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>
88 #include <sys/pctrie.h>
90 #include <sys/racct.h>
91 #include <sys/refcount.h>
92 #include <sys/resourcevar.h>
93 #include <sys/rwlock.h>
94 #include <sys/signalvar.h>
95 #include <sys/sysctl.h>
96 #include <sys/sysent.h>
97 #include <sys/vmmeter.h>
98 #include <sys/vnode.h>
100 #include <sys/ktrace.h>
104 #include <vm/vm_param.h>
106 #include <vm/vm_map.h>
107 #include <vm/vm_object.h>
108 #include <vm/vm_page.h>
109 #include <vm/vm_pageout.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_pager.h>
112 #include <vm/vm_extern.h>
113 #include <vm/vm_reserv.h>
118 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
120 #define VM_FAULT_DONTNEED_MIN 1048576
123 /* Fault parameters. */
126 vm_prot_t fault_type;
132 struct timeval oom_start_time;
137 /* Page reference for cow. */
140 /* Current object. */
145 /* Top-level map object. */
146 vm_object_t first_object;
147 vm_pindex_t first_pindex;
152 vm_map_entry_t entry;
154 bool lookup_still_valid;
156 /* Vnode if locked. */
161 * Return codes for internal fault routines.
164 FAULT_SUCCESS = 1, /* Return success to user. */
165 FAULT_FAILURE, /* Return failure to user. */
166 FAULT_CONTINUE, /* Continue faulting. */
167 FAULT_RESTART, /* Restart fault. */
168 FAULT_OUT_OF_BOUNDS, /* Invalid address for pager. */
169 FAULT_HARD, /* Performed I/O. */
170 FAULT_SOFT, /* Found valid page. */
171 FAULT_PROTECTION_FAILURE, /* Invalid access. */
174 enum fault_next_status {
175 FAULT_NEXT_GOTOBJ = 1,
180 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
182 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
183 int backward, int forward, bool obj_locked);
185 static int vm_pfault_oom_attempts = 3;
186 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
187 &vm_pfault_oom_attempts, 0,
188 "Number of page allocation attempts in page fault handler before it "
189 "triggers OOM handling");
191 static int vm_pfault_oom_wait = 10;
192 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
193 &vm_pfault_oom_wait, 0,
194 "Number of seconds to wait for free pages before retrying "
195 "the page fault handler");
198 vm_fault_page_release(vm_page_t *mp)
205 * We are likely to loop around again and attempt to busy
206 * this page. Deactivating it leaves it available for
207 * pageout while optimizing fault restarts.
209 vm_page_deactivate(m);
216 vm_fault_page_free(vm_page_t *mp)
222 VM_OBJECT_ASSERT_WLOCKED(m->object);
223 if (!vm_page_wired(m))
232 * Return true if a vm_pager_get_pages() call is needed in order to check
233 * whether the pager might have a particular page, false if it can be determined
234 * immediately that the pager can not have a copy. For swap objects, this can
235 * be checked quickly.
238 vm_fault_object_needs_getpages(vm_object_t object)
240 VM_OBJECT_ASSERT_LOCKED(object);
242 return ((object->flags & OBJ_SWAP) == 0 ||
243 !pctrie_is_empty(&object->un_pager.swp.swp_blks));
247 vm_fault_unlock_map(struct faultstate *fs)
250 if (fs->lookup_still_valid) {
251 vm_map_lookup_done(fs->map, fs->entry);
252 fs->lookup_still_valid = false;
257 vm_fault_unlock_vp(struct faultstate *fs)
260 if (fs->vp != NULL) {
267 vm_fault_deallocate(struct faultstate *fs)
270 vm_fault_page_release(&fs->m_cow);
271 vm_fault_page_release(&fs->m);
272 vm_object_pip_wakeup(fs->object);
273 if (fs->object != fs->first_object) {
274 VM_OBJECT_WLOCK(fs->first_object);
275 vm_fault_page_free(&fs->first_m);
276 VM_OBJECT_WUNLOCK(fs->first_object);
277 vm_object_pip_wakeup(fs->first_object);
279 vm_object_deallocate(fs->first_object);
280 vm_fault_unlock_map(fs);
281 vm_fault_unlock_vp(fs);
285 vm_fault_unlock_and_deallocate(struct faultstate *fs)
288 VM_OBJECT_UNLOCK(fs->object);
289 vm_fault_deallocate(fs);
293 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
297 if (((fs->prot & VM_PROT_WRITE) == 0 &&
298 (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
299 (m->oflags & VPO_UNMANAGED) != 0)
302 VM_PAGE_OBJECT_BUSY_ASSERT(m);
304 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
305 (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
306 (fs->fault_flags & VM_FAULT_DIRTY) != 0;
308 vm_object_set_writeable_dirty(m->object);
311 * If the fault is a write, we know that this page is being
312 * written NOW so dirty it explicitly to save on
313 * pmap_is_modified() calls later.
315 * Also, since the page is now dirty, we can possibly tell
316 * the pager to release any swap backing the page.
318 if (need_dirty && vm_page_set_dirty(m) == 0) {
320 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
321 * if the page is already dirty to prevent data written with
322 * the expectation of being synced from not being synced.
323 * Likewise if this entry does not request NOSYNC then make
324 * sure the page isn't marked NOSYNC. Applications sharing
325 * data should use the same flags to avoid ping ponging.
327 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
328 vm_page_aflag_set(m, PGA_NOSYNC);
330 vm_page_aflag_clear(m, PGA_NOSYNC);
336 * Unlocks fs.first_object and fs.map on success.
338 static enum fault_status
339 vm_fault_soft_fast(struct faultstate *fs)
342 #if VM_NRESERVLEVEL > 0
349 MPASS(fs->vp == NULL);
352 * If we fail, vast majority of the time it is because the page is not
353 * there to begin with. Opportunistically perform the lookup and
354 * subsequent checks without the object lock, revalidate later.
356 * Note: a busy page can be mapped for read|execute access.
358 m = vm_page_lookup_unlocked(fs->first_object, fs->first_pindex);
359 if (m == NULL || !vm_page_all_valid(m) ||
360 ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m))) {
361 VM_OBJECT_WLOCK(fs->first_object);
362 return (FAULT_FAILURE);
367 VM_OBJECT_RLOCK(fs->first_object);
370 * Now that we stabilized the state, revalidate the page is in the shape
371 * we encountered above.
374 if (m->object != fs->first_object || m->pindex != fs->first_pindex)
377 vm_object_busy(fs->first_object);
379 if (!vm_page_all_valid(m) ||
380 ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m)))
385 #if VM_NRESERVLEVEL > 0
386 if ((m->flags & PG_FICTITIOUS) == 0 &&
387 (m_super = vm_reserv_to_superpage(m)) != NULL &&
388 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
389 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
390 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
391 (pagesizes[m_super->psind] - 1)) && !fs->wired &&
392 pmap_ps_enabled(fs->map->pmap)) {
393 flags = PS_ALL_VALID;
394 if ((fs->prot & VM_PROT_WRITE) != 0) {
396 * Create a superpage mapping allowing write access
397 * only if none of the constituent pages are busy and
398 * all of them are already dirty (except possibly for
399 * the page that was faulted on).
401 flags |= PS_NONE_BUSY;
402 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
403 flags |= PS_ALL_DIRTY;
405 if (vm_page_ps_test(m_super, flags, m)) {
407 psind = m_super->psind;
408 vaddr = rounddown2(vaddr, pagesizes[psind]);
409 /* Preset the modified bit for dirty superpages. */
410 if ((flags & PS_ALL_DIRTY) != 0)
411 fs->fault_type |= VM_PROT_WRITE;
415 if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
416 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) !=
419 if (fs->m_hold != NULL) {
423 if (psind == 0 && !fs->wired)
424 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
425 VM_OBJECT_RUNLOCK(fs->first_object);
426 vm_fault_dirty(fs, m);
427 vm_object_unbusy(fs->first_object);
428 vm_map_lookup_done(fs->map, fs->entry);
429 curthread->td_ru.ru_minflt++;
430 return (FAULT_SUCCESS);
432 vm_object_unbusy(fs->first_object);
434 if (!VM_OBJECT_TRYUPGRADE(fs->first_object)) {
435 VM_OBJECT_RUNLOCK(fs->first_object);
436 VM_OBJECT_WLOCK(fs->first_object);
438 return (FAULT_FAILURE);
442 vm_fault_restore_map_lock(struct faultstate *fs)
445 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
446 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
448 if (!vm_map_trylock_read(fs->map)) {
449 VM_OBJECT_WUNLOCK(fs->first_object);
450 vm_map_lock_read(fs->map);
451 VM_OBJECT_WLOCK(fs->first_object);
453 fs->lookup_still_valid = true;
457 vm_fault_populate_check_page(vm_page_t m)
461 * Check each page to ensure that the pager is obeying the
462 * interface: the page must be installed in the object, fully
463 * valid, and exclusively busied.
466 MPASS(vm_page_all_valid(m));
467 MPASS(vm_page_xbusied(m));
471 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
477 VM_OBJECT_ASSERT_WLOCKED(object);
478 MPASS(first <= last);
479 for (pidx = first, m = vm_page_lookup(object, pidx);
480 pidx <= last; pidx++, m = vm_page_next(m)) {
481 vm_fault_populate_check_page(m);
482 vm_page_deactivate(m);
487 static enum fault_status
488 vm_fault_populate(struct faultstate *fs)
492 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
493 int bdry_idx, i, npages, psind, rv;
494 enum fault_status res;
496 MPASS(fs->object == fs->first_object);
497 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
498 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
499 MPASS(fs->first_object->backing_object == NULL);
500 MPASS(fs->lookup_still_valid);
502 pager_first = OFF_TO_IDX(fs->entry->offset);
503 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
504 vm_fault_unlock_map(fs);
505 vm_fault_unlock_vp(fs);
510 * Call the pager (driver) populate() method.
512 * There is no guarantee that the method will be called again
513 * if the current fault is for read, and a future fault is
514 * for write. Report the entry's maximum allowed protection
517 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
518 fs->fault_type, fs->entry->max_protection, &pager_first,
521 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
522 if (rv == VM_PAGER_BAD) {
524 * VM_PAGER_BAD is the backdoor for a pager to request
525 * normal fault handling.
527 vm_fault_restore_map_lock(fs);
528 if (fs->map->timestamp != fs->map_generation)
529 return (FAULT_RESTART);
530 return (FAULT_CONTINUE);
532 if (rv != VM_PAGER_OK)
533 return (FAULT_FAILURE); /* AKA SIGSEGV */
535 /* Ensure that the driver is obeying the interface. */
536 MPASS(pager_first <= pager_last);
537 MPASS(fs->first_pindex <= pager_last);
538 MPASS(fs->first_pindex >= pager_first);
539 MPASS(pager_last < fs->first_object->size);
541 vm_fault_restore_map_lock(fs);
542 bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
543 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
544 if (fs->map->timestamp != fs->map_generation) {
546 vm_fault_populate_cleanup(fs->first_object, pager_first,
549 m = vm_page_lookup(fs->first_object, pager_first);
553 return (FAULT_RESTART);
557 * The map is unchanged after our last unlock. Process the fault.
559 * First, the special case of largepage mappings, where
560 * populate only busies the first page in superpage run.
563 KASSERT(PMAP_HAS_LARGEPAGES,
564 ("missing pmap support for large pages"));
565 m = vm_page_lookup(fs->first_object, pager_first);
566 vm_fault_populate_check_page(m);
567 VM_OBJECT_WUNLOCK(fs->first_object);
568 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
570 /* assert alignment for entry */
571 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
572 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
573 (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
574 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
575 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
576 ("unaligned superpage m %p %#jx", m,
577 (uintmax_t)VM_PAGE_TO_PHYS(m)));
578 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
579 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
580 PMAP_ENTER_LARGEPAGE, bdry_idx);
581 VM_OBJECT_WLOCK(fs->first_object);
583 if (rv != KERN_SUCCESS) {
587 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
588 for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
591 if (fs->m_hold != NULL) {
592 *fs->m_hold = m + (fs->first_pindex - pager_first);
593 vm_page_wire(*fs->m_hold);
599 * The range [pager_first, pager_last] that is given to the
600 * pager is only a hint. The pager may populate any range
601 * within the object that includes the requested page index.
602 * In case the pager expanded the range, clip it to fit into
605 map_first = OFF_TO_IDX(fs->entry->offset);
606 if (map_first > pager_first) {
607 vm_fault_populate_cleanup(fs->first_object, pager_first,
609 pager_first = map_first;
611 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
612 if (map_last < pager_last) {
613 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
615 pager_last = map_last;
617 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
619 pidx += npages, m = vm_page_next(&m[npages - 1])) {
620 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
623 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
624 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
625 !pmap_ps_enabled(fs->map->pmap) || fs->wired))
628 npages = atop(pagesizes[psind]);
629 for (i = 0; i < npages; i++) {
630 vm_fault_populate_check_page(&m[i]);
631 vm_fault_dirty(fs, &m[i]);
633 VM_OBJECT_WUNLOCK(fs->first_object);
634 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
635 (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
638 * pmap_enter() may fail for a superpage mapping if additional
639 * protection policies prevent the full mapping.
640 * For example, this will happen on amd64 if the entire
641 * address range does not share the same userspace protection
642 * key. Revert to single-page mappings if this happens.
644 MPASS(rv == KERN_SUCCESS ||
645 (psind > 0 && rv == KERN_PROTECTION_FAILURE));
646 if (__predict_false(psind > 0 &&
647 rv == KERN_PROTECTION_FAILURE)) {
649 for (i = 0; i < npages; i++) {
650 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
651 &m[i], fs->prot, fs->fault_type, 0);
652 MPASS(rv == KERN_SUCCESS);
656 VM_OBJECT_WLOCK(fs->first_object);
657 for (i = 0; i < npages; i++) {
658 if ((fs->fault_flags & VM_FAULT_WIRE) != 0 &&
659 m[i].pindex == fs->first_pindex)
662 vm_page_activate(&m[i]);
663 if (fs->m_hold != NULL &&
664 m[i].pindex == fs->first_pindex) {
665 (*fs->m_hold) = &m[i];
668 vm_page_xunbusy(&m[i]);
672 curthread->td_ru.ru_majflt++;
676 static int prot_fault_translation;
677 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
678 &prot_fault_translation, 0,
679 "Control signal to deliver on protection fault");
681 /* compat definition to keep common code for signal translation */
682 #define UCODE_PAGEFLT 12
684 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
690 * Handle a page fault occurring at the given address,
691 * requiring the given permissions, in the map specified.
692 * If successful, the page is inserted into the
693 * associated physical map.
695 * NOTE: the given address should be truncated to the
696 * proper page address.
698 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
699 * a standard error specifying why the fault is fatal is returned.
701 * The map in question must be referenced, and remains so.
702 * Caller may hold no locks.
705 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
706 int fault_flags, int *signo, int *ucode)
710 MPASS(signo == NULL || ucode != NULL);
712 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
713 ktrfault(vaddr, fault_type);
715 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
717 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
718 result == KERN_INVALID_ADDRESS ||
719 result == KERN_RESOURCE_SHORTAGE ||
720 result == KERN_PROTECTION_FAILURE ||
721 result == KERN_OUT_OF_BOUNDS,
722 ("Unexpected Mach error %d from vm_fault()", result));
724 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
727 if (result != KERN_SUCCESS && signo != NULL) {
730 case KERN_INVALID_ADDRESS:
732 *ucode = SEGV_MAPERR;
734 case KERN_RESOURCE_SHORTAGE:
738 case KERN_OUT_OF_BOUNDS:
742 case KERN_PROTECTION_FAILURE:
743 if (prot_fault_translation == 0) {
745 * Autodetect. This check also covers
746 * the images without the ABI-tag ELF
749 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
750 curproc->p_osrel >= P_OSREL_SIGSEGV) {
752 *ucode = SEGV_ACCERR;
755 *ucode = UCODE_PAGEFLT;
757 } else if (prot_fault_translation == 1) {
758 /* Always compat mode. */
760 *ucode = UCODE_PAGEFLT;
762 /* Always SIGSEGV mode. */
764 *ucode = SEGV_ACCERR;
768 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
777 vm_fault_object_ensure_wlocked(struct faultstate *fs)
779 if (fs->object == fs->first_object)
780 VM_OBJECT_ASSERT_WLOCKED(fs->object);
782 if (!fs->can_read_lock) {
783 VM_OBJECT_ASSERT_WLOCKED(fs->object);
787 if (VM_OBJECT_WOWNED(fs->object))
790 if (VM_OBJECT_TRYUPGRADE(fs->object))
796 static enum fault_status
797 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
802 if (fs->object->type != OBJT_VNODE)
803 return (FAULT_CONTINUE);
804 vp = fs->object->handle;
806 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
807 return (FAULT_CONTINUE);
811 * Perform an unlock in case the desired vnode changed while
812 * the map was unlocked during a retry.
814 vm_fault_unlock_vp(fs);
816 locked = VOP_ISLOCKED(vp);
817 if (locked != LK_EXCLUSIVE)
821 * We must not sleep acquiring the vnode lock while we have
822 * the page exclusive busied or the object's
823 * paging-in-progress count incremented. Otherwise, we could
826 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
829 return (FAULT_CONTINUE);
834 vm_fault_unlock_and_deallocate(fs);
836 vm_fault_deallocate(fs);
837 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
840 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
841 return (FAULT_RESTART);
845 * Calculate the desired readahead. Handle drop-behind.
847 * Returns the number of readahead blocks to pass to the pager.
850 vm_fault_readahead(struct faultstate *fs)
855 KASSERT(fs->lookup_still_valid, ("map unlocked"));
856 era = fs->entry->read_ahead;
857 behavior = vm_map_entry_behavior(fs->entry);
858 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
860 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
861 nera = VM_FAULT_READ_AHEAD_MAX;
862 if (fs->vaddr == fs->entry->next_read)
863 vm_fault_dontneed(fs, fs->vaddr, nera);
864 } else if (fs->vaddr == fs->entry->next_read) {
866 * This is a sequential fault. Arithmetically
867 * increase the requested number of pages in
868 * the read-ahead window. The requested
869 * number of pages is "# of sequential faults
870 * x (read ahead min + 1) + read ahead min"
872 nera = VM_FAULT_READ_AHEAD_MIN;
875 if (nera > VM_FAULT_READ_AHEAD_MAX)
876 nera = VM_FAULT_READ_AHEAD_MAX;
878 if (era == VM_FAULT_READ_AHEAD_MAX)
879 vm_fault_dontneed(fs, fs->vaddr, nera);
882 * This is a non-sequential fault.
888 * A read lock on the map suffices to update
889 * the read ahead count safely.
891 fs->entry->read_ahead = nera;
898 vm_fault_lookup(struct faultstate *fs)
902 KASSERT(!fs->lookup_still_valid,
903 ("vm_fault_lookup: Map already locked."));
904 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
905 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
906 &fs->first_pindex, &fs->prot, &fs->wired);
907 if (result != KERN_SUCCESS) {
908 vm_fault_unlock_vp(fs);
912 fs->map_generation = fs->map->timestamp;
914 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
915 panic("%s: fault on nofault entry, addr: %#lx",
916 __func__, (u_long)fs->vaddr);
919 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
920 fs->entry->wiring_thread != curthread) {
921 vm_map_unlock_read(fs->map);
922 vm_map_lock(fs->map);
923 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
924 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
925 vm_fault_unlock_vp(fs);
926 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
927 vm_map_unlock_and_wait(fs->map, 0);
929 vm_map_unlock(fs->map);
930 return (KERN_RESOURCE_SHORTAGE);
933 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
936 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
938 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
939 ("!fs->wired && VM_FAULT_WIRE"));
940 fs->lookup_still_valid = true;
942 return (KERN_SUCCESS);
946 vm_fault_relookup(struct faultstate *fs)
948 vm_object_t retry_object;
949 vm_pindex_t retry_pindex;
950 vm_prot_t retry_prot;
953 if (!vm_map_trylock_read(fs->map))
954 return (KERN_RESTART);
956 fs->lookup_still_valid = true;
957 if (fs->map->timestamp == fs->map_generation)
958 return (KERN_SUCCESS);
960 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
961 &fs->entry, &retry_object, &retry_pindex, &retry_prot,
963 if (result != KERN_SUCCESS) {
965 * If retry of map lookup would have blocked then
966 * retry fault from start.
968 if (result == KERN_FAILURE)
969 return (KERN_RESTART);
972 if (retry_object != fs->first_object ||
973 retry_pindex != fs->first_pindex)
974 return (KERN_RESTART);
977 * Check whether the protection has changed or the object has
978 * been copied while we left the map unlocked. Changing from
979 * read to write permission is OK - we leave the page
980 * write-protected, and catch the write fault. Changing from
981 * write to read permission means that we can't mark the page
982 * write-enabled after all.
984 fs->prot &= retry_prot;
985 fs->fault_type &= retry_prot;
987 return (KERN_RESTART);
989 /* Reassert because wired may have changed. */
990 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
991 ("!wired && VM_FAULT_WIRE"));
993 return (KERN_SUCCESS);
997 vm_fault_cow(struct faultstate *fs)
999 bool is_first_object_locked;
1001 KASSERT(fs->object != fs->first_object,
1002 ("source and target COW objects are identical"));
1005 * This allows pages to be virtually copied from a backing_object
1006 * into the first_object, where the backing object has no other
1007 * refs to it, and cannot gain any more refs. Instead of a bcopy,
1008 * we just move the page from the backing object to the first
1009 * object. Note that we must mark the page dirty in the first
1010 * object so that it will go out to swap when needed.
1012 is_first_object_locked = false;
1015 * Only one shadow object and no other refs.
1017 fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
1019 * No other ways to look the object up
1021 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
1023 * We don't chase down the shadow chain and we can acquire locks.
1025 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
1026 fs->object == fs->first_object->backing_object &&
1027 VM_OBJECT_TRYWLOCK(fs->object)) {
1029 * Remove but keep xbusy for replace. fs->m is moved into
1030 * fs->first_object and left busy while fs->first_m is
1031 * conditionally freed.
1033 vm_page_remove_xbusy(fs->m);
1034 vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
1036 vm_page_dirty(fs->m);
1037 #if VM_NRESERVLEVEL > 0
1039 * Rename the reservation.
1041 vm_reserv_rename(fs->m, fs->first_object, fs->object,
1042 OFF_TO_IDX(fs->first_object->backing_object_offset));
1044 VM_OBJECT_WUNLOCK(fs->object);
1045 VM_OBJECT_WUNLOCK(fs->first_object);
1046 fs->first_m = fs->m;
1048 VM_CNT_INC(v_cow_optim);
1050 if (is_first_object_locked)
1051 VM_OBJECT_WUNLOCK(fs->first_object);
1053 * Oh, well, lets copy it.
1055 pmap_copy_page(fs->m, fs->first_m);
1056 vm_page_valid(fs->first_m);
1057 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
1058 vm_page_wire(fs->first_m);
1059 vm_page_unwire(fs->m, PQ_INACTIVE);
1062 * Save the cow page to be released after
1063 * pmap_enter is complete.
1069 * Typically, the shadow object is either private to this
1070 * address space (OBJ_ONEMAPPING) or its pages are read only.
1071 * In the highly unusual case where the pages of a shadow object
1072 * are read/write shared between this and other address spaces,
1073 * we need to ensure that any pmap-level mappings to the
1074 * original, copy-on-write page from the backing object are
1075 * removed from those other address spaces.
1077 * The flag check is racy, but this is tolerable: if
1078 * OBJ_ONEMAPPING is cleared after the check, the busy state
1079 * ensures that new mappings of m_cow can't be created.
1080 * pmap_enter() will replace an existing mapping in the current
1081 * address space. If OBJ_ONEMAPPING is set after the check,
1082 * removing mappings will at worse trigger some unnecessary page
1085 vm_page_assert_xbusied(fs->m_cow);
1086 if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
1087 pmap_remove_all(fs->m_cow);
1090 vm_object_pip_wakeup(fs->object);
1093 * Only use the new page below...
1095 fs->object = fs->first_object;
1096 fs->pindex = fs->first_pindex;
1097 fs->m = fs->first_m;
1098 VM_CNT_INC(v_cow_faults);
1099 curthread->td_cow++;
1102 static enum fault_next_status
1103 vm_fault_next(struct faultstate *fs)
1105 vm_object_t next_object;
1107 if (fs->object == fs->first_object || !fs->can_read_lock)
1108 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1110 VM_OBJECT_ASSERT_LOCKED(fs->object);
1113 * The requested page does not exist at this object/
1114 * offset. Remove the invalid page from the object,
1115 * waking up anyone waiting for it, and continue on to
1116 * the next object. However, if this is the top-level
1117 * object, we must leave the busy page in place to
1118 * prevent another process from rushing past us, and
1119 * inserting the page in that object at the same time
1122 if (fs->object == fs->first_object) {
1123 fs->first_m = fs->m;
1125 } else if (fs->m != NULL) {
1126 if (!vm_fault_object_ensure_wlocked(fs)) {
1127 fs->can_read_lock = false;
1128 vm_fault_unlock_and_deallocate(fs);
1129 return (FAULT_NEXT_RESTART);
1131 vm_fault_page_free(&fs->m);
1135 * Move on to the next object. Lock the next object before
1136 * unlocking the current one.
1138 next_object = fs->object->backing_object;
1139 if (next_object == NULL)
1140 return (FAULT_NEXT_NOOBJ);
1141 MPASS(fs->first_m != NULL);
1142 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1143 if (fs->can_read_lock)
1144 VM_OBJECT_RLOCK(next_object);
1146 VM_OBJECT_WLOCK(next_object);
1147 vm_object_pip_add(next_object, 1);
1148 if (fs->object != fs->first_object)
1149 vm_object_pip_wakeup(fs->object);
1150 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1151 VM_OBJECT_UNLOCK(fs->object);
1152 fs->object = next_object;
1154 return (FAULT_NEXT_GOTOBJ);
1158 vm_fault_zerofill(struct faultstate *fs)
1162 * If there's no object left, fill the page in the top
1163 * object with zeros.
1165 if (fs->object != fs->first_object) {
1166 vm_object_pip_wakeup(fs->object);
1167 fs->object = fs->first_object;
1168 fs->pindex = fs->first_pindex;
1170 MPASS(fs->first_m != NULL);
1171 MPASS(fs->m == NULL);
1172 fs->m = fs->first_m;
1176 * Zero the page if necessary and mark it valid.
1178 if ((fs->m->flags & PG_ZERO) == 0) {
1179 pmap_zero_page(fs->m);
1181 VM_CNT_INC(v_ozfod);
1184 vm_page_valid(fs->m);
1188 * Initiate page fault after timeout. Returns true if caller should
1189 * do vm_waitpfault() after the call.
1192 vm_fault_allocate_oom(struct faultstate *fs)
1196 vm_fault_unlock_and_deallocate(fs);
1197 if (vm_pfault_oom_attempts < 0)
1199 if (!fs->oom_started) {
1200 fs->oom_started = true;
1201 getmicrotime(&fs->oom_start_time);
1206 timevalsub(&now, &fs->oom_start_time);
1207 if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait)
1212 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1213 curproc->p_pid, curproc->p_comm);
1214 vm_pageout_oom(VM_OOM_MEM_PF);
1215 fs->oom_started = false;
1220 * Allocate a page directly or via the object populate method.
1222 static enum fault_status
1223 vm_fault_allocate(struct faultstate *fs)
1225 struct domainset *dset;
1226 enum fault_status res;
1228 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1229 res = vm_fault_lock_vnode(fs, true);
1230 MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART);
1231 if (res == FAULT_RESTART)
1235 if (fs->pindex >= fs->object->size) {
1236 vm_fault_unlock_and_deallocate(fs);
1237 return (FAULT_OUT_OF_BOUNDS);
1240 if (fs->object == fs->first_object &&
1241 (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1242 fs->first_object->shadow_count == 0) {
1243 res = vm_fault_populate(fs);
1248 vm_fault_unlock_and_deallocate(fs);
1250 case FAULT_CONTINUE:
1252 * Pager's populate() method
1253 * returned VM_PAGER_BAD.
1257 panic("inconsistent return codes");
1262 * Allocate a new page for this object/offset pair.
1264 * If the process has a fatal signal pending, prioritize the allocation
1265 * with the expectation that the process will exit shortly and free some
1266 * pages. In particular, the signal may have been posted by the page
1267 * daemon in an attempt to resolve an out-of-memory condition.
1269 * The unlocked read of the p_flag is harmless. At worst, the P_KILLED
1270 * might be not observed here, and allocation fails, causing a restart
1271 * and new reading of the p_flag.
1273 dset = fs->object->domain.dr_policy;
1275 dset = curthread->td_domain.dr_policy;
1276 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1277 #if VM_NRESERVLEVEL > 0
1278 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1280 if (!vm_pager_can_alloc_page(fs->object, fs->pindex)) {
1281 vm_fault_unlock_and_deallocate(fs);
1282 return (FAULT_FAILURE);
1284 fs->m = vm_page_alloc(fs->object, fs->pindex,
1285 P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0);
1287 if (fs->m == NULL) {
1288 if (vm_fault_allocate_oom(fs))
1289 vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1290 return (FAULT_RESTART);
1292 fs->oom_started = false;
1294 return (FAULT_CONTINUE);
1298 * Call the pager to retrieve the page if there is a chance
1299 * that the pager has it, and potentially retrieve additional
1300 * pages at the same time.
1302 static enum fault_status
1303 vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp)
1305 vm_offset_t e_end, e_start;
1306 int ahead, behind, cluster_offset, rv;
1307 enum fault_status status;
1311 * Prepare for unlocking the map. Save the map
1312 * entry's start and end addresses, which are used to
1313 * optimize the size of the pager operation below.
1314 * Even if the map entry's addresses change after
1315 * unlocking the map, using the saved addresses is
1318 e_start = fs->entry->start;
1319 e_end = fs->entry->end;
1320 behavior = vm_map_entry_behavior(fs->entry);
1323 * If the pager for the current object might have
1324 * the page, then determine the number of additional
1325 * pages to read and potentially reprioritize
1326 * previously read pages for earlier reclamation.
1327 * These operations should only be performed once per
1328 * page fault. Even if the current pager doesn't
1329 * have the page, the number of additional pages to
1330 * read will apply to subsequent objects in the
1333 if (fs->nera == -1 && !P_KILLED(curproc))
1334 fs->nera = vm_fault_readahead(fs);
1337 * Release the map lock before locking the vnode or
1338 * sleeping in the pager. (If the current object has
1339 * a shadow, then an earlier iteration of this loop
1340 * may have already unlocked the map.)
1342 vm_fault_unlock_map(fs);
1344 status = vm_fault_lock_vnode(fs, false);
1345 MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART);
1346 if (status == FAULT_RESTART)
1348 KASSERT(fs->vp == NULL || !fs->map->system_map,
1349 ("vm_fault: vnode-backed object mapped by system map"));
1352 * Page in the requested page and hint the pager,
1353 * that it may bring up surrounding pages.
1355 if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1356 P_KILLED(curproc)) {
1360 /* Is this a sequential fault? */
1366 * Request a cluster of pages that is
1367 * aligned to a VM_FAULT_READ_DEFAULT
1368 * page offset boundary within the
1369 * object. Alignment to a page offset
1370 * boundary is more likely to coincide
1371 * with the underlying file system
1372 * block than alignment to a virtual
1375 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1376 behind = ulmin(cluster_offset,
1377 atop(fs->vaddr - e_start));
1378 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1380 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1384 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1385 if (rv == VM_PAGER_OK)
1386 return (FAULT_HARD);
1387 if (rv == VM_PAGER_ERROR)
1388 printf("vm_fault: pager read error, pid %d (%s)\n",
1389 curproc->p_pid, curproc->p_comm);
1391 * If an I/O error occurred or the requested page was
1392 * outside the range of the pager, clean up and return
1395 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1396 VM_OBJECT_WLOCK(fs->object);
1397 vm_fault_page_free(&fs->m);
1398 vm_fault_unlock_and_deallocate(fs);
1399 return (FAULT_OUT_OF_BOUNDS);
1401 KASSERT(rv == VM_PAGER_FAIL,
1402 ("%s: unexpected pager error %d", __func__, rv));
1403 return (FAULT_CONTINUE);
1407 * Wait/Retry if the page is busy. We have to do this if the page is
1408 * either exclusive or shared busy because the vm_pager may be using
1409 * read busy for pageouts (and even pageins if it is the vnode pager),
1410 * and we could end up trying to pagein and pageout the same page
1413 * We can theoretically allow the busy case on a read fault if the page
1414 * is marked valid, but since such pages are typically already pmap'd,
1415 * putting that special case in might be more effort then it is worth.
1416 * We cannot under any circumstances mess around with a shared busied
1417 * page except, perhaps, to pmap it.
1420 vm_fault_busy_sleep(struct faultstate *fs)
1423 * Reference the page before unlocking and
1424 * sleeping so that the page daemon is less
1425 * likely to reclaim it.
1427 vm_page_aflag_set(fs->m, PGA_REFERENCED);
1428 if (fs->object != fs->first_object) {
1429 vm_fault_page_release(&fs->first_m);
1430 vm_object_pip_wakeup(fs->first_object);
1432 vm_object_pip_wakeup(fs->object);
1433 vm_fault_unlock_map(fs);
1434 if (fs->m != vm_page_lookup(fs->object, fs->pindex) ||
1435 !vm_page_busy_sleep(fs->m, "vmpfw", 0))
1436 VM_OBJECT_UNLOCK(fs->object);
1437 VM_CNT_INC(v_intrans);
1438 vm_object_deallocate(fs->first_object);
1442 * Handle page lookup, populate, allocate, page-in for the current
1445 * The object is locked on entry and will remain locked with a return
1446 * code of FAULT_CONTINUE so that fault may follow the shadow chain.
1447 * Otherwise, the object will be unlocked upon return.
1449 static enum fault_status
1450 vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp)
1452 enum fault_status res;
1455 if (fs->object == fs->first_object || !fs->can_read_lock)
1456 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1458 VM_OBJECT_ASSERT_LOCKED(fs->object);
1461 * If the object is marked for imminent termination, we retry
1462 * here, since the collapse pass has raced with us. Otherwise,
1463 * if we see terminally dead object, return fail.
1465 if ((fs->object->flags & OBJ_DEAD) != 0) {
1466 dead = fs->object->type == OBJT_DEAD;
1467 vm_fault_unlock_and_deallocate(fs);
1469 return (FAULT_PROTECTION_FAILURE);
1471 return (FAULT_RESTART);
1475 * See if the page is resident.
1477 fs->m = vm_page_lookup(fs->object, fs->pindex);
1478 if (fs->m != NULL) {
1479 if (!vm_page_tryxbusy(fs->m)) {
1480 vm_fault_busy_sleep(fs);
1481 return (FAULT_RESTART);
1485 * The page is marked busy for other processes and the
1486 * pagedaemon. If it is still completely valid we are
1489 if (vm_page_all_valid(fs->m)) {
1490 VM_OBJECT_UNLOCK(fs->object);
1491 return (FAULT_SOFT);
1496 * Page is not resident. If the pager might contain the page
1497 * or this is the beginning of the search, allocate a new
1500 if (fs->m == NULL && (vm_fault_object_needs_getpages(fs->object) ||
1501 fs->object == fs->first_object)) {
1502 if (!vm_fault_object_ensure_wlocked(fs)) {
1503 fs->can_read_lock = false;
1504 vm_fault_unlock_and_deallocate(fs);
1505 return (FAULT_RESTART);
1507 res = vm_fault_allocate(fs);
1508 if (res != FAULT_CONTINUE)
1513 * Check to see if the pager can possibly satisfy this fault.
1514 * If not, skip to the next object without dropping the lock to
1515 * preserve atomicity of shadow faults.
1517 if (vm_fault_object_needs_getpages(fs->object)) {
1519 * At this point, we have either allocated a new page
1520 * or found an existing page that is only partially
1523 * We hold a reference on the current object and the
1524 * page is exclusive busied. The exclusive busy
1525 * prevents simultaneous faults and collapses while
1526 * the object lock is dropped.
1528 VM_OBJECT_UNLOCK(fs->object);
1529 res = vm_fault_getpages(fs, behindp, aheadp);
1530 if (res == FAULT_CONTINUE)
1531 VM_OBJECT_WLOCK(fs->object);
1533 res = FAULT_CONTINUE;
1539 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1540 int fault_flags, vm_page_t *m_hold)
1542 struct faultstate fs;
1543 int ahead, behind, faultcount, rv;
1544 enum fault_status res;
1545 enum fault_next_status res_next;
1548 VM_CNT_INC(v_vm_faults);
1550 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1551 return (KERN_PROTECTION_FAILURE);
1556 fs.fault_flags = fault_flags;
1558 fs.lookup_still_valid = false;
1559 fs.oom_started = false;
1561 fs.can_read_lock = true;
1566 fs.fault_type = fault_type;
1569 * Find the backing store object and offset into it to begin the
1572 rv = vm_fault_lookup(&fs);
1573 if (rv != KERN_SUCCESS) {
1574 if (rv == KERN_RESOURCE_SHORTAGE)
1580 * Try to avoid lock contention on the top-level object through
1581 * special-case handling of some types of page faults, specifically,
1582 * those that are mapping an existing page from the top-level object.
1583 * Under this condition, a read lock on the object suffices, allowing
1584 * multiple page faults of a similar type to run in parallel.
1586 if (fs.vp == NULL /* avoid locked vnode leak */ &&
1587 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1588 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1589 res = vm_fault_soft_fast(&fs);
1590 if (res == FAULT_SUCCESS) {
1591 VM_OBJECT_ASSERT_UNLOCKED(fs.first_object);
1592 return (KERN_SUCCESS);
1594 VM_OBJECT_ASSERT_WLOCKED(fs.first_object);
1596 VM_OBJECT_WLOCK(fs.first_object);
1600 * Make a reference to this object to prevent its disposal while we
1601 * are messing with it. Once we have the reference, the map is free
1602 * to be diddled. Since objects reference their shadows (and copies),
1603 * they will stay around as well.
1605 * Bump the paging-in-progress count to prevent size changes (e.g.
1606 * truncation operations) during I/O.
1608 vm_object_reference_locked(fs.first_object);
1609 vm_object_pip_add(fs.first_object, 1);
1611 fs.m_cow = fs.m = fs.first_m = NULL;
1614 * Search for the page at object/offset.
1616 fs.object = fs.first_object;
1617 fs.pindex = fs.first_pindex;
1619 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1620 res = vm_fault_allocate(&fs);
1625 return (KERN_SUCCESS);
1627 return (KERN_FAILURE);
1628 case FAULT_OUT_OF_BOUNDS:
1629 return (KERN_OUT_OF_BOUNDS);
1630 case FAULT_CONTINUE:
1633 panic("vm_fault: Unhandled status %d", res);
1638 KASSERT(fs.m == NULL,
1639 ("page still set %p at loop start", fs.m));
1641 res = vm_fault_object(&fs, &behind, &ahead);
1646 faultcount = behind + 1 + ahead;
1652 return (KERN_SUCCESS);
1654 return (KERN_FAILURE);
1655 case FAULT_OUT_OF_BOUNDS:
1656 return (KERN_OUT_OF_BOUNDS);
1657 case FAULT_PROTECTION_FAILURE:
1658 return (KERN_PROTECTION_FAILURE);
1659 case FAULT_CONTINUE:
1662 panic("vm_fault: Unhandled status %d", res);
1666 * The page was not found in the current object. Try to
1667 * traverse into a backing object or zero fill if none is
1670 res_next = vm_fault_next(&fs);
1671 if (res_next == FAULT_NEXT_RESTART)
1673 else if (res_next == FAULT_NEXT_GOTOBJ)
1675 MPASS(res_next == FAULT_NEXT_NOOBJ);
1676 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1677 if (fs.first_object == fs.object)
1678 vm_fault_page_free(&fs.first_m);
1679 vm_fault_unlock_and_deallocate(&fs);
1680 return (KERN_OUT_OF_BOUNDS);
1682 VM_OBJECT_UNLOCK(fs.object);
1683 vm_fault_zerofill(&fs);
1684 /* Don't try to prefault neighboring pages. */
1691 * A valid page has been found and exclusively busied. The
1692 * object lock must no longer be held.
1694 vm_page_assert_xbusied(fs.m);
1695 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1698 * If the page is being written, but isn't already owned by the
1699 * top-level object, we have to copy it into a new page owned by the
1702 if (fs.object != fs.first_object) {
1704 * We only really need to copy if we want to write it.
1706 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1709 * We only try to prefault read-only mappings to the
1710 * neighboring pages when this copy-on-write fault is
1711 * a hard fault. In other cases, trying to prefault
1712 * is typically wasted effort.
1714 if (faultcount == 0)
1718 fs.prot &= ~VM_PROT_WRITE;
1723 * We must verify that the maps have not changed since our last
1726 if (!fs.lookup_still_valid) {
1727 rv = vm_fault_relookup(&fs);
1728 if (rv != KERN_SUCCESS) {
1729 vm_fault_deallocate(&fs);
1730 if (rv == KERN_RESTART)
1735 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1738 * If the page was filled by a pager, save the virtual address that
1739 * should be faulted on next under a sequential access pattern to the
1740 * map entry. A read lock on the map suffices to update this address
1744 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1747 * Page must be completely valid or it is not fit to
1748 * map into user space. vm_pager_get_pages() ensures this.
1750 vm_page_assert_xbusied(fs.m);
1751 KASSERT(vm_page_all_valid(fs.m),
1752 ("vm_fault: page %p partially invalid", fs.m));
1754 vm_fault_dirty(&fs, fs.m);
1757 * Put this page into the physical map. We had to do the unlock above
1758 * because pmap_enter() may sleep. We don't put the page
1759 * back on the active queue until later so that the pageout daemon
1760 * won't find it (yet).
1762 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1763 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1764 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1766 vm_fault_prefault(&fs, vaddr,
1767 faultcount > 0 ? behind : PFBAK,
1768 faultcount > 0 ? ahead : PFFOR, false);
1771 * If the page is not wired down, then put it where the pageout daemon
1774 if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1777 vm_page_activate(fs.m);
1778 if (fs.m_hold != NULL) {
1779 (*fs.m_hold) = fs.m;
1782 vm_page_xunbusy(fs.m);
1786 * Unlock everything, and return
1788 vm_fault_deallocate(&fs);
1790 VM_CNT_INC(v_io_faults);
1791 curthread->td_ru.ru_majflt++;
1793 if (racct_enable && fs.object->type == OBJT_VNODE) {
1795 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1796 racct_add_force(curproc, RACCT_WRITEBPS,
1797 PAGE_SIZE + behind * PAGE_SIZE);
1798 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1800 racct_add_force(curproc, RACCT_READBPS,
1801 PAGE_SIZE + ahead * PAGE_SIZE);
1802 racct_add_force(curproc, RACCT_READIOPS, 1);
1804 PROC_UNLOCK(curproc);
1808 curthread->td_ru.ru_minflt++;
1810 return (KERN_SUCCESS);
1814 * Speed up the reclamation of pages that precede the faulting pindex within
1815 * the first object of the shadow chain. Essentially, perform the equivalent
1816 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1817 * the faulting pindex by the cluster size when the pages read by vm_fault()
1818 * cross a cluster-size boundary. The cluster size is the greater of the
1819 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1821 * When "fs->first_object" is a shadow object, the pages in the backing object
1822 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1823 * function must only be concerned with pages in the first object.
1826 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1828 vm_map_entry_t entry;
1829 vm_object_t first_object;
1830 vm_offset_t end, start;
1831 vm_page_t m, m_next;
1832 vm_pindex_t pend, pstart;
1835 VM_OBJECT_ASSERT_UNLOCKED(fs->object);
1836 first_object = fs->first_object;
1837 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1838 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1839 VM_OBJECT_RLOCK(first_object);
1840 size = VM_FAULT_DONTNEED_MIN;
1841 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1842 size = pagesizes[1];
1843 end = rounddown2(vaddr, size);
1844 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1845 (entry = fs->entry)->start < end) {
1846 if (end - entry->start < size)
1847 start = entry->start;
1850 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1851 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1853 m_next = vm_page_find_least(first_object, pstart);
1854 pend = OFF_TO_IDX(entry->offset) + atop(end -
1856 while ((m = m_next) != NULL && m->pindex < pend) {
1857 m_next = TAILQ_NEXT(m, listq);
1858 if (!vm_page_all_valid(m) ||
1863 * Don't clear PGA_REFERENCED, since it would
1864 * likely represent a reference by a different
1867 * Typically, at this point, prefetched pages
1868 * are still in the inactive queue. Only
1869 * pages that triggered page faults are in the
1870 * active queue. The test for whether the page
1871 * is in the inactive queue is racy; in the
1872 * worst case we will requeue the page
1875 if (!vm_page_inactive(m))
1876 vm_page_deactivate(m);
1879 VM_OBJECT_RUNLOCK(first_object);
1884 * vm_fault_prefault provides a quick way of clustering
1885 * pagefaults into a processes address space. It is a "cousin"
1886 * of vm_map_pmap_enter, except it runs at page fault time instead
1890 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1891 int backward, int forward, bool obj_locked)
1894 vm_map_entry_t entry;
1895 vm_object_t backing_object, lobject;
1896 vm_offset_t addr, starta;
1901 pmap = fs->map->pmap;
1902 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1907 if (addra < backward * PAGE_SIZE) {
1908 starta = entry->start;
1910 starta = addra - backward * PAGE_SIZE;
1911 if (starta < entry->start)
1912 starta = entry->start;
1916 * Generate the sequence of virtual addresses that are candidates for
1917 * prefaulting in an outward spiral from the faulting virtual address,
1918 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1919 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1920 * If the candidate address doesn't have a backing physical page, then
1921 * the loop immediately terminates.
1923 for (i = 0; i < 2 * imax(backward, forward); i++) {
1924 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1926 if (addr > addra + forward * PAGE_SIZE)
1929 if (addr < starta || addr >= entry->end)
1932 if (!pmap_is_prefaultable(pmap, addr))
1935 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1936 lobject = entry->object.vm_object;
1938 VM_OBJECT_RLOCK(lobject);
1939 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1940 !vm_fault_object_needs_getpages(lobject) &&
1941 (backing_object = lobject->backing_object) != NULL) {
1942 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1943 0, ("vm_fault_prefault: unaligned object offset"));
1944 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1945 VM_OBJECT_RLOCK(backing_object);
1946 if (!obj_locked || lobject != entry->object.vm_object)
1947 VM_OBJECT_RUNLOCK(lobject);
1948 lobject = backing_object;
1951 if (!obj_locked || lobject != entry->object.vm_object)
1952 VM_OBJECT_RUNLOCK(lobject);
1955 if (vm_page_all_valid(m) &&
1956 (m->flags & PG_FICTITIOUS) == 0)
1957 pmap_enter_quick(pmap, addr, m, entry->protection);
1958 if (!obj_locked || lobject != entry->object.vm_object)
1959 VM_OBJECT_RUNLOCK(lobject);
1964 * Hold each of the physical pages that are mapped by the specified range of
1965 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1966 * and allow the specified types of access, "prot". If all of the implied
1967 * pages are successfully held, then the number of held pages is returned
1968 * together with pointers to those pages in the array "ma". However, if any
1969 * of the pages cannot be held, -1 is returned.
1972 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1973 vm_prot_t prot, vm_page_t *ma, int max_count)
1975 vm_offset_t end, va;
1978 boolean_t pmap_failed;
1982 end = round_page(addr + len);
1983 addr = trunc_page(addr);
1985 if (!vm_map_range_valid(map, addr, end))
1988 if (atop(end - addr) > max_count)
1989 panic("vm_fault_quick_hold_pages: count > max_count");
1990 count = atop(end - addr);
1993 * Most likely, the physical pages are resident in the pmap, so it is
1994 * faster to try pmap_extract_and_hold() first.
1996 pmap_failed = FALSE;
1997 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1998 *mp = pmap_extract_and_hold(map->pmap, va, prot);
2001 else if ((prot & VM_PROT_WRITE) != 0 &&
2002 (*mp)->dirty != VM_PAGE_BITS_ALL) {
2004 * Explicitly dirty the physical page. Otherwise, the
2005 * caller's changes may go unnoticed because they are
2006 * performed through an unmanaged mapping or by a DMA
2009 * The object lock is not held here.
2010 * See vm_page_clear_dirty_mask().
2017 * One or more pages could not be held by the pmap. Either no
2018 * page was mapped at the specified virtual address or that
2019 * mapping had insufficient permissions. Attempt to fault in
2020 * and hold these pages.
2022 * If vm_fault_disable_pagefaults() was called,
2023 * i.e., TDP_NOFAULTING is set, we must not sleep nor
2024 * acquire MD VM locks, which means we must not call
2025 * vm_fault(). Some (out of tree) callers mark
2026 * too wide a code area with vm_fault_disable_pagefaults()
2027 * already, use the VM_PROT_QUICK_NOFAULT flag to request
2028 * the proper behaviour explicitly.
2030 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
2031 (curthread->td_pflags & TDP_NOFAULTING) != 0)
2033 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
2034 if (*mp == NULL && vm_fault(map, va, prot,
2035 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
2040 for (mp = ma; mp < ma + count; mp++)
2042 vm_page_unwire(*mp, PQ_INACTIVE);
2048 * vm_fault_copy_entry
2050 * Create new object backing dst_entry with private copy of all
2051 * underlying pages. When src_entry is equal to dst_entry, function
2052 * implements COW for wired-down map entry. Otherwise, it forks
2053 * wired entry into dst_map.
2055 * In/out conditions:
2056 * The source and destination maps must be locked for write.
2057 * The source map entry must be wired down (or be a sharing map
2058 * entry corresponding to a main map entry that is wired down).
2061 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map __unused,
2062 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
2063 vm_ooffset_t *fork_charge)
2065 vm_object_t backing_object, dst_object, object, src_object;
2066 vm_pindex_t dst_pindex, pindex, src_pindex;
2067 vm_prot_t access, prot;
2073 upgrade = src_entry == dst_entry;
2074 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
2075 ("vm_fault_copy_entry: vm_object not NULL"));
2078 * If not an upgrade, then enter the mappings in the pmap as
2079 * read and/or execute accesses. Otherwise, enter them as
2082 * A writeable large page mapping is only created if all of
2083 * the constituent small page mappings are modified. Marking
2084 * PTEs as modified on inception allows promotion to happen
2085 * without taking potentially large number of soft faults.
2087 access = prot = dst_entry->protection;
2089 access &= ~VM_PROT_WRITE;
2091 src_object = src_entry->object.vm_object;
2092 src_pindex = OFF_TO_IDX(src_entry->offset);
2094 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2095 dst_object = src_object;
2096 vm_object_reference(dst_object);
2099 * Create the top-level object for the destination entry.
2100 * Doesn't actually shadow anything - we copy the pages
2103 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
2104 dst_entry->start), NULL, NULL, 0);
2105 #if VM_NRESERVLEVEL > 0
2106 dst_object->flags |= OBJ_COLORED;
2107 dst_object->pg_color = atop(dst_entry->start);
2109 dst_object->domain = src_object->domain;
2110 dst_object->charge = dst_entry->end - dst_entry->start;
2112 dst_entry->object.vm_object = dst_object;
2113 dst_entry->offset = 0;
2114 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
2117 VM_OBJECT_WLOCK(dst_object);
2118 if (fork_charge != NULL) {
2119 KASSERT(dst_entry->cred == NULL,
2120 ("vm_fault_copy_entry: leaked swp charge"));
2121 dst_object->cred = curthread->td_ucred;
2122 crhold(dst_object->cred);
2123 *fork_charge += dst_object->charge;
2124 } else if ((dst_object->flags & OBJ_SWAP) != 0 &&
2125 dst_object->cred == NULL) {
2126 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
2128 dst_object->cred = dst_entry->cred;
2129 dst_entry->cred = NULL;
2133 * Loop through all of the virtual pages within the entry's
2134 * range, copying each page from the source object to the
2135 * destination object. Since the source is wired, those pages
2136 * must exist. In contrast, the destination is pageable.
2137 * Since the destination object doesn't share any backing storage
2138 * with the source object, all of its pages must be dirtied,
2139 * regardless of whether they can be written.
2141 for (vaddr = dst_entry->start, dst_pindex = 0;
2142 vaddr < dst_entry->end;
2143 vaddr += PAGE_SIZE, dst_pindex++) {
2146 * Find the page in the source object, and copy it in.
2147 * Because the source is wired down, the page will be
2150 if (src_object != dst_object)
2151 VM_OBJECT_RLOCK(src_object);
2152 object = src_object;
2153 pindex = src_pindex + dst_pindex;
2154 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
2155 (backing_object = object->backing_object) != NULL) {
2157 * Unless the source mapping is read-only or
2158 * it is presently being upgraded from
2159 * read-only, the first object in the shadow
2160 * chain should provide all of the pages. In
2161 * other words, this loop body should never be
2162 * executed when the source mapping is already
2165 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
2167 ("vm_fault_copy_entry: main object missing page"));
2169 VM_OBJECT_RLOCK(backing_object);
2170 pindex += OFF_TO_IDX(object->backing_object_offset);
2171 if (object != dst_object)
2172 VM_OBJECT_RUNLOCK(object);
2173 object = backing_object;
2175 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2177 if (object != dst_object) {
2179 * Allocate a page in the destination object.
2181 dst_m = vm_page_alloc(dst_object, (src_object ==
2182 dst_object ? src_pindex : 0) + dst_pindex,
2184 if (dst_m == NULL) {
2185 VM_OBJECT_WUNLOCK(dst_object);
2186 VM_OBJECT_RUNLOCK(object);
2187 vm_wait(dst_object);
2188 VM_OBJECT_WLOCK(dst_object);
2193 * See the comment in vm_fault_cow().
2195 if (src_object == dst_object &&
2196 (object->flags & OBJ_ONEMAPPING) == 0)
2197 pmap_remove_all(src_m);
2198 pmap_copy_page(src_m, dst_m);
2201 * The object lock does not guarantee that "src_m" will
2202 * transition from invalid to valid, but it does ensure
2203 * that "src_m" will not transition from valid to
2206 dst_m->dirty = dst_m->valid = src_m->valid;
2207 VM_OBJECT_RUNLOCK(object);
2210 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2212 if (dst_m->pindex >= dst_object->size) {
2214 * We are upgrading. Index can occur
2215 * out of bounds if the object type is
2216 * vnode and the file was truncated.
2218 vm_page_xunbusy(dst_m);
2224 * Enter it in the pmap. If a wired, copy-on-write
2225 * mapping is being replaced by a write-enabled
2226 * mapping, then wire that new mapping.
2228 * The page can be invalid if the user called
2229 * msync(MS_INVALIDATE) or truncated the backing vnode
2230 * or shared memory object. In this case, do not
2231 * insert it into pmap, but still do the copy so that
2232 * all copies of the wired map entry have similar
2235 if (vm_page_all_valid(dst_m)) {
2236 VM_OBJECT_WUNLOCK(dst_object);
2237 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2238 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2239 VM_OBJECT_WLOCK(dst_object);
2243 * Mark it no longer busy, and put it on the active list.
2246 if (src_m != dst_m) {
2247 vm_page_unwire(src_m, PQ_INACTIVE);
2248 vm_page_wire(dst_m);
2250 KASSERT(vm_page_wired(dst_m),
2251 ("dst_m %p is not wired", dst_m));
2254 vm_page_activate(dst_m);
2256 vm_page_xunbusy(dst_m);
2258 VM_OBJECT_WUNLOCK(dst_object);
2260 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2261 vm_object_deallocate(src_object);
2266 * Block entry into the machine-independent layer's page fault handler by
2267 * the calling thread. Subsequent calls to vm_fault() by that thread will
2268 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
2269 * spurious page faults.
2272 vm_fault_disable_pagefaults(void)
2275 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2279 vm_fault_enable_pagefaults(int save)
2282 curthread_pflags_restore(save);