; va += PAGE_SIZE) {
1725 rv = vm_fault(map, va, VM_PROT_READ,
1726 VM_FAULT_USER_WIRE);
1728 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1729 VM_FAULT_CHANGE_WIRING);
1732 while (va > start) {
1734 if ((pa = pmap_extract(pmap, va)) == 0)
1736 pmap_change_wiring(pmap, va, FALSE);
1738 m = PHYS_TO_VM_PAGE(pa);
1739 vm_page_busy_wait(m, FALSE, "vmwrpg");
1740 vm_page_unwire(m, 1);
1750 lwkt_reltoken(&map->token);
1755 * Unwire a range of virtual addresses in a map. The map should be
1759 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1761 boolean_t fictitious;
1769 lwkt_gettoken(&map->token);
1771 pmap = vm_map_pmap(map);
1772 start = entry->start;
1774 fictitious = entry->object.vm_object &&
1775 (entry->object.vm_object->type == OBJT_DEVICE);
1776 if (entry->eflags & MAP_ENTRY_KSTACK)
1780 * Since the pages are wired down, we must be able to get their
1781 * mappings from the physical map system.
1783 for (va = start; va < end; va += PAGE_SIZE) {
1784 pa = pmap_extract(pmap, va);
1786 pmap_change_wiring(pmap, va, FALSE);
1788 m = PHYS_TO_VM_PAGE(pa);
1789 vm_page_busy_wait(m, FALSE, "vmwupg");
1790 vm_page_unwire(m, 1);
1795 lwkt_reltoken(&map->token);
1799 * Reduce the rate at which memory is allocated to a process based
1800 * on the perceived load on the VM system. As the load increases
1801 * the allocation burst rate goes down and the delay increases.
1803 * Rate limiting does not apply when faulting active or inactive
1804 * pages. When faulting 'cache' pages, rate limiting only applies
1805 * if the system currently has a severe page deficit.
1807 * XXX vm_pagesupply should be increased when a page is freed.
1809 * We sleep up to 1/10 of a second.
1812 vm_fault_ratelimit(struct vmspace *vmspace)
1814 if (vm_load_enable == 0)
1816 if (vmspace->vm_pagesupply > 0) {
1817 --vmspace->vm_pagesupply; /* SMP race ok */
1821 if (vm_load_debug) {
1822 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1824 (1000 - vm_load ) / 10, vm_load * hz / 10000,
1825 curproc->p_pid, curproc->p_comm);
1828 vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1829 return(vm_load * hz / 10000);
1833 * Copy all of the pages from a wired-down map entry to another.
1835 * The source and destination maps must be locked for write.
1836 * The source and destination maps token must be held
1837 * The source map entry must be wired down (or be a sharing map
1838 * entry corresponding to a main map entry that is wired down).
1840 * No other requirements.
1843 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1844 vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1846 vm_object_t dst_object;
1847 vm_object_t src_object;
1848 vm_ooffset_t dst_offset;
1849 vm_ooffset_t src_offset;
1855 src_object = src_entry->object.vm_object;
1856 src_offset = src_entry->offset;
1859 * Create the top-level object for the destination entry. (Doesn't
1860 * actually shadow anything - we copy the pages directly.)
1862 vm_map_entry_allocate_object(dst_entry);
1863 dst_object = dst_entry->object.vm_object;
1865 prot = dst_entry->max_protection;
1868 * Loop through all of the pages in the entry's range, copying each
1869 * one from the source object (it should be there) to the destination
1872 for (vaddr = dst_entry->start, dst_offset = 0;
1873 vaddr < dst_entry->end;
1874 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1877 * Allocate a page in the destination object
1880 dst_m = vm_page_alloc(dst_object,
1881 OFF_TO_IDX(dst_offset),
1883 if (dst_m == NULL) {
1886 } while (dst_m == NULL);
1889 * Find the page in the source object, and copy it in.
1890 * (Because the source is wired down, the page will be in
1893 src_m = vm_page_lookup(src_object,
1894 OFF_TO_IDX(dst_offset + src_offset));
1896 panic("vm_fault_copy_wired: page missing");
1898 vm_page_copy(src_m, dst_m);
1899 vm_page_event(src_m, VMEVENT_COW);
1902 * Enter it in the pmap...
1905 vm_page_flag_clear(dst_m, PG_ZERO);
1906 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1909 * Mark it no longer busy, and put it on the active list.
1911 vm_page_activate(dst_m);
1912 vm_page_wakeup(dst_m);
1919 * This routine checks around the requested page for other pages that
1920 * might be able to be faulted in. This routine brackets the viable
1921 * pages for the pages to be paged in.
1924 * m, rbehind, rahead
1927 * marray (array of vm_page_t), reqpage (index of requested page)
1930 * number of pages in marray
1933 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1934 vm_page_t *marray, int *reqpage)
1938 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1940 int cbehind, cahead;
1946 * we don't fault-ahead for device pager
1948 if (object->type == OBJT_DEVICE) {
1955 * if the requested page is not available, then give up now
1957 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1958 *reqpage = 0; /* not used by caller, fix compiler warn */
1962 if ((cbehind == 0) && (cahead == 0)) {
1968 if (rahead > cahead) {
1972 if (rbehind > cbehind) {
1977 * Do not do any readahead if we have insufficient free memory.
1979 * XXX code was broken disabled before and has instability
1980 * with this conditonal fixed, so shortcut for now.
1982 if (burst_fault == 0 || vm_page_count_severe()) {
1989 * scan backward for the read behind pages -- in memory
1991 * Assume that if the page is not found an interrupt will not
1992 * create it. Theoretically interrupts can only remove (busy)
1993 * pages, not create new associations.
1996 if (rbehind > pindex) {
2000 startpindex = pindex - rbehind;
2003 vm_object_hold(object);
2004 for (tpindex = pindex; tpindex > startpindex; --tpindex) {
2005 if (vm_page_lookup(object, tpindex - 1))
2010 while (tpindex < pindex) {
2011 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2014 for (j = 0; j < i; j++) {
2015 vm_page_free(marray[j]);
2017 vm_object_drop(object);
2026 vm_object_drop(object);
2032 * Assign requested page
2039 * Scan forwards for read-ahead pages
2041 tpindex = pindex + 1;
2042 endpindex = tpindex + rahead;
2043 if (endpindex > object->size)
2044 endpindex = object->size;
2046 vm_object_hold(object);
2047 while (tpindex < endpindex) {
2048 if (vm_page_lookup(object, tpindex))
2050 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2058 vm_object_drop(object);
2066 * vm_prefault() provides a quick way of clustering pagefaults into a
2067 * processes address space. It is a "cousin" of pmap_object_init_pt,
2068 * except it runs at page fault time instead of mmap time.
2070 * vm.fast_fault Enables pre-faulting zero-fill pages
2072 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to
2073 * prefault. Scan stops in either direction when
2074 * a page is found to already exist.
2076 * This code used to be per-platform pmap_prefault(). It is now
2077 * machine-independent and enhanced to also pre-fault zero-fill pages
2078 * (see vm.fast_fault) as well as make them writable, which greatly
2079 * reduces the number of page faults programs incur.
2081 * Application performance when pre-faulting zero-fill pages is heavily
2082 * dependent on the application. Very tiny applications like /bin/echo
2083 * lose a little performance while applications of any appreciable size
2084 * gain performance. Prefaulting multiple pages also reduces SMP
2085 * congestion and can improve SMP performance significantly.
2087 * NOTE! prot may allow writing but this only applies to the top level
2088 * object. If we wind up mapping a page extracted from a backing
2089 * object we have to make sure it is read-only.
2091 * NOTE! The caller has already handled any COW operations on the
2092 * vm_map_entry via the normal fault code. Do NOT call this
2093 * shortcut unless the normal fault code has run on this entry.
2095 * The related map must be locked.
2096 * No other requirements.
2098 static int vm_prefault_pages = 8;
2099 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0,
2100 "Maximum number of pages to pre-fault");
2101 static int vm_fast_fault = 1;
2102 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
2103 "Burst fault zero-fill regions");
2106 * Set PG_NOSYNC if the map entry indicates so, but only if the page
2107 * is not already dirty by other means. This will prevent passive
2108 * filesystem syncing as well as 'sync' from writing out the page.
2111 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
2113 if (entry->eflags & MAP_ENTRY_NOSYNC) {
2115 vm_page_flag_set(m, PG_NOSYNC);
2117 vm_page_flag_clear(m, PG_NOSYNC);
2122 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot)
2137 * Get stable max count value, disabled if set to 0
2139 maxpages = vm_prefault_pages;
2145 * We do not currently prefault mappings that use virtual page
2146 * tables. We do not prefault foreign pmaps.
2148 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2150 lp = curthread->td_lwp;
2151 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2155 * Limit pre-fault count to 1024 pages.
2157 if (maxpages > 1024)
2160 object = entry->object.vm_object;
2161 KKASSERT(object != NULL);
2162 vm_object_hold(object);
2163 KKASSERT(object == entry->object.vm_object);
2164 vm_object_chain_acquire(object);
2168 for (i = 0; i < maxpages; ++i) {
2169 vm_object_t lobject;
2170 vm_object_t nobject;
2175 * This can eat a lot of time on a heavily contended
2176 * machine so yield on the tick if needed.
2182 * Calculate the page to pre-fault, stopping the scan in
2183 * each direction separately if the limit is reached.
2188 addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2192 addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2194 if (addr < entry->start) {
2200 if (addr >= entry->end) {
2208 * Skip pages already mapped, and stop scanning in that
2209 * direction. When the scan terminates in both directions
2212 if (pmap_prefault_ok(pmap, addr) == 0) {
2223 * Follow the VM object chain to obtain the page to be mapped
2226 * If we reach the terminal object without finding a page
2227 * and we determine it would be advantageous, then allocate
2228 * a zero-fill page for the base object. The base object
2229 * is guaranteed to be OBJT_DEFAULT for this case.
2231 * In order to not have to check the pager via *haspage*()
2232 * we stop if any non-default object is encountered. e.g.
2233 * a vnode or swap object would stop the loop.
2235 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2240 KKASSERT(lobject == entry->object.vm_object);
2241 /*vm_object_hold(lobject); implied */
2243 while ((m = vm_page_lookup_busy_try(lobject, pindex,
2244 TRUE, &error)) == NULL) {
2245 if (lobject->type != OBJT_DEFAULT)
2247 if (lobject->backing_object == NULL) {
2248 if (vm_fast_fault == 0)
2250 if ((prot & VM_PROT_WRITE) == 0 ||
2251 vm_page_count_min(0)) {
2256 * NOTE: Allocated from base object
2258 m = vm_page_alloc(object, index,
2265 if ((m->flags & PG_ZERO) == 0) {
2266 vm_page_zero_fill(m);
2269 pmap_page_assertzero(
2270 VM_PAGE_TO_PHYS(m));
2272 vm_page_flag_clear(m, PG_ZERO);
2273 mycpu->gd_cnt.v_ozfod++;
2275 mycpu->gd_cnt.v_zfod++;
2276 m->valid = VM_PAGE_BITS_ALL;
2279 /* lobject = object .. not needed */
2282 if (lobject->backing_object_offset & PAGE_MASK)
2284 nobject = lobject->backing_object;
2285 vm_object_hold(nobject);
2286 KKASSERT(nobject == lobject->backing_object);
2287 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2288 if (lobject != object) {
2289 vm_object_lock_swap();
2290 vm_object_drop(lobject);
2293 pprot &= ~VM_PROT_WRITE;
2294 vm_object_chain_acquire(lobject);
2298 * NOTE: A non-NULL (m) will be associated with lobject if
2299 * it was found there, otherwise it is probably a
2300 * zero-fill page associated with the base object.
2302 * Give-up if no page is available.
2305 if (lobject != object) {
2306 if (object->backing_object != lobject)
2307 vm_object_hold(object->backing_object);
2308 vm_object_chain_release_all(
2309 object->backing_object, lobject);
2310 if (object->backing_object != lobject)
2311 vm_object_drop(object->backing_object);
2312 vm_object_drop(lobject);
2318 * Do not conditionalize on PG_RAM. If pages are present in
2319 * the VM system we assume optimal caching. If caching is
2320 * not optimal the I/O gravy train will be restarted when we
2321 * hit an unavailable page. We do not want to try to restart
2322 * the gravy train now because we really don't know how much
2323 * of the object has been cached. The cost for restarting
2324 * the gravy train should be low (since accesses will likely
2325 * be I/O bound anyway).
2327 * The object must be marked dirty if we are mapping a
2328 * writable page. m->object is either lobject or object,
2329 * both of which are still held.
2331 if (pprot & VM_PROT_WRITE)
2332 vm_object_set_writeable_dirty(m->object);
2334 if (lobject != object) {
2335 if (object->backing_object != lobject)
2336 vm_object_hold(object->backing_object);
2337 vm_object_chain_release_all(object->backing_object,
2339 if (object->backing_object != lobject)
2340 vm_object_drop(object->backing_object);
2341 vm_object_drop(lobject);
2345 * Enter the page into the pmap if appropriate. If we had
2346 * allocated the page we have to place it on a queue. If not
2347 * we just have to make sure it isn't on the cache queue
2348 * (pages on the cache queue are not allowed to be mapped).
2351 if (pprot & VM_PROT_WRITE)
2352 vm_set_nosync(m, entry);
2353 pmap_enter(pmap, addr, m, pprot, 0);
2354 vm_page_deactivate(m);
2357 /* couldn't busy page, no wakeup */
2359 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2360 (m->flags & PG_FICTITIOUS) == 0) {
2362 * A fully valid page not undergoing soft I/O can
2363 * be immediately entered into the pmap.
2365 if ((m->queue - m->pc) == PQ_CACHE)
2366 vm_page_deactivate(m);
2367 if (pprot & VM_PROT_WRITE)
2368 vm_set_nosync(m, entry);
2369 pmap_enter(pmap, addr, m, pprot, 0);
2375 vm_object_chain_release(object);
2376 vm_object_drop(object);