2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
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.
73 * Page fault handling module.
76 #include <sys/cdefs.h>
77 __FBSDID("$FreeBSD$");
79 #include "opt_ktrace.h"
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/kernel.h>
87 #include <sys/mutex.h>
89 #include <sys/racct.h>
90 #include <sys/refcount.h>
91 #include <sys/resourcevar.h>
92 #include <sys/rwlock.h>
93 #include <sys/signalvar.h>
94 #include <sys/sysctl.h>
95 #include <sys/sysent.h>
96 #include <sys/vmmeter.h>
97 #include <sys/vnode.h>
99 #include <sys/ktrace.h>
103 #include <vm/vm_param.h>
105 #include <vm/vm_map.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_pageout.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_pager.h>
111 #include <vm/vm_extern.h>
112 #include <vm/vm_reserv.h>
117 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
119 #define VM_FAULT_DONTNEED_MIN 1048576
122 /* Fault parameters. */
125 vm_prot_t fault_type;
131 /* Page reference for cow. */
134 /* Current object. */
139 /* Top-level map object. */
140 vm_object_t first_object;
141 vm_pindex_t first_pindex;
146 vm_map_entry_t entry;
148 bool lookup_still_valid;
150 /* Vnode if locked. */
154 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
156 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
157 int backward, int forward, bool obj_locked);
159 static int vm_pfault_oom_attempts = 3;
160 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
161 &vm_pfault_oom_attempts, 0,
162 "Number of page allocation attempts in page fault handler before it "
163 "triggers OOM handling");
165 static int vm_pfault_oom_wait = 10;
166 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
167 &vm_pfault_oom_wait, 0,
168 "Number of seconds to wait for free pages before retrying "
169 "the page fault handler");
172 fault_page_release(vm_page_t *mp)
179 * We are likely to loop around again and attempt to busy
180 * this page. Deactivating it leaves it available for
181 * pageout while optimizing fault restarts.
183 vm_page_deactivate(m);
190 fault_page_free(vm_page_t *mp)
196 VM_OBJECT_ASSERT_WLOCKED(m->object);
197 if (!vm_page_wired(m))
206 unlock_map(struct faultstate *fs)
209 if (fs->lookup_still_valid) {
210 vm_map_lookup_done(fs->map, fs->entry);
211 fs->lookup_still_valid = false;
216 unlock_vp(struct faultstate *fs)
219 if (fs->vp != NULL) {
226 fault_deallocate(struct faultstate *fs)
229 fault_page_release(&fs->m_cow);
230 fault_page_release(&fs->m);
231 vm_object_pip_wakeup(fs->object);
232 if (fs->object != fs->first_object) {
233 VM_OBJECT_WLOCK(fs->first_object);
234 fault_page_free(&fs->first_m);
235 VM_OBJECT_WUNLOCK(fs->first_object);
236 vm_object_pip_wakeup(fs->first_object);
238 vm_object_deallocate(fs->first_object);
244 unlock_and_deallocate(struct faultstate *fs)
247 VM_OBJECT_WUNLOCK(fs->object);
248 fault_deallocate(fs);
252 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
256 if (((fs->prot & VM_PROT_WRITE) == 0 &&
257 (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
258 (m->oflags & VPO_UNMANAGED) != 0)
261 VM_PAGE_OBJECT_BUSY_ASSERT(m);
263 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
264 (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
265 (fs->fault_flags & VM_FAULT_DIRTY) != 0;
267 vm_object_set_writeable_dirty(m->object);
270 * If the fault is a write, we know that this page is being
271 * written NOW so dirty it explicitly to save on
272 * pmap_is_modified() calls later.
274 * Also, since the page is now dirty, we can possibly tell
275 * the pager to release any swap backing the page.
277 if (need_dirty && vm_page_set_dirty(m) == 0) {
279 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
280 * if the page is already dirty to prevent data written with
281 * the expectation of being synced from not being synced.
282 * Likewise if this entry does not request NOSYNC then make
283 * sure the page isn't marked NOSYNC. Applications sharing
284 * data should use the same flags to avoid ping ponging.
286 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
287 vm_page_aflag_set(m, PGA_NOSYNC);
289 vm_page_aflag_clear(m, PGA_NOSYNC);
295 * Unlocks fs.first_object and fs.map on success.
298 vm_fault_soft_fast(struct faultstate *fs)
301 #if VM_NRESERVLEVEL > 0
308 MPASS(fs->vp == NULL);
310 vm_object_busy(fs->first_object);
311 m = vm_page_lookup(fs->first_object, fs->first_pindex);
312 /* A busy page can be mapped for read|execute access. */
313 if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 &&
314 vm_page_busied(m)) || !vm_page_all_valid(m)) {
320 #if VM_NRESERVLEVEL > 0
321 if ((m->flags & PG_FICTITIOUS) == 0 &&
322 (m_super = vm_reserv_to_superpage(m)) != NULL &&
323 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
324 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
325 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
326 (pagesizes[m_super->psind] - 1)) && !fs->wired &&
327 pmap_ps_enabled(fs->map->pmap)) {
328 flags = PS_ALL_VALID;
329 if ((fs->prot & VM_PROT_WRITE) != 0) {
331 * Create a superpage mapping allowing write access
332 * only if none of the constituent pages are busy and
333 * all of them are already dirty (except possibly for
334 * the page that was faulted on).
336 flags |= PS_NONE_BUSY;
337 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
338 flags |= PS_ALL_DIRTY;
340 if (vm_page_ps_test(m_super, flags, m)) {
342 psind = m_super->psind;
343 vaddr = rounddown2(vaddr, pagesizes[psind]);
344 /* Preset the modified bit for dirty superpages. */
345 if ((flags & PS_ALL_DIRTY) != 0)
346 fs->fault_type |= VM_PROT_WRITE;
350 rv = pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
351 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
352 if (rv != KERN_SUCCESS)
354 if (fs->m_hold != NULL) {
358 if (psind == 0 && !fs->wired)
359 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
360 VM_OBJECT_RUNLOCK(fs->first_object);
361 vm_fault_dirty(fs, m);
362 vm_map_lookup_done(fs->map, fs->entry);
363 curthread->td_ru.ru_minflt++;
366 vm_object_unbusy(fs->first_object);
371 vm_fault_restore_map_lock(struct faultstate *fs)
374 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
375 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
377 if (!vm_map_trylock_read(fs->map)) {
378 VM_OBJECT_WUNLOCK(fs->first_object);
379 vm_map_lock_read(fs->map);
380 VM_OBJECT_WLOCK(fs->first_object);
382 fs->lookup_still_valid = true;
386 vm_fault_populate_check_page(vm_page_t m)
390 * Check each page to ensure that the pager is obeying the
391 * interface: the page must be installed in the object, fully
392 * valid, and exclusively busied.
395 MPASS(vm_page_all_valid(m));
396 MPASS(vm_page_xbusied(m));
400 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
406 VM_OBJECT_ASSERT_WLOCKED(object);
407 MPASS(first <= last);
408 for (pidx = first, m = vm_page_lookup(object, pidx);
409 pidx <= last; pidx++, m = vm_page_next(m)) {
410 vm_fault_populate_check_page(m);
411 vm_page_deactivate(m);
417 vm_fault_populate(struct faultstate *fs)
421 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
422 int bdry_idx, i, npages, psind, rv;
424 MPASS(fs->object == fs->first_object);
425 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
426 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
427 MPASS(fs->first_object->backing_object == NULL);
428 MPASS(fs->lookup_still_valid);
430 pager_first = OFF_TO_IDX(fs->entry->offset);
431 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
436 * Call the pager (driver) populate() method.
438 * There is no guarantee that the method will be called again
439 * if the current fault is for read, and a future fault is
440 * for write. Report the entry's maximum allowed protection
443 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
444 fs->fault_type, fs->entry->max_protection, &pager_first,
447 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
448 if (rv == VM_PAGER_BAD) {
450 * VM_PAGER_BAD is the backdoor for a pager to request
451 * normal fault handling.
453 vm_fault_restore_map_lock(fs);
454 if (fs->map->timestamp != fs->map_generation)
455 return (KERN_RESTART);
456 return (KERN_NOT_RECEIVER);
458 if (rv != VM_PAGER_OK)
459 return (KERN_FAILURE); /* AKA SIGSEGV */
461 /* Ensure that the driver is obeying the interface. */
462 MPASS(pager_first <= pager_last);
463 MPASS(fs->first_pindex <= pager_last);
464 MPASS(fs->first_pindex >= pager_first);
465 MPASS(pager_last < fs->first_object->size);
467 vm_fault_restore_map_lock(fs);
468 bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
469 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
470 if (fs->map->timestamp != fs->map_generation) {
472 vm_fault_populate_cleanup(fs->first_object, pager_first,
475 m = vm_page_lookup(fs->first_object, pager_first);
479 return (KERN_RESTART);
483 * The map is unchanged after our last unlock. Process the fault.
485 * First, the special case of largepage mappings, where
486 * populate only busies the first page in superpage run.
489 KASSERT(PMAP_HAS_LARGEPAGES,
490 ("missing pmap support for large pages"));
491 m = vm_page_lookup(fs->first_object, pager_first);
492 vm_fault_populate_check_page(m);
493 VM_OBJECT_WUNLOCK(fs->first_object);
494 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
496 /* assert alignment for entry */
497 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
498 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
499 (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
500 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
501 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
502 ("unaligned superpage m %p %#jx", m,
503 (uintmax_t)VM_PAGE_TO_PHYS(m)));
504 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
505 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
506 PMAP_ENTER_LARGEPAGE, bdry_idx);
507 VM_OBJECT_WLOCK(fs->first_object);
509 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
510 for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
513 if (fs->m_hold != NULL) {
514 *fs->m_hold = m + (fs->first_pindex - pager_first);
515 vm_page_wire(*fs->m_hold);
521 * The range [pager_first, pager_last] that is given to the
522 * pager is only a hint. The pager may populate any range
523 * within the object that includes the requested page index.
524 * In case the pager expanded the range, clip it to fit into
527 map_first = OFF_TO_IDX(fs->entry->offset);
528 if (map_first > pager_first) {
529 vm_fault_populate_cleanup(fs->first_object, pager_first,
531 pager_first = map_first;
533 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
534 if (map_last < pager_last) {
535 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
537 pager_last = map_last;
539 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
541 pidx += npages, m = vm_page_next(&m[npages - 1])) {
542 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
543 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
544 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) || \
545 defined(__powerpc64__)
547 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
548 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
549 !pmap_ps_enabled(fs->map->pmap) || fs->wired))
554 npages = atop(pagesizes[psind]);
555 for (i = 0; i < npages; i++) {
556 vm_fault_populate_check_page(&m[i]);
557 vm_fault_dirty(fs, &m[i]);
559 VM_OBJECT_WUNLOCK(fs->first_object);
560 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
561 (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
562 #if defined(__amd64__)
563 if (psind > 0 && rv == KERN_FAILURE) {
564 for (i = 0; i < npages; i++) {
565 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
566 &m[i], fs->prot, fs->fault_type |
567 (fs->wired ? PMAP_ENTER_WIRED : 0), 0);
568 MPASS(rv == KERN_SUCCESS);
572 MPASS(rv == KERN_SUCCESS);
574 VM_OBJECT_WLOCK(fs->first_object);
575 for (i = 0; i < npages; i++) {
576 if ((fs->fault_flags & VM_FAULT_WIRE) != 0)
579 vm_page_activate(&m[i]);
580 if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) {
581 (*fs->m_hold) = &m[i];
584 vm_page_xunbusy(&m[i]);
588 curthread->td_ru.ru_majflt++;
589 return (KERN_SUCCESS);
592 static int prot_fault_translation;
593 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
594 &prot_fault_translation, 0,
595 "Control signal to deliver on protection fault");
597 /* compat definition to keep common code for signal translation */
598 #define UCODE_PAGEFLT 12
600 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
606 * Handle a page fault occurring at the given address,
607 * requiring the given permissions, in the map specified.
608 * If successful, the page is inserted into the
609 * associated physical map.
611 * NOTE: the given address should be truncated to the
612 * proper page address.
614 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
615 * a standard error specifying why the fault is fatal is returned.
617 * The map in question must be referenced, and remains so.
618 * Caller may hold no locks.
621 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
622 int fault_flags, int *signo, int *ucode)
626 MPASS(signo == NULL || ucode != NULL);
628 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
629 ktrfault(vaddr, fault_type);
631 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
633 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
634 result == KERN_INVALID_ADDRESS ||
635 result == KERN_RESOURCE_SHORTAGE ||
636 result == KERN_PROTECTION_FAILURE ||
637 result == KERN_OUT_OF_BOUNDS,
638 ("Unexpected Mach error %d from vm_fault()", result));
640 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
643 if (result != KERN_SUCCESS && signo != NULL) {
646 case KERN_INVALID_ADDRESS:
648 *ucode = SEGV_MAPERR;
650 case KERN_RESOURCE_SHORTAGE:
654 case KERN_OUT_OF_BOUNDS:
658 case KERN_PROTECTION_FAILURE:
659 if (prot_fault_translation == 0) {
661 * Autodetect. This check also covers
662 * the images without the ABI-tag ELF
665 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
666 curproc->p_osrel >= P_OSREL_SIGSEGV) {
668 *ucode = SEGV_ACCERR;
671 *ucode = UCODE_PAGEFLT;
673 } else if (prot_fault_translation == 1) {
674 /* Always compat mode. */
676 *ucode = UCODE_PAGEFLT;
678 /* Always SIGSEGV mode. */
680 *ucode = SEGV_ACCERR;
684 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
693 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
698 if (fs->object->type != OBJT_VNODE)
699 return (KERN_SUCCESS);
700 vp = fs->object->handle;
702 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
703 return (KERN_SUCCESS);
707 * Perform an unlock in case the desired vnode changed while
708 * the map was unlocked during a retry.
712 locked = VOP_ISLOCKED(vp);
713 if (locked != LK_EXCLUSIVE)
717 * We must not sleep acquiring the vnode lock while we have
718 * the page exclusive busied or the object's
719 * paging-in-progress count incremented. Otherwise, we could
722 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
725 return (KERN_SUCCESS);
730 unlock_and_deallocate(fs);
732 fault_deallocate(fs);
733 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
736 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
737 return (KERN_RESOURCE_SHORTAGE);
741 * Calculate the desired readahead. Handle drop-behind.
743 * Returns the number of readahead blocks to pass to the pager.
746 vm_fault_readahead(struct faultstate *fs)
751 KASSERT(fs->lookup_still_valid, ("map unlocked"));
752 era = fs->entry->read_ahead;
753 behavior = vm_map_entry_behavior(fs->entry);
754 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
756 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
757 nera = VM_FAULT_READ_AHEAD_MAX;
758 if (fs->vaddr == fs->entry->next_read)
759 vm_fault_dontneed(fs, fs->vaddr, nera);
760 } else if (fs->vaddr == fs->entry->next_read) {
762 * This is a sequential fault. Arithmetically
763 * increase the requested number of pages in
764 * the read-ahead window. The requested
765 * number of pages is "# of sequential faults
766 * x (read ahead min + 1) + read ahead min"
768 nera = VM_FAULT_READ_AHEAD_MIN;
771 if (nera > VM_FAULT_READ_AHEAD_MAX)
772 nera = VM_FAULT_READ_AHEAD_MAX;
774 if (era == VM_FAULT_READ_AHEAD_MAX)
775 vm_fault_dontneed(fs, fs->vaddr, nera);
778 * This is a non-sequential fault.
784 * A read lock on the map suffices to update
785 * the read ahead count safely.
787 fs->entry->read_ahead = nera;
794 vm_fault_lookup(struct faultstate *fs)
798 KASSERT(!fs->lookup_still_valid,
799 ("vm_fault_lookup: Map already locked."));
800 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
801 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
802 &fs->first_pindex, &fs->prot, &fs->wired);
803 if (result != KERN_SUCCESS) {
808 fs->map_generation = fs->map->timestamp;
810 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
811 panic("%s: fault on nofault entry, addr: %#lx",
812 __func__, (u_long)fs->vaddr);
815 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
816 fs->entry->wiring_thread != curthread) {
817 vm_map_unlock_read(fs->map);
818 vm_map_lock(fs->map);
819 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
820 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
822 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
823 vm_map_unlock_and_wait(fs->map, 0);
825 vm_map_unlock(fs->map);
826 return (KERN_RESOURCE_SHORTAGE);
829 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
832 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
834 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
835 ("!fs->wired && VM_FAULT_WIRE"));
836 fs->lookup_still_valid = true;
838 return (KERN_SUCCESS);
842 vm_fault_relookup(struct faultstate *fs)
844 vm_object_t retry_object;
845 vm_pindex_t retry_pindex;
846 vm_prot_t retry_prot;
849 if (!vm_map_trylock_read(fs->map))
850 return (KERN_RESTART);
852 fs->lookup_still_valid = true;
853 if (fs->map->timestamp == fs->map_generation)
854 return (KERN_SUCCESS);
856 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
857 &fs->entry, &retry_object, &retry_pindex, &retry_prot,
859 if (result != KERN_SUCCESS) {
861 * If retry of map lookup would have blocked then
862 * retry fault from start.
864 if (result == KERN_FAILURE)
865 return (KERN_RESTART);
868 if (retry_object != fs->first_object ||
869 retry_pindex != fs->first_pindex)
870 return (KERN_RESTART);
873 * Check whether the protection has changed or the object has
874 * been copied while we left the map unlocked. Changing from
875 * read to write permission is OK - we leave the page
876 * write-protected, and catch the write fault. Changing from
877 * write to read permission means that we can't mark the page
878 * write-enabled after all.
880 fs->prot &= retry_prot;
881 fs->fault_type &= retry_prot;
883 return (KERN_RESTART);
885 /* Reassert because wired may have changed. */
886 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
887 ("!wired && VM_FAULT_WIRE"));
889 return (KERN_SUCCESS);
893 vm_fault_cow(struct faultstate *fs)
895 bool is_first_object_locked;
897 KASSERT(fs->object != fs->first_object,
898 ("source and target COW objects are identical"));
901 * This allows pages to be virtually copied from a backing_object
902 * into the first_object, where the backing object has no other
903 * refs to it, and cannot gain any more refs. Instead of a bcopy,
904 * we just move the page from the backing object to the first
905 * object. Note that we must mark the page dirty in the first
906 * object so that it will go out to swap when needed.
908 is_first_object_locked = false;
911 * Only one shadow object and no other refs.
913 fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
915 * No other ways to look the object up
917 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
919 * We don't chase down the shadow chain and we can acquire locks.
921 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
922 fs->object == fs->first_object->backing_object &&
923 VM_OBJECT_TRYWLOCK(fs->object)) {
925 * Remove but keep xbusy for replace. fs->m is moved into
926 * fs->first_object and left busy while fs->first_m is
927 * conditionally freed.
929 vm_page_remove_xbusy(fs->m);
930 vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
932 vm_page_dirty(fs->m);
933 #if VM_NRESERVLEVEL > 0
935 * Rename the reservation.
937 vm_reserv_rename(fs->m, fs->first_object, fs->object,
938 OFF_TO_IDX(fs->first_object->backing_object_offset));
940 VM_OBJECT_WUNLOCK(fs->object);
941 VM_OBJECT_WUNLOCK(fs->first_object);
944 VM_CNT_INC(v_cow_optim);
946 if (is_first_object_locked)
947 VM_OBJECT_WUNLOCK(fs->first_object);
949 * Oh, well, lets copy it.
951 pmap_copy_page(fs->m, fs->first_m);
952 vm_page_valid(fs->first_m);
953 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
954 vm_page_wire(fs->first_m);
955 vm_page_unwire(fs->m, PQ_INACTIVE);
958 * Save the cow page to be released after
959 * pmap_enter is complete.
965 * Typically, the shadow object is either private to this
966 * address space (OBJ_ONEMAPPING) or its pages are read only.
967 * In the highly unusual case where the pages of a shadow object
968 * are read/write shared between this and other address spaces,
969 * we need to ensure that any pmap-level mappings to the
970 * original, copy-on-write page from the backing object are
971 * removed from those other address spaces.
973 * The flag check is racy, but this is tolerable: if
974 * OBJ_ONEMAPPING is cleared after the check, the busy state
975 * ensures that new mappings of m_cow can't be created.
976 * pmap_enter() will replace an existing mapping in the current
977 * address space. If OBJ_ONEMAPPING is set after the check,
978 * removing mappings will at worse trigger some unnecessary page
981 vm_page_assert_xbusied(fs->m_cow);
982 if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
983 pmap_remove_all(fs->m_cow);
986 vm_object_pip_wakeup(fs->object);
989 * Only use the new page below...
991 fs->object = fs->first_object;
992 fs->pindex = fs->first_pindex;
994 VM_CNT_INC(v_cow_faults);
999 vm_fault_next(struct faultstate *fs)
1001 vm_object_t next_object;
1004 * The requested page does not exist at this object/
1005 * offset. Remove the invalid page from the object,
1006 * waking up anyone waiting for it, and continue on to
1007 * the next object. However, if this is the top-level
1008 * object, we must leave the busy page in place to
1009 * prevent another process from rushing past us, and
1010 * inserting the page in that object at the same time
1013 if (fs->object == fs->first_object) {
1014 fs->first_m = fs->m;
1017 fault_page_free(&fs->m);
1020 * Move on to the next object. Lock the next object before
1021 * unlocking the current one.
1023 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1024 next_object = fs->object->backing_object;
1025 if (next_object == NULL)
1027 MPASS(fs->first_m != NULL);
1028 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1029 VM_OBJECT_WLOCK(next_object);
1030 vm_object_pip_add(next_object, 1);
1031 if (fs->object != fs->first_object)
1032 vm_object_pip_wakeup(fs->object);
1033 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1034 VM_OBJECT_WUNLOCK(fs->object);
1035 fs->object = next_object;
1041 vm_fault_zerofill(struct faultstate *fs)
1045 * If there's no object left, fill the page in the top
1046 * object with zeros.
1048 if (fs->object != fs->first_object) {
1049 vm_object_pip_wakeup(fs->object);
1050 fs->object = fs->first_object;
1051 fs->pindex = fs->first_pindex;
1053 MPASS(fs->first_m != NULL);
1054 MPASS(fs->m == NULL);
1055 fs->m = fs->first_m;
1059 * Zero the page if necessary and mark it valid.
1061 if ((fs->m->flags & PG_ZERO) == 0) {
1062 pmap_zero_page(fs->m);
1064 VM_CNT_INC(v_ozfod);
1067 vm_page_valid(fs->m);
1071 * Allocate a page directly or via the object populate method.
1074 vm_fault_allocate(struct faultstate *fs)
1076 struct domainset *dset;
1080 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1081 rv = vm_fault_lock_vnode(fs, true);
1082 MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1083 if (rv == KERN_RESOURCE_SHORTAGE)
1087 if (fs->pindex >= fs->object->size)
1088 return (KERN_OUT_OF_BOUNDS);
1090 if (fs->object == fs->first_object &&
1091 (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1092 fs->first_object->shadow_count == 0) {
1093 rv = vm_fault_populate(fs);
1099 case KERN_NOT_RECEIVER:
1101 * Pager's populate() method
1102 * returned VM_PAGER_BAD.
1106 panic("inconsistent return codes");
1111 * Allocate a new page for this object/offset pair.
1113 * Unlocked read of the p_flag is harmless. At worst, the P_KILLED
1114 * might be not observed there, and allocation can fail, causing
1115 * restart and new reading of the p_flag.
1117 dset = fs->object->domain.dr_policy;
1119 dset = curthread->td_domain.dr_policy;
1120 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1121 #if VM_NRESERVLEVEL > 0
1122 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1124 alloc_req = P_KILLED(curproc) ?
1125 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
1126 if (fs->object->type != OBJT_VNODE &&
1127 fs->object->backing_object == NULL)
1128 alloc_req |= VM_ALLOC_ZERO;
1129 fs->m = vm_page_alloc(fs->object, fs->pindex, alloc_req);
1131 if (fs->m == NULL) {
1132 unlock_and_deallocate(fs);
1133 if (vm_pfault_oom_attempts < 0 ||
1134 fs->oom < vm_pfault_oom_attempts) {
1136 vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1140 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1141 curproc->p_pid, curproc->p_comm);
1142 vm_pageout_oom(VM_OOM_MEM_PF);
1145 return (KERN_RESOURCE_SHORTAGE);
1149 return (KERN_NOT_RECEIVER);
1153 * Call the pager to retrieve the page if there is a chance
1154 * that the pager has it, and potentially retrieve additional
1155 * pages at the same time.
1158 vm_fault_getpages(struct faultstate *fs, int nera, int *behindp, int *aheadp)
1160 vm_offset_t e_end, e_start;
1161 int ahead, behind, cluster_offset, rv;
1165 * Prepare for unlocking the map. Save the map
1166 * entry's start and end addresses, which are used to
1167 * optimize the size of the pager operation below.
1168 * Even if the map entry's addresses change after
1169 * unlocking the map, using the saved addresses is
1172 e_start = fs->entry->start;
1173 e_end = fs->entry->end;
1174 behavior = vm_map_entry_behavior(fs->entry);
1177 * Release the map lock before locking the vnode or
1178 * sleeping in the pager. (If the current object has
1179 * a shadow, then an earlier iteration of this loop
1180 * may have already unlocked the map.)
1184 rv = vm_fault_lock_vnode(fs, false);
1185 MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1186 if (rv == KERN_RESOURCE_SHORTAGE)
1188 KASSERT(fs->vp == NULL || !fs->map->system_map,
1189 ("vm_fault: vnode-backed object mapped by system map"));
1192 * Page in the requested page and hint the pager,
1193 * that it may bring up surrounding pages.
1195 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1196 P_KILLED(curproc)) {
1200 /* Is this a sequential fault? */
1206 * Request a cluster of pages that is
1207 * aligned to a VM_FAULT_READ_DEFAULT
1208 * page offset boundary within the
1209 * object. Alignment to a page offset
1210 * boundary is more likely to coincide
1211 * with the underlying file system
1212 * block than alignment to a virtual
1215 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1216 behind = ulmin(cluster_offset,
1217 atop(fs->vaddr - e_start));
1218 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1220 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1224 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1225 if (rv == VM_PAGER_OK)
1226 return (KERN_SUCCESS);
1227 if (rv == VM_PAGER_ERROR)
1228 printf("vm_fault: pager read error, pid %d (%s)\n",
1229 curproc->p_pid, curproc->p_comm);
1231 * If an I/O error occurred or the requested page was
1232 * outside the range of the pager, clean up and return
1235 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD)
1236 return (KERN_OUT_OF_BOUNDS);
1237 return (KERN_NOT_RECEIVER);
1241 * Wait/Retry if the page is busy. We have to do this if the page is
1242 * either exclusive or shared busy because the vm_pager may be using
1243 * read busy for pageouts (and even pageins if it is the vnode pager),
1244 * and we could end up trying to pagein and pageout the same page
1247 * We can theoretically allow the busy case on a read fault if the page
1248 * is marked valid, but since such pages are typically already pmap'd,
1249 * putting that special case in might be more effort then it is worth.
1250 * We cannot under any circumstances mess around with a shared busied
1251 * page except, perhaps, to pmap it.
1254 vm_fault_busy_sleep(struct faultstate *fs)
1257 * Reference the page before unlocking and
1258 * sleeping so that the page daemon is less
1259 * likely to reclaim it.
1261 vm_page_aflag_set(fs->m, PGA_REFERENCED);
1262 if (fs->object != fs->first_object) {
1263 fault_page_release(&fs->first_m);
1264 vm_object_pip_wakeup(fs->first_object);
1266 vm_object_pip_wakeup(fs->object);
1268 if (fs->m == vm_page_lookup(fs->object, fs->pindex))
1269 vm_page_busy_sleep(fs->m, "vmpfw", false);
1271 VM_OBJECT_WUNLOCK(fs->object);
1272 VM_CNT_INC(v_intrans);
1273 vm_object_deallocate(fs->first_object);
1277 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1278 int fault_flags, vm_page_t *m_hold)
1280 struct faultstate fs;
1281 int ahead, behind, faultcount;
1282 int nera, result, rv;
1283 bool dead, hardfault;
1285 VM_CNT_INC(v_vm_faults);
1287 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1288 return (KERN_PROTECTION_FAILURE);
1293 fs.fault_flags = fault_flags;
1295 fs.lookup_still_valid = false;
1302 fs.fault_type = fault_type;
1305 * Find the backing store object and offset into it to begin the
1308 result = vm_fault_lookup(&fs);
1309 if (result != KERN_SUCCESS) {
1310 if (result == KERN_RESOURCE_SHORTAGE)
1316 * Try to avoid lock contention on the top-level object through
1317 * special-case handling of some types of page faults, specifically,
1318 * those that are mapping an existing page from the top-level object.
1319 * Under this condition, a read lock on the object suffices, allowing
1320 * multiple page faults of a similar type to run in parallel.
1322 if (fs.vp == NULL /* avoid locked vnode leak */ &&
1323 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1324 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1325 VM_OBJECT_RLOCK(fs.first_object);
1326 rv = vm_fault_soft_fast(&fs);
1327 if (rv == KERN_SUCCESS)
1329 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
1330 VM_OBJECT_RUNLOCK(fs.first_object);
1331 VM_OBJECT_WLOCK(fs.first_object);
1334 VM_OBJECT_WLOCK(fs.first_object);
1338 * Make a reference to this object to prevent its disposal while we
1339 * are messing with it. Once we have the reference, the map is free
1340 * to be diddled. Since objects reference their shadows (and copies),
1341 * they will stay around as well.
1343 * Bump the paging-in-progress count to prevent size changes (e.g.
1344 * truncation operations) during I/O.
1346 vm_object_reference_locked(fs.first_object);
1347 vm_object_pip_add(fs.first_object, 1);
1349 fs.m_cow = fs.m = fs.first_m = NULL;
1352 * Search for the page at object/offset.
1354 fs.object = fs.first_object;
1355 fs.pindex = fs.first_pindex;
1357 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1358 rv = vm_fault_allocate(&fs);
1361 unlock_and_deallocate(&fs);
1363 case KERN_RESOURCE_SHORTAGE:
1367 case KERN_OUT_OF_BOUNDS:
1368 unlock_and_deallocate(&fs);
1370 case KERN_NOT_RECEIVER:
1373 panic("vm_fault: Unhandled rv %d", rv);
1378 KASSERT(fs.m == NULL,
1379 ("page still set %p at loop start", fs.m));
1381 * If the object is marked for imminent termination,
1382 * we retry here, since the collapse pass has raced
1383 * with us. Otherwise, if we see terminally dead
1384 * object, return fail.
1386 if ((fs.object->flags & OBJ_DEAD) != 0) {
1387 dead = fs.object->type == OBJT_DEAD;
1388 unlock_and_deallocate(&fs);
1390 return (KERN_PROTECTION_FAILURE);
1396 * See if page is resident
1398 fs.m = vm_page_lookup(fs.object, fs.pindex);
1400 if (vm_page_tryxbusy(fs.m) == 0) {
1401 vm_fault_busy_sleep(&fs);
1406 * The page is marked busy for other processes and the
1407 * pagedaemon. If it still is completely valid we
1410 if (vm_page_all_valid(fs.m)) {
1411 VM_OBJECT_WUNLOCK(fs.object);
1412 break; /* break to PAGE HAS BEEN FOUND. */
1415 VM_OBJECT_ASSERT_WLOCKED(fs.object);
1418 * Page is not resident. If the pager might contain the page
1419 * or this is the beginning of the search, allocate a new
1420 * page. (Default objects are zero-fill, so there is no real
1423 if (fs.m == NULL && (fs.object->type != OBJT_DEFAULT ||
1424 fs.object == fs.first_object)) {
1425 rv = vm_fault_allocate(&fs);
1428 unlock_and_deallocate(&fs);
1430 case KERN_RESOURCE_SHORTAGE:
1434 case KERN_OUT_OF_BOUNDS:
1435 unlock_and_deallocate(&fs);
1437 case KERN_NOT_RECEIVER:
1440 panic("vm_fault: Unhandled rv %d", rv);
1445 * Default objects have no pager so no exclusive busy exists
1446 * to protect this page in the chain. Skip to the next
1447 * object without dropping the lock to preserve atomicity of
1450 if (fs.object->type != OBJT_DEFAULT) {
1452 * At this point, we have either allocated a new page
1453 * or found an existing page that is only partially
1456 * We hold a reference on the current object and the
1457 * page is exclusive busied. The exclusive busy
1458 * prevents simultaneous faults and collapses while
1459 * the object lock is dropped.
1461 VM_OBJECT_WUNLOCK(fs.object);
1464 * If the pager for the current object might have
1465 * the page, then determine the number of additional
1466 * pages to read and potentially reprioritize
1467 * previously read pages for earlier reclamation.
1468 * These operations should only be performed once per
1469 * page fault. Even if the current pager doesn't
1470 * have the page, the number of additional pages to
1471 * read will apply to subsequent objects in the
1474 if (nera == -1 && !P_KILLED(curproc))
1475 nera = vm_fault_readahead(&fs);
1477 rv = vm_fault_getpages(&fs, nera, &behind, &ahead);
1478 if (rv == KERN_SUCCESS) {
1479 faultcount = behind + 1 + ahead;
1481 break; /* break to PAGE HAS BEEN FOUND. */
1483 if (rv == KERN_RESOURCE_SHORTAGE)
1485 VM_OBJECT_WLOCK(fs.object);
1486 if (rv == KERN_OUT_OF_BOUNDS) {
1487 fault_page_free(&fs.m);
1488 unlock_and_deallocate(&fs);
1494 * The page was not found in the current object. Try to
1495 * traverse into a backing object or zero fill if none is
1498 if (vm_fault_next(&fs))
1500 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1501 if (fs.first_object == fs.object)
1502 fault_page_free(&fs.first_m);
1503 unlock_and_deallocate(&fs);
1504 return (KERN_OUT_OF_BOUNDS);
1506 VM_OBJECT_WUNLOCK(fs.object);
1507 vm_fault_zerofill(&fs);
1508 /* Don't try to prefault neighboring pages. */
1510 break; /* break to PAGE HAS BEEN FOUND. */
1514 * PAGE HAS BEEN FOUND. A valid page has been found and exclusively
1515 * busied. The object lock must no longer be held.
1517 vm_page_assert_xbusied(fs.m);
1518 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1521 * If the page is being written, but isn't already owned by the
1522 * top-level object, we have to copy it into a new page owned by the
1525 if (fs.object != fs.first_object) {
1527 * We only really need to copy if we want to write it.
1529 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1532 * We only try to prefault read-only mappings to the
1533 * neighboring pages when this copy-on-write fault is
1534 * a hard fault. In other cases, trying to prefault
1535 * is typically wasted effort.
1537 if (faultcount == 0)
1541 fs.prot &= ~VM_PROT_WRITE;
1546 * We must verify that the maps have not changed since our last
1549 if (!fs.lookup_still_valid) {
1550 result = vm_fault_relookup(&fs);
1551 if (result != KERN_SUCCESS) {
1552 fault_deallocate(&fs);
1553 if (result == KERN_RESTART)
1558 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1561 * If the page was filled by a pager, save the virtual address that
1562 * should be faulted on next under a sequential access pattern to the
1563 * map entry. A read lock on the map suffices to update this address
1567 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1570 * Page must be completely valid or it is not fit to
1571 * map into user space. vm_pager_get_pages() ensures this.
1573 vm_page_assert_xbusied(fs.m);
1574 KASSERT(vm_page_all_valid(fs.m),
1575 ("vm_fault: page %p partially invalid", fs.m));
1577 vm_fault_dirty(&fs, fs.m);
1580 * Put this page into the physical map. We had to do the unlock above
1581 * because pmap_enter() may sleep. We don't put the page
1582 * back on the active queue until later so that the pageout daemon
1583 * won't find it (yet).
1585 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1586 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1587 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1589 vm_fault_prefault(&fs, vaddr,
1590 faultcount > 0 ? behind : PFBAK,
1591 faultcount > 0 ? ahead : PFFOR, false);
1594 * If the page is not wired down, then put it where the pageout daemon
1597 if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1600 vm_page_activate(fs.m);
1601 if (fs.m_hold != NULL) {
1602 (*fs.m_hold) = fs.m;
1605 vm_page_xunbusy(fs.m);
1609 * Unlock everything, and return
1611 fault_deallocate(&fs);
1613 VM_CNT_INC(v_io_faults);
1614 curthread->td_ru.ru_majflt++;
1616 if (racct_enable && fs.object->type == OBJT_VNODE) {
1618 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1619 racct_add_force(curproc, RACCT_WRITEBPS,
1620 PAGE_SIZE + behind * PAGE_SIZE);
1621 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1623 racct_add_force(curproc, RACCT_READBPS,
1624 PAGE_SIZE + ahead * PAGE_SIZE);
1625 racct_add_force(curproc, RACCT_READIOPS, 1);
1627 PROC_UNLOCK(curproc);
1631 curthread->td_ru.ru_minflt++;
1633 return (KERN_SUCCESS);
1637 * Speed up the reclamation of pages that precede the faulting pindex within
1638 * the first object of the shadow chain. Essentially, perform the equivalent
1639 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1640 * the faulting pindex by the cluster size when the pages read by vm_fault()
1641 * cross a cluster-size boundary. The cluster size is the greater of the
1642 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1644 * When "fs->first_object" is a shadow object, the pages in the backing object
1645 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1646 * function must only be concerned with pages in the first object.
1649 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1651 vm_map_entry_t entry;
1652 vm_object_t first_object, object;
1653 vm_offset_t end, start;
1654 vm_page_t m, m_next;
1655 vm_pindex_t pend, pstart;
1658 object = fs->object;
1659 VM_OBJECT_ASSERT_UNLOCKED(object);
1660 first_object = fs->first_object;
1661 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1662 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1663 VM_OBJECT_RLOCK(first_object);
1664 size = VM_FAULT_DONTNEED_MIN;
1665 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1666 size = pagesizes[1];
1667 end = rounddown2(vaddr, size);
1668 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1669 (entry = fs->entry)->start < end) {
1670 if (end - entry->start < size)
1671 start = entry->start;
1674 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1675 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1677 m_next = vm_page_find_least(first_object, pstart);
1678 pend = OFF_TO_IDX(entry->offset) + atop(end -
1680 while ((m = m_next) != NULL && m->pindex < pend) {
1681 m_next = TAILQ_NEXT(m, listq);
1682 if (!vm_page_all_valid(m) ||
1687 * Don't clear PGA_REFERENCED, since it would
1688 * likely represent a reference by a different
1691 * Typically, at this point, prefetched pages
1692 * are still in the inactive queue. Only
1693 * pages that triggered page faults are in the
1694 * active queue. The test for whether the page
1695 * is in the inactive queue is racy; in the
1696 * worst case we will requeue the page
1699 if (!vm_page_inactive(m))
1700 vm_page_deactivate(m);
1703 VM_OBJECT_RUNLOCK(first_object);
1708 * vm_fault_prefault provides a quick way of clustering
1709 * pagefaults into a processes address space. It is a "cousin"
1710 * of vm_map_pmap_enter, except it runs at page fault time instead
1714 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1715 int backward, int forward, bool obj_locked)
1718 vm_map_entry_t entry;
1719 vm_object_t backing_object, lobject;
1720 vm_offset_t addr, starta;
1725 pmap = fs->map->pmap;
1726 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1731 if (addra < backward * PAGE_SIZE) {
1732 starta = entry->start;
1734 starta = addra - backward * PAGE_SIZE;
1735 if (starta < entry->start)
1736 starta = entry->start;
1740 * Generate the sequence of virtual addresses that are candidates for
1741 * prefaulting in an outward spiral from the faulting virtual address,
1742 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1743 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1744 * If the candidate address doesn't have a backing physical page, then
1745 * the loop immediately terminates.
1747 for (i = 0; i < 2 * imax(backward, forward); i++) {
1748 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1750 if (addr > addra + forward * PAGE_SIZE)
1753 if (addr < starta || addr >= entry->end)
1756 if (!pmap_is_prefaultable(pmap, addr))
1759 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1760 lobject = entry->object.vm_object;
1762 VM_OBJECT_RLOCK(lobject);
1763 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1764 lobject->type == OBJT_DEFAULT &&
1765 (backing_object = lobject->backing_object) != NULL) {
1766 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1767 0, ("vm_fault_prefault: unaligned object offset"));
1768 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1769 VM_OBJECT_RLOCK(backing_object);
1770 if (!obj_locked || lobject != entry->object.vm_object)
1771 VM_OBJECT_RUNLOCK(lobject);
1772 lobject = backing_object;
1775 if (!obj_locked || lobject != entry->object.vm_object)
1776 VM_OBJECT_RUNLOCK(lobject);
1779 if (vm_page_all_valid(m) &&
1780 (m->flags & PG_FICTITIOUS) == 0)
1781 pmap_enter_quick(pmap, addr, m, entry->protection);
1782 if (!obj_locked || lobject != entry->object.vm_object)
1783 VM_OBJECT_RUNLOCK(lobject);
1788 * Hold each of the physical pages that are mapped by the specified range of
1789 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1790 * and allow the specified types of access, "prot". If all of the implied
1791 * pages are successfully held, then the number of held pages is returned
1792 * together with pointers to those pages in the array "ma". However, if any
1793 * of the pages cannot be held, -1 is returned.
1796 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1797 vm_prot_t prot, vm_page_t *ma, int max_count)
1799 vm_offset_t end, va;
1802 boolean_t pmap_failed;
1806 end = round_page(addr + len);
1807 addr = trunc_page(addr);
1809 if (!vm_map_range_valid(map, addr, end))
1812 if (atop(end - addr) > max_count)
1813 panic("vm_fault_quick_hold_pages: count > max_count");
1814 count = atop(end - addr);
1817 * Most likely, the physical pages are resident in the pmap, so it is
1818 * faster to try pmap_extract_and_hold() first.
1820 pmap_failed = FALSE;
1821 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1822 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1825 else if ((prot & VM_PROT_WRITE) != 0 &&
1826 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1828 * Explicitly dirty the physical page. Otherwise, the
1829 * caller's changes may go unnoticed because they are
1830 * performed through an unmanaged mapping or by a DMA
1833 * The object lock is not held here.
1834 * See vm_page_clear_dirty_mask().
1841 * One or more pages could not be held by the pmap. Either no
1842 * page was mapped at the specified virtual address or that
1843 * mapping had insufficient permissions. Attempt to fault in
1844 * and hold these pages.
1846 * If vm_fault_disable_pagefaults() was called,
1847 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1848 * acquire MD VM locks, which means we must not call
1849 * vm_fault(). Some (out of tree) callers mark
1850 * too wide a code area with vm_fault_disable_pagefaults()
1851 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1852 * the proper behaviour explicitly.
1854 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1855 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1857 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1858 if (*mp == NULL && vm_fault(map, va, prot,
1859 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1864 for (mp = ma; mp < ma + count; mp++)
1866 vm_page_unwire(*mp, PQ_INACTIVE);
1872 * vm_fault_copy_entry
1874 * Create new shadow object backing dst_entry with private copy of
1875 * all underlying pages. When src_entry is equal to dst_entry,
1876 * function implements COW for wired-down map entry. Otherwise,
1877 * it forks wired entry into dst_map.
1879 * In/out conditions:
1880 * The source and destination maps must be locked for write.
1881 * The source map entry must be wired down (or be a sharing map
1882 * entry corresponding to a main map entry that is wired down).
1885 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1886 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1887 vm_ooffset_t *fork_charge)
1889 vm_object_t backing_object, dst_object, object, src_object;
1890 vm_pindex_t dst_pindex, pindex, src_pindex;
1891 vm_prot_t access, prot;
1901 upgrade = src_entry == dst_entry;
1902 access = prot = dst_entry->protection;
1904 src_object = src_entry->object.vm_object;
1905 src_pindex = OFF_TO_IDX(src_entry->offset);
1907 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1908 dst_object = src_object;
1909 vm_object_reference(dst_object);
1912 * Create the top-level object for the destination entry.
1913 * Doesn't actually shadow anything - we copy the pages
1916 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1917 dst_entry->start), NULL, NULL, 0);
1918 #if VM_NRESERVLEVEL > 0
1919 dst_object->flags |= OBJ_COLORED;
1920 dst_object->pg_color = atop(dst_entry->start);
1922 dst_object->domain = src_object->domain;
1923 dst_object->charge = dst_entry->end - dst_entry->start;
1926 VM_OBJECT_WLOCK(dst_object);
1927 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1928 ("vm_fault_copy_entry: vm_object not NULL"));
1929 if (src_object != dst_object) {
1930 dst_entry->object.vm_object = dst_object;
1931 dst_entry->offset = 0;
1932 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1934 if (fork_charge != NULL) {
1935 KASSERT(dst_entry->cred == NULL,
1936 ("vm_fault_copy_entry: leaked swp charge"));
1937 dst_object->cred = curthread->td_ucred;
1938 crhold(dst_object->cred);
1939 *fork_charge += dst_object->charge;
1940 } else if ((dst_object->type == OBJT_DEFAULT ||
1941 dst_object->type == OBJT_SWAP) &&
1942 dst_object->cred == NULL) {
1943 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1945 dst_object->cred = dst_entry->cred;
1946 dst_entry->cred = NULL;
1950 * If not an upgrade, then enter the mappings in the pmap as
1951 * read and/or execute accesses. Otherwise, enter them as
1954 * A writeable large page mapping is only created if all of
1955 * the constituent small page mappings are modified. Marking
1956 * PTEs as modified on inception allows promotion to happen
1957 * without taking potentially large number of soft faults.
1960 access &= ~VM_PROT_WRITE;
1963 * Loop through all of the virtual pages within the entry's
1964 * range, copying each page from the source object to the
1965 * destination object. Since the source is wired, those pages
1966 * must exist. In contrast, the destination is pageable.
1967 * Since the destination object doesn't share any backing storage
1968 * with the source object, all of its pages must be dirtied,
1969 * regardless of whether they can be written.
1971 for (vaddr = dst_entry->start, dst_pindex = 0;
1972 vaddr < dst_entry->end;
1973 vaddr += PAGE_SIZE, dst_pindex++) {
1976 * Find the page in the source object, and copy it in.
1977 * Because the source is wired down, the page will be
1980 if (src_object != dst_object)
1981 VM_OBJECT_RLOCK(src_object);
1982 object = src_object;
1983 pindex = src_pindex + dst_pindex;
1984 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1985 (backing_object = object->backing_object) != NULL) {
1987 * Unless the source mapping is read-only or
1988 * it is presently being upgraded from
1989 * read-only, the first object in the shadow
1990 * chain should provide all of the pages. In
1991 * other words, this loop body should never be
1992 * executed when the source mapping is already
1995 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1997 ("vm_fault_copy_entry: main object missing page"));
1999 VM_OBJECT_RLOCK(backing_object);
2000 pindex += OFF_TO_IDX(object->backing_object_offset);
2001 if (object != dst_object)
2002 VM_OBJECT_RUNLOCK(object);
2003 object = backing_object;
2005 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2007 if (object != dst_object) {
2009 * Allocate a page in the destination object.
2011 dst_m = vm_page_alloc(dst_object, (src_object ==
2012 dst_object ? src_pindex : 0) + dst_pindex,
2014 if (dst_m == NULL) {
2015 VM_OBJECT_WUNLOCK(dst_object);
2016 VM_OBJECT_RUNLOCK(object);
2017 vm_wait(dst_object);
2018 VM_OBJECT_WLOCK(dst_object);
2023 * See the comment in vm_fault_cow().
2025 if (src_object == dst_object &&
2026 (object->flags & OBJ_ONEMAPPING) == 0)
2027 pmap_remove_all(src_m);
2028 pmap_copy_page(src_m, dst_m);
2029 VM_OBJECT_RUNLOCK(object);
2030 dst_m->dirty = dst_m->valid = src_m->valid;
2033 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2035 if (dst_m->pindex >= dst_object->size) {
2037 * We are upgrading. Index can occur
2038 * out of bounds if the object type is
2039 * vnode and the file was truncated.
2041 vm_page_xunbusy(dst_m);
2045 VM_OBJECT_WUNLOCK(dst_object);
2048 * Enter it in the pmap. If a wired, copy-on-write
2049 * mapping is being replaced by a write-enabled
2050 * mapping, then wire that new mapping.
2052 * The page can be invalid if the user called
2053 * msync(MS_INVALIDATE) or truncated the backing vnode
2054 * or shared memory object. In this case, do not
2055 * insert it into pmap, but still do the copy so that
2056 * all copies of the wired map entry have similar
2059 if (vm_page_all_valid(dst_m)) {
2060 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2061 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2065 * Mark it no longer busy, and put it on the active list.
2067 VM_OBJECT_WLOCK(dst_object);
2070 if (src_m != dst_m) {
2071 vm_page_unwire(src_m, PQ_INACTIVE);
2072 vm_page_wire(dst_m);
2074 KASSERT(vm_page_wired(dst_m),
2075 ("dst_m %p is not wired", dst_m));
2078 vm_page_activate(dst_m);
2080 vm_page_xunbusy(dst_m);
2082 VM_OBJECT_WUNLOCK(dst_object);
2084 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2085 vm_object_deallocate(src_object);
2090 * Block entry into the machine-independent layer's page fault handler by
2091 * the calling thread. Subsequent calls to vm_fault() by that thread will
2092 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
2093 * spurious page faults.
2096 vm_fault_disable_pagefaults(void)
2099 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2103 vm_fault_enable_pagefaults(int save)
2106 curthread_pflags_restore(save);