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, "");
128 int vm_shared_fault = 1;
129 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, &vm_shared_fault, 0,
130 "Allow shared token on vm_object");
131 static long vm_shared_hit = 0;
132 SYSCTL_LONG(_vm, OID_AUTO, shared_hit, CTLFLAG_RW, &vm_shared_hit, 0,
133 "Successful shared faults");
134 static long vm_shared_miss = 0;
135 SYSCTL_LONG(_vm, OID_AUTO, shared_miss, CTLFLAG_RW, &vm_shared_miss, 0,
136 "Unsuccessful shared faults");
138 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
139 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
141 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
143 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry);
144 static void vm_prefault(pmap_t pmap, vm_offset_t addra,
145 vm_map_entry_t entry, int prot, int fault_flags);
146 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
147 vm_map_entry_t entry, int prot, int fault_flags);
150 release_page(struct faultstate *fs)
152 vm_page_deactivate(fs->m);
153 vm_page_wakeup(fs->m);
158 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse
159 * requires relocking and then checking the timestamp.
161 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do
162 * not have to update fs->map_generation here.
164 * NOTE: This function can fail due to a deadlock against the caller's
165 * holding of a vm_page BUSY.
168 relock_map(struct faultstate *fs)
172 if (fs->lookup_still_valid == FALSE && fs->map) {
173 error = vm_map_lock_read_to(fs->map);
175 fs->lookup_still_valid = TRUE;
183 unlock_map(struct faultstate *fs)
185 if (fs->lookup_still_valid && fs->map) {
186 vm_map_lookup_done(fs->map, fs->entry, 0);
187 fs->lookup_still_valid = FALSE;
192 * Clean up after a successful call to vm_fault_object() so another call
193 * to vm_fault_object() can be made.
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 _unlock_things(struct faultstate *fs, int dealloc)
214 _cleanup_successful_fault(fs, 0);
216 /*vm_object_deallocate(fs->first_object);*/
217 /*fs->first_object = NULL; drop used later on */
220 if (fs->vp != NULL) {
226 #define unlock_things(fs) _unlock_things(fs, 0)
227 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
228 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
233 * Determine if the pager for the current object *might* contain the page.
235 * We only need to try the pager if this is not a default object (default
236 * objects are zero-fill and have no real pager), and if we are not taking
237 * a wiring fault or if the FS entry is wired.
239 #define TRYPAGER(fs) \
240 (fs->object->type != OBJT_DEFAULT && \
241 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
246 * Handle a page fault occuring at the given address, requiring the given
247 * permissions, in the map specified. If successful, the page is inserted
248 * into the associated physical map.
250 * NOTE: The given address should be truncated to the proper page address.
252 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
253 * a standard error specifying why the fault is fatal is returned.
255 * The map in question must be referenced, and remains so.
256 * The caller may hold no locks.
257 * No other requirements.
260 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
263 vm_pindex_t first_pindex;
264 struct faultstate fs;
268 vm_page_pcpu_cache();
270 fs.fault_flags = fault_flags;
274 if ((lp = curthread->td_lwp) != NULL)
275 lp->lwp_flags |= LWP_PAGING;
277 lwkt_gettoken(&map->token);
281 * Find the vm_map_entry representing the backing store and resolve
282 * the top level object and page index. This may have the side
283 * effect of executing a copy-on-write on the map entry and/or
284 * creating a shadow object, but will not COW any actual VM pages.
286 * On success fs.map is left read-locked and various other fields
287 * are initialized but not otherwise referenced or locked.
289 * NOTE! vm_map_lookup will try to upgrade the fault_type to
290 * VM_FAULT_WRITE if the map entry is a virtual page table and also
291 * writable, so we can set the 'A'accessed bit in the virtual page
295 result = vm_map_lookup(&fs.map, vaddr, fault_type,
296 &fs.entry, &fs.first_object,
297 &first_pindex, &fs.first_prot, &fs.wired);
300 * If the lookup failed or the map protections are incompatible,
301 * the fault generally fails. However, if the caller is trying
302 * to do a user wiring we have more work to do.
304 if (result != KERN_SUCCESS) {
305 if (result != KERN_PROTECTION_FAILURE ||
306 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
308 if (result == KERN_INVALID_ADDRESS && growstack &&
309 map != &kernel_map && curproc != NULL) {
310 result = vm_map_growstack(curproc, vaddr);
311 if (result == KERN_SUCCESS) {
315 result = KERN_FAILURE;
321 * If we are user-wiring a r/w segment, and it is COW, then
322 * we need to do the COW operation. Note that we don't
323 * currently COW RO sections now, because it is NOT desirable
324 * to COW .text. We simply keep .text from ever being COW'ed
325 * and take the heat that one cannot debug wired .text sections.
327 result = vm_map_lookup(&fs.map, vaddr,
328 VM_PROT_READ|VM_PROT_WRITE|
329 VM_PROT_OVERRIDE_WRITE,
330 &fs.entry, &fs.first_object,
331 &first_pindex, &fs.first_prot,
333 if (result != KERN_SUCCESS) {
334 result = KERN_FAILURE;
339 * If we don't COW now, on a user wire, the user will never
340 * be able to write to the mapping. If we don't make this
341 * restriction, the bookkeeping would be nearly impossible.
343 * XXX We have a shared lock, this will have a MP race but
344 * I don't see how it can hurt anything.
346 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
347 fs.entry->max_protection &= ~VM_PROT_WRITE;
351 * fs.map is read-locked
353 * Misc checks. Save the map generation number to detect races.
355 fs.map_generation = fs.map->timestamp;
357 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) {
358 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
359 panic("vm_fault: fault on nofault entry, addr: %p",
362 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) &&
363 vaddr >= fs.entry->start &&
364 vaddr < fs.entry->start + PAGE_SIZE) {
365 panic("vm_fault: fault on stack guard, addr: %p",
371 * A system map entry may return a NULL object. No object means
372 * no pager means an unrecoverable kernel fault.
374 if (fs.first_object == NULL) {
375 panic("vm_fault: unrecoverable fault at %p in entry %p",
376 (void *)vaddr, fs.entry);
380 * Attempt to shortcut the fault if the lookup returns a
381 * terminal object and the page is present. This allows us
382 * to obtain a shared token on the object instead of an exclusive
383 * token, which theoretically should allow concurrent faults.
385 * We cannot acquire a shared token on kernel_map, at least not
386 * on i386, because the i386 pmap code uses the kernel_object for
387 * its page table page management, resulting in a shared->exclusive
388 * sequence which will deadlock. This will not happen normally
389 * anyway, except on well cached pageable kmem (like pipe buffers),
390 * so it should not impact performance.
392 if (vm_shared_fault &&
393 fs.first_object->backing_object == NULL &&
394 fs.entry->maptype == VM_MAPTYPE_NORMAL &&
395 fs.map != &kernel_map) {
397 vm_object_hold_shared(fs.first_object);
398 /*fs.vp = vnode_pager_lock(fs.first_object);*/
399 fs.m = vm_page_lookup_busy_try(fs.first_object,
402 if (error == 0 && fs.m) {
404 * Activate the page and figure out if we can
405 * short-cut a quick mapping.
407 * WARNING! We cannot call swap_pager_unswapped()
408 * with a shared token! Note that we
409 * have to test fs.first_prot here.
411 vm_page_activate(fs.m);
412 if (fs.m->valid == VM_PAGE_BITS_ALL &&
413 ((fs.m->flags & PG_SWAPPED) == 0 ||
414 (fs.first_prot & VM_PROT_WRITE) == 0 ||
415 (fs.fault_flags & VM_FAULT_DIRTY) == 0)) {
416 fs.lookup_still_valid = TRUE;
418 fs.object = fs.first_object;
419 fs.prot = fs.first_prot;
421 fault_type = fs.first_prot;
422 if (fs.prot & VM_PROT_WRITE) {
423 vm_object_set_writeable_dirty(
425 vm_set_nosync(fs.m, fs.entry);
426 if (fs.fault_flags & VM_FAULT_DIRTY) {
429 swap_pager_unswapped(fs.m);
432 result = KERN_SUCCESS;
433 fault_flags |= VM_FAULT_BURST_QUICK;
434 fault_flags &= ~VM_FAULT_BURST;
438 vm_page_wakeup(fs.m);
441 vm_object_drop(fs.first_object); /* XXX drop on shared tok?*/
446 * Bump the paging-in-progress count to prevent size changes (e.g.
447 * truncation operations) during I/O. This must be done after
448 * obtaining the vnode lock in order to avoid possible deadlocks.
450 vm_object_hold(fs.first_object);
452 fs.vp = vnode_pager_lock(fs.first_object);
454 fs.lookup_still_valid = TRUE;
456 fs.object = fs.first_object; /* so unlock_and_deallocate works */
460 * If the entry is wired we cannot change the page protection.
463 fault_type = fs.first_prot;
466 * The page we want is at (first_object, first_pindex), but if the
467 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
468 * page table to figure out the actual pindex.
470 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
473 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
474 result = vm_fault_vpagetable(&fs, &first_pindex,
475 fs.entry->aux.master_pde,
477 if (result == KERN_TRY_AGAIN) {
478 vm_object_drop(fs.first_object);
481 if (result != KERN_SUCCESS)
486 * Now we have the actual (object, pindex), fault in the page. If
487 * vm_fault_object() fails it will unlock and deallocate the FS
488 * data. If it succeeds everything remains locked and fs->object
489 * will have an additional PIP count if it is not equal to
492 * vm_fault_object will set fs->prot for the pmap operation. It is
493 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
494 * page can be safely written. However, it will force a read-only
495 * mapping for a read fault if the memory is managed by a virtual
498 * If the fault code uses the shared object lock shortcut
499 * we must not try to burst (we can't allocate VM pages).
501 result = vm_fault_object(&fs, first_pindex, fault_type);
503 fault_flags &= ~VM_FAULT_BURST;
505 if (result == KERN_TRY_AGAIN) {
506 vm_object_drop(fs.first_object);
509 if (result != KERN_SUCCESS)
514 * On success vm_fault_object() does not unlock or deallocate, and fs.m
515 * will contain a busied page.
517 * Enter the page into the pmap and do pmap-related adjustments.
519 vm_page_flag_set(fs.m, PG_REFERENCED);
520 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, fs.entry);
521 mycpu->gd_cnt.v_vm_faults++;
522 if (curthread->td_lwp)
523 ++curthread->td_lwp->lwp_ru.ru_minflt;
525 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */
526 KKASSERT(fs.m->flags & PG_BUSY);
529 * If the page is not wired down, then put it where the pageout daemon
532 if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
536 vm_page_unwire(fs.m, 1);
538 vm_page_activate(fs.m);
540 vm_page_wakeup(fs.m);
543 * Burst in a few more pages if possible. The fs.map should still
544 * be locked. To avoid interlocking against a vnode->getblk
545 * operation we had to be sure to unbusy our primary vm_page above
548 if (fault_flags & VM_FAULT_BURST) {
549 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0
551 vm_prefault(fs.map->pmap, vaddr,
552 fs.entry, fs.prot, fault_flags);
555 if (fault_flags & VM_FAULT_BURST_QUICK) {
556 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0
558 vm_prefault_quick(fs.map->pmap, vaddr,
559 fs.entry, fs.prot, fault_flags);
564 * Unlock everything, and return
568 if (curthread->td_lwp) {
570 curthread->td_lwp->lwp_ru.ru_majflt++;
572 curthread->td_lwp->lwp_ru.ru_minflt++;
576 /*vm_object_deallocate(fs.first_object);*/
578 /*fs.first_object = NULL; must still drop later */
580 result = KERN_SUCCESS;
583 vm_object_drop(fs.first_object);
584 lwkt_reltoken(&map->token);
586 lp->lwp_flags &= ~LWP_PAGING;
591 * Fault in the specified virtual address in the current process map,
592 * returning a held VM page or NULL. See vm_fault_page() for more
598 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
600 struct lwp *lp = curthread->td_lwp;
603 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
604 fault_type, VM_FAULT_NORMAL, errorp);
609 * Fault in the specified virtual address in the specified map, doing all
610 * necessary manipulation of the object store and all necessary I/O. Return
611 * a held VM page or NULL, and set *errorp. The related pmap is not
614 * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
615 * and marked PG_REFERENCED as well.
617 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
618 * error will be returned.
623 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
624 int fault_flags, int *errorp)
626 vm_pindex_t first_pindex;
627 struct faultstate fs;
629 vm_prot_t orig_fault_type = fault_type;
632 fs.fault_flags = fault_flags;
633 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
635 lwkt_gettoken(&map->token);
639 * Find the vm_map_entry representing the backing store and resolve
640 * the top level object and page index. This may have the side
641 * effect of executing a copy-on-write on the map entry and/or
642 * creating a shadow object, but will not COW any actual VM pages.
644 * On success fs.map is left read-locked and various other fields
645 * are initialized but not otherwise referenced or locked.
647 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
648 * if the map entry is a virtual page table and also writable,
649 * so we can set the 'A'accessed bit in the virtual page table entry.
652 result = vm_map_lookup(&fs.map, vaddr, fault_type,
653 &fs.entry, &fs.first_object,
654 &first_pindex, &fs.first_prot, &fs.wired);
656 if (result != KERN_SUCCESS) {
663 * fs.map is read-locked
665 * Misc checks. Save the map generation number to detect races.
667 fs.map_generation = fs.map->timestamp;
669 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
670 panic("vm_fault: fault on nofault entry, addr: %lx",
675 * A system map entry may return a NULL object. No object means
676 * no pager means an unrecoverable kernel fault.
678 if (fs.first_object == NULL) {
679 panic("vm_fault: unrecoverable fault at %p in entry %p",
680 (void *)vaddr, fs.entry);
684 * Make a reference to this object to prevent its disposal while we
685 * are messing with it. Once we have the reference, the map is free
686 * to be diddled. Since objects reference their shadows (and copies),
687 * they will stay around as well.
689 * The reference should also prevent an unexpected collapse of the
690 * parent that might move pages from the current object into the
691 * parent unexpectedly, resulting in corruption.
693 * Bump the paging-in-progress count to prevent size changes (e.g.
694 * truncation operations) during I/O. This must be done after
695 * obtaining the vnode lock in order to avoid possible deadlocks.
697 vm_object_hold(fs.first_object);
698 fs.vp = vnode_pager_lock(fs.first_object);
700 fs.lookup_still_valid = TRUE;
702 fs.object = fs.first_object; /* so unlock_and_deallocate works */
706 * If the entry is wired we cannot change the page protection.
709 fault_type = fs.first_prot;
712 * The page we want is at (first_object, first_pindex), but if the
713 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
714 * page table to figure out the actual pindex.
716 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
719 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
720 result = vm_fault_vpagetable(&fs, &first_pindex,
721 fs.entry->aux.master_pde,
723 if (result == KERN_TRY_AGAIN) {
724 vm_object_drop(fs.first_object);
727 if (result != KERN_SUCCESS) {
735 * Now we have the actual (object, pindex), fault in the page. If
736 * vm_fault_object() fails it will unlock and deallocate the FS
737 * data. If it succeeds everything remains locked and fs->object
738 * will have an additinal PIP count if it is not equal to
742 result = vm_fault_object(&fs, first_pindex, fault_type);
744 if (result == KERN_TRY_AGAIN) {
745 vm_object_drop(fs.first_object);
748 if (result != KERN_SUCCESS) {
754 if ((orig_fault_type & VM_PROT_WRITE) &&
755 (fs.prot & VM_PROT_WRITE) == 0) {
756 *errorp = KERN_PROTECTION_FAILURE;
757 unlock_and_deallocate(&fs);
763 * DO NOT UPDATE THE PMAP!!! This function may be called for
764 * a pmap unrelated to the current process pmap, in which case
765 * the current cpu core will not be listed in the pmap's pm_active
766 * mask. Thus invalidation interlocks will fail to work properly.
768 * (for example, 'ps' uses procfs to read program arguments from
769 * each process's stack).
771 * In addition to the above this function will be called to acquire
772 * a page that might already be faulted in, re-faulting it
773 * continuously is a waste of time.
775 * XXX could this have been the cause of our random seg-fault
776 * issues? procfs accesses user stacks.
778 vm_page_flag_set(fs.m, PG_REFERENCED);
780 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL);
781 mycpu->gd_cnt.v_vm_faults++;
782 if (curthread->td_lwp)
783 ++curthread->td_lwp->lwp_ru.ru_minflt;
787 * On success vm_fault_object() does not unlock or deallocate, and fs.m
788 * will contain a busied page. So we must unlock here after having
789 * messed with the pmap.
794 * Return a held page. We are not doing any pmap manipulation so do
795 * not set PG_MAPPED. However, adjust the page flags according to
796 * the fault type because the caller may not use a managed pmapping
797 * (so we don't want to lose the fact that the page will be dirtied
798 * if a write fault was specified).
801 vm_page_activate(fs.m);
802 if (fault_type & VM_PROT_WRITE)
805 if (curthread->td_lwp) {
807 curthread->td_lwp->lwp_ru.ru_majflt++;
809 curthread->td_lwp->lwp_ru.ru_minflt++;
814 * Unlock everything, and return the held page.
816 vm_page_wakeup(fs.m);
817 /*vm_object_deallocate(fs.first_object);*/
818 /*fs.first_object = NULL; */
823 vm_object_drop(fs.first_object);
824 lwkt_reltoken(&map->token);
829 * Fault in the specified (object,offset), dirty the returned page as
830 * needed. If the requested fault_type cannot be done NULL and an
833 * A held (but not busied) page is returned.
838 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
839 vm_prot_t fault_type, int fault_flags,
840 int shared, int *errorp)
843 vm_pindex_t first_pindex;
844 struct faultstate fs;
845 struct vm_map_entry entry;
847 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
848 bzero(&entry, sizeof(entry));
849 entry.object.vm_object = object;
850 entry.maptype = VM_MAPTYPE_NORMAL;
851 entry.protection = entry.max_protection = fault_type;
854 fs.fault_flags = fault_flags;
856 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
860 fs.first_object = object;
861 first_pindex = OFF_TO_IDX(offset);
863 fs.first_prot = fault_type;
866 /*fs.map_generation = 0; unused */
869 * Make a reference to this object to prevent its disposal while we
870 * are messing with it. Once we have the reference, the map is free
871 * to be diddled. Since objects reference their shadows (and copies),
872 * they will stay around as well.
874 * The reference should also prevent an unexpected collapse of the
875 * parent that might move pages from the current object into the
876 * parent unexpectedly, resulting in corruption.
878 * Bump the paging-in-progress count to prevent size changes (e.g.
879 * truncation operations) during I/O. This must be done after
880 * obtaining the vnode lock in order to avoid possible deadlocks.
882 fs.vp = vnode_pager_lock(fs.first_object);
884 fs.lookup_still_valid = TRUE;
886 fs.object = fs.first_object; /* so unlock_and_deallocate works */
889 /* XXX future - ability to operate on VM object using vpagetable */
890 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
891 result = vm_fault_vpagetable(&fs, &first_pindex,
892 fs.entry->aux.master_pde,
894 if (result == KERN_TRY_AGAIN)
896 if (result != KERN_SUCCESS) {
904 * Now we have the actual (object, pindex), fault in the page. If
905 * vm_fault_object() fails it will unlock and deallocate the FS
906 * data. If it succeeds everything remains locked and fs->object
907 * will have an additinal PIP count if it is not equal to
910 result = vm_fault_object(&fs, first_pindex, fault_type);
912 if (result == KERN_TRY_AGAIN)
914 if (result != KERN_SUCCESS) {
919 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
920 *errorp = KERN_PROTECTION_FAILURE;
921 unlock_and_deallocate(&fs);
926 * On success vm_fault_object() does not unlock or deallocate, so we
927 * do it here. Note that the returned fs.m will be busied.
932 * Return a held page. We are not doing any pmap manipulation so do
933 * not set PG_MAPPED. However, adjust the page flags according to
934 * the fault type because the caller may not use a managed pmapping
935 * (so we don't want to lose the fact that the page will be dirtied
936 * if a write fault was specified).
939 vm_page_activate(fs.m);
940 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY))
942 if (fault_flags & VM_FAULT_UNSWAP)
943 swap_pager_unswapped(fs.m);
946 * Indicate that the page was accessed.
948 vm_page_flag_set(fs.m, PG_REFERENCED);
950 if (curthread->td_lwp) {
952 curthread->td_lwp->lwp_ru.ru_majflt++;
954 curthread->td_lwp->lwp_ru.ru_minflt++;
959 * Unlock everything, and return the held page.
961 vm_page_wakeup(fs.m);
962 /*vm_object_deallocate(fs.first_object);*/
963 /*fs.first_object = NULL; */
970 * Translate the virtual page number (first_pindex) that is relative
971 * to the address space into a logical page number that is relative to the
972 * backing object. Use the virtual page table pointed to by (vpte).
974 * This implements an N-level page table. Any level can terminate the
975 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
976 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
980 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
981 vpte_t vpte, int fault_type)
984 struct lwbuf lwb_cache;
985 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */
986 int result = KERN_SUCCESS;
989 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
992 * We cannot proceed if the vpte is not valid, not readable
993 * for a read fault, or not writable for a write fault.
995 if ((vpte & VPTE_V) == 0) {
996 unlock_and_deallocate(fs);
997 return (KERN_FAILURE);
999 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
1000 unlock_and_deallocate(fs);
1001 return (KERN_FAILURE);
1003 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
1004 unlock_and_deallocate(fs);
1005 return (KERN_FAILURE);
1007 if ((vpte & VPTE_PS) || vshift == 0)
1009 KKASSERT(vshift >= VPTE_PAGE_BITS);
1012 * Get the page table page. Nominally we only read the page
1013 * table, but since we are actively setting VPTE_M and VPTE_A,
1014 * tell vm_fault_object() that we are writing it.
1016 * There is currently no real need to optimize this.
1018 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT,
1019 VM_PROT_READ|VM_PROT_WRITE);
1020 if (result != KERN_SUCCESS)
1024 * Process the returned fs.m and look up the page table
1025 * entry in the page table page.
1027 vshift -= VPTE_PAGE_BITS;
1028 lwb = lwbuf_alloc(fs->m, &lwb_cache);
1029 ptep = ((vpte_t *)lwbuf_kva(lwb) +
1030 ((*pindex >> vshift) & VPTE_PAGE_MASK));
1034 * Page table write-back. If the vpte is valid for the
1035 * requested operation, do a write-back to the page table.
1037 * XXX VPTE_M is not set properly for page directory pages.
1038 * It doesn't get set in the page directory if the page table
1039 * is modified during a read access.
1041 vm_page_activate(fs->m);
1042 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
1044 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
1045 atomic_set_long(ptep, VPTE_M | VPTE_A);
1046 vm_page_dirty(fs->m);
1049 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
1051 if ((vpte & VPTE_A) == 0) {
1052 atomic_set_long(ptep, VPTE_A);
1053 vm_page_dirty(fs->m);
1057 vm_page_flag_set(fs->m, PG_REFERENCED);
1058 vm_page_wakeup(fs->m);
1060 cleanup_successful_fault(fs);
1063 * Combine remaining address bits with the vpte.
1065 /* JG how many bits from each? */
1066 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) +
1067 (*pindex & ((1L << vshift) - 1));
1068 return (KERN_SUCCESS);
1073 * This is the core of the vm_fault code.
1075 * Do all operations required to fault-in (fs.first_object, pindex). Run
1076 * through the shadow chain as necessary and do required COW or virtual
1077 * copy operations. The caller has already fully resolved the vm_map_entry
1078 * and, if appropriate, has created a copy-on-write layer. All we need to
1079 * do is iterate the object chain.
1081 * On failure (fs) is unlocked and deallocated and the caller may return or
1082 * retry depending on the failure code. On success (fs) is NOT unlocked or
1083 * deallocated, fs.m will contained a resolved, busied page, and fs.object
1084 * will have an additional PIP count if it is not equal to fs.first_object.
1086 * fs->first_object must be held on call.
1090 vm_fault_object(struct faultstate *fs,
1091 vm_pindex_t first_pindex, vm_prot_t fault_type)
1093 vm_object_t next_object;
1097 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object));
1098 fs->prot = fs->first_prot;
1099 fs->object = fs->first_object;
1100 pindex = first_pindex;
1102 vm_object_chain_acquire(fs->first_object);
1103 vm_object_pip_add(fs->first_object, 1);
1106 * If a read fault occurs we try to make the page writable if
1107 * possible. There are three cases where we cannot make the
1108 * page mapping writable:
1110 * (1) The mapping is read-only or the VM object is read-only,
1111 * fs->prot above will simply not have VM_PROT_WRITE set.
1113 * (2) If the mapping is a virtual page table we need to be able
1114 * to detect writes so we can set VPTE_M in the virtual page
1117 * (3) If the VM page is read-only or copy-on-write, upgrading would
1118 * just result in an unnecessary COW fault.
1120 * VM_PROT_VPAGED is set if faulting via a virtual page table and
1121 * causes adjustments to the 'M'odify bit to also turn off write
1122 * access to force a re-fault.
1124 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
1125 if ((fault_type & VM_PROT_WRITE) == 0)
1126 fs->prot &= ~VM_PROT_WRITE;
1129 /* vm_object_hold(fs->object); implied b/c object == first_object */
1133 * The entire backing chain from first_object to object
1134 * inclusive is chainlocked.
1136 * If the object is dead, we stop here
1138 * vm_shared_fault (fs->shared != 0) case: nothing special.
1140 if (fs->object->flags & OBJ_DEAD) {
1141 vm_object_pip_wakeup(fs->first_object);
1142 vm_object_chain_release_all(fs->first_object,
1144 if (fs->object != fs->first_object)
1145 vm_object_drop(fs->object);
1146 unlock_and_deallocate(fs);
1147 return (KERN_PROTECTION_FAILURE);
1151 * See if the page is resident. Wait/Retry if the page is
1152 * busy (lots of stuff may have changed so we can't continue
1155 * We can theoretically allow the soft-busy case on a read
1156 * fault if the page is marked valid, but since such
1157 * pages are typically already pmap'd, putting that
1158 * special case in might be more effort then it is
1159 * worth. We cannot under any circumstances mess
1160 * around with a vm_page_t->busy page except, perhaps,
1163 * vm_shared_fault (fs->shared != 0) case:
1164 * error nothing special
1165 * fs->m relock excl if I/O needed
1168 fs->m = vm_page_lookup_busy_try(fs->object, pindex,
1171 vm_object_pip_wakeup(fs->first_object);
1172 vm_object_chain_release_all(fs->first_object,
1174 if (fs->object != fs->first_object)
1175 vm_object_drop(fs->object);
1177 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
1178 mycpu->gd_cnt.v_intrans++;
1179 /*vm_object_deallocate(fs->first_object);*/
1180 /*fs->first_object = NULL;*/
1182 return (KERN_TRY_AGAIN);
1186 * The page is busied for us.
1188 * If reactivating a page from PQ_CACHE we may have
1191 int queue = fs->m->queue;
1192 vm_page_unqueue_nowakeup(fs->m);
1194 if ((queue - fs->m->pc) == PQ_CACHE &&
1195 vm_page_count_severe()) {
1196 vm_page_activate(fs->m);
1197 vm_page_wakeup(fs->m);
1199 vm_object_pip_wakeup(fs->first_object);
1200 vm_object_chain_release_all(fs->first_object,
1202 if (fs->object != fs->first_object)
1203 vm_object_drop(fs->object);
1204 unlock_and_deallocate(fs);
1206 return (KERN_TRY_AGAIN);
1210 * If it still isn't completely valid (readable),
1211 * or if a read-ahead-mark is set on the VM page,
1212 * jump to readrest, else we found the page and
1215 * We can release the spl once we have marked the
1218 if (fs->m->object != &kernel_object) {
1219 if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1222 vm_object_drop(fs->object);
1223 vm_object_hold(fs->object);
1228 if (fs->m->flags & PG_RAM) {
1231 vm_page_flag_clear(fs->m, PG_RAM);
1233 vm_object_drop(fs->object);
1234 vm_object_hold(fs->object);
1240 break; /* break to PAGE HAS BEEN FOUND */
1244 vm_object_drop(fs->object);
1245 vm_object_hold(fs->object);
1250 * Page is not resident, If this is the search termination
1251 * or the pager might contain the page, allocate a new page.
1253 if (TRYPAGER(fs) || fs->object == fs->first_object) {
1255 * If the page is beyond the object size we fail
1257 if (pindex >= fs->object->size) {
1258 vm_object_pip_wakeup(fs->first_object);
1259 vm_object_chain_release_all(fs->first_object,
1261 if (fs->object != fs->first_object)
1262 vm_object_drop(fs->object);
1263 unlock_and_deallocate(fs);
1264 return (KERN_PROTECTION_FAILURE);
1268 * Allocate a new page for this object/offset pair.
1270 * It is possible for the allocation to race, so
1274 if (!vm_page_count_severe()) {
1275 fs->m = vm_page_alloc(fs->object, pindex,
1276 ((fs->vp || fs->object->backing_object) ?
1277 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL :
1278 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1279 VM_ALLOC_USE_GD | VM_ALLOC_ZERO));
1281 if (fs->m == NULL) {
1282 vm_object_pip_wakeup(fs->first_object);
1283 vm_object_chain_release_all(fs->first_object,
1285 if (fs->object != fs->first_object)
1286 vm_object_drop(fs->object);
1287 unlock_and_deallocate(fs);
1289 return (KERN_TRY_AGAIN);
1293 * Fall through to readrest. We have a new page which
1294 * will have to be paged (since m->valid will be 0).
1300 * We have found an invalid or partially valid page, a
1301 * page with a read-ahead mark which might be partially or
1302 * fully valid (and maybe dirty too), or we have allocated
1305 * Attempt to fault-in the page if there is a chance that the
1306 * pager has it, and potentially fault in additional pages
1309 * If TRYPAGER is true then fs.m will be non-NULL and busied
1315 u_char behavior = vm_map_entry_behavior(fs->entry);
1317 if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1324 * If sequential access is detected then attempt
1325 * to deactivate/cache pages behind the scan to
1326 * prevent resource hogging.
1328 * Use of PG_RAM to detect sequential access
1329 * also simulates multi-zone sequential access
1330 * detection for free.
1332 * NOTE: Partially valid dirty pages cannot be
1333 * deactivated without causing NFS picemeal
1336 if ((fs->first_object->type != OBJT_DEVICE) &&
1337 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1338 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1339 (fs->m->flags & PG_RAM)))
1341 vm_pindex_t scan_pindex;
1342 int scan_count = 16;
1344 if (first_pindex < 16) {
1348 scan_pindex = first_pindex - 16;
1349 if (scan_pindex < 16)
1350 scan_count = scan_pindex;
1355 while (scan_count) {
1358 mt = vm_page_lookup(fs->first_object,
1362 if (vm_page_busy_try(mt, TRUE))
1365 if (mt->valid != VM_PAGE_BITS_ALL) {
1370 (PG_FICTITIOUS | PG_UNMANAGED |
1378 vm_page_test_dirty(mt);
1382 vm_page_deactivate(mt);
1397 * Avoid deadlocking against the map when doing I/O.
1398 * fs.object and the page is PG_BUSY'd.
1400 * NOTE: Once unlocked, fs->entry can become stale
1401 * so this will NULL it out.
1403 * NOTE: fs->entry is invalid until we relock the
1404 * map and verify that the timestamp has not
1410 * Acquire the page data. We still hold a ref on
1411 * fs.object and the page has been PG_BUSY's.
1413 * The pager may replace the page (for example, in
1414 * order to enter a fictitious page into the
1415 * object). If it does so it is responsible for
1416 * cleaning up the passed page and properly setting
1417 * the new page PG_BUSY.
1419 * If we got here through a PG_RAM read-ahead
1420 * mark the page may be partially dirty and thus
1421 * not freeable. Don't bother checking to see
1422 * if the pager has the page because we can't free
1423 * it anyway. We have to depend on the get_page
1424 * operation filling in any gaps whether there is
1425 * backing store or not.
1427 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
1429 if (rv == VM_PAGER_OK) {
1431 * Relookup in case pager changed page. Pager
1432 * is responsible for disposition of old page
1435 * XXX other code segments do relookups too.
1436 * It's a bad abstraction that needs to be
1439 fs->m = vm_page_lookup(fs->object, pindex);
1440 if (fs->m == NULL) {
1441 vm_object_pip_wakeup(fs->first_object);
1442 vm_object_chain_release_all(
1443 fs->first_object, fs->object);
1444 if (fs->object != fs->first_object)
1445 vm_object_drop(fs->object);
1446 unlock_and_deallocate(fs);
1447 return (KERN_TRY_AGAIN);
1451 break; /* break to PAGE HAS BEEN FOUND */
1455 * Remove the bogus page (which does not exist at this
1456 * object/offset); before doing so, we must get back
1457 * our object lock to preserve our invariant.
1459 * Also wake up any other process that may want to bring
1462 * If this is the top-level object, we must leave the
1463 * busy page to prevent another process from rushing
1464 * past us, and inserting the page in that object at
1465 * the same time that we are.
1467 if (rv == VM_PAGER_ERROR) {
1469 kprintf("vm_fault: pager read error, "
1474 kprintf("vm_fault: pager read error, "
1482 * Data outside the range of the pager or an I/O error
1484 * The page may have been wired during the pagein,
1485 * e.g. by the buffer cache, and cannot simply be
1486 * freed. Call vnode_pager_freepage() to deal with it.
1489 * XXX - the check for kernel_map is a kludge to work
1490 * around having the machine panic on a kernel space
1491 * fault w/ I/O error.
1493 if (((fs->map != &kernel_map) &&
1494 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
1495 vnode_pager_freepage(fs->m);
1497 vm_object_pip_wakeup(fs->first_object);
1498 vm_object_chain_release_all(fs->first_object,
1500 if (fs->object != fs->first_object)
1501 vm_object_drop(fs->object);
1502 unlock_and_deallocate(fs);
1503 if (rv == VM_PAGER_ERROR)
1504 return (KERN_FAILURE);
1506 return (KERN_PROTECTION_FAILURE);
1509 if (fs->object != fs->first_object) {
1510 vnode_pager_freepage(fs->m);
1513 * XXX - we cannot just fall out at this
1514 * point, m has been freed and is invalid!
1520 * We get here if the object has a default pager (or unwiring)
1521 * or the pager doesn't have the page.
1523 if (fs->object == fs->first_object)
1524 fs->first_m = fs->m;
1527 * Move on to the next object. The chain lock should prevent
1528 * the backing_object from getting ripped out from under us.
1530 * vm_shared_fault case:
1532 * If the next object is the last object and
1533 * vnode-backed (thus possibly shared), we can try a
1534 * shared object lock. There is no 'chain' for this
1535 * last object if vnode-backed (otherwise we would
1536 * need an exclusive lock).
1538 * fs->shared mode is very fragile and only works
1539 * under certain specific conditions, and is only
1540 * handled for those conditions in our loop. Essentially
1541 * it is designed only to be able to 'dip into' the
1542 * vnode's object and extract an already-cached page.
1545 if ((next_object = fs->object->backing_object) != NULL) {
1546 fs->shared = vm_object_hold_maybe_shared(next_object);
1547 vm_object_chain_acquire(next_object);
1548 KKASSERT(next_object == fs->object->backing_object);
1549 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1552 if (next_object == NULL) {
1554 * If there's no object left, fill the page in the top
1555 * object with zeros.
1557 if (fs->object != fs->first_object) {
1558 if (fs->first_object->backing_object !=
1560 vm_object_hold(fs->first_object->backing_object);
1562 vm_object_chain_release_all(
1563 fs->first_object->backing_object,
1565 if (fs->first_object->backing_object !=
1567 vm_object_drop(fs->first_object->backing_object);
1569 vm_object_pip_wakeup(fs->object);
1570 vm_object_drop(fs->object);
1571 fs->object = fs->first_object;
1572 pindex = first_pindex;
1573 fs->m = fs->first_m;
1578 * Zero the page if necessary and mark it valid.
1580 if ((fs->m->flags & PG_ZERO) == 0) {
1581 vm_page_zero_fill(fs->m);
1584 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m));
1586 vm_page_flag_clear(fs->m, PG_ZERO);
1587 mycpu->gd_cnt.v_ozfod++;
1589 mycpu->gd_cnt.v_zfod++;
1590 fs->m->valid = VM_PAGE_BITS_ALL;
1591 break; /* break to PAGE HAS BEEN FOUND */
1593 if (fs->object != fs->first_object) {
1594 vm_object_pip_wakeup(fs->object);
1595 vm_object_lock_swap();
1596 vm_object_drop(fs->object);
1598 KASSERT(fs->object != next_object,
1599 ("object loop %p", next_object));
1600 fs->object = next_object;
1601 vm_object_pip_add(fs->object, 1);
1605 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1608 * object still held.
1610 * If the page is being written, but isn't already owned by the
1611 * top-level object, we have to copy it into a new page owned by the
1614 KASSERT((fs->m->flags & PG_BUSY) != 0,
1615 ("vm_fault: not busy after main loop"));
1617 if (fs->object != fs->first_object) {
1619 * We only really need to copy if we want to write it.
1621 if (fault_type & VM_PROT_WRITE) {
1623 * This allows pages to be virtually copied from a
1624 * backing_object into the first_object, where the
1625 * backing object has no other refs to it, and cannot
1626 * gain any more refs. Instead of a bcopy, we just
1627 * move the page from the backing object to the
1628 * first object. Note that we must mark the page
1629 * dirty in the first object so that it will go out
1630 * to swap when needed.
1634 * Map, if present, has not changed
1637 fs->map_generation == fs->map->timestamp) &&
1639 * Only one shadow object
1641 (fs->object->shadow_count == 1) &&
1643 * No COW refs, except us
1645 (fs->object->ref_count == 1) &&
1647 * No one else can look this object up
1649 (fs->object->handle == NULL) &&
1651 * No other ways to look the object up
1653 ((fs->object->type == OBJT_DEFAULT) ||
1654 (fs->object->type == OBJT_SWAP)) &&
1656 * We don't chase down the shadow chain
1658 (fs->object == fs->first_object->backing_object) &&
1661 * grab the lock if we need to
1663 (fs->lookup_still_valid ||
1665 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1668 * (first_m) and (m) are both busied. We have
1669 * move (m) into (first_m)'s object/pindex
1670 * in an atomic fashion, then free (first_m).
1672 * first_object is held so second remove
1673 * followed by the rename should wind
1674 * up being atomic. vm_page_free() might
1675 * block so we don't do it until after the
1678 fs->lookup_still_valid = 1;
1679 vm_page_protect(fs->first_m, VM_PROT_NONE);
1680 vm_page_remove(fs->first_m);
1681 vm_page_rename(fs->m, fs->first_object,
1683 vm_page_free(fs->first_m);
1684 fs->first_m = fs->m;
1686 mycpu->gd_cnt.v_cow_optim++;
1689 * Oh, well, lets copy it.
1691 * Why are we unmapping the original page
1692 * here? Well, in short, not all accessors
1693 * of user memory go through the pmap. The
1694 * procfs code doesn't have access user memory
1695 * via a local pmap, so vm_fault_page*()
1696 * can't call pmap_enter(). And the umtx*()
1697 * code may modify the COW'd page via a DMAP
1698 * or kernel mapping and not via the pmap,
1699 * leaving the original page still mapped
1700 * read-only into the pmap.
1702 * So we have to remove the page from at
1703 * least the current pmap if it is in it.
1704 * Just remove it from all pmaps.
1706 vm_page_copy(fs->m, fs->first_m);
1707 vm_page_protect(fs->m, VM_PROT_NONE);
1708 vm_page_event(fs->m, VMEVENT_COW);
1713 * We no longer need the old page or object.
1719 * We intend to revert to first_object, undo the
1720 * chain lock through to that.
1722 if (fs->first_object->backing_object != fs->object)
1723 vm_object_hold(fs->first_object->backing_object);
1724 vm_object_chain_release_all(
1725 fs->first_object->backing_object,
1727 if (fs->first_object->backing_object != fs->object)
1728 vm_object_drop(fs->first_object->backing_object);
1731 * fs->object != fs->first_object due to above
1734 vm_object_pip_wakeup(fs->object);
1735 vm_object_drop(fs->object);
1738 * Only use the new page below...
1741 mycpu->gd_cnt.v_cow_faults++;
1742 fs->m = fs->first_m;
1743 fs->object = fs->first_object;
1744 pindex = first_pindex;
1747 * If it wasn't a write fault avoid having to copy
1748 * the page by mapping it read-only.
1750 fs->prot &= ~VM_PROT_WRITE;
1755 * Relock the map if necessary, then check the generation count.
1756 * relock_map() will update fs->timestamp to account for the
1757 * relocking if necessary.
1759 * If the count has changed after relocking then all sorts of
1760 * crap may have happened and we have to retry.
1762 * NOTE: The relock_map() can fail due to a deadlock against
1763 * the vm_page we are holding BUSY.
1765 if (fs->lookup_still_valid == FALSE && fs->map) {
1766 if (relock_map(fs) ||
1767 fs->map->timestamp != fs->map_generation) {
1769 vm_object_pip_wakeup(fs->first_object);
1770 vm_object_chain_release_all(fs->first_object,
1772 if (fs->object != fs->first_object)
1773 vm_object_drop(fs->object);
1774 unlock_and_deallocate(fs);
1775 return (KERN_TRY_AGAIN);
1780 * If the fault is a write, we know that this page is being
1781 * written NOW so dirty it explicitly to save on pmap_is_modified()
1784 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1785 * if the page is already dirty to prevent data written with
1786 * the expectation of being synced from not being synced.
1787 * Likewise if this entry does not request NOSYNC then make
1788 * sure the page isn't marked NOSYNC. Applications sharing
1789 * data should use the same flags to avoid ping ponging.
1791 * Also tell the backing pager, if any, that it should remove
1792 * any swap backing since the page is now dirty.
1794 vm_page_activate(fs->m);
1795 if (fs->prot & VM_PROT_WRITE) {
1796 vm_object_set_writeable_dirty(fs->m->object);
1797 vm_set_nosync(fs->m, fs->entry);
1798 if (fs->fault_flags & VM_FAULT_DIRTY) {
1799 vm_page_dirty(fs->m);
1800 swap_pager_unswapped(fs->m);
1804 vm_object_pip_wakeup(fs->first_object);
1805 vm_object_chain_release_all(fs->first_object, fs->object);
1806 if (fs->object != fs->first_object)
1807 vm_object_drop(fs->object);
1810 * Page had better still be busy. We are still locked up and
1811 * fs->object will have another PIP reference if it is not equal
1812 * to fs->first_object.
1814 KASSERT(fs->m->flags & PG_BUSY,
1815 ("vm_fault: page %p not busy!", fs->m));
1818 * Sanity check: page must be completely valid or it is not fit to
1819 * map into user space. vm_pager_get_pages() ensures this.
1821 if (fs->m->valid != VM_PAGE_BITS_ALL) {
1822 vm_page_zero_invalid(fs->m, TRUE);
1823 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1825 vm_page_flag_clear(fs->m, PG_ZERO);
1827 return (KERN_SUCCESS);
1831 * Wire down a range of virtual addresses in a map. The entry in question
1832 * should be marked in-transition and the map must be locked. We must
1833 * release the map temporarily while faulting-in the page to avoid a
1834 * deadlock. Note that the entry may be clipped while we are blocked but
1835 * will never be freed.
1840 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1842 boolean_t fictitious;
1851 lwkt_gettoken(&map->token);
1853 pmap = vm_map_pmap(map);
1854 start = entry->start;
1856 fictitious = entry->object.vm_object &&
1857 (entry->object.vm_object->type == OBJT_DEVICE);
1858 if (entry->eflags & MAP_ENTRY_KSTACK)
1864 * We simulate a fault to get the page and enter it in the physical
1867 for (va = start; va < end; va += PAGE_SIZE) {
1869 rv = vm_fault(map, va, VM_PROT_READ,
1870 VM_FAULT_USER_WIRE);
1872 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1873 VM_FAULT_CHANGE_WIRING);
1876 while (va > start) {
1878 if ((pa = pmap_extract(pmap, va)) == 0)
1880 pmap_change_wiring(pmap, va, FALSE, entry);
1882 m = PHYS_TO_VM_PAGE(pa);
1883 vm_page_busy_wait(m, FALSE, "vmwrpg");
1884 vm_page_unwire(m, 1);
1894 lwkt_reltoken(&map->token);
1899 * Unwire a range of virtual addresses in a map. The map should be
1903 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1905 boolean_t fictitious;
1913 lwkt_gettoken(&map->token);
1915 pmap = vm_map_pmap(map);
1916 start = entry->start;
1918 fictitious = entry->object.vm_object &&
1919 (entry->object.vm_object->type == OBJT_DEVICE);
1920 if (entry->eflags & MAP_ENTRY_KSTACK)
1924 * Since the pages are wired down, we must be able to get their
1925 * mappings from the physical map system.
1927 for (va = start; va < end; va += PAGE_SIZE) {
1928 pa = pmap_extract(pmap, va);
1930 pmap_change_wiring(pmap, va, FALSE, entry);
1932 m = PHYS_TO_VM_PAGE(pa);
1933 vm_page_busy_wait(m, FALSE, "vmwupg");
1934 vm_page_unwire(m, 1);
1939 lwkt_reltoken(&map->token);
1943 * Copy all of the pages from a wired-down map entry to another.
1945 * The source and destination maps must be locked for write.
1946 * The source and destination maps token must be held
1947 * The source map entry must be wired down (or be a sharing map
1948 * entry corresponding to a main map entry that is wired down).
1950 * No other requirements.
1952 * XXX do segment optimization
1955 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1956 vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1958 vm_object_t dst_object;
1959 vm_object_t src_object;
1960 vm_ooffset_t dst_offset;
1961 vm_ooffset_t src_offset;
1967 src_object = src_entry->object.vm_object;
1968 src_offset = src_entry->offset;
1971 * Create the top-level object for the destination entry. (Doesn't
1972 * actually shadow anything - we copy the pages directly.)
1974 vm_map_entry_allocate_object(dst_entry);
1975 dst_object = dst_entry->object.vm_object;
1977 prot = dst_entry->max_protection;
1980 * Loop through all of the pages in the entry's range, copying each
1981 * one from the source object (it should be there) to the destination
1984 vm_object_hold(src_object);
1985 vm_object_hold(dst_object);
1986 for (vaddr = dst_entry->start, dst_offset = 0;
1987 vaddr < dst_entry->end;
1988 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1991 * Allocate a page in the destination object
1994 dst_m = vm_page_alloc(dst_object,
1995 OFF_TO_IDX(dst_offset),
1997 if (dst_m == NULL) {
2000 } while (dst_m == NULL);
2003 * Find the page in the source object, and copy it in.
2004 * (Because the source is wired down, the page will be in
2007 src_m = vm_page_lookup(src_object,
2008 OFF_TO_IDX(dst_offset + src_offset));
2010 panic("vm_fault_copy_wired: page missing");
2012 vm_page_copy(src_m, dst_m);
2013 vm_page_event(src_m, VMEVENT_COW);
2016 * Enter it in the pmap...
2019 vm_page_flag_clear(dst_m, PG_ZERO);
2020 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry);
2023 * Mark it no longer busy, and put it on the active list.
2025 vm_page_activate(dst_m);
2026 vm_page_wakeup(dst_m);
2028 vm_object_drop(dst_object);
2029 vm_object_drop(src_object);
2035 * This routine checks around the requested page for other pages that
2036 * might be able to be faulted in. This routine brackets the viable
2037 * pages for the pages to be paged in.
2040 * m, rbehind, rahead
2043 * marray (array of vm_page_t), reqpage (index of requested page)
2046 * number of pages in marray
2049 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
2050 vm_page_t *marray, int *reqpage)
2054 vm_pindex_t pindex, startpindex, endpindex, tpindex;
2056 int cbehind, cahead;
2062 * we don't fault-ahead for device pager
2064 if (object->type == OBJT_DEVICE) {
2071 * if the requested page is not available, then give up now
2073 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
2074 *reqpage = 0; /* not used by caller, fix compiler warn */
2078 if ((cbehind == 0) && (cahead == 0)) {
2084 if (rahead > cahead) {
2088 if (rbehind > cbehind) {
2093 * Do not do any readahead if we have insufficient free memory.
2095 * XXX code was broken disabled before and has instability
2096 * with this conditonal fixed, so shortcut for now.
2098 if (burst_fault == 0 || vm_page_count_severe()) {
2105 * scan backward for the read behind pages -- in memory
2107 * Assume that if the page is not found an interrupt will not
2108 * create it. Theoretically interrupts can only remove (busy)
2109 * pages, not create new associations.
2112 if (rbehind > pindex) {
2116 startpindex = pindex - rbehind;
2119 vm_object_hold(object);
2120 for (tpindex = pindex; tpindex > startpindex; --tpindex) {
2121 if (vm_page_lookup(object, tpindex - 1))
2126 while (tpindex < pindex) {
2127 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2130 for (j = 0; j < i; j++) {
2131 vm_page_free(marray[j]);
2133 vm_object_drop(object);
2142 vm_object_drop(object);
2148 * Assign requested page
2155 * Scan forwards for read-ahead pages
2157 tpindex = pindex + 1;
2158 endpindex = tpindex + rahead;
2159 if (endpindex > object->size)
2160 endpindex = object->size;
2162 vm_object_hold(object);
2163 while (tpindex < endpindex) {
2164 if (vm_page_lookup(object, tpindex))
2166 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM |
2174 vm_object_drop(object);
2182 * vm_prefault() provides a quick way of clustering pagefaults into a
2183 * processes address space. It is a "cousin" of pmap_object_init_pt,
2184 * except it runs at page fault time instead of mmap time.
2186 * vm.fast_fault Enables pre-faulting zero-fill pages
2188 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to
2189 * prefault. Scan stops in either direction when
2190 * a page is found to already exist.
2192 * This code used to be per-platform pmap_prefault(). It is now
2193 * machine-independent and enhanced to also pre-fault zero-fill pages
2194 * (see vm.fast_fault) as well as make them writable, which greatly
2195 * reduces the number of page faults programs incur.
2197 * Application performance when pre-faulting zero-fill pages is heavily
2198 * dependent on the application. Very tiny applications like /bin/echo
2199 * lose a little performance while applications of any appreciable size
2200 * gain performance. Prefaulting multiple pages also reduces SMP
2201 * congestion and can improve SMP performance significantly.
2203 * NOTE! prot may allow writing but this only applies to the top level
2204 * object. If we wind up mapping a page extracted from a backing
2205 * object we have to make sure it is read-only.
2207 * NOTE! The caller has already handled any COW operations on the
2208 * vm_map_entry via the normal fault code. Do NOT call this
2209 * shortcut unless the normal fault code has run on this entry.
2211 * The related map must be locked.
2212 * No other requirements.
2214 static int vm_prefault_pages = 8;
2215 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0,
2216 "Maximum number of pages to pre-fault");
2217 static int vm_fast_fault = 1;
2218 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0,
2219 "Burst fault zero-fill regions");
2222 * Set PG_NOSYNC if the map entry indicates so, but only if the page
2223 * is not already dirty by other means. This will prevent passive
2224 * filesystem syncing as well as 'sync' from writing out the page.
2227 vm_set_nosync(vm_page_t m, vm_map_entry_t entry)
2229 if (entry->eflags & MAP_ENTRY_NOSYNC) {
2231 vm_page_flag_set(m, PG_NOSYNC);
2233 vm_page_flag_clear(m, PG_NOSYNC);
2238 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot,
2254 * Get stable max count value, disabled if set to 0
2256 maxpages = vm_prefault_pages;
2262 * We do not currently prefault mappings that use virtual page
2263 * tables. We do not prefault foreign pmaps.
2265 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2267 lp = curthread->td_lwp;
2268 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2272 * Limit pre-fault count to 1024 pages.
2274 if (maxpages > 1024)
2277 object = entry->object.vm_object;
2278 KKASSERT(object != NULL);
2279 KKASSERT(object == entry->object.vm_object);
2280 vm_object_hold(object);
2281 vm_object_chain_acquire(object);
2285 for (i = 0; i < maxpages; ++i) {
2286 vm_object_t lobject;
2287 vm_object_t nobject;
2292 * This can eat a lot of time on a heavily contended
2293 * machine so yield on the tick if needed.
2299 * Calculate the page to pre-fault, stopping the scan in
2300 * each direction separately if the limit is reached.
2305 addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2309 addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2311 if (addr < entry->start) {
2317 if (addr >= entry->end) {
2325 * Skip pages already mapped, and stop scanning in that
2326 * direction. When the scan terminates in both directions
2329 if (pmap_prefault_ok(pmap, addr) == 0) {
2340 * Follow the VM object chain to obtain the page to be mapped
2343 * If we reach the terminal object without finding a page
2344 * and we determine it would be advantageous, then allocate
2345 * a zero-fill page for the base object. The base object
2346 * is guaranteed to be OBJT_DEFAULT for this case.
2348 * In order to not have to check the pager via *haspage*()
2349 * we stop if any non-default object is encountered. e.g.
2350 * a vnode or swap object would stop the loop.
2352 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2357 KKASSERT(lobject == entry->object.vm_object);
2358 /*vm_object_hold(lobject); implied */
2360 while ((m = vm_page_lookup_busy_try(lobject, pindex,
2361 TRUE, &error)) == NULL) {
2362 if (lobject->type != OBJT_DEFAULT)
2364 if (lobject->backing_object == NULL) {
2365 if (vm_fast_fault == 0)
2367 if ((prot & VM_PROT_WRITE) == 0 ||
2368 vm_page_count_min(0)) {
2373 * NOTE: Allocated from base object
2375 m = vm_page_alloc(object, index,
2384 /* lobject = object .. not needed */
2387 if (lobject->backing_object_offset & PAGE_MASK)
2389 nobject = lobject->backing_object;
2390 vm_object_hold(nobject);
2391 KKASSERT(nobject == lobject->backing_object);
2392 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2393 if (lobject != object) {
2394 vm_object_lock_swap();
2395 vm_object_drop(lobject);
2398 pprot &= ~VM_PROT_WRITE;
2399 vm_object_chain_acquire(lobject);
2403 * NOTE: A non-NULL (m) will be associated with lobject if
2404 * it was found there, otherwise it is probably a
2405 * zero-fill page associated with the base object.
2407 * Give-up if no page is available.
2410 if (lobject != object) {
2411 if (object->backing_object != lobject)
2412 vm_object_hold(object->backing_object);
2413 vm_object_chain_release_all(
2414 object->backing_object, lobject);
2415 if (object->backing_object != lobject)
2416 vm_object_drop(object->backing_object);
2417 vm_object_drop(lobject);
2423 * The object must be marked dirty if we are mapping a
2424 * writable page. m->object is either lobject or object,
2425 * both of which are still held. Do this before we
2426 * potentially drop the object.
2428 if (pprot & VM_PROT_WRITE)
2429 vm_object_set_writeable_dirty(m->object);
2432 * Do not conditionalize on PG_RAM. If pages are present in
2433 * the VM system we assume optimal caching. If caching is
2434 * not optimal the I/O gravy train will be restarted when we
2435 * hit an unavailable page. We do not want to try to restart
2436 * the gravy train now because we really don't know how much
2437 * of the object has been cached. The cost for restarting
2438 * the gravy train should be low (since accesses will likely
2439 * be I/O bound anyway).
2441 if (lobject != object) {
2442 if (object->backing_object != lobject)
2443 vm_object_hold(object->backing_object);
2444 vm_object_chain_release_all(object->backing_object,
2446 if (object->backing_object != lobject)
2447 vm_object_drop(object->backing_object);
2448 vm_object_drop(lobject);
2452 * Enter the page into the pmap if appropriate. If we had
2453 * allocated the page we have to place it on a queue. If not
2454 * we just have to make sure it isn't on the cache queue
2455 * (pages on the cache queue are not allowed to be mapped).
2459 * Page must be zerod.
2461 if ((m->flags & PG_ZERO) == 0) {
2462 vm_page_zero_fill(m);
2465 pmap_page_assertzero(
2466 VM_PAGE_TO_PHYS(m));
2468 vm_page_flag_clear(m, PG_ZERO);
2469 mycpu->gd_cnt.v_ozfod++;
2471 mycpu->gd_cnt.v_zfod++;
2472 m->valid = VM_PAGE_BITS_ALL;
2475 * Handle dirty page case
2477 if (pprot & VM_PROT_WRITE)
2478 vm_set_nosync(m, entry);
2479 pmap_enter(pmap, addr, m, pprot, 0, entry);
2480 mycpu->gd_cnt.v_vm_faults++;
2481 if (curthread->td_lwp)
2482 ++curthread->td_lwp->lwp_ru.ru_minflt;
2483 vm_page_deactivate(m);
2484 if (pprot & VM_PROT_WRITE) {
2485 /*vm_object_set_writeable_dirty(m->object);*/
2486 vm_set_nosync(m, entry);
2487 if (fault_flags & VM_FAULT_DIRTY) {
2490 swap_pager_unswapped(m);
2495 /* couldn't busy page, no wakeup */
2497 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2498 (m->flags & PG_FICTITIOUS) == 0) {
2500 * A fully valid page not undergoing soft I/O can
2501 * be immediately entered into the pmap.
2503 if ((m->queue - m->pc) == PQ_CACHE)
2504 vm_page_deactivate(m);
2505 if (pprot & VM_PROT_WRITE) {
2506 /*vm_object_set_writeable_dirty(m->object);*/
2507 vm_set_nosync(m, entry);
2508 if (fault_flags & VM_FAULT_DIRTY) {
2511 swap_pager_unswapped(m);
2514 if (pprot & VM_PROT_WRITE)
2515 vm_set_nosync(m, entry);
2516 pmap_enter(pmap, addr, m, pprot, 0, entry);
2517 mycpu->gd_cnt.v_vm_faults++;
2518 if (curthread->td_lwp)
2519 ++curthread->td_lwp->lwp_ru.ru_minflt;
2525 vm_object_chain_release(object);
2526 vm_object_drop(object);
2530 vm_prefault_quick(pmap_t pmap, vm_offset_t addra,
2531 vm_map_entry_t entry, int prot, int fault_flags)
2544 * Get stable max count value, disabled if set to 0
2546 maxpages = vm_prefault_pages;
2552 * We do not currently prefault mappings that use virtual page
2553 * tables. We do not prefault foreign pmaps.
2555 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2557 lp = curthread->td_lwp;
2558 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2562 * Limit pre-fault count to 1024 pages.
2564 if (maxpages > 1024)
2567 object = entry->object.vm_object;
2568 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
2569 KKASSERT(object->backing_object == NULL);
2573 for (i = 0; i < maxpages; ++i) {
2577 * Calculate the page to pre-fault, stopping the scan in
2578 * each direction separately if the limit is reached.
2583 addr = addra - ((i + 1) >> 1) * PAGE_SIZE;
2587 addr = addra + ((i + 2) >> 1) * PAGE_SIZE;
2589 if (addr < entry->start) {
2595 if (addr >= entry->end) {
2603 * Skip pages already mapped, and stop scanning in that
2604 * direction. When the scan terminates in both directions
2607 if (pmap_prefault_ok(pmap, addr) == 0) {
2618 * Follow the VM object chain to obtain the page to be mapped
2619 * into the pmap. This version of the prefault code only
2620 * works with terminal objects.
2622 * WARNING! We cannot call swap_pager_unswapped() with a
2625 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2627 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2628 if (m == NULL || error)
2631 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2632 (m->flags & PG_FICTITIOUS) == 0 &&
2633 ((m->flags & PG_SWAPPED) == 0 ||
2634 (prot & VM_PROT_WRITE) == 0 ||
2635 (fault_flags & VM_FAULT_DIRTY) == 0)) {
2637 * A fully valid page not undergoing soft I/O can
2638 * be immediately entered into the pmap.
2640 if ((m->queue - m->pc) == PQ_CACHE)
2641 vm_page_deactivate(m);
2642 if (prot & VM_PROT_WRITE) {
2643 vm_object_set_writeable_dirty(m->object);
2644 vm_set_nosync(m, entry);
2645 if (fault_flags & VM_FAULT_DIRTY) {
2648 swap_pager_unswapped(m);
2651 pmap_enter(pmap, addr, m, prot, 0, entry);
2652 mycpu->gd_cnt.v_vm_faults++;
2653 if (curthread->td_lwp)
2654 ++curthread->td_lwp->lwp_ru.ru_minflt;