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 vm_fast_fault = 1;
127 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
128 "Burst fault zero-fill regions");
129 static int debug_cluster = 0;
130 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
132 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
133 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
135 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
137 static int vm_fault_ratelimit(struct vmspace *);
138 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry);
139 static void vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry,
143 * The caller must hold vm_token.
146 release_page(struct faultstate *fs)
148 vm_page_deactivate(fs->m);
149 vm_page_wakeup(fs->m);
154 * The caller must hold vm_token.
156 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse
157 * requires relocking and then checking the timestamp.
159 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do
160 * not have to update fs->map_generation here.
162 * NOTE: This function can fail due to a deadlock against the caller's
163 * holding of a vm_page BUSY.
166 relock_map(struct faultstate *fs)
170 if (fs->lookup_still_valid == FALSE && fs->map) {
171 error = vm_map_lock_read_to(fs->map);
173 fs->lookup_still_valid = TRUE;
181 unlock_map(struct faultstate *fs)
183 if (fs->lookup_still_valid && fs->map) {
184 vm_map_lookup_done(fs->map, fs->entry, 0);
185 fs->lookup_still_valid = FALSE;
190 * Clean up after a successful call to vm_fault_object() so another call
191 * to vm_fault_object() can be made.
193 * The caller must hold vm_token.
196 _cleanup_successful_fault(struct faultstate *fs, int relock)
198 if (fs->object != fs->first_object) {
199 vm_page_free(fs->first_m);
200 vm_object_pip_wakeup(fs->object);
203 fs->object = fs->first_object;
204 if (relock && fs->lookup_still_valid == FALSE) {
206 vm_map_lock_read(fs->map);
207 fs->lookup_still_valid = TRUE;
212 * The caller must hold vm_token.
215 _unlock_things(struct faultstate *fs, int dealloc)
217 vm_object_pip_wakeup(fs->first_object);
218 _cleanup_successful_fault(fs, 0);
220 vm_object_deallocate(fs->first_object);
221 fs->first_object = NULL;
224 if (fs->vp != NULL) {
230 #define unlock_things(fs) _unlock_things(fs, 0)
231 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
232 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
237 * Determine if the pager for the current object *might* contain the page.
239 * We only need to try the pager if this is not a default object (default
240 * objects are zero-fill and have no real pager), and if we are not taking
241 * a wiring fault or if the FS entry is wired.
243 #define TRYPAGER(fs) \
244 (fs->object->type != OBJT_DEFAULT && \
245 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
250 * Handle a page fault occuring at the given address, requiring the given
251 * permissions, in the map specified. If successful, the page is inserted
252 * into the associated physical map.
254 * NOTE: The given address should be truncated to the proper page address.
256 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
257 * a standard error specifying why the fault is fatal is returned.
259 * The map in question must be referenced, and remains so.
260 * The caller may hold no locks.
261 * No other requirements.
264 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
267 vm_pindex_t first_pindex;
268 struct faultstate fs;
271 mycpu->gd_cnt.v_vm_faults++;
275 fs.fault_flags = fault_flags;
280 * Find the vm_map_entry representing the backing store and resolve
281 * the top level object and page index. This may have the side
282 * effect of executing a copy-on-write on the map entry and/or
283 * creating a shadow object, but will not COW any actual VM pages.
285 * On success fs.map is left read-locked and various other fields
286 * are initialized but not otherwise referenced or locked.
288 * NOTE! vm_map_lookup will try to upgrade the fault_type to
289 * VM_FAULT_WRITE if the map entry is a virtual page table and also
290 * writable, so we can set the 'A'accessed bit in the virtual page
294 result = vm_map_lookup(&fs.map, vaddr, fault_type,
295 &fs.entry, &fs.first_object,
296 &first_pindex, &fs.first_prot, &fs.wired);
299 * If the lookup failed or the map protections are incompatible,
300 * the fault generally fails. However, if the caller is trying
301 * to do a user wiring we have more work to do.
303 if (result != KERN_SUCCESS) {
304 if (result != KERN_PROTECTION_FAILURE ||
305 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
307 if (result == KERN_INVALID_ADDRESS && growstack &&
308 map != &kernel_map && curproc != NULL) {
309 result = vm_map_growstack(curproc, vaddr);
310 if (result != KERN_SUCCESS)
311 return (KERN_FAILURE);
319 * If we are user-wiring a r/w segment, and it is COW, then
320 * we need to do the COW operation. Note that we don't
321 * currently COW RO sections now, because it is NOT desirable
322 * to COW .text. We simply keep .text from ever being COW'ed
323 * and take the heat that one cannot debug wired .text sections.
325 result = vm_map_lookup(&fs.map, vaddr,
326 VM_PROT_READ|VM_PROT_WRITE|
327 VM_PROT_OVERRIDE_WRITE,
328 &fs.entry, &fs.first_object,
329 &first_pindex, &fs.first_prot,
331 if (result != KERN_SUCCESS)
335 * If we don't COW now, on a user wire, the user will never
336 * be able to write to the mapping. If we don't make this
337 * restriction, the bookkeeping would be nearly impossible.
339 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
340 fs.entry->max_protection &= ~VM_PROT_WRITE;
344 * fs.map is read-locked
346 * Misc checks. Save the map generation number to detect races.
348 fs.map_generation = fs.map->timestamp;
350 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) {
351 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
352 panic("vm_fault: fault on nofault entry, addr: %p",
355 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) &&
356 vaddr >= fs.entry->start &&
357 vaddr < fs.entry->start + PAGE_SIZE) {
358 panic("vm_fault: fault on stack guard, addr: %p",
364 * A system map entry may return a NULL object. No object means
365 * no pager means an unrecoverable kernel fault.
367 if (fs.first_object == NULL) {
368 panic("vm_fault: unrecoverable fault at %p in entry %p",
369 (void *)vaddr, fs.entry);
373 * Make a reference to this object to prevent its disposal while we
374 * are messing with it. Once we have the reference, the map is free
375 * to be diddled. Since objects reference their shadows (and copies),
376 * they will stay around as well.
378 * Bump the paging-in-progress count to prevent size changes (e.g.
379 * truncation operations) during I/O. This must be done after
380 * obtaining the vnode lock in order to avoid possible deadlocks.
382 * The vm_object must be held before manipulation.
384 lwkt_gettoken(&vm_token);
385 vm_object_hold(fs.first_object);
386 vm_object_reference(fs.first_object);
387 fs.vp = vnode_pager_lock(fs.first_object);
388 vm_object_pip_add(fs.first_object, 1);
389 vm_object_drop(fs.first_object);
390 lwkt_reltoken(&vm_token);
392 fs.lookup_still_valid = TRUE;
394 fs.object = fs.first_object; /* so unlock_and_deallocate works */
397 * If the entry is wired we cannot change the page protection.
400 fault_type = fs.first_prot;
403 * The page we want is at (first_object, first_pindex), but if the
404 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
405 * page table to figure out the actual pindex.
407 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
410 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
411 result = vm_fault_vpagetable(&fs, &first_pindex,
412 fs.entry->aux.master_pde,
414 if (result == KERN_TRY_AGAIN)
416 if (result != KERN_SUCCESS)
421 * Now we have the actual (object, pindex), fault in the page. If
422 * vm_fault_object() fails it will unlock and deallocate the FS
423 * data. If it succeeds everything remains locked and fs->object
424 * will have an additional PIP count if it is not equal to
427 * vm_fault_object will set fs->prot for the pmap operation. It is
428 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
429 * page can be safely written. However, it will force a read-only
430 * mapping for a read fault if the memory is managed by a virtual
434 result = vm_fault_object(&fs, first_pindex, fault_type);
436 if (result == KERN_TRY_AGAIN) {
437 /*lwkt_reltoken(&vm_token);*/
440 if (result != KERN_SUCCESS) {
441 /*lwkt_reltoken(&vm_token);*/
446 * On success vm_fault_object() does not unlock or deallocate, and fs.m
447 * will contain a busied page.
449 * Enter the page into the pmap and do pmap-related adjustments.
451 vm_page_flag_set(fs.m, PG_REFERENCED);
452 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
455 * Burst in a few more pages if possible. The fs.map should still
458 if (fault_flags & VM_FAULT_BURST) {
459 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
461 vm_prefault(fs.map->pmap, vaddr, fs.entry, fs.prot);
464 lwkt_gettoken(&vm_token);
467 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
468 KKASSERT(fs.m->flags & PG_BUSY);
471 * If the page is not wired down, then put it where the pageout daemon
474 * We do not really need to get vm_token here but since all the
475 * vm_*() calls have to doing it here improves efficiency.
477 /*lwkt_gettoken(&vm_token);*/
479 if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
480 lwkt_reltoken(&vm_token); /* before wire activate does not */
484 vm_page_unwire(fs.m, 1);
486 vm_page_activate(fs.m);
487 lwkt_reltoken(&vm_token); /* before wire activate does not */
489 /*lwkt_reltoken(&vm_token); after wire/activate works */
491 if (curthread->td_lwp) {
493 curthread->td_lwp->lwp_ru.ru_majflt++;
495 curthread->td_lwp->lwp_ru.ru_minflt++;
500 * Unlock everything, and return
502 vm_page_wakeup(fs.m);
503 vm_object_deallocate(fs.first_object);
505 /*fs.first_object = NULL; */
506 /*lwkt_reltoken(&vm_token);*/
508 return (KERN_SUCCESS);
512 * Fault in the specified virtual address in the current process map,
513 * returning a held VM page or NULL. See vm_fault_page() for more
519 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
521 struct lwp *lp = curthread->td_lwp;
524 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
525 fault_type, VM_FAULT_NORMAL, errorp);
530 * Fault in the specified virtual address in the specified map, doing all
531 * necessary manipulation of the object store and all necessary I/O. Return
532 * a held VM page or NULL, and set *errorp. The related pmap is not
535 * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
536 * and marked PG_REFERENCED as well.
538 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
539 * error will be returned.
544 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
545 int fault_flags, int *errorp)
547 vm_pindex_t first_pindex;
548 struct faultstate fs;
550 vm_prot_t orig_fault_type = fault_type;
552 mycpu->gd_cnt.v_vm_faults++;
556 fs.fault_flags = fault_flags;
557 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
561 * Find the vm_map_entry representing the backing store and resolve
562 * the top level object and page index. This may have the side
563 * effect of executing a copy-on-write on the map entry and/or
564 * creating a shadow object, but will not COW any actual VM pages.
566 * On success fs.map is left read-locked and various other fields
567 * are initialized but not otherwise referenced or locked.
569 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
570 * if the map entry is a virtual page table and also writable,
571 * so we can set the 'A'accessed bit in the virtual page table entry.
574 result = vm_map_lookup(&fs.map, vaddr, fault_type,
575 &fs.entry, &fs.first_object,
576 &first_pindex, &fs.first_prot, &fs.wired);
578 if (result != KERN_SUCCESS) {
584 * fs.map is read-locked
586 * Misc checks. Save the map generation number to detect races.
588 fs.map_generation = fs.map->timestamp;
590 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
591 panic("vm_fault: fault on nofault entry, addr: %lx",
596 * A system map entry may return a NULL object. No object means
597 * no pager means an unrecoverable kernel fault.
599 if (fs.first_object == NULL) {
600 panic("vm_fault: unrecoverable fault at %p in entry %p",
601 (void *)vaddr, fs.entry);
605 * Make a reference to this object to prevent its disposal while we
606 * are messing with it. Once we have the reference, the map is free
607 * to be diddled. Since objects reference their shadows (and copies),
608 * they will stay around as well.
610 * Bump the paging-in-progress count to prevent size changes (e.g.
611 * truncation operations) during I/O. This must be done after
612 * obtaining the vnode lock in order to avoid possible deadlocks.
614 * The vm_object must be held before manipulation.
616 lwkt_gettoken(&vm_token);
617 vm_object_hold(fs.first_object);
618 vm_object_reference(fs.first_object);
619 fs.vp = vnode_pager_lock(fs.first_object);
620 vm_object_pip_add(fs.first_object, 1);
621 vm_object_drop(fs.first_object);
622 lwkt_reltoken(&vm_token);
624 fs.lookup_still_valid = TRUE;
626 fs.object = fs.first_object; /* so unlock_and_deallocate works */
629 * If the entry is wired we cannot change the page protection.
632 fault_type = fs.first_prot;
635 * The page we want is at (first_object, first_pindex), but if the
636 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
637 * page table to figure out the actual pindex.
639 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
642 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
643 result = vm_fault_vpagetable(&fs, &first_pindex,
644 fs.entry->aux.master_pde,
646 if (result == KERN_TRY_AGAIN)
648 if (result != KERN_SUCCESS) {
655 * Now we have the actual (object, pindex), fault in the page. If
656 * vm_fault_object() fails it will unlock and deallocate the FS
657 * data. If it succeeds everything remains locked and fs->object
658 * will have an additinal PIP count if it is not equal to
661 result = vm_fault_object(&fs, first_pindex, fault_type);
663 if (result == KERN_TRY_AGAIN)
665 if (result != KERN_SUCCESS) {
670 if ((orig_fault_type & VM_PROT_WRITE) &&
671 (fs.prot & VM_PROT_WRITE) == 0) {
672 *errorp = KERN_PROTECTION_FAILURE;
673 unlock_and_deallocate(&fs);
678 * On success vm_fault_object() does not unlock or deallocate, and fs.m
679 * will contain a busied page.
684 * Return a held page. We are not doing any pmap manipulation so do
685 * not set PG_MAPPED. However, adjust the page flags according to
686 * the fault type because the caller may not use a managed pmapping
687 * (so we don't want to lose the fact that the page will be dirtied
688 * if a write fault was specified).
690 lwkt_gettoken(&vm_token);
692 if (fault_type & VM_PROT_WRITE)
696 * Update the pmap. We really only have to do this if a COW
697 * occured to replace the read-only page with the new page. For
698 * now just do it unconditionally. XXX
700 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
701 vm_page_flag_set(fs.m, PG_REFERENCED);
704 * Unbusy the page by activating it. It remains held and will not
707 vm_page_activate(fs.m);
709 if (curthread->td_lwp) {
711 curthread->td_lwp->lwp_ru.ru_majflt++;
713 curthread->td_lwp->lwp_ru.ru_minflt++;
718 * Unlock everything, and return the held page.
720 vm_page_wakeup(fs.m);
721 vm_object_deallocate(fs.first_object);
722 /*fs.first_object = NULL; */
723 lwkt_reltoken(&vm_token);
730 * Fault in the specified (object,offset), dirty the returned page as
731 * needed. If the requested fault_type cannot be done NULL and an
734 * A held (but not busied) page is returned.
739 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
740 vm_prot_t fault_type, int fault_flags, int *errorp)
743 vm_pindex_t first_pindex;
744 struct faultstate fs;
745 struct vm_map_entry entry;
747 bzero(&entry, sizeof(entry));
748 entry.object.vm_object = object;
749 entry.maptype = VM_MAPTYPE_NORMAL;
750 entry.protection = entry.max_protection = fault_type;
754 fs.fault_flags = fault_flags;
756 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
760 fs.first_object = object;
761 first_pindex = OFF_TO_IDX(offset);
763 fs.first_prot = fault_type;
765 /*fs.map_generation = 0; unused */
768 * Make a reference to this object to prevent its disposal while we
769 * are messing with it. Once we have the reference, the map is free
770 * to be diddled. Since objects reference their shadows (and copies),
771 * they will stay around as well.
773 * Bump the paging-in-progress count to prevent size changes (e.g.
774 * truncation operations) during I/O. This must be done after
775 * obtaining the vnode lock in order to avoid possible deadlocks.
777 lwkt_gettoken(&vm_token);
778 vm_object_hold(fs.first_object);
779 vm_object_reference(fs.first_object);
780 fs.vp = vnode_pager_lock(fs.first_object);
781 vm_object_pip_add(fs.first_object, 1);
782 vm_object_drop(fs.first_object);
783 lwkt_reltoken(&vm_token);
785 fs.lookup_still_valid = TRUE;
787 fs.object = fs.first_object; /* so unlock_and_deallocate works */
790 /* XXX future - ability to operate on VM object using vpagetable */
791 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
792 result = vm_fault_vpagetable(&fs, &first_pindex,
793 fs.entry->aux.master_pde,
795 if (result == KERN_TRY_AGAIN)
797 if (result != KERN_SUCCESS) {
805 * Now we have the actual (object, pindex), fault in the page. If
806 * vm_fault_object() fails it will unlock and deallocate the FS
807 * data. If it succeeds everything remains locked and fs->object
808 * will have an additinal PIP count if it is not equal to
811 result = vm_fault_object(&fs, first_pindex, fault_type);
813 if (result == KERN_TRY_AGAIN)
815 if (result != KERN_SUCCESS) {
820 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
821 *errorp = KERN_PROTECTION_FAILURE;
822 unlock_and_deallocate(&fs);
827 * On success vm_fault_object() does not unlock or deallocate, and fs.m
828 * will contain a busied page.
833 * Return a held page. We are not doing any pmap manipulation so do
834 * not set PG_MAPPED. However, adjust the page flags according to
835 * the fault type because the caller may not use a managed pmapping
836 * (so we don't want to lose the fact that the page will be dirtied
837 * if a write fault was specified).
839 lwkt_gettoken(&vm_token);
841 if (fault_type & VM_PROT_WRITE)
844 if (fault_flags & VM_FAULT_DIRTY)
846 if (fault_flags & VM_FAULT_UNSWAP)
847 swap_pager_unswapped(fs.m);
850 * Indicate that the page was accessed.
852 vm_page_flag_set(fs.m, PG_REFERENCED);
855 * Unbusy the page by activating it. It remains held and will not
858 vm_page_activate(fs.m);
860 if (curthread->td_lwp) {
862 mycpu->gd_cnt.v_vm_faults++;
863 curthread->td_lwp->lwp_ru.ru_majflt++;
865 curthread->td_lwp->lwp_ru.ru_minflt++;
870 * Unlock everything, and return the held page.
872 vm_page_wakeup(fs.m);
873 vm_object_deallocate(fs.first_object);
874 /*fs.first_object = NULL; */
875 lwkt_reltoken(&vm_token);
882 * Translate the virtual page number (first_pindex) that is relative
883 * to the address space into a logical page number that is relative to the
884 * backing object. Use the virtual page table pointed to by (vpte).
886 * This implements an N-level page table. Any level can terminate the
887 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
888 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
890 * No requirements (vm_token need not be held).
894 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
895 vpte_t vpte, int fault_type)
898 struct lwbuf lwb_cache;
899 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
900 int result = KERN_SUCCESS;
905 * We cannot proceed if the vpte is not valid, not readable
906 * for a read fault, or not writable for a write fault.
908 if ((vpte & VPTE_V) == 0) {
909 unlock_and_deallocate(fs);
910 return (KERN_FAILURE);
912 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
913 unlock_and_deallocate(fs);
914 return (KERN_FAILURE);
916 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
917 unlock_and_deallocate(fs);
918 return (KERN_FAILURE);
920 if ((vpte & VPTE_PS) || vshift == 0)
922 KKASSERT(vshift >= VPTE_PAGE_BITS);
925 * Get the page table page. Nominally we only read the page
926 * table, but since we are actively setting VPTE_M and VPTE_A,
927 * tell vm_fault_object() that we are writing it.
929 * There is currently no real need to optimize this.
931 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
932 VM_PROT_READ|VM_PROT_WRITE);
933 if (result != KERN_SUCCESS)
937 * Process the returned fs.m and look up the page table
938 * entry in the page table page.
940 vshift -= VPTE_PAGE_BITS;
941 lwb = lwbuf_alloc(fs->m, &lwb_cache);
942 ptep = ((vpte_t *)lwbuf_kva(lwb) +
943 ((*pindex >> vshift) & VPTE_PAGE_MASK));
947 * Page table write-back. If the vpte is valid for the
948 * requested operation, do a write-back to the page table.
950 * XXX VPTE_M is not set properly for page directory pages.
951 * It doesn't get set in the page directory if the page table
952 * is modified during a read access.
954 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
956 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
957 atomic_set_long(ptep, VPTE_M | VPTE_A);
958 vm_page_dirty(fs->m);
961 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
963 if ((vpte & VPTE_A) == 0) {
964 atomic_set_long(ptep, VPTE_A);
965 vm_page_dirty(fs->m);
969 vm_page_flag_set(fs->m, PG_REFERENCED);
970 vm_page_activate(fs->m);
971 vm_page_wakeup(fs->m);
973 cleanup_successful_fault(fs);
976 * Combine remaining address bits with the vpte.
978 /* JG how many bits from each? */
979 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
980 (*pindex & ((1L << vshift) - 1));
981 return (KERN_SUCCESS);
986 * This is the core of the vm_fault code.
988 * Do all operations required to fault-in (fs.first_object, pindex). Run
989 * through the shadow chain as necessary and do required COW or virtual
990 * copy operations. The caller has already fully resolved the vm_map_entry
991 * and, if appropriate, has created a copy-on-write layer. All we need to
992 * do is iterate the object chain.
994 * On failure (fs) is unlocked and deallocated and the caller may return or
995 * retry depending on the failure code. On success (fs) is NOT unlocked or
996 * deallocated, fs.m will contained a resolved, busied page, and fs.object
997 * will have an additional PIP count if it is not equal to fs.first_object.
1003 vm_fault_object(struct faultstate *fs,
1004 vm_pindex_t first_pindex, vm_prot_t fault_type)
1006 vm_object_t next_object;
1009 fs->prot = fs->first_prot;
1010 fs->object = fs->first_object;
1011 pindex = first_pindex;
1014 * If a read fault occurs we try to make the page writable if
1015 * possible. There are three cases where we cannot make the
1016 * page mapping writable:
1018 * (1) The mapping is read-only or the VM object is read-only,
1019 * fs->prot above will simply not have VM_PROT_WRITE set.
1021 * (2) If the mapping is a virtual page table we need to be able
1022 * to detect writes so we can set VPTE_M in the virtual page
1025 * (3) If the VM page is read-only or copy-on-write, upgrading would
1026 * just result in an unnecessary COW fault.
1028 * VM_PROT_VPAGED is set if faulting via a virtual page table and
1029 * causes adjustments to the 'M'odify bit to also turn off write
1030 * access to force a re-fault.
1032 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1033 if ((fault_type & VM_PROT_WRITE) == 0)
1034 fs->prot &= ~VM_PROT_WRITE;
1037 lwkt_gettoken(&vm_token);
1041 * If the object is dead, we stop here
1043 if (fs->object->flags & OBJ_DEAD) {
1044 unlock_and_deallocate(fs);
1045 lwkt_reltoken(&vm_token);
1046 return (KERN_PROTECTION_FAILURE);
1050 * See if the page is resident.
1052 fs->m = vm_page_lookup(fs->object, pindex);
1053 if (fs->m != NULL) {
1056 * Wait/Retry if the page is busy. We have to do this
1057 * if the page is busy via either PG_BUSY or
1058 * vm_page_t->busy because the vm_pager may be using
1059 * vm_page_t->busy for pageouts ( and even pageins if
1060 * it is the vnode pager ), and we could end up trying
1061 * to pagein and pageout the same page simultaneously.
1063 * We can theoretically allow the busy case on a read
1064 * fault if the page is marked valid, but since such
1065 * pages are typically already pmap'd, putting that
1066 * special case in might be more effort then it is
1067 * worth. We cannot under any circumstances mess
1068 * around with a vm_page_t->busy page except, perhaps,
1071 if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
1073 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1074 mycpu->gd_cnt.v_intrans++;
1075 vm_object_deallocate(fs->first_object);
1076 fs->first_object = NULL;
1077 lwkt_reltoken(&vm_token);
1078 return (KERN_TRY_AGAIN);
1082 * If reactivating a page from PQ_CACHE we may have
1085 queue = fs->m->queue;
1086 vm_page_unqueue_nowakeup(fs->m);
1088 if ((queue - fs->m->pc) == PQ_CACHE &&
1089 vm_page_count_severe()) {
1090 vm_page_activate(fs->m);
1091 unlock_and_deallocate(fs);
1093 lwkt_reltoken(&vm_token);
1094 return (KERN_TRY_AGAIN);
1098 * Mark page busy for other processes, and the
1099 * pagedaemon. If it still isn't completely valid
1100 * (readable), or if a read-ahead-mark is set on
1101 * the VM page, jump to readrest, else we found the
1102 * page and can return.
1104 * We can release the spl once we have marked the
1107 vm_page_busy(fs->m);
1109 if (fs->m->object != &kernel_object) {
1110 if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1114 if (fs->m->flags & PG_RAM) {
1117 vm_page_flag_clear(fs->m, PG_RAM);
1121 break; /* break to PAGE HAS BEEN FOUND */
1125 * Page is not resident, If this is the search termination
1126 * or the pager might contain the page, allocate a new page.
1128 if (TRYPAGER(fs) || fs->object == fs->first_object) {
1130 * If the page is beyond the object size we fail
1132 if (pindex >= fs->object->size) {
1133 lwkt_reltoken(&vm_token);
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 lwkt_reltoken(&vm_token);
1147 unlock_and_deallocate(fs);
1148 tsleep(curproc, 0, "vmrate", limticks);
1150 return (KERN_TRY_AGAIN);
1155 * Allocate a new page for this object/offset pair.
1158 if (!vm_page_count_severe()) {
1159 fs->m = vm_page_alloc(fs->object, pindex,
1160 (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
1162 if (fs->m == NULL) {
1163 lwkt_reltoken(&vm_token);
1164 unlock_and_deallocate(fs);
1166 return (KERN_TRY_AGAIN);
1172 * We have found an invalid or partially valid page, a
1173 * page with a read-ahead mark which might be partially or
1174 * fully valid (and maybe dirty too), or we have allocated
1177 * Attempt to fault-in the page if there is a chance that the
1178 * pager has it, and potentially fault in additional pages
1181 * We are NOT in splvm here and if TRYPAGER is true then
1182 * fs.m will be non-NULL and will be PG_BUSY for us.
1187 u_char behavior = vm_map_entry_behavior(fs->entry);
1189 if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1195 * If sequential access is detected then attempt
1196 * to deactivate/cache pages behind the scan to
1197 * prevent resource hogging.
1199 * Use of PG_RAM to detect sequential access
1200 * also simulates multi-zone sequential access
1201 * detection for free.
1203 * NOTE: Partially valid dirty pages cannot be
1204 * deactivated without causing NFS picemeal
1207 if ((fs->first_object->type != OBJT_DEVICE) &&
1208 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1209 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1210 (fs->m->flags & PG_RAM)))
1212 vm_pindex_t scan_pindex;
1213 int scan_count = 16;
1215 if (first_pindex < 16) {
1219 scan_pindex = first_pindex - 16;
1220 if (scan_pindex < 16)
1221 scan_count = scan_pindex;
1226 while (scan_count) {
1229 mt = vm_page_lookup(fs->first_object,
1232 (mt->valid != VM_PAGE_BITS_ALL)) {
1236 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
1243 vm_page_test_dirty(mt);
1247 vm_page_deactivate(mt);
1261 * Avoid deadlocking against the map when doing I/O.
1262 * fs.object and the page is PG_BUSY'd.
1264 * NOTE: Once unlocked, fs->entry can become stale
1265 * so this will NULL it out.
1267 * NOTE: fs->entry is invalid until we relock the
1268 * map and verify that the timestamp has not
1274 * Acquire the page data. We still hold a ref on
1275 * fs.object and the page has been PG_BUSY's.
1277 * The pager may replace the page (for example, in
1278 * order to enter a fictitious page into the
1279 * object). If it does so it is responsible for
1280 * cleaning up the passed page and properly setting
1281 * the new page PG_BUSY.
1283 * If we got here through a PG_RAM read-ahead
1284 * mark the page may be partially dirty and thus
1285 * not freeable. Don't bother checking to see
1286 * if the pager has the page because we can't free
1287 * it anyway. We have to depend on the get_page
1288 * operation filling in any gaps whether there is
1289 * backing store or not.
1291 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1293 if (rv == VM_PAGER_OK) {
1295 * Relookup in case pager changed page. Pager
1296 * is responsible for disposition of old page
1299 * XXX other code segments do relookups too.
1300 * It's a bad abstraction that needs to be
1303 fs->m = vm_page_lookup(fs->object, pindex);
1304 if (fs->m == NULL) {
1305 lwkt_reltoken(&vm_token);
1306 unlock_and_deallocate(fs);
1307 return (KERN_TRY_AGAIN);
1311 break; /* break to PAGE HAS BEEN FOUND */
1315 * Remove the bogus page (which does not exist at this
1316 * object/offset); before doing so, we must get back
1317 * our object lock to preserve our invariant.
1319 * Also wake up any other process that may want to bring
1322 * If this is the top-level object, we must leave the
1323 * busy page to prevent another process from rushing
1324 * past us, and inserting the page in that object at
1325 * the same time that we are.
1327 if (rv == VM_PAGER_ERROR) {
1329 kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
1331 kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
1335 * Data outside the range of the pager or an I/O error
1337 * The page may have been wired during the pagein,
1338 * e.g. by the buffer cache, and cannot simply be
1339 * freed. Call vnode_pager_freepage() to deal with it.
1342 * XXX - the check for kernel_map is a kludge to work
1343 * around having the machine panic on a kernel space
1344 * fault w/ I/O error.
1346 if (((fs->map != &kernel_map) &&
1347 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1348 vnode_pager_freepage(fs->m);
1349 lwkt_reltoken(&vm_token);
1351 unlock_and_deallocate(fs);
1352 if (rv == VM_PAGER_ERROR)
1353 return (KERN_FAILURE);
1355 return (KERN_PROTECTION_FAILURE);
1358 if (fs->object != fs->first_object) {
1359 vnode_pager_freepage(fs->m);
1362 * XXX - we cannot just fall out at this
1363 * point, m has been freed and is invalid!
1369 * We get here if the object has a default pager (or unwiring)
1370 * or the pager doesn't have the page.
1372 if (fs->object == fs->first_object)
1373 fs->first_m = fs->m;
1376 * Move on to the next object. Lock the next object before
1377 * unlocking the current one.
1379 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1380 next_object = fs->object->backing_object;
1381 if (next_object == NULL) {
1383 * If there's no object left, fill the page in the top
1384 * object with zeros.
1386 if (fs->object != fs->first_object) {
1387 vm_object_pip_wakeup(fs->object);
1389 fs->object = fs->first_object;
1390 pindex = first_pindex;
1391 fs->m = fs->first_m;
1396 * Zero the page if necessary and mark it valid.
1398 if ((fs->m->flags & PG_ZERO) == 0) {
1399 vm_page_zero_fill(fs->m);
1402 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m));
1404 vm_page_flag_clear(fs->m, PG_ZERO);
1405 mycpu->gd_cnt.v_ozfod++;
1407 mycpu->gd_cnt.v_zfod++;
1408 fs->m->valid = VM_PAGE_BITS_ALL;
1409 break; /* break to PAGE HAS BEEN FOUND */
1411 if (fs->object != fs->first_object) {
1412 vm_object_pip_wakeup(fs->object);
1414 KASSERT(fs->object != next_object,
1415 ("object loop %p", next_object));
1416 fs->object = next_object;
1417 vm_object_pip_add(fs->object, 1);
1421 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1424 * vm_token is still held
1426 * If the page is being written, but isn't already owned by the
1427 * top-level object, we have to copy it into a new page owned by the
1430 KASSERT((fs->m->flags & PG_BUSY) != 0,
1431 ("vm_fault: not busy after main loop"));
1433 if (fs->object != fs->first_object) {
1435 * We only really need to copy if we want to write it.
1437 if (fault_type & VM_PROT_WRITE) {
1439 * This allows pages to be virtually copied from a
1440 * backing_object into the first_object, where the
1441 * backing object has no other refs to it, and cannot
1442 * gain any more refs. Instead of a bcopy, we just
1443 * move the page from the backing object to the
1444 * first object. Note that we must mark the page
1445 * dirty in the first object so that it will go out
1446 * to swap when needed.
1450 * Map, if present, has not changed
1453 fs->map_generation == fs->map->timestamp) &&
1455 * Only one shadow object
1457 (fs->object->shadow_count == 1) &&
1459 * No COW refs, except us
1461 (fs->object->ref_count == 1) &&
1463 * No one else can look this object up
1465 (fs->object->handle == NULL) &&
1467 * No other ways to look the object up
1469 ((fs->object->type == OBJT_DEFAULT) ||
1470 (fs->object->type == OBJT_SWAP)) &&
1472 * We don't chase down the shadow chain
1474 (fs->object == fs->first_object->backing_object) &&
1477 * grab the lock if we need to
1479 (fs->lookup_still_valid ||
1481 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1484 fs->lookup_still_valid = 1;
1486 * get rid of the unnecessary page
1488 vm_page_protect(fs->first_m, VM_PROT_NONE);
1489 vm_page_free(fs->first_m);
1493 * grab the page and put it into the
1494 * process'es object. The page is
1495 * automatically made dirty.
1497 vm_page_rename(fs->m, fs->first_object, first_pindex);
1498 fs->first_m = fs->m;
1499 vm_page_busy(fs->first_m);
1501 mycpu->gd_cnt.v_cow_optim++;
1504 * Oh, well, lets copy it.
1506 vm_page_copy(fs->m, fs->first_m);
1507 vm_page_event(fs->m, VMEVENT_COW);
1512 * We no longer need the old page or object.
1518 * fs->object != fs->first_object due to above
1521 vm_object_pip_wakeup(fs->object);
1524 * Only use the new page below...
1527 mycpu->gd_cnt.v_cow_faults++;
1528 fs->m = fs->first_m;
1529 fs->object = fs->first_object;
1530 pindex = first_pindex;
1533 * If it wasn't a write fault avoid having to copy
1534 * the page by mapping it read-only.
1536 fs->prot &= ~VM_PROT_WRITE;
1541 * Relock the map if necessary, then check the generation count.
1542 * relock_map() will update fs->timestamp to account for the
1543 * relocking if necessary.
1545 * If the count has changed after relocking then all sorts of
1546 * crap may have happened and we have to retry.
1548 * NOTE: The relock_map() can fail due to a deadlock against
1549 * the vm_page we are holding BUSY.
1551 if (fs->lookup_still_valid == FALSE && fs->map) {
1552 if (relock_map(fs) ||
1553 fs->map->timestamp != fs->map_generation) {
1555 lwkt_reltoken(&vm_token);
1556 unlock_and_deallocate(fs);
1557 return (KERN_TRY_AGAIN);
1562 * If the fault is a write, we know that this page is being
1563 * written NOW so dirty it explicitly to save on pmap_is_modified()
1566 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1567 * if the page is already dirty to prevent data written with
1568 * the expectation of being synced from not being synced.
1569 * Likewise if this entry does not request NOSYNC then make
1570 * sure the page isn't marked NOSYNC. Applications sharing
1571 * data should use the same flags to avoid ping ponging.
1573 * Also tell the backing pager, if any, that it should remove
1574 * any swap backing since the page is now dirty.
1576 if (fs->prot & VM_PROT_WRITE) {
1577 vm_object_set_writeable_dirty(fs->m->object);
1578 vm_set_nosync(fs->m, fs->entry);
1579 if (fs->fault_flags & VM_FAULT_DIRTY) {
1580 vm_page_dirty(fs->m);
1581 swap_pager_unswapped(fs->m);
1585 lwkt_reltoken(&vm_token);
1588 * Page had better still be busy. We are still locked up and
1589 * fs->object will have another PIP reference if it is not equal
1590 * to fs->first_object.
1592 KASSERT(fs->m->flags & PG_BUSY,
1593 ("vm_fault: page %p not busy!", fs->m));
1596 * Sanity check: page must be completely valid or it is not fit to
1597 * map into user space. vm_pager_get_pages() ensures this.
1599 if (fs->m->valid != VM_PAGE_BITS_ALL) {
1600 vm_page_zero_invalid(fs->m, TRUE);
1601 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1603 vm_page_flag_clear(fs->m, PG_ZERO);
1605 return (KERN_SUCCESS);
1609 * Wire down a range of virtual addresses in a map. The entry in question
1610 * should be marked in-transition and the map must be locked. We must
1611 * release the map temporarily while faulting-in the page to avoid a
1612 * deadlock. Note that the entry may be clipped while we are blocked but
1613 * will never be freed.
1618 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1620 boolean_t fictitious;
1628 pmap = vm_map_pmap(map);
1629 start = entry->start;
1631 fictitious = entry->object.vm_object &&
1632 (entry->object.vm_object->type == OBJT_DEVICE);
1633 if (entry->eflags & MAP_ENTRY_KSTACK)
1635 lwkt_gettoken(&vm_token);
1640 * We simulate a fault to get the page and enter it in the physical
1643 for (va = start; va < end; va += PAGE_SIZE) {
1645 rv = vm_fault(map, va, VM_PROT_READ,
1646 VM_FAULT_USER_WIRE);
1648 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1649 VM_FAULT_CHANGE_WIRING);
1652 while (va > start) {
1654 if ((pa = pmap_extract(pmap, va)) == 0)
1656 pmap_change_wiring(pmap, va, FALSE);
1658 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1661 lwkt_reltoken(&vm_token);
1666 lwkt_reltoken(&vm_token);
1667 return (KERN_SUCCESS);
1671 * Unwire a range of virtual addresses in a map. The map should be
1675 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1677 boolean_t fictitious;
1684 pmap = vm_map_pmap(map);
1685 start = entry->start;
1687 fictitious = entry->object.vm_object &&
1688 (entry->object.vm_object->type == OBJT_DEVICE);
1689 if (entry->eflags & MAP_ENTRY_KSTACK)
1693 * Since the pages are wired down, we must be able to get their
1694 * mappings from the physical map system.
1696 lwkt_gettoken(&vm_token);
1697 for (va = start; va < end; va += PAGE_SIZE) {
1698 pa = pmap_extract(pmap, va);
1700 pmap_change_wiring(pmap, va, FALSE);
1702 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1705 lwkt_reltoken(&vm_token);
1709 * Reduce the rate at which memory is allocated to a process based
1710 * on the perceived load on the VM system. As the load increases
1711 * the allocation burst rate goes down and the delay increases.
1713 * Rate limiting does not apply when faulting active or inactive
1714 * pages. When faulting 'cache' pages, rate limiting only applies
1715 * if the system currently has a severe page deficit.
1717 * XXX vm_pagesupply should be increased when a page is freed.
1719 * We sleep up to 1/10 of a second.
1722 vm_fault_ratelimit(struct vmspace *vmspace)
1724 if (vm_load_enable == 0)
1726 if (vmspace->vm_pagesupply > 0) {
1727 --vmspace->vm_pagesupply; /* SMP race ok */
1731 if (vm_load_debug) {
1732 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1734 (1000 - vm_load ) / 10, vm_load * hz / 10000,
1735 curproc->p_pid, curproc->p_comm);
1738 vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1739 return(vm_load * hz / 10000);
1743 * Copy all of the pages from a wired-down map entry to another.
1745 * The source and destination maps must be locked for write.
1746 * The source map entry must be wired down (or be a sharing map
1747 * entry corresponding to a main map entry that is wired down).
1749 * No other requirements.
1752 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1753 vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1755 vm_object_t dst_object;
1756 vm_object_t src_object;
1757 vm_ooffset_t dst_offset;
1758 vm_ooffset_t src_offset;
1768 src_object = src_entry->object.vm_object;
1769 src_offset = src_entry->offset;
1772 * Create the top-level object for the destination entry. (Doesn't
1773 * actually shadow anything - we copy the pages directly.)
1775 vm_map_entry_allocate_object(dst_entry);
1776 dst_object = dst_entry->object.vm_object;
1778 prot = dst_entry->max_protection;
1781 * Loop through all of the pages in the entry's range, copying each
1782 * one from the source object (it should be there) to the destination
1785 for (vaddr = dst_entry->start, dst_offset = 0;
1786 vaddr < dst_entry->end;
1787 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1790 * Allocate a page in the destination object
1793 dst_m = vm_page_alloc(dst_object,
1794 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1795 if (dst_m == NULL) {
1798 } while (dst_m == NULL);
1801 * Find the page in the source object, and copy it in.
1802 * (Because the source is wired down, the page will be in
1805 src_m = vm_page_lookup(src_object,
1806 OFF_TO_IDX(dst_offset + src_offset));
1808 panic("vm_fault_copy_wired: page missing");
1810 vm_page_copy(src_m, dst_m);
1811 vm_page_event(src_m, VMEVENT_COW);
1814 * Enter it in the pmap...
1817 vm_page_flag_clear(dst_m, PG_ZERO);
1818 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1821 * Mark it no longer busy, and put it on the active list.
1823 vm_page_activate(dst_m);
1824 vm_page_wakeup(dst_m);
1831 * This routine checks around the requested page for other pages that
1832 * might be able to be faulted in. This routine brackets the viable
1833 * pages for the pages to be paged in.
1836 * m, rbehind, rahead
1839 * marray (array of vm_page_t), reqpage (index of requested page)
1842 * number of pages in marray
1845 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1846 vm_page_t *marray, int *reqpage)
1850 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1852 int cbehind, cahead;
1858 * we don't fault-ahead for device pager
1860 if (object->type == OBJT_DEVICE) {
1867 * if the requested page is not available, then give up now
1869 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1870 *reqpage = 0; /* not used by caller, fix compiler warn */
1874 if ((cbehind == 0) && (cahead == 0)) {
1880 if (rahead > cahead) {
1884 if (rbehind > cbehind) {
1889 * Do not do any readahead if we have insufficient free memory.
1891 * XXX code was broken disabled before and has instability
1892 * with this conditonal fixed, so shortcut for now.
1894 if (burst_fault == 0 || vm_page_count_severe()) {
1901 * scan backward for the read behind pages -- in memory
1903 * Assume that if the page is not found an interrupt will not
1904 * create it. Theoretically interrupts can only remove (busy)
1905 * pages, not create new associations.
1908 if (rbehind > pindex) {
1912 startpindex = pindex - rbehind;
1915 lwkt_gettoken(&vm_token);
1916 for (tpindex = pindex; tpindex > startpindex; --tpindex) {
1917 if (vm_page_lookup(object, tpindex - 1))
1922 while (tpindex < pindex) {
1923 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1925 lwkt_reltoken(&vm_token);
1926 for (j = 0; j < i; j++) {
1927 vm_page_free(marray[j]);
1937 lwkt_reltoken(&vm_token);
1943 * Assign requested page
1950 * Scan forwards for read-ahead pages
1952 tpindex = pindex + 1;
1953 endpindex = tpindex + rahead;
1954 if (endpindex > object->size)
1955 endpindex = object->size;
1957 lwkt_gettoken(&vm_token);
1958 while (tpindex < endpindex) {
1959 if (vm_page_lookup(object, tpindex))
1961 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1968 lwkt_reltoken(&vm_token);
1976 * vm_prefault() provides a quick way of clustering pagefaults into a
1977 * processes address space. It is a "cousin" of pmap_object_init_pt,
1978 * except it runs at page fault time instead of mmap time.
1980 * This code used to be per-platform pmap_prefault(). It is now
1981 * machine-independent and enhanced to also pre-fault zero-fill pages
1982 * (see vm.fast_fault) as well as make them writable, which greatly
1983 * reduces the number of page faults programs incur.
1985 * Application performance when pre-faulting zero-fill pages is heavily
1986 * dependent on the application. Very tiny applications like /bin/echo
1987 * lose a little performance while applications of any appreciable size
1988 * gain performance. Prefaulting multiple pages also reduces SMP
1989 * congestion and can improve SMP performance significantly.
1991 * NOTE! prot may allow writing but this only applies to the top level
1992 * object. If we wind up mapping a page extracted from a backing
1993 * object we have to make sure it is read-only.
1995 * NOTE! The caller has already handled any COW operations on the
1996 * vm_map_entry via the normal fault code. Do NOT call this
1997 * shortcut unless the normal fault code has run on this entry.
1999 * No other requirements.
2003 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2005 static int vm_prefault_pageorder[] = {
2006 -PAGE_SIZE, PAGE_SIZE,
2007 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2008 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2009 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2013 * Set PG_NOSYNC if the map entry indicates so, but only if the page
2014 * is not already dirty by other means. This will prevent passive
2015 * filesystem syncing as well as 'sync' from writing out the page.
2018 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
2020 if (entry->eflags & MAP_ENTRY_NOSYNC) {
2022 vm_page_flag_set(m, PG_NOSYNC);
2024 vm_page_flag_clear(m, PG_NOSYNC);
2029 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot)
2042 * We do not currently prefault mappings that use virtual page
2043 * tables. We do not prefault foreign pmaps.
2045 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2047 lp = curthread->td_lwp;
2048 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2051 lwkt_gettoken(&vm_token);
2053 object = entry->object.vm_object;
2054 KKASSERT(object != NULL);
2055 vm_object_hold(object);
2057 starta = addra - PFBAK * PAGE_SIZE;
2058 if (starta < entry->start)
2059 starta = entry->start;
2060 else if (starta > addra)
2063 KKASSERT(object == entry->object.vm_object);
2064 for (i = 0; i < PAGEORDER_SIZE; i++) {
2065 vm_object_t lobject;
2066 vm_object_t nobject;
2069 addr = addra + vm_prefault_pageorder[i];
2070 if (addr > addra + (PFFOR * PAGE_SIZE))
2073 if (addr < starta || addr >= entry->end)
2076 if (pmap_prefault_ok(pmap, addr) == 0)
2080 * Follow the VM object chain to obtain the page to be mapped
2083 * If we reach the terminal object without finding a page
2084 * and we determine it would be advantageous, then allocate
2085 * a zero-fill page for the base object. The base object
2086 * is guaranteed to be OBJT_DEFAULT for this case.
2088 * In order to not have to check the pager via *haspage*()
2089 * we stop if any non-default object is encountered. e.g.
2090 * a vnode or swap object would stop the loop.
2092 * XXX It is unclear whether hold chaining is sufficient
2093 * to maintain the validity of the backing object chain.
2095 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2100 KKASSERT(lobject == entry->object.vm_object);
2101 vm_object_hold(lobject);
2103 while ((m = vm_page_lookup(lobject, pindex)) == NULL) {
2104 if (lobject->type != OBJT_DEFAULT)
2106 if (lobject->backing_object == NULL) {
2107 if (vm_fast_fault == 0)
2109 if (vm_prefault_pageorder[i] < 0 ||
2110 (prot & VM_PROT_WRITE) == 0 ||
2111 vm_page_count_min(0)) {
2115 /* NOTE: allocated from base object */
2116 m = vm_page_alloc(object, index,
2117 VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
2119 if ((m->flags & PG_ZERO) == 0) {
2120 vm_page_zero_fill(m);
2123 pmap_page_assertzero(
2124 VM_PAGE_TO_PHYS(m));
2126 vm_page_flag_clear(m, PG_ZERO);
2127 mycpu->gd_cnt.v_ozfod++;
2129 mycpu->gd_cnt.v_zfod++;
2130 m->valid = VM_PAGE_BITS_ALL;
2133 /* lobject = object .. not needed */
2136 if (lobject->backing_object_offset & PAGE_MASK)
2138 while ((nobject = lobject->backing_object) != NULL) {
2139 vm_object_hold(nobject);
2140 if (nobject == lobject->backing_object) {
2142 lobject->backing_object_offset >>
2144 vm_object_lock_swap();
2145 vm_object_drop(lobject);
2149 vm_object_drop(nobject);
2151 if (nobject == NULL) {
2152 kprintf("vm_prefault: Warning, backing object "
2153 "race averted lobject %p\n",
2157 pprot &= ~VM_PROT_WRITE;
2159 vm_object_drop(lobject);
2162 * NOTE: lobject now invalid (if we did a zero-fill we didn't
2163 * bother assigning lobject = object).
2165 * Give-up if the page is not available.
2171 * Do not conditionalize on PG_RAM. If pages are present in
2172 * the VM system we assume optimal caching. If caching is
2173 * not optimal the I/O gravy train will be restarted when we
2174 * hit an unavailable page. We do not want to try to restart
2175 * the gravy train now because we really don't know how much
2176 * of the object has been cached. The cost for restarting
2177 * the gravy train should be low (since accesses will likely
2178 * be I/O bound anyway).
2180 * The object must be marked dirty if we are mapping a
2183 if (pprot & VM_PROT_WRITE)
2184 vm_object_set_writeable_dirty(m->object);
2187 * Enter the page into the pmap if appropriate. If we had
2188 * allocated the page we have to place it on a queue. If not
2189 * we just have to make sure it isn't on the cache queue
2190 * (pages on the cache queue are not allowed to be mapped).
2193 if (pprot & VM_PROT_WRITE)
2194 vm_set_nosync(m, entry);
2195 pmap_enter(pmap, addr, m, pprot, 0);
2196 vm_page_deactivate(m);
2199 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2201 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2203 * A fully valid page not undergoing soft I/O can
2204 * be immediately entered into the pmap.
2207 if ((m->queue - m->pc) == PQ_CACHE) {
2208 vm_page_deactivate(m);
2210 if (pprot & VM_PROT_WRITE)
2211 vm_set_nosync(m, entry);
2212 pmap_enter(pmap, addr, m, pprot, 0);
2216 vm_object_drop(object);
2217 lwkt_reltoken(&vm_token);