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 #include "opt_ktrace.h"
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
85 #include <sys/mutex.h>
86 #include <sys/pctrie.h>
88 #include <sys/racct.h>
89 #include <sys/refcount.h>
90 #include <sys/resourcevar.h>
91 #include <sys/rwlock.h>
92 #include <sys/signalvar.h>
93 #include <sys/sysctl.h>
94 #include <sys/sysent.h>
95 #include <sys/vmmeter.h>
96 #include <sys/vnode.h>
98 #include <sys/ktrace.h>
102 #include <vm/vm_param.h>
104 #include <vm/vm_map.h>
105 #include <vm/vm_object.h>
106 #include <vm/vm_page.h>
107 #include <vm/vm_pageout.h>
108 #include <vm/vm_kern.h>
109 #include <vm/vm_pager.h>
110 #include <vm/vm_extern.h>
111 #include <vm/vm_reserv.h>
116 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
118 #define VM_FAULT_DONTNEED_MIN 1048576
121 /* Fault parameters. */
124 vm_prot_t fault_type;
130 struct timeval oom_start_time;
135 /* Page reference for cow. */
138 /* Current object. */
143 /* Top-level map object. */
144 vm_object_t first_object;
145 vm_pindex_t first_pindex;
150 vm_map_entry_t entry;
152 bool lookup_still_valid;
154 /* Vnode if locked. */
159 * Return codes for internal fault routines.
162 FAULT_SUCCESS = 10000, /* Return success to user. */
163 FAULT_FAILURE, /* Return failure to user. */
164 FAULT_CONTINUE, /* Continue faulting. */
165 FAULT_RESTART, /* Restart fault. */
166 FAULT_OUT_OF_BOUNDS, /* Invalid address for pager. */
167 FAULT_HARD, /* Performed I/O. */
168 FAULT_SOFT, /* Found valid page. */
169 FAULT_PROTECTION_FAILURE, /* Invalid access. */
172 enum fault_next_status {
173 FAULT_NEXT_GOTOBJ = 1,
178 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
180 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
181 int backward, int forward, bool obj_locked);
183 static int vm_pfault_oom_attempts = 3;
184 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
185 &vm_pfault_oom_attempts, 0,
186 "Number of page allocation attempts in page fault handler before it "
187 "triggers OOM handling");
189 static int vm_pfault_oom_wait = 10;
190 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
191 &vm_pfault_oom_wait, 0,
192 "Number of seconds to wait for free pages before retrying "
193 "the page fault handler");
196 vm_fault_page_release(vm_page_t *mp)
203 * We are likely to loop around again and attempt to busy
204 * this page. Deactivating it leaves it available for
205 * pageout while optimizing fault restarts.
207 vm_page_deactivate(m);
214 vm_fault_page_free(vm_page_t *mp)
220 VM_OBJECT_ASSERT_WLOCKED(m->object);
221 if (!vm_page_wired(m))
230 * Return true if a vm_pager_get_pages() call is needed in order to check
231 * whether the pager might have a particular page, false if it can be determined
232 * immediately that the pager can not have a copy. For swap objects, this can
233 * be checked quickly.
236 vm_fault_object_needs_getpages(vm_object_t object)
238 VM_OBJECT_ASSERT_LOCKED(object);
240 return ((object->flags & OBJ_SWAP) == 0 ||
241 !pctrie_is_empty(&object->un_pager.swp.swp_blks));
245 vm_fault_unlock_map(struct faultstate *fs)
248 if (fs->lookup_still_valid) {
249 vm_map_lookup_done(fs->map, fs->entry);
250 fs->lookup_still_valid = false;
255 vm_fault_unlock_vp(struct faultstate *fs)
258 if (fs->vp != NULL) {
265 vm_fault_deallocate(struct faultstate *fs)
268 vm_fault_page_release(&fs->m_cow);
269 vm_fault_page_release(&fs->m);
270 vm_object_pip_wakeup(fs->object);
271 if (fs->object != fs->first_object) {
272 VM_OBJECT_WLOCK(fs->first_object);
273 vm_fault_page_free(&fs->first_m);
274 VM_OBJECT_WUNLOCK(fs->first_object);
275 vm_object_pip_wakeup(fs->first_object);
277 vm_object_deallocate(fs->first_object);
278 vm_fault_unlock_map(fs);
279 vm_fault_unlock_vp(fs);
283 vm_fault_unlock_and_deallocate(struct faultstate *fs)
286 VM_OBJECT_UNLOCK(fs->object);
287 vm_fault_deallocate(fs);
291 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
295 if (((fs->prot & VM_PROT_WRITE) == 0 &&
296 (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
297 (m->oflags & VPO_UNMANAGED) != 0)
300 VM_PAGE_OBJECT_BUSY_ASSERT(m);
302 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
303 (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
304 (fs->fault_flags & VM_FAULT_DIRTY) != 0;
306 vm_object_set_writeable_dirty(m->object);
309 * If the fault is a write, we know that this page is being
310 * written NOW so dirty it explicitly to save on
311 * pmap_is_modified() calls later.
313 * Also, since the page is now dirty, we can possibly tell
314 * the pager to release any swap backing the page.
316 if (need_dirty && vm_page_set_dirty(m) == 0) {
318 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
319 * if the page is already dirty to prevent data written with
320 * the expectation of being synced from not being synced.
321 * Likewise if this entry does not request NOSYNC then make
322 * sure the page isn't marked NOSYNC. Applications sharing
323 * data should use the same flags to avoid ping ponging.
325 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
326 vm_page_aflag_set(m, PGA_NOSYNC);
328 vm_page_aflag_clear(m, PGA_NOSYNC);
334 * Unlocks fs.first_object and fs.map on success.
336 static enum fault_status
337 vm_fault_soft_fast(struct faultstate *fs)
340 #if VM_NRESERVLEVEL > 0
347 MPASS(fs->vp == NULL);
350 * If we fail, vast majority of the time it is because the page is not
351 * there to begin with. Opportunistically perform the lookup and
352 * subsequent checks without the object lock, revalidate later.
354 * Note: a busy page can be mapped for read|execute access.
356 m = vm_page_lookup_unlocked(fs->first_object, fs->first_pindex);
357 if (m == NULL || !vm_page_all_valid(m) ||
358 ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m))) {
359 VM_OBJECT_WLOCK(fs->first_object);
360 return (FAULT_FAILURE);
365 VM_OBJECT_RLOCK(fs->first_object);
368 * Now that we stabilized the state, revalidate the page is in the shape
369 * we encountered above.
372 if (m->object != fs->first_object || m->pindex != fs->first_pindex)
375 vm_object_busy(fs->first_object);
377 if (!vm_page_all_valid(m) ||
378 ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m)))
383 #if VM_NRESERVLEVEL > 0
384 if ((m->flags & PG_FICTITIOUS) == 0 &&
385 (m_super = vm_reserv_to_superpage(m)) != NULL &&
386 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
387 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
388 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
389 (pagesizes[m_super->psind] - 1)) && !fs->wired &&
390 pmap_ps_enabled(fs->map->pmap)) {
391 flags = PS_ALL_VALID;
392 if ((fs->prot & VM_PROT_WRITE) != 0) {
394 * Create a superpage mapping allowing write access
395 * only if none of the constituent pages are busy and
396 * all of them are already dirty (except possibly for
397 * the page that was faulted on).
399 flags |= PS_NONE_BUSY;
400 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
401 flags |= PS_ALL_DIRTY;
403 if (vm_page_ps_test(m_super, flags, m)) {
405 psind = m_super->psind;
406 vaddr = rounddown2(vaddr, pagesizes[psind]);
407 /* Preset the modified bit for dirty superpages. */
408 if ((flags & PS_ALL_DIRTY) != 0)
409 fs->fault_type |= VM_PROT_WRITE;
413 if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
414 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) !=
417 if (fs->m_hold != NULL) {
421 if (psind == 0 && !fs->wired)
422 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
423 VM_OBJECT_RUNLOCK(fs->first_object);
424 vm_fault_dirty(fs, m);
425 vm_object_unbusy(fs->first_object);
426 vm_map_lookup_done(fs->map, fs->entry);
427 curthread->td_ru.ru_minflt++;
428 return (FAULT_SUCCESS);
430 vm_object_unbusy(fs->first_object);
432 if (!VM_OBJECT_TRYUPGRADE(fs->first_object)) {
433 VM_OBJECT_RUNLOCK(fs->first_object);
434 VM_OBJECT_WLOCK(fs->first_object);
436 return (FAULT_FAILURE);
440 vm_fault_restore_map_lock(struct faultstate *fs)
443 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
444 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
446 if (!vm_map_trylock_read(fs->map)) {
447 VM_OBJECT_WUNLOCK(fs->first_object);
448 vm_map_lock_read(fs->map);
449 VM_OBJECT_WLOCK(fs->first_object);
451 fs->lookup_still_valid = true;
455 vm_fault_populate_check_page(vm_page_t m)
459 * Check each page to ensure that the pager is obeying the
460 * interface: the page must be installed in the object, fully
461 * valid, and exclusively busied.
464 MPASS(vm_page_all_valid(m));
465 MPASS(vm_page_xbusied(m));
469 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
475 VM_OBJECT_ASSERT_WLOCKED(object);
476 MPASS(first <= last);
477 for (pidx = first, m = vm_page_lookup(object, pidx);
478 pidx <= last; pidx++, m = vm_page_next(m)) {
479 vm_fault_populate_check_page(m);
480 vm_page_deactivate(m);
485 static enum fault_status
486 vm_fault_populate(struct faultstate *fs)
490 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
491 int bdry_idx, i, npages, psind, rv;
492 enum fault_status res;
494 MPASS(fs->object == fs->first_object);
495 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
496 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
497 MPASS(fs->first_object->backing_object == NULL);
498 MPASS(fs->lookup_still_valid);
500 pager_first = OFF_TO_IDX(fs->entry->offset);
501 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
502 vm_fault_unlock_map(fs);
503 vm_fault_unlock_vp(fs);
508 * Call the pager (driver) populate() method.
510 * There is no guarantee that the method will be called again
511 * if the current fault is for read, and a future fault is
512 * for write. Report the entry's maximum allowed protection
515 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
516 fs->fault_type, fs->entry->max_protection, &pager_first,
519 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
520 if (rv == VM_PAGER_BAD) {
522 * VM_PAGER_BAD is the backdoor for a pager to request
523 * normal fault handling.
525 vm_fault_restore_map_lock(fs);
526 if (fs->map->timestamp != fs->map_generation)
527 return (FAULT_RESTART);
528 return (FAULT_CONTINUE);
530 if (rv != VM_PAGER_OK)
531 return (FAULT_FAILURE); /* AKA SIGSEGV */
533 /* Ensure that the driver is obeying the interface. */
534 MPASS(pager_first <= pager_last);
535 MPASS(fs->first_pindex <= pager_last);
536 MPASS(fs->first_pindex >= pager_first);
537 MPASS(pager_last < fs->first_object->size);
539 vm_fault_restore_map_lock(fs);
540 bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(fs->entry);
541 if (fs->map->timestamp != fs->map_generation) {
543 vm_fault_populate_cleanup(fs->first_object, pager_first,
546 m = vm_page_lookup(fs->first_object, pager_first);
550 return (FAULT_RESTART);
554 * The map is unchanged after our last unlock. Process the fault.
556 * First, the special case of largepage mappings, where
557 * populate only busies the first page in superpage run.
560 KASSERT(PMAP_HAS_LARGEPAGES,
561 ("missing pmap support for large pages"));
562 m = vm_page_lookup(fs->first_object, pager_first);
563 vm_fault_populate_check_page(m);
564 VM_OBJECT_WUNLOCK(fs->first_object);
565 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
567 /* assert alignment for entry */
568 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
569 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
570 (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
571 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
572 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
573 ("unaligned superpage m %p %#jx", m,
574 (uintmax_t)VM_PAGE_TO_PHYS(m)));
575 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
576 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
577 PMAP_ENTER_LARGEPAGE, bdry_idx);
578 VM_OBJECT_WLOCK(fs->first_object);
580 if (rv != KERN_SUCCESS) {
584 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
585 for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
588 if (fs->m_hold != NULL) {
589 *fs->m_hold = m + (fs->first_pindex - pager_first);
590 vm_page_wire(*fs->m_hold);
596 * The range [pager_first, pager_last] that is given to the
597 * pager is only a hint. The pager may populate any range
598 * within the object that includes the requested page index.
599 * In case the pager expanded the range, clip it to fit into
602 map_first = OFF_TO_IDX(fs->entry->offset);
603 if (map_first > pager_first) {
604 vm_fault_populate_cleanup(fs->first_object, pager_first,
606 pager_first = map_first;
608 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
609 if (map_last < pager_last) {
610 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
612 pager_last = map_last;
614 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
616 pidx += npages, m = vm_page_next(&m[npages - 1])) {
617 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
620 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
621 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
622 !pmap_ps_enabled(fs->map->pmap) || fs->wired))
625 npages = atop(pagesizes[psind]);
626 for (i = 0; i < npages; i++) {
627 vm_fault_populate_check_page(&m[i]);
628 vm_fault_dirty(fs, &m[i]);
630 VM_OBJECT_WUNLOCK(fs->first_object);
631 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
632 (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
635 * pmap_enter() may fail for a superpage mapping if additional
636 * protection policies prevent the full mapping.
637 * For example, this will happen on amd64 if the entire
638 * address range does not share the same userspace protection
639 * key. Revert to single-page mappings if this happens.
641 MPASS(rv == KERN_SUCCESS ||
642 (psind > 0 && rv == KERN_PROTECTION_FAILURE));
643 if (__predict_false(psind > 0 &&
644 rv == KERN_PROTECTION_FAILURE)) {
646 for (i = 0; i < npages; i++) {
647 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
648 &m[i], fs->prot, fs->fault_type, 0);
649 MPASS(rv == KERN_SUCCESS);
653 VM_OBJECT_WLOCK(fs->first_object);
654 for (i = 0; i < npages; i++) {
655 if ((fs->fault_flags & VM_FAULT_WIRE) != 0 &&
656 m[i].pindex == fs->first_pindex)
659 vm_page_activate(&m[i]);
660 if (fs->m_hold != NULL &&
661 m[i].pindex == fs->first_pindex) {
662 (*fs->m_hold) = &m[i];
665 vm_page_xunbusy(&m[i]);
669 curthread->td_ru.ru_majflt++;
673 static int prot_fault_translation;
674 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
675 &prot_fault_translation, 0,
676 "Control signal to deliver on protection fault");
678 /* compat definition to keep common code for signal translation */
679 #define UCODE_PAGEFLT 12
681 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
687 * Handle a page fault occurring at the given address,
688 * requiring the given permissions, in the map specified.
689 * If successful, the page is inserted into the
690 * associated physical map.
692 * NOTE: the given address should be truncated to the
693 * proper page address.
695 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
696 * a standard error specifying why the fault is fatal is returned.
698 * The map in question must be referenced, and remains so.
699 * Caller may hold no locks.
702 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
703 int fault_flags, int *signo, int *ucode)
707 MPASS(signo == NULL || ucode != NULL);
709 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
710 ktrfault(vaddr, fault_type);
712 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
714 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
715 result == KERN_INVALID_ADDRESS ||
716 result == KERN_RESOURCE_SHORTAGE ||
717 result == KERN_PROTECTION_FAILURE ||
718 result == KERN_OUT_OF_BOUNDS,
719 ("Unexpected Mach error %d from vm_fault()", result));
721 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
724 if (result != KERN_SUCCESS && signo != NULL) {
727 case KERN_INVALID_ADDRESS:
729 *ucode = SEGV_MAPERR;
731 case KERN_RESOURCE_SHORTAGE:
735 case KERN_OUT_OF_BOUNDS:
739 case KERN_PROTECTION_FAILURE:
740 if (prot_fault_translation == 0) {
742 * Autodetect. This check also covers
743 * the images without the ABI-tag ELF
746 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
747 curproc->p_osrel >= P_OSREL_SIGSEGV) {
749 *ucode = SEGV_ACCERR;
752 *ucode = UCODE_PAGEFLT;
754 } else if (prot_fault_translation == 1) {
755 /* Always compat mode. */
757 *ucode = UCODE_PAGEFLT;
759 /* Always SIGSEGV mode. */
761 *ucode = SEGV_ACCERR;
765 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
774 vm_fault_object_ensure_wlocked(struct faultstate *fs)
776 if (fs->object == fs->first_object)
777 VM_OBJECT_ASSERT_WLOCKED(fs->object);
779 if (!fs->can_read_lock) {
780 VM_OBJECT_ASSERT_WLOCKED(fs->object);
784 if (VM_OBJECT_WOWNED(fs->object))
787 if (VM_OBJECT_TRYUPGRADE(fs->object))
793 static enum fault_status
794 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
799 if (fs->object->type != OBJT_VNODE)
800 return (FAULT_CONTINUE);
801 vp = fs->object->handle;
803 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
804 return (FAULT_CONTINUE);
808 * Perform an unlock in case the desired vnode changed while
809 * the map was unlocked during a retry.
811 vm_fault_unlock_vp(fs);
813 locked = VOP_ISLOCKED(vp);
814 if (locked != LK_EXCLUSIVE)
818 * We must not sleep acquiring the vnode lock while we have
819 * the page exclusive busied or the object's
820 * paging-in-progress count incremented. Otherwise, we could
823 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
826 return (FAULT_CONTINUE);
831 vm_fault_unlock_and_deallocate(fs);
833 vm_fault_deallocate(fs);
834 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
837 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
838 return (FAULT_RESTART);
842 * Calculate the desired readahead. Handle drop-behind.
844 * Returns the number of readahead blocks to pass to the pager.
847 vm_fault_readahead(struct faultstate *fs)
852 KASSERT(fs->lookup_still_valid, ("map unlocked"));
853 era = fs->entry->read_ahead;
854 behavior = vm_map_entry_behavior(fs->entry);
855 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
857 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
858 nera = VM_FAULT_READ_AHEAD_MAX;
859 if (fs->vaddr == fs->entry->next_read)
860 vm_fault_dontneed(fs, fs->vaddr, nera);
861 } else if (fs->vaddr == fs->entry->next_read) {
863 * This is a sequential fault. Arithmetically
864 * increase the requested number of pages in
865 * the read-ahead window. The requested
866 * number of pages is "# of sequential faults
867 * x (read ahead min + 1) + read ahead min"
869 nera = VM_FAULT_READ_AHEAD_MIN;
872 if (nera > VM_FAULT_READ_AHEAD_MAX)
873 nera = VM_FAULT_READ_AHEAD_MAX;
875 if (era == VM_FAULT_READ_AHEAD_MAX)
876 vm_fault_dontneed(fs, fs->vaddr, nera);
879 * This is a non-sequential fault.
885 * A read lock on the map suffices to update
886 * the read ahead count safely.
888 fs->entry->read_ahead = nera;
895 vm_fault_lookup(struct faultstate *fs)
899 KASSERT(!fs->lookup_still_valid,
900 ("vm_fault_lookup: Map already locked."));
901 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
902 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
903 &fs->first_pindex, &fs->prot, &fs->wired);
904 if (result != KERN_SUCCESS) {
905 vm_fault_unlock_vp(fs);
909 fs->map_generation = fs->map->timestamp;
911 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
912 panic("%s: fault on nofault entry, addr: %#lx",
913 __func__, (u_long)fs->vaddr);
916 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
917 fs->entry->wiring_thread != curthread) {
918 vm_map_unlock_read(fs->map);
919 vm_map_lock(fs->map);
920 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
921 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
922 vm_fault_unlock_vp(fs);
923 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
924 vm_map_unlock_and_wait(fs->map, 0);
926 vm_map_unlock(fs->map);
927 return (KERN_RESOURCE_SHORTAGE);
930 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
933 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
935 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
936 ("!fs->wired && VM_FAULT_WIRE"));
937 fs->lookup_still_valid = true;
939 return (KERN_SUCCESS);
943 vm_fault_relookup(struct faultstate *fs)
945 vm_object_t retry_object;
946 vm_pindex_t retry_pindex;
947 vm_prot_t retry_prot;
950 if (!vm_map_trylock_read(fs->map))
951 return (KERN_RESTART);
953 fs->lookup_still_valid = true;
954 if (fs->map->timestamp == fs->map_generation)
955 return (KERN_SUCCESS);
957 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
958 &fs->entry, &retry_object, &retry_pindex, &retry_prot,
960 if (result != KERN_SUCCESS) {
962 * If retry of map lookup would have blocked then
963 * retry fault from start.
965 if (result == KERN_FAILURE)
966 return (KERN_RESTART);
969 if (retry_object != fs->first_object ||
970 retry_pindex != fs->first_pindex)
971 return (KERN_RESTART);
974 * Check whether the protection has changed or the object has
975 * been copied while we left the map unlocked. Changing from
976 * read to write permission is OK - we leave the page
977 * write-protected, and catch the write fault. Changing from
978 * write to read permission means that we can't mark the page
979 * write-enabled after all.
981 fs->prot &= retry_prot;
982 fs->fault_type &= retry_prot;
984 return (KERN_RESTART);
986 /* Reassert because wired may have changed. */
987 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
988 ("!wired && VM_FAULT_WIRE"));
990 return (KERN_SUCCESS);
994 vm_fault_cow(struct faultstate *fs)
996 bool is_first_object_locked;
998 KASSERT(fs->object != fs->first_object,
999 ("source and target COW objects are identical"));
1002 * This allows pages to be virtually copied from a backing_object
1003 * into the first_object, where the backing object has no other
1004 * refs to it, and cannot gain any more refs. Instead of a bcopy,
1005 * we just move the page from the backing object to the first
1006 * object. Note that we must mark the page dirty in the first
1007 * object so that it will go out to swap when needed.
1009 is_first_object_locked = false;
1012 * Only one shadow object and no other refs.
1014 fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
1016 * No other ways to look the object up
1018 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
1020 * We don't chase down the shadow chain and we can acquire locks.
1022 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
1023 fs->object == fs->first_object->backing_object &&
1024 VM_OBJECT_TRYWLOCK(fs->object)) {
1026 * Remove but keep xbusy for replace. fs->m is moved into
1027 * fs->first_object and left busy while fs->first_m is
1028 * conditionally freed.
1030 vm_page_remove_xbusy(fs->m);
1031 vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
1033 vm_page_dirty(fs->m);
1034 #if VM_NRESERVLEVEL > 0
1036 * Rename the reservation.
1038 vm_reserv_rename(fs->m, fs->first_object, fs->object,
1039 OFF_TO_IDX(fs->first_object->backing_object_offset));
1041 VM_OBJECT_WUNLOCK(fs->object);
1042 VM_OBJECT_WUNLOCK(fs->first_object);
1043 fs->first_m = fs->m;
1045 VM_CNT_INC(v_cow_optim);
1047 if (is_first_object_locked)
1048 VM_OBJECT_WUNLOCK(fs->first_object);
1050 * Oh, well, lets copy it.
1052 pmap_copy_page(fs->m, fs->first_m);
1053 vm_page_valid(fs->first_m);
1054 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
1055 vm_page_wire(fs->first_m);
1056 vm_page_unwire(fs->m, PQ_INACTIVE);
1059 * Save the cow page to be released after
1060 * pmap_enter is complete.
1066 * Typically, the shadow object is either private to this
1067 * address space (OBJ_ONEMAPPING) or its pages are read only.
1068 * In the highly unusual case where the pages of a shadow object
1069 * are read/write shared between this and other address spaces,
1070 * we need to ensure that any pmap-level mappings to the
1071 * original, copy-on-write page from the backing object are
1072 * removed from those other address spaces.
1074 * The flag check is racy, but this is tolerable: if
1075 * OBJ_ONEMAPPING is cleared after the check, the busy state
1076 * ensures that new mappings of m_cow can't be created.
1077 * pmap_enter() will replace an existing mapping in the current
1078 * address space. If OBJ_ONEMAPPING is set after the check,
1079 * removing mappings will at worse trigger some unnecessary page
1082 vm_page_assert_xbusied(fs->m_cow);
1083 if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
1084 pmap_remove_all(fs->m_cow);
1087 vm_object_pip_wakeup(fs->object);
1090 * Only use the new page below...
1092 fs->object = fs->first_object;
1093 fs->pindex = fs->first_pindex;
1094 fs->m = fs->first_m;
1095 VM_CNT_INC(v_cow_faults);
1096 curthread->td_cow++;
1099 static enum fault_next_status
1100 vm_fault_next(struct faultstate *fs)
1102 vm_object_t next_object;
1104 if (fs->object == fs->first_object || !fs->can_read_lock)
1105 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1107 VM_OBJECT_ASSERT_LOCKED(fs->object);
1110 * The requested page does not exist at this object/
1111 * offset. Remove the invalid page from the object,
1112 * waking up anyone waiting for it, and continue on to
1113 * the next object. However, if this is the top-level
1114 * object, we must leave the busy page in place to
1115 * prevent another process from rushing past us, and
1116 * inserting the page in that object at the same time
1119 if (fs->object == fs->first_object) {
1120 fs->first_m = fs->m;
1122 } else if (fs->m != NULL) {
1123 if (!vm_fault_object_ensure_wlocked(fs)) {
1124 fs->can_read_lock = false;
1125 vm_fault_unlock_and_deallocate(fs);
1126 return (FAULT_NEXT_RESTART);
1128 vm_fault_page_free(&fs->m);
1132 * Move on to the next object. Lock the next object before
1133 * unlocking the current one.
1135 next_object = fs->object->backing_object;
1136 if (next_object == NULL)
1137 return (FAULT_NEXT_NOOBJ);
1138 MPASS(fs->first_m != NULL);
1139 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1140 if (fs->can_read_lock)
1141 VM_OBJECT_RLOCK(next_object);
1143 VM_OBJECT_WLOCK(next_object);
1144 vm_object_pip_add(next_object, 1);
1145 if (fs->object != fs->first_object)
1146 vm_object_pip_wakeup(fs->object);
1147 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1148 VM_OBJECT_UNLOCK(fs->object);
1149 fs->object = next_object;
1151 return (FAULT_NEXT_GOTOBJ);
1155 vm_fault_zerofill(struct faultstate *fs)
1159 * If there's no object left, fill the page in the top
1160 * object with zeros.
1162 if (fs->object != fs->first_object) {
1163 vm_object_pip_wakeup(fs->object);
1164 fs->object = fs->first_object;
1165 fs->pindex = fs->first_pindex;
1167 MPASS(fs->first_m != NULL);
1168 MPASS(fs->m == NULL);
1169 fs->m = fs->first_m;
1173 * Zero the page if necessary and mark it valid.
1175 if ((fs->m->flags & PG_ZERO) == 0) {
1176 pmap_zero_page(fs->m);
1178 VM_CNT_INC(v_ozfod);
1181 vm_page_valid(fs->m);
1185 * Initiate page fault after timeout. Returns true if caller should
1186 * do vm_waitpfault() after the call.
1189 vm_fault_allocate_oom(struct faultstate *fs)
1193 vm_fault_unlock_and_deallocate(fs);
1194 if (vm_pfault_oom_attempts < 0)
1196 if (!fs->oom_started) {
1197 fs->oom_started = true;
1198 getmicrotime(&fs->oom_start_time);
1203 timevalsub(&now, &fs->oom_start_time);
1204 if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait)
1209 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1210 curproc->p_pid, curproc->p_comm);
1211 vm_pageout_oom(VM_OOM_MEM_PF);
1212 fs->oom_started = false;
1217 * Allocate a page directly or via the object populate method.
1219 static enum fault_status
1220 vm_fault_allocate(struct faultstate *fs)
1222 struct domainset *dset;
1223 enum fault_status res;
1225 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1226 res = vm_fault_lock_vnode(fs, true);
1227 MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART);
1228 if (res == FAULT_RESTART)
1232 if (fs->pindex >= fs->object->size) {
1233 vm_fault_unlock_and_deallocate(fs);
1234 return (FAULT_OUT_OF_BOUNDS);
1237 if (fs->object == fs->first_object &&
1238 (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1239 fs->first_object->shadow_count == 0) {
1240 res = vm_fault_populate(fs);
1245 vm_fault_unlock_and_deallocate(fs);
1247 case FAULT_CONTINUE:
1249 * Pager's populate() method
1250 * returned VM_PAGER_BAD.
1254 panic("inconsistent return codes");
1259 * Allocate a new page for this object/offset pair.
1261 * If the process has a fatal signal pending, prioritize the allocation
1262 * with the expectation that the process will exit shortly and free some
1263 * pages. In particular, the signal may have been posted by the page
1264 * daemon in an attempt to resolve an out-of-memory condition.
1266 * The unlocked read of the p_flag is harmless. At worst, the P_KILLED
1267 * might be not observed here, and allocation fails, causing a restart
1268 * and new reading of the p_flag.
1270 dset = fs->object->domain.dr_policy;
1272 dset = curthread->td_domain.dr_policy;
1273 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1274 #if VM_NRESERVLEVEL > 0
1275 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1277 if (!vm_pager_can_alloc_page(fs->object, fs->pindex)) {
1278 vm_fault_unlock_and_deallocate(fs);
1279 return (FAULT_FAILURE);
1281 fs->m = vm_page_alloc(fs->object, fs->pindex,
1282 P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0);
1284 if (fs->m == NULL) {
1285 if (vm_fault_allocate_oom(fs))
1286 vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1287 return (FAULT_RESTART);
1289 fs->oom_started = false;
1291 return (FAULT_CONTINUE);
1295 * Call the pager to retrieve the page if there is a chance
1296 * that the pager has it, and potentially retrieve additional
1297 * pages at the same time.
1299 static enum fault_status
1300 vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp)
1302 vm_offset_t e_end, e_start;
1303 int ahead, behind, cluster_offset, rv;
1304 enum fault_status status;
1308 * Prepare for unlocking the map. Save the map
1309 * entry's start and end addresses, which are used to
1310 * optimize the size of the pager operation below.
1311 * Even if the map entry's addresses change after
1312 * unlocking the map, using the saved addresses is
1315 e_start = fs->entry->start;
1316 e_end = fs->entry->end;
1317 behavior = vm_map_entry_behavior(fs->entry);
1320 * If the pager for the current object might have
1321 * the page, then determine the number of additional
1322 * pages to read and potentially reprioritize
1323 * previously read pages for earlier reclamation.
1324 * These operations should only be performed once per
1325 * page fault. Even if the current pager doesn't
1326 * have the page, the number of additional pages to
1327 * read will apply to subsequent objects in the
1330 if (fs->nera == -1 && !P_KILLED(curproc))
1331 fs->nera = vm_fault_readahead(fs);
1334 * Release the map lock before locking the vnode or
1335 * sleeping in the pager. (If the current object has
1336 * a shadow, then an earlier iteration of this loop
1337 * may have already unlocked the map.)
1339 vm_fault_unlock_map(fs);
1341 status = vm_fault_lock_vnode(fs, false);
1342 MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART);
1343 if (status == FAULT_RESTART)
1345 KASSERT(fs->vp == NULL || !fs->map->system_map,
1346 ("vm_fault: vnode-backed object mapped by system map"));
1349 * Page in the requested page and hint the pager,
1350 * that it may bring up surrounding pages.
1352 if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1353 P_KILLED(curproc)) {
1357 /* Is this a sequential fault? */
1363 * Request a cluster of pages that is
1364 * aligned to a VM_FAULT_READ_DEFAULT
1365 * page offset boundary within the
1366 * object. Alignment to a page offset
1367 * boundary is more likely to coincide
1368 * with the underlying file system
1369 * block than alignment to a virtual
1372 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1373 behind = ulmin(cluster_offset,
1374 atop(fs->vaddr - e_start));
1375 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1377 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1381 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1382 if (rv == VM_PAGER_OK)
1383 return (FAULT_HARD);
1384 if (rv == VM_PAGER_ERROR)
1385 printf("vm_fault: pager read error, pid %d (%s)\n",
1386 curproc->p_pid, curproc->p_comm);
1388 * If an I/O error occurred or the requested page was
1389 * outside the range of the pager, clean up and return
1392 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1393 VM_OBJECT_WLOCK(fs->object);
1394 vm_fault_page_free(&fs->m);
1395 vm_fault_unlock_and_deallocate(fs);
1396 return (FAULT_OUT_OF_BOUNDS);
1398 KASSERT(rv == VM_PAGER_FAIL,
1399 ("%s: unexpected pager error %d", __func__, rv));
1400 return (FAULT_CONTINUE);
1404 * Wait/Retry if the page is busy. We have to do this if the page is
1405 * either exclusive or shared busy because the vm_pager may be using
1406 * read busy for pageouts (and even pageins if it is the vnode pager),
1407 * and we could end up trying to pagein and pageout the same page
1410 * We can theoretically allow the busy case on a read fault if the page
1411 * is marked valid, but since such pages are typically already pmap'd,
1412 * putting that special case in might be more effort then it is worth.
1413 * We cannot under any circumstances mess around with a shared busied
1414 * page except, perhaps, to pmap it.
1417 vm_fault_busy_sleep(struct faultstate *fs)
1420 * Reference the page before unlocking and
1421 * sleeping so that the page daemon is less
1422 * likely to reclaim it.
1424 vm_page_aflag_set(fs->m, PGA_REFERENCED);
1425 if (fs->object != fs->first_object) {
1426 vm_fault_page_release(&fs->first_m);
1427 vm_object_pip_wakeup(fs->first_object);
1429 vm_object_pip_wakeup(fs->object);
1430 vm_fault_unlock_map(fs);
1431 if (fs->m != vm_page_lookup(fs->object, fs->pindex) ||
1432 !vm_page_busy_sleep(fs->m, "vmpfw", 0))
1433 VM_OBJECT_UNLOCK(fs->object);
1434 VM_CNT_INC(v_intrans);
1435 vm_object_deallocate(fs->first_object);
1439 * Handle page lookup, populate, allocate, page-in for the current
1442 * The object is locked on entry and will remain locked with a return
1443 * code of FAULT_CONTINUE so that fault may follow the shadow chain.
1444 * Otherwise, the object will be unlocked upon return.
1446 static enum fault_status
1447 vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp)
1449 enum fault_status res;
1452 if (fs->object == fs->first_object || !fs->can_read_lock)
1453 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1455 VM_OBJECT_ASSERT_LOCKED(fs->object);
1458 * If the object is marked for imminent termination, we retry
1459 * here, since the collapse pass has raced with us. Otherwise,
1460 * if we see terminally dead object, return fail.
1462 if ((fs->object->flags & OBJ_DEAD) != 0) {
1463 dead = fs->object->type == OBJT_DEAD;
1464 vm_fault_unlock_and_deallocate(fs);
1466 return (FAULT_PROTECTION_FAILURE);
1468 return (FAULT_RESTART);
1472 * See if the page is resident.
1474 fs->m = vm_page_lookup(fs->object, fs->pindex);
1475 if (fs->m != NULL) {
1476 if (!vm_page_tryxbusy(fs->m)) {
1477 vm_fault_busy_sleep(fs);
1478 return (FAULT_RESTART);
1482 * The page is marked busy for other processes and the
1483 * pagedaemon. If it is still completely valid we are
1486 if (vm_page_all_valid(fs->m)) {
1487 VM_OBJECT_UNLOCK(fs->object);
1488 return (FAULT_SOFT);
1493 * Page is not resident. If the pager might contain the page
1494 * or this is the beginning of the search, allocate a new
1497 if (fs->m == NULL && (vm_fault_object_needs_getpages(fs->object) ||
1498 fs->object == fs->first_object)) {
1499 if (!vm_fault_object_ensure_wlocked(fs)) {
1500 fs->can_read_lock = false;
1501 vm_fault_unlock_and_deallocate(fs);
1502 return (FAULT_RESTART);
1504 res = vm_fault_allocate(fs);
1505 if (res != FAULT_CONTINUE)
1510 * Check to see if the pager can possibly satisfy this fault.
1511 * If not, skip to the next object without dropping the lock to
1512 * preserve atomicity of shadow faults.
1514 if (vm_fault_object_needs_getpages(fs->object)) {
1516 * At this point, we have either allocated a new page
1517 * or found an existing page that is only partially
1520 * We hold a reference on the current object and the
1521 * page is exclusive busied. The exclusive busy
1522 * prevents simultaneous faults and collapses while
1523 * the object lock is dropped.
1525 VM_OBJECT_UNLOCK(fs->object);
1526 res = vm_fault_getpages(fs, behindp, aheadp);
1527 if (res == FAULT_CONTINUE)
1528 VM_OBJECT_WLOCK(fs->object);
1530 res = FAULT_CONTINUE;
1536 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1537 int fault_flags, vm_page_t *m_hold)
1539 struct faultstate fs;
1540 int ahead, behind, faultcount, rv;
1541 enum fault_status res;
1542 enum fault_next_status res_next;
1545 VM_CNT_INC(v_vm_faults);
1547 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1548 return (KERN_PROTECTION_FAILURE);
1553 fs.fault_flags = fault_flags;
1555 fs.lookup_still_valid = false;
1556 fs.oom_started = false;
1558 fs.can_read_lock = true;
1563 fs.fault_type = fault_type;
1566 * Find the backing store object and offset into it to begin the
1569 rv = vm_fault_lookup(&fs);
1570 if (rv != KERN_SUCCESS) {
1571 if (rv == KERN_RESOURCE_SHORTAGE)
1577 * Try to avoid lock contention on the top-level object through
1578 * special-case handling of some types of page faults, specifically,
1579 * those that are mapping an existing page from the top-level object.
1580 * Under this condition, a read lock on the object suffices, allowing
1581 * multiple page faults of a similar type to run in parallel.
1583 if (fs.vp == NULL /* avoid locked vnode leak */ &&
1584 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1585 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1586 res = vm_fault_soft_fast(&fs);
1587 if (res == FAULT_SUCCESS) {
1588 VM_OBJECT_ASSERT_UNLOCKED(fs.first_object);
1589 return (KERN_SUCCESS);
1591 VM_OBJECT_ASSERT_WLOCKED(fs.first_object);
1593 VM_OBJECT_WLOCK(fs.first_object);
1597 * Make a reference to this object to prevent its disposal while we
1598 * are messing with it. Once we have the reference, the map is free
1599 * to be diddled. Since objects reference their shadows (and copies),
1600 * they will stay around as well.
1602 * Bump the paging-in-progress count to prevent size changes (e.g.
1603 * truncation operations) during I/O.
1605 vm_object_reference_locked(fs.first_object);
1606 vm_object_pip_add(fs.first_object, 1);
1608 fs.m_cow = fs.m = fs.first_m = NULL;
1611 * Search for the page at object/offset.
1613 fs.object = fs.first_object;
1614 fs.pindex = fs.first_pindex;
1616 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1617 res = vm_fault_allocate(&fs);
1622 return (KERN_SUCCESS);
1624 return (KERN_FAILURE);
1625 case FAULT_OUT_OF_BOUNDS:
1626 return (KERN_OUT_OF_BOUNDS);
1627 case FAULT_CONTINUE:
1630 panic("vm_fault: Unhandled status %d", res);
1635 KASSERT(fs.m == NULL,
1636 ("page still set %p at loop start", fs.m));
1638 res = vm_fault_object(&fs, &behind, &ahead);
1643 faultcount = behind + 1 + ahead;
1649 return (KERN_SUCCESS);
1651 return (KERN_FAILURE);
1652 case FAULT_OUT_OF_BOUNDS:
1653 return (KERN_OUT_OF_BOUNDS);
1654 case FAULT_PROTECTION_FAILURE:
1655 return (KERN_PROTECTION_FAILURE);
1656 case FAULT_CONTINUE:
1659 panic("vm_fault: Unhandled status %d", res);
1663 * The page was not found in the current object. Try to
1664 * traverse into a backing object or zero fill if none is
1667 res_next = vm_fault_next(&fs);
1668 if (res_next == FAULT_NEXT_RESTART)
1670 else if (res_next == FAULT_NEXT_GOTOBJ)
1672 MPASS(res_next == FAULT_NEXT_NOOBJ);
1673 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1674 if (fs.first_object == fs.object)
1675 vm_fault_page_free(&fs.first_m);
1676 vm_fault_unlock_and_deallocate(&fs);
1677 return (KERN_OUT_OF_BOUNDS);
1679 VM_OBJECT_UNLOCK(fs.object);
1680 vm_fault_zerofill(&fs);
1681 /* Don't try to prefault neighboring pages. */
1688 * A valid page has been found and exclusively busied. The
1689 * object lock must no longer be held.
1691 vm_page_assert_xbusied(fs.m);
1692 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1695 * If the page is being written, but isn't already owned by the
1696 * top-level object, we have to copy it into a new page owned by the
1699 if (fs.object != fs.first_object) {
1701 * We only really need to copy if we want to write it.
1703 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1706 * We only try to prefault read-only mappings to the
1707 * neighboring pages when this copy-on-write fault is
1708 * a hard fault. In other cases, trying to prefault
1709 * is typically wasted effort.
1711 if (faultcount == 0)
1715 fs.prot &= ~VM_PROT_WRITE;
1720 * We must verify that the maps have not changed since our last
1723 if (!fs.lookup_still_valid) {
1724 rv = vm_fault_relookup(&fs);
1725 if (rv != KERN_SUCCESS) {
1726 vm_fault_deallocate(&fs);
1727 if (rv == KERN_RESTART)
1732 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1735 * If the page was filled by a pager, save the virtual address that
1736 * should be faulted on next under a sequential access pattern to the
1737 * map entry. A read lock on the map suffices to update this address
1741 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1744 * Page must be completely valid or it is not fit to
1745 * map into user space. vm_pager_get_pages() ensures this.
1747 vm_page_assert_xbusied(fs.m);
1748 KASSERT(vm_page_all_valid(fs.m),
1749 ("vm_fault: page %p partially invalid", fs.m));
1751 vm_fault_dirty(&fs, fs.m);
1754 * Put this page into the physical map. We had to do the unlock above
1755 * because pmap_enter() may sleep. We don't put the page
1756 * back on the active queue until later so that the pageout daemon
1757 * won't find it (yet).
1759 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1760 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1761 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1763 vm_fault_prefault(&fs, vaddr,
1764 faultcount > 0 ? behind : PFBAK,
1765 faultcount > 0 ? ahead : PFFOR, false);
1768 * If the page is not wired down, then put it where the pageout daemon
1771 if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1774 vm_page_activate(fs.m);
1775 if (fs.m_hold != NULL) {
1776 (*fs.m_hold) = fs.m;
1779 vm_page_xunbusy(fs.m);
1783 * Unlock everything, and return
1785 vm_fault_deallocate(&fs);
1787 VM_CNT_INC(v_io_faults);
1788 curthread->td_ru.ru_majflt++;
1790 if (racct_enable && fs.object->type == OBJT_VNODE) {
1792 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1793 racct_add_force(curproc, RACCT_WRITEBPS,
1794 PAGE_SIZE + behind * PAGE_SIZE);
1795 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1797 racct_add_force(curproc, RACCT_READBPS,
1798 PAGE_SIZE + ahead * PAGE_SIZE);
1799 racct_add_force(curproc, RACCT_READIOPS, 1);
1801 PROC_UNLOCK(curproc);
1805 curthread->td_ru.ru_minflt++;
1807 return (KERN_SUCCESS);
1811 * Speed up the reclamation of pages that precede the faulting pindex within
1812 * the first object of the shadow chain. Essentially, perform the equivalent
1813 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1814 * the faulting pindex by the cluster size when the pages read by vm_fault()
1815 * cross a cluster-size boundary. The cluster size is the greater of the
1816 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1818 * When "fs->first_object" is a shadow object, the pages in the backing object
1819 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1820 * function must only be concerned with pages in the first object.
1823 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1825 vm_map_entry_t entry;
1826 vm_object_t first_object;
1827 vm_offset_t end, start;
1828 vm_page_t m, m_next;
1829 vm_pindex_t pend, pstart;
1832 VM_OBJECT_ASSERT_UNLOCKED(fs->object);
1833 first_object = fs->first_object;
1834 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1835 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1836 VM_OBJECT_RLOCK(first_object);
1837 size = VM_FAULT_DONTNEED_MIN;
1838 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1839 size = pagesizes[1];
1840 end = rounddown2(vaddr, size);
1841 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1842 (entry = fs->entry)->start < end) {
1843 if (end - entry->start < size)
1844 start = entry->start;
1847 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1848 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1850 m_next = vm_page_find_least(first_object, pstart);
1851 pend = OFF_TO_IDX(entry->offset) + atop(end -
1853 while ((m = m_next) != NULL && m->pindex < pend) {
1854 m_next = TAILQ_NEXT(m, listq);
1855 if (!vm_page_all_valid(m) ||
1860 * Don't clear PGA_REFERENCED, since it would
1861 * likely represent a reference by a different
1864 * Typically, at this point, prefetched pages
1865 * are still in the inactive queue. Only
1866 * pages that triggered page faults are in the
1867 * active queue. The test for whether the page
1868 * is in the inactive queue is racy; in the
1869 * worst case we will requeue the page
1872 if (!vm_page_inactive(m))
1873 vm_page_deactivate(m);
1876 VM_OBJECT_RUNLOCK(first_object);
1881 * vm_fault_prefault provides a quick way of clustering
1882 * pagefaults into a processes address space. It is a "cousin"
1883 * of vm_map_pmap_enter, except it runs at page fault time instead
1887 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1888 int backward, int forward, bool obj_locked)
1891 vm_map_entry_t entry;
1892 vm_object_t backing_object, lobject;
1893 vm_offset_t addr, starta;
1898 pmap = fs->map->pmap;
1899 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1904 if (addra < backward * PAGE_SIZE) {
1905 starta = entry->start;
1907 starta = addra - backward * PAGE_SIZE;
1908 if (starta < entry->start)
1909 starta = entry->start;
1913 * Generate the sequence of virtual addresses that are candidates for
1914 * prefaulting in an outward spiral from the faulting virtual address,
1915 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1916 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1917 * If the candidate address doesn't have a backing physical page, then
1918 * the loop immediately terminates.
1920 for (i = 0; i < 2 * imax(backward, forward); i++) {
1921 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1923 if (addr > addra + forward * PAGE_SIZE)
1926 if (addr < starta || addr >= entry->end)
1929 if (!pmap_is_prefaultable(pmap, addr))
1932 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1933 lobject = entry->object.vm_object;
1935 VM_OBJECT_RLOCK(lobject);
1936 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1937 !vm_fault_object_needs_getpages(lobject) &&
1938 (backing_object = lobject->backing_object) != NULL) {
1939 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1940 0, ("vm_fault_prefault: unaligned object offset"));
1941 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1942 VM_OBJECT_RLOCK(backing_object);
1943 if (!obj_locked || lobject != entry->object.vm_object)
1944 VM_OBJECT_RUNLOCK(lobject);
1945 lobject = backing_object;
1948 if (!obj_locked || lobject != entry->object.vm_object)
1949 VM_OBJECT_RUNLOCK(lobject);
1952 if (vm_page_all_valid(m) &&
1953 (m->flags & PG_FICTITIOUS) == 0)
1954 pmap_enter_quick(pmap, addr, m, entry->protection);
1955 if (!obj_locked || lobject != entry->object.vm_object)
1956 VM_OBJECT_RUNLOCK(lobject);
1961 * Hold each of the physical pages that are mapped by the specified range of
1962 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1963 * and allow the specified types of access, "prot". If all of the implied
1964 * pages are successfully held, then the number of held pages is returned
1965 * together with pointers to those pages in the array "ma". However, if any
1966 * of the pages cannot be held, -1 is returned.
1969 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1970 vm_prot_t prot, vm_page_t *ma, int max_count)
1972 vm_offset_t end, va;
1975 boolean_t pmap_failed;
1979 end = round_page(addr + len);
1980 addr = trunc_page(addr);
1982 if (!vm_map_range_valid(map, addr, end))
1985 if (atop(end - addr) > max_count)
1986 panic("vm_fault_quick_hold_pages: count > max_count");
1987 count = atop(end - addr);
1990 * Most likely, the physical pages are resident in the pmap, so it is
1991 * faster to try pmap_extract_and_hold() first.
1993 pmap_failed = FALSE;
1994 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1995 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1998 else if ((prot & VM_PROT_WRITE) != 0 &&
1999 (*mp)->dirty != VM_PAGE_BITS_ALL) {
2001 * Explicitly dirty the physical page. Otherwise, the
2002 * caller's changes may go unnoticed because they are
2003 * performed through an unmanaged mapping or by a DMA
2006 * The object lock is not held here.
2007 * See vm_page_clear_dirty_mask().
2014 * One or more pages could not be held by the pmap. Either no
2015 * page was mapped at the specified virtual address or that
2016 * mapping had insufficient permissions. Attempt to fault in
2017 * and hold these pages.
2019 * If vm_fault_disable_pagefaults() was called,
2020 * i.e., TDP_NOFAULTING is set, we must not sleep nor
2021 * acquire MD VM locks, which means we must not call
2022 * vm_fault(). Some (out of tree) callers mark
2023 * too wide a code area with vm_fault_disable_pagefaults()
2024 * already, use the VM_PROT_QUICK_NOFAULT flag to request
2025 * the proper behaviour explicitly.
2027 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
2028 (curthread->td_pflags & TDP_NOFAULTING) != 0)
2030 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
2031 if (*mp == NULL && vm_fault(map, va, prot,
2032 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
2037 for (mp = ma; mp < ma + count; mp++)
2039 vm_page_unwire(*mp, PQ_INACTIVE);
2045 * vm_fault_copy_entry
2047 * Create new object backing dst_entry with private copy of all
2048 * underlying pages. When src_entry is equal to dst_entry, function
2049 * implements COW for wired-down map entry. Otherwise, it forks
2050 * wired entry into dst_map.
2052 * In/out conditions:
2053 * The source and destination maps must be locked for write.
2054 * The source map entry must be wired down (or be a sharing map
2055 * entry corresponding to a main map entry that is wired down).
2058 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map __unused,
2059 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
2060 vm_ooffset_t *fork_charge)
2062 vm_object_t backing_object, dst_object, object, src_object;
2063 vm_pindex_t dst_pindex, pindex, src_pindex;
2064 vm_prot_t access, prot;
2070 upgrade = src_entry == dst_entry;
2071 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
2072 ("vm_fault_copy_entry: vm_object not NULL"));
2075 * If not an upgrade, then enter the mappings in the pmap as
2076 * read and/or execute accesses. Otherwise, enter them as
2079 * A writeable large page mapping is only created if all of
2080 * the constituent small page mappings are modified. Marking
2081 * PTEs as modified on inception allows promotion to happen
2082 * without taking potentially large number of soft faults.
2084 access = prot = dst_entry->protection;
2086 access &= ~VM_PROT_WRITE;
2088 src_object = src_entry->object.vm_object;
2089 src_pindex = OFF_TO_IDX(src_entry->offset);
2091 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2092 dst_object = src_object;
2093 vm_object_reference(dst_object);
2096 * Create the top-level object for the destination entry.
2097 * Doesn't actually shadow anything - we copy the pages
2100 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
2101 dst_entry->start), NULL, NULL, 0);
2102 #if VM_NRESERVLEVEL > 0
2103 dst_object->flags |= OBJ_COLORED;
2104 dst_object->pg_color = atop(dst_entry->start);
2106 dst_object->domain = src_object->domain;
2107 dst_object->charge = dst_entry->end - dst_entry->start;
2109 dst_entry->object.vm_object = dst_object;
2110 dst_entry->offset = 0;
2111 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
2114 VM_OBJECT_WLOCK(dst_object);
2115 if (fork_charge != NULL) {
2116 KASSERT(dst_entry->cred == NULL,
2117 ("vm_fault_copy_entry: leaked swp charge"));
2118 dst_object->cred = curthread->td_ucred;
2119 crhold(dst_object->cred);
2120 *fork_charge += dst_object->charge;
2121 } else if ((dst_object->flags & OBJ_SWAP) != 0 &&
2122 dst_object->cred == NULL) {
2123 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
2125 dst_object->cred = dst_entry->cred;
2126 dst_entry->cred = NULL;
2130 * Loop through all of the virtual pages within the entry's
2131 * range, copying each page from the source object to the
2132 * destination object. Since the source is wired, those pages
2133 * must exist. In contrast, the destination is pageable.
2134 * Since the destination object doesn't share any backing storage
2135 * with the source object, all of its pages must be dirtied,
2136 * regardless of whether they can be written.
2138 for (vaddr = dst_entry->start, dst_pindex = 0;
2139 vaddr < dst_entry->end;
2140 vaddr += PAGE_SIZE, dst_pindex++) {
2143 * Find the page in the source object, and copy it in.
2144 * Because the source is wired down, the page will be
2147 if (src_object != dst_object)
2148 VM_OBJECT_RLOCK(src_object);
2149 object = src_object;
2150 pindex = src_pindex + dst_pindex;
2151 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
2152 (backing_object = object->backing_object) != NULL) {
2154 * Unless the source mapping is read-only or
2155 * it is presently being upgraded from
2156 * read-only, the first object in the shadow
2157 * chain should provide all of the pages. In
2158 * other words, this loop body should never be
2159 * executed when the source mapping is already
2162 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
2164 ("vm_fault_copy_entry: main object missing page"));
2166 VM_OBJECT_RLOCK(backing_object);
2167 pindex += OFF_TO_IDX(object->backing_object_offset);
2168 if (object != dst_object)
2169 VM_OBJECT_RUNLOCK(object);
2170 object = backing_object;
2172 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2174 if (object != dst_object) {
2176 * Allocate a page in the destination object.
2178 dst_m = vm_page_alloc(dst_object, (src_object ==
2179 dst_object ? src_pindex : 0) + dst_pindex,
2181 if (dst_m == NULL) {
2182 VM_OBJECT_WUNLOCK(dst_object);
2183 VM_OBJECT_RUNLOCK(object);
2184 vm_wait(dst_object);
2185 VM_OBJECT_WLOCK(dst_object);
2190 * See the comment in vm_fault_cow().
2192 if (src_object == dst_object &&
2193 (object->flags & OBJ_ONEMAPPING) == 0)
2194 pmap_remove_all(src_m);
2195 pmap_copy_page(src_m, dst_m);
2198 * The object lock does not guarantee that "src_m" will
2199 * transition from invalid to valid, but it does ensure
2200 * that "src_m" will not transition from valid to
2203 dst_m->dirty = dst_m->valid = src_m->valid;
2204 VM_OBJECT_RUNLOCK(object);
2207 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2209 if (dst_m->pindex >= dst_object->size) {
2211 * We are upgrading. Index can occur
2212 * out of bounds if the object type is
2213 * vnode and the file was truncated.
2215 vm_page_xunbusy(dst_m);
2221 * Enter it in the pmap. If a wired, copy-on-write
2222 * mapping is being replaced by a write-enabled
2223 * mapping, then wire that new mapping.
2225 * The page can be invalid if the user called
2226 * msync(MS_INVALIDATE) or truncated the backing vnode
2227 * or shared memory object. In this case, do not
2228 * insert it into pmap, but still do the copy so that
2229 * all copies of the wired map entry have similar
2232 if (vm_page_all_valid(dst_m)) {
2233 VM_OBJECT_WUNLOCK(dst_object);
2234 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2235 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2236 VM_OBJECT_WLOCK(dst_object);
2240 * Mark it no longer busy, and put it on the active list.
2243 if (src_m != dst_m) {
2244 vm_page_unwire(src_m, PQ_INACTIVE);
2245 vm_page_wire(dst_m);
2247 KASSERT(vm_page_wired(dst_m),
2248 ("dst_m %p is not wired", dst_m));
2251 vm_page_activate(dst_m);
2253 vm_page_xunbusy(dst_m);
2255 VM_OBJECT_WUNLOCK(dst_object);
2257 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2258 vm_object_deallocate(src_object);
2263 * Block entry into the machine-independent layer's page fault handler by
2264 * the calling thread. Subsequent calls to vm_fault() by that thread will
2265 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
2266 * spurious page faults.
2269 vm_fault_disable_pagefaults(void)
2272 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2276 vm_fault_enable_pagefaults(int save)
2279 curthread_pflags_restore(save);