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 vm_object_busy(fs->first_object);
353 m = vm_page_lookup(fs->first_object, fs->first_pindex);
354 /* A busy page can be mapped for read|execute access. */
355 if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 &&
356 vm_page_busied(m)) || !vm_page_all_valid(m))
360 #if VM_NRESERVLEVEL > 0
361 if ((m->flags & PG_FICTITIOUS) == 0 &&
362 (m_super = vm_reserv_to_superpage(m)) != NULL &&
363 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
364 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
365 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
366 (pagesizes[m_super->psind] - 1)) && !fs->wired &&
367 pmap_ps_enabled(fs->map->pmap)) {
368 flags = PS_ALL_VALID;
369 if ((fs->prot & VM_PROT_WRITE) != 0) {
371 * Create a superpage mapping allowing write access
372 * only if none of the constituent pages are busy and
373 * all of them are already dirty (except possibly for
374 * the page that was faulted on).
376 flags |= PS_NONE_BUSY;
377 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
378 flags |= PS_ALL_DIRTY;
380 if (vm_page_ps_test(m_super, flags, m)) {
382 psind = m_super->psind;
383 vaddr = rounddown2(vaddr, pagesizes[psind]);
384 /* Preset the modified bit for dirty superpages. */
385 if ((flags & PS_ALL_DIRTY) != 0)
386 fs->fault_type |= VM_PROT_WRITE;
390 if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
391 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) !=
394 if (fs->m_hold != NULL) {
398 if (psind == 0 && !fs->wired)
399 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
400 VM_OBJECT_RUNLOCK(fs->first_object);
401 vm_fault_dirty(fs, m);
402 vm_object_unbusy(fs->first_object);
403 vm_map_lookup_done(fs->map, fs->entry);
404 curthread->td_ru.ru_minflt++;
405 return (FAULT_SUCCESS);
407 vm_object_unbusy(fs->first_object);
408 return (FAULT_FAILURE);
412 vm_fault_restore_map_lock(struct faultstate *fs)
415 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
416 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
418 if (!vm_map_trylock_read(fs->map)) {
419 VM_OBJECT_WUNLOCK(fs->first_object);
420 vm_map_lock_read(fs->map);
421 VM_OBJECT_WLOCK(fs->first_object);
423 fs->lookup_still_valid = true;
427 vm_fault_populate_check_page(vm_page_t m)
431 * Check each page to ensure that the pager is obeying the
432 * interface: the page must be installed in the object, fully
433 * valid, and exclusively busied.
436 MPASS(vm_page_all_valid(m));
437 MPASS(vm_page_xbusied(m));
441 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
447 VM_OBJECT_ASSERT_WLOCKED(object);
448 MPASS(first <= last);
449 for (pidx = first, m = vm_page_lookup(object, pidx);
450 pidx <= last; pidx++, m = vm_page_next(m)) {
451 vm_fault_populate_check_page(m);
452 vm_page_deactivate(m);
457 static enum fault_status
458 vm_fault_populate(struct faultstate *fs)
462 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
463 int bdry_idx, i, npages, psind, rv;
464 enum fault_status res;
466 MPASS(fs->object == fs->first_object);
467 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
468 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
469 MPASS(fs->first_object->backing_object == NULL);
470 MPASS(fs->lookup_still_valid);
472 pager_first = OFF_TO_IDX(fs->entry->offset);
473 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
474 vm_fault_unlock_map(fs);
475 vm_fault_unlock_vp(fs);
480 * Call the pager (driver) populate() method.
482 * There is no guarantee that the method will be called again
483 * if the current fault is for read, and a future fault is
484 * for write. Report the entry's maximum allowed protection
487 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
488 fs->fault_type, fs->entry->max_protection, &pager_first,
491 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
492 if (rv == VM_PAGER_BAD) {
494 * VM_PAGER_BAD is the backdoor for a pager to request
495 * normal fault handling.
497 vm_fault_restore_map_lock(fs);
498 if (fs->map->timestamp != fs->map_generation)
499 return (FAULT_RESTART);
500 return (FAULT_CONTINUE);
502 if (rv != VM_PAGER_OK)
503 return (FAULT_FAILURE); /* AKA SIGSEGV */
505 /* Ensure that the driver is obeying the interface. */
506 MPASS(pager_first <= pager_last);
507 MPASS(fs->first_pindex <= pager_last);
508 MPASS(fs->first_pindex >= pager_first);
509 MPASS(pager_last < fs->first_object->size);
511 vm_fault_restore_map_lock(fs);
512 bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
513 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
514 if (fs->map->timestamp != fs->map_generation) {
516 vm_fault_populate_cleanup(fs->first_object, pager_first,
519 m = vm_page_lookup(fs->first_object, pager_first);
523 return (FAULT_RESTART);
527 * The map is unchanged after our last unlock. Process the fault.
529 * First, the special case of largepage mappings, where
530 * populate only busies the first page in superpage run.
533 KASSERT(PMAP_HAS_LARGEPAGES,
534 ("missing pmap support for large pages"));
535 m = vm_page_lookup(fs->first_object, pager_first);
536 vm_fault_populate_check_page(m);
537 VM_OBJECT_WUNLOCK(fs->first_object);
538 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
540 /* assert alignment for entry */
541 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
542 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
543 (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
544 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
545 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
546 ("unaligned superpage m %p %#jx", m,
547 (uintmax_t)VM_PAGE_TO_PHYS(m)));
548 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
549 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
550 PMAP_ENTER_LARGEPAGE, bdry_idx);
551 VM_OBJECT_WLOCK(fs->first_object);
553 if (rv != KERN_SUCCESS) {
557 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
558 for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
561 if (fs->m_hold != NULL) {
562 *fs->m_hold = m + (fs->first_pindex - pager_first);
563 vm_page_wire(*fs->m_hold);
569 * The range [pager_first, pager_last] that is given to the
570 * pager is only a hint. The pager may populate any range
571 * within the object that includes the requested page index.
572 * In case the pager expanded the range, clip it to fit into
575 map_first = OFF_TO_IDX(fs->entry->offset);
576 if (map_first > pager_first) {
577 vm_fault_populate_cleanup(fs->first_object, pager_first,
579 pager_first = map_first;
581 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
582 if (map_last < pager_last) {
583 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
585 pager_last = map_last;
587 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
589 pidx += npages, m = vm_page_next(&m[npages - 1])) {
590 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
593 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
594 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
595 !pmap_ps_enabled(fs->map->pmap) || fs->wired))
598 npages = atop(pagesizes[psind]);
599 for (i = 0; i < npages; i++) {
600 vm_fault_populate_check_page(&m[i]);
601 vm_fault_dirty(fs, &m[i]);
603 VM_OBJECT_WUNLOCK(fs->first_object);
604 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
605 (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
608 * pmap_enter() may fail for a superpage mapping if additional
609 * protection policies prevent the full mapping.
610 * For example, this will happen on amd64 if the entire
611 * address range does not share the same userspace protection
612 * key. Revert to single-page mappings if this happens.
614 MPASS(rv == KERN_SUCCESS ||
615 (psind > 0 && rv == KERN_PROTECTION_FAILURE));
616 if (__predict_false(psind > 0 &&
617 rv == KERN_PROTECTION_FAILURE)) {
619 for (i = 0; i < npages; i++) {
620 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
621 &m[i], fs->prot, fs->fault_type, 0);
622 MPASS(rv == KERN_SUCCESS);
626 VM_OBJECT_WLOCK(fs->first_object);
627 for (i = 0; i < npages; i++) {
628 if ((fs->fault_flags & VM_FAULT_WIRE) != 0 &&
629 m[i].pindex == fs->first_pindex)
632 vm_page_activate(&m[i]);
633 if (fs->m_hold != NULL &&
634 m[i].pindex == fs->first_pindex) {
635 (*fs->m_hold) = &m[i];
638 vm_page_xunbusy(&m[i]);
642 curthread->td_ru.ru_majflt++;
646 static int prot_fault_translation;
647 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
648 &prot_fault_translation, 0,
649 "Control signal to deliver on protection fault");
651 /* compat definition to keep common code for signal translation */
652 #define UCODE_PAGEFLT 12
654 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
660 * Handle a page fault occurring at the given address,
661 * requiring the given permissions, in the map specified.
662 * If successful, the page is inserted into the
663 * associated physical map.
665 * NOTE: the given address should be truncated to the
666 * proper page address.
668 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
669 * a standard error specifying why the fault is fatal is returned.
671 * The map in question must be referenced, and remains so.
672 * Caller may hold no locks.
675 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
676 int fault_flags, int *signo, int *ucode)
680 MPASS(signo == NULL || ucode != NULL);
682 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
683 ktrfault(vaddr, fault_type);
685 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
687 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
688 result == KERN_INVALID_ADDRESS ||
689 result == KERN_RESOURCE_SHORTAGE ||
690 result == KERN_PROTECTION_FAILURE ||
691 result == KERN_OUT_OF_BOUNDS,
692 ("Unexpected Mach error %d from vm_fault()", result));
694 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
697 if (result != KERN_SUCCESS && signo != NULL) {
700 case KERN_INVALID_ADDRESS:
702 *ucode = SEGV_MAPERR;
704 case KERN_RESOURCE_SHORTAGE:
708 case KERN_OUT_OF_BOUNDS:
712 case KERN_PROTECTION_FAILURE:
713 if (prot_fault_translation == 0) {
715 * Autodetect. This check also covers
716 * the images without the ABI-tag ELF
719 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
720 curproc->p_osrel >= P_OSREL_SIGSEGV) {
722 *ucode = SEGV_ACCERR;
725 *ucode = UCODE_PAGEFLT;
727 } else if (prot_fault_translation == 1) {
728 /* Always compat mode. */
730 *ucode = UCODE_PAGEFLT;
732 /* Always SIGSEGV mode. */
734 *ucode = SEGV_ACCERR;
738 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
747 vm_fault_object_ensure_wlocked(struct faultstate *fs)
749 if (fs->object == fs->first_object)
750 VM_OBJECT_ASSERT_WLOCKED(fs->object);
752 if (!fs->can_read_lock) {
753 VM_OBJECT_ASSERT_WLOCKED(fs->object);
757 if (VM_OBJECT_WOWNED(fs->object))
760 if (VM_OBJECT_TRYUPGRADE(fs->object))
766 static enum fault_status
767 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
772 if (fs->object->type != OBJT_VNODE)
773 return (FAULT_CONTINUE);
774 vp = fs->object->handle;
776 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
777 return (FAULT_CONTINUE);
781 * Perform an unlock in case the desired vnode changed while
782 * the map was unlocked during a retry.
784 vm_fault_unlock_vp(fs);
786 locked = VOP_ISLOCKED(vp);
787 if (locked != LK_EXCLUSIVE)
791 * We must not sleep acquiring the vnode lock while we have
792 * the page exclusive busied or the object's
793 * paging-in-progress count incremented. Otherwise, we could
796 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
799 return (FAULT_CONTINUE);
804 vm_fault_unlock_and_deallocate(fs);
806 vm_fault_deallocate(fs);
807 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
810 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
811 return (FAULT_RESTART);
815 * Calculate the desired readahead. Handle drop-behind.
817 * Returns the number of readahead blocks to pass to the pager.
820 vm_fault_readahead(struct faultstate *fs)
825 KASSERT(fs->lookup_still_valid, ("map unlocked"));
826 era = fs->entry->read_ahead;
827 behavior = vm_map_entry_behavior(fs->entry);
828 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
830 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
831 nera = VM_FAULT_READ_AHEAD_MAX;
832 if (fs->vaddr == fs->entry->next_read)
833 vm_fault_dontneed(fs, fs->vaddr, nera);
834 } else if (fs->vaddr == fs->entry->next_read) {
836 * This is a sequential fault. Arithmetically
837 * increase the requested number of pages in
838 * the read-ahead window. The requested
839 * number of pages is "# of sequential faults
840 * x (read ahead min + 1) + read ahead min"
842 nera = VM_FAULT_READ_AHEAD_MIN;
845 if (nera > VM_FAULT_READ_AHEAD_MAX)
846 nera = VM_FAULT_READ_AHEAD_MAX;
848 if (era == VM_FAULT_READ_AHEAD_MAX)
849 vm_fault_dontneed(fs, fs->vaddr, nera);
852 * This is a non-sequential fault.
858 * A read lock on the map suffices to update
859 * the read ahead count safely.
861 fs->entry->read_ahead = nera;
868 vm_fault_lookup(struct faultstate *fs)
872 KASSERT(!fs->lookup_still_valid,
873 ("vm_fault_lookup: Map already locked."));
874 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
875 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
876 &fs->first_pindex, &fs->prot, &fs->wired);
877 if (result != KERN_SUCCESS) {
878 vm_fault_unlock_vp(fs);
882 fs->map_generation = fs->map->timestamp;
884 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
885 panic("%s: fault on nofault entry, addr: %#lx",
886 __func__, (u_long)fs->vaddr);
889 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
890 fs->entry->wiring_thread != curthread) {
891 vm_map_unlock_read(fs->map);
892 vm_map_lock(fs->map);
893 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
894 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
895 vm_fault_unlock_vp(fs);
896 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
897 vm_map_unlock_and_wait(fs->map, 0);
899 vm_map_unlock(fs->map);
900 return (KERN_RESOURCE_SHORTAGE);
903 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
906 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
908 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
909 ("!fs->wired && VM_FAULT_WIRE"));
910 fs->lookup_still_valid = true;
912 return (KERN_SUCCESS);
916 vm_fault_relookup(struct faultstate *fs)
918 vm_object_t retry_object;
919 vm_pindex_t retry_pindex;
920 vm_prot_t retry_prot;
923 if (!vm_map_trylock_read(fs->map))
924 return (KERN_RESTART);
926 fs->lookup_still_valid = true;
927 if (fs->map->timestamp == fs->map_generation)
928 return (KERN_SUCCESS);
930 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
931 &fs->entry, &retry_object, &retry_pindex, &retry_prot,
933 if (result != KERN_SUCCESS) {
935 * If retry of map lookup would have blocked then
936 * retry fault from start.
938 if (result == KERN_FAILURE)
939 return (KERN_RESTART);
942 if (retry_object != fs->first_object ||
943 retry_pindex != fs->first_pindex)
944 return (KERN_RESTART);
947 * Check whether the protection has changed or the object has
948 * been copied while we left the map unlocked. Changing from
949 * read to write permission is OK - we leave the page
950 * write-protected, and catch the write fault. Changing from
951 * write to read permission means that we can't mark the page
952 * write-enabled after all.
954 fs->prot &= retry_prot;
955 fs->fault_type &= retry_prot;
957 return (KERN_RESTART);
959 /* Reassert because wired may have changed. */
960 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
961 ("!wired && VM_FAULT_WIRE"));
963 return (KERN_SUCCESS);
967 vm_fault_cow(struct faultstate *fs)
969 bool is_first_object_locked;
971 KASSERT(fs->object != fs->first_object,
972 ("source and target COW objects are identical"));
975 * This allows pages to be virtually copied from a backing_object
976 * into the first_object, where the backing object has no other
977 * refs to it, and cannot gain any more refs. Instead of a bcopy,
978 * we just move the page from the backing object to the first
979 * object. Note that we must mark the page dirty in the first
980 * object so that it will go out to swap when needed.
982 is_first_object_locked = false;
985 * Only one shadow object and no other refs.
987 fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
989 * No other ways to look the object up
991 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
993 * We don't chase down the shadow chain and we can acquire locks.
995 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
996 fs->object == fs->first_object->backing_object &&
997 VM_OBJECT_TRYWLOCK(fs->object)) {
999 * Remove but keep xbusy for replace. fs->m is moved into
1000 * fs->first_object and left busy while fs->first_m is
1001 * conditionally freed.
1003 vm_page_remove_xbusy(fs->m);
1004 vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
1006 vm_page_dirty(fs->m);
1007 #if VM_NRESERVLEVEL > 0
1009 * Rename the reservation.
1011 vm_reserv_rename(fs->m, fs->first_object, fs->object,
1012 OFF_TO_IDX(fs->first_object->backing_object_offset));
1014 VM_OBJECT_WUNLOCK(fs->object);
1015 VM_OBJECT_WUNLOCK(fs->first_object);
1016 fs->first_m = fs->m;
1018 VM_CNT_INC(v_cow_optim);
1020 if (is_first_object_locked)
1021 VM_OBJECT_WUNLOCK(fs->first_object);
1023 * Oh, well, lets copy it.
1025 pmap_copy_page(fs->m, fs->first_m);
1026 vm_page_valid(fs->first_m);
1027 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
1028 vm_page_wire(fs->first_m);
1029 vm_page_unwire(fs->m, PQ_INACTIVE);
1032 * Save the cow page to be released after
1033 * pmap_enter is complete.
1039 * Typically, the shadow object is either private to this
1040 * address space (OBJ_ONEMAPPING) or its pages are read only.
1041 * In the highly unusual case where the pages of a shadow object
1042 * are read/write shared between this and other address spaces,
1043 * we need to ensure that any pmap-level mappings to the
1044 * original, copy-on-write page from the backing object are
1045 * removed from those other address spaces.
1047 * The flag check is racy, but this is tolerable: if
1048 * OBJ_ONEMAPPING is cleared after the check, the busy state
1049 * ensures that new mappings of m_cow can't be created.
1050 * pmap_enter() will replace an existing mapping in the current
1051 * address space. If OBJ_ONEMAPPING is set after the check,
1052 * removing mappings will at worse trigger some unnecessary page
1055 vm_page_assert_xbusied(fs->m_cow);
1056 if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
1057 pmap_remove_all(fs->m_cow);
1060 vm_object_pip_wakeup(fs->object);
1063 * Only use the new page below...
1065 fs->object = fs->first_object;
1066 fs->pindex = fs->first_pindex;
1067 fs->m = fs->first_m;
1068 VM_CNT_INC(v_cow_faults);
1069 curthread->td_cow++;
1072 static enum fault_next_status
1073 vm_fault_next(struct faultstate *fs)
1075 vm_object_t next_object;
1077 if (fs->object == fs->first_object || !fs->can_read_lock)
1078 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1080 VM_OBJECT_ASSERT_LOCKED(fs->object);
1083 * The requested page does not exist at this object/
1084 * offset. Remove the invalid page from the object,
1085 * waking up anyone waiting for it, and continue on to
1086 * the next object. However, if this is the top-level
1087 * object, we must leave the busy page in place to
1088 * prevent another process from rushing past us, and
1089 * inserting the page in that object at the same time
1092 if (fs->object == fs->first_object) {
1093 fs->first_m = fs->m;
1095 } else if (fs->m != NULL) {
1096 if (!vm_fault_object_ensure_wlocked(fs)) {
1097 fs->can_read_lock = false;
1098 vm_fault_unlock_and_deallocate(fs);
1099 return (FAULT_NEXT_RESTART);
1101 vm_fault_page_free(&fs->m);
1105 * Move on to the next object. Lock the next object before
1106 * unlocking the current one.
1108 next_object = fs->object->backing_object;
1109 if (next_object == NULL)
1110 return (FAULT_NEXT_NOOBJ);
1111 MPASS(fs->first_m != NULL);
1112 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1113 if (fs->can_read_lock)
1114 VM_OBJECT_RLOCK(next_object);
1116 VM_OBJECT_WLOCK(next_object);
1117 vm_object_pip_add(next_object, 1);
1118 if (fs->object != fs->first_object)
1119 vm_object_pip_wakeup(fs->object);
1120 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1121 VM_OBJECT_UNLOCK(fs->object);
1122 fs->object = next_object;
1124 return (FAULT_NEXT_GOTOBJ);
1128 vm_fault_zerofill(struct faultstate *fs)
1132 * If there's no object left, fill the page in the top
1133 * object with zeros.
1135 if (fs->object != fs->first_object) {
1136 vm_object_pip_wakeup(fs->object);
1137 fs->object = fs->first_object;
1138 fs->pindex = fs->first_pindex;
1140 MPASS(fs->first_m != NULL);
1141 MPASS(fs->m == NULL);
1142 fs->m = fs->first_m;
1146 * Zero the page if necessary and mark it valid.
1148 if ((fs->m->flags & PG_ZERO) == 0) {
1149 pmap_zero_page(fs->m);
1151 VM_CNT_INC(v_ozfod);
1154 vm_page_valid(fs->m);
1158 * Initiate page fault after timeout. Returns true if caller should
1159 * do vm_waitpfault() after the call.
1162 vm_fault_allocate_oom(struct faultstate *fs)
1166 vm_fault_unlock_and_deallocate(fs);
1167 if (vm_pfault_oom_attempts < 0)
1169 if (!fs->oom_started) {
1170 fs->oom_started = true;
1171 getmicrotime(&fs->oom_start_time);
1176 timevalsub(&now, &fs->oom_start_time);
1177 if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait)
1182 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1183 curproc->p_pid, curproc->p_comm);
1184 vm_pageout_oom(VM_OOM_MEM_PF);
1185 fs->oom_started = false;
1190 * Allocate a page directly or via the object populate method.
1192 static enum fault_status
1193 vm_fault_allocate(struct faultstate *fs)
1195 struct domainset *dset;
1196 enum fault_status res;
1198 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1199 res = vm_fault_lock_vnode(fs, true);
1200 MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART);
1201 if (res == FAULT_RESTART)
1205 if (fs->pindex >= fs->object->size) {
1206 vm_fault_unlock_and_deallocate(fs);
1207 return (FAULT_OUT_OF_BOUNDS);
1210 if (fs->object == fs->first_object &&
1211 (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1212 fs->first_object->shadow_count == 0) {
1213 res = vm_fault_populate(fs);
1218 vm_fault_unlock_and_deallocate(fs);
1220 case FAULT_CONTINUE:
1222 * Pager's populate() method
1223 * returned VM_PAGER_BAD.
1227 panic("inconsistent return codes");
1232 * Allocate a new page for this object/offset pair.
1234 * If the process has a fatal signal pending, prioritize the allocation
1235 * with the expectation that the process will exit shortly and free some
1236 * pages. In particular, the signal may have been posted by the page
1237 * daemon in an attempt to resolve an out-of-memory condition.
1239 * The unlocked read of the p_flag is harmless. At worst, the P_KILLED
1240 * might be not observed here, and allocation fails, causing a restart
1241 * and new reading of the p_flag.
1243 dset = fs->object->domain.dr_policy;
1245 dset = curthread->td_domain.dr_policy;
1246 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1247 #if VM_NRESERVLEVEL > 0
1248 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1250 if (!vm_pager_can_alloc_page(fs->object, fs->pindex)) {
1251 vm_fault_unlock_and_deallocate(fs);
1252 return (FAULT_FAILURE);
1254 fs->m = vm_page_alloc(fs->object, fs->pindex,
1255 P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0);
1257 if (fs->m == NULL) {
1258 if (vm_fault_allocate_oom(fs))
1259 vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1260 return (FAULT_RESTART);
1262 fs->oom_started = false;
1264 return (FAULT_CONTINUE);
1268 * Call the pager to retrieve the page if there is a chance
1269 * that the pager has it, and potentially retrieve additional
1270 * pages at the same time.
1272 static enum fault_status
1273 vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp)
1275 vm_offset_t e_end, e_start;
1276 int ahead, behind, cluster_offset, rv;
1277 enum fault_status status;
1281 * Prepare for unlocking the map. Save the map
1282 * entry's start and end addresses, which are used to
1283 * optimize the size of the pager operation below.
1284 * Even if the map entry's addresses change after
1285 * unlocking the map, using the saved addresses is
1288 e_start = fs->entry->start;
1289 e_end = fs->entry->end;
1290 behavior = vm_map_entry_behavior(fs->entry);
1293 * If the pager for the current object might have
1294 * the page, then determine the number of additional
1295 * pages to read and potentially reprioritize
1296 * previously read pages for earlier reclamation.
1297 * These operations should only be performed once per
1298 * page fault. Even if the current pager doesn't
1299 * have the page, the number of additional pages to
1300 * read will apply to subsequent objects in the
1303 if (fs->nera == -1 && !P_KILLED(curproc))
1304 fs->nera = vm_fault_readahead(fs);
1307 * Release the map lock before locking the vnode or
1308 * sleeping in the pager. (If the current object has
1309 * a shadow, then an earlier iteration of this loop
1310 * may have already unlocked the map.)
1312 vm_fault_unlock_map(fs);
1314 status = vm_fault_lock_vnode(fs, false);
1315 MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART);
1316 if (status == FAULT_RESTART)
1318 KASSERT(fs->vp == NULL || !fs->map->system_map,
1319 ("vm_fault: vnode-backed object mapped by system map"));
1322 * Page in the requested page and hint the pager,
1323 * that it may bring up surrounding pages.
1325 if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1326 P_KILLED(curproc)) {
1330 /* Is this a sequential fault? */
1336 * Request a cluster of pages that is
1337 * aligned to a VM_FAULT_READ_DEFAULT
1338 * page offset boundary within the
1339 * object. Alignment to a page offset
1340 * boundary is more likely to coincide
1341 * with the underlying file system
1342 * block than alignment to a virtual
1345 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1346 behind = ulmin(cluster_offset,
1347 atop(fs->vaddr - e_start));
1348 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1350 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1354 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1355 if (rv == VM_PAGER_OK)
1356 return (FAULT_HARD);
1357 if (rv == VM_PAGER_ERROR)
1358 printf("vm_fault: pager read error, pid %d (%s)\n",
1359 curproc->p_pid, curproc->p_comm);
1361 * If an I/O error occurred or the requested page was
1362 * outside the range of the pager, clean up and return
1365 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1366 VM_OBJECT_WLOCK(fs->object);
1367 vm_fault_page_free(&fs->m);
1368 vm_fault_unlock_and_deallocate(fs);
1369 return (FAULT_OUT_OF_BOUNDS);
1371 KASSERT(rv == VM_PAGER_FAIL,
1372 ("%s: unexpected pager error %d", __func__, rv));
1373 return (FAULT_CONTINUE);
1377 * Wait/Retry if the page is busy. We have to do this if the page is
1378 * either exclusive or shared busy because the vm_pager may be using
1379 * read busy for pageouts (and even pageins if it is the vnode pager),
1380 * and we could end up trying to pagein and pageout the same page
1383 * We can theoretically allow the busy case on a read fault if the page
1384 * is marked valid, but since such pages are typically already pmap'd,
1385 * putting that special case in might be more effort then it is worth.
1386 * We cannot under any circumstances mess around with a shared busied
1387 * page except, perhaps, to pmap it.
1390 vm_fault_busy_sleep(struct faultstate *fs)
1393 * Reference the page before unlocking and
1394 * sleeping so that the page daemon is less
1395 * likely to reclaim it.
1397 vm_page_aflag_set(fs->m, PGA_REFERENCED);
1398 if (fs->object != fs->first_object) {
1399 vm_fault_page_release(&fs->first_m);
1400 vm_object_pip_wakeup(fs->first_object);
1402 vm_object_pip_wakeup(fs->object);
1403 vm_fault_unlock_map(fs);
1404 if (fs->m != vm_page_lookup(fs->object, fs->pindex) ||
1405 !vm_page_busy_sleep(fs->m, "vmpfw", 0))
1406 VM_OBJECT_UNLOCK(fs->object);
1407 VM_CNT_INC(v_intrans);
1408 vm_object_deallocate(fs->first_object);
1412 * Handle page lookup, populate, allocate, page-in for the current
1415 * The object is locked on entry and will remain locked with a return
1416 * code of FAULT_CONTINUE so that fault may follow the shadow chain.
1417 * Otherwise, the object will be unlocked upon return.
1419 static enum fault_status
1420 vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp)
1422 enum fault_status res;
1425 if (fs->object == fs->first_object || !fs->can_read_lock)
1426 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1428 VM_OBJECT_ASSERT_LOCKED(fs->object);
1431 * If the object is marked for imminent termination, we retry
1432 * here, since the collapse pass has raced with us. Otherwise,
1433 * if we see terminally dead object, return fail.
1435 if ((fs->object->flags & OBJ_DEAD) != 0) {
1436 dead = fs->object->type == OBJT_DEAD;
1437 vm_fault_unlock_and_deallocate(fs);
1439 return (FAULT_PROTECTION_FAILURE);
1441 return (FAULT_RESTART);
1445 * See if the page is resident.
1447 fs->m = vm_page_lookup(fs->object, fs->pindex);
1448 if (fs->m != NULL) {
1449 if (!vm_page_tryxbusy(fs->m)) {
1450 vm_fault_busy_sleep(fs);
1451 return (FAULT_RESTART);
1455 * The page is marked busy for other processes and the
1456 * pagedaemon. If it is still completely valid we are
1459 if (vm_page_all_valid(fs->m)) {
1460 VM_OBJECT_UNLOCK(fs->object);
1461 return (FAULT_SOFT);
1466 * Page is not resident. If the pager might contain the page
1467 * or this is the beginning of the search, allocate a new
1470 if (fs->m == NULL && (vm_fault_object_needs_getpages(fs->object) ||
1471 fs->object == fs->first_object)) {
1472 if (!vm_fault_object_ensure_wlocked(fs)) {
1473 fs->can_read_lock = false;
1474 vm_fault_unlock_and_deallocate(fs);
1475 return (FAULT_RESTART);
1477 res = vm_fault_allocate(fs);
1478 if (res != FAULT_CONTINUE)
1483 * Check to see if the pager can possibly satisfy this fault.
1484 * If not, skip to the next object without dropping the lock to
1485 * preserve atomicity of shadow faults.
1487 if (vm_fault_object_needs_getpages(fs->object)) {
1489 * At this point, we have either allocated a new page
1490 * or found an existing page that is only partially
1493 * We hold a reference on the current object and the
1494 * page is exclusive busied. The exclusive busy
1495 * prevents simultaneous faults and collapses while
1496 * the object lock is dropped.
1498 VM_OBJECT_UNLOCK(fs->object);
1499 res = vm_fault_getpages(fs, behindp, aheadp);
1500 if (res == FAULT_CONTINUE)
1501 VM_OBJECT_WLOCK(fs->object);
1503 res = FAULT_CONTINUE;
1509 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1510 int fault_flags, vm_page_t *m_hold)
1512 struct faultstate fs;
1513 int ahead, behind, faultcount, rv;
1514 enum fault_status res;
1515 enum fault_next_status res_next;
1518 VM_CNT_INC(v_vm_faults);
1520 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1521 return (KERN_PROTECTION_FAILURE);
1526 fs.fault_flags = fault_flags;
1528 fs.lookup_still_valid = false;
1529 fs.oom_started = false;
1531 fs.can_read_lock = true;
1536 fs.fault_type = fault_type;
1539 * Find the backing store object and offset into it to begin the
1542 rv = vm_fault_lookup(&fs);
1543 if (rv != KERN_SUCCESS) {
1544 if (rv == KERN_RESOURCE_SHORTAGE)
1550 * Try to avoid lock contention on the top-level object through
1551 * special-case handling of some types of page faults, specifically,
1552 * those that are mapping an existing page from the top-level object.
1553 * Under this condition, a read lock on the object suffices, allowing
1554 * multiple page faults of a similar type to run in parallel.
1556 if (fs.vp == NULL /* avoid locked vnode leak */ &&
1557 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1558 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1559 VM_OBJECT_RLOCK(fs.first_object);
1560 res = vm_fault_soft_fast(&fs);
1561 if (res == FAULT_SUCCESS)
1562 return (KERN_SUCCESS);
1563 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
1564 VM_OBJECT_RUNLOCK(fs.first_object);
1565 VM_OBJECT_WLOCK(fs.first_object);
1568 VM_OBJECT_WLOCK(fs.first_object);
1572 * Make a reference to this object to prevent its disposal while we
1573 * are messing with it. Once we have the reference, the map is free
1574 * to be diddled. Since objects reference their shadows (and copies),
1575 * they will stay around as well.
1577 * Bump the paging-in-progress count to prevent size changes (e.g.
1578 * truncation operations) during I/O.
1580 vm_object_reference_locked(fs.first_object);
1581 vm_object_pip_add(fs.first_object, 1);
1583 fs.m_cow = fs.m = fs.first_m = NULL;
1586 * Search for the page at object/offset.
1588 fs.object = fs.first_object;
1589 fs.pindex = fs.first_pindex;
1591 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1592 res = vm_fault_allocate(&fs);
1597 return (KERN_SUCCESS);
1599 return (KERN_FAILURE);
1600 case FAULT_OUT_OF_BOUNDS:
1601 return (KERN_OUT_OF_BOUNDS);
1602 case FAULT_CONTINUE:
1605 panic("vm_fault: Unhandled status %d", res);
1610 KASSERT(fs.m == NULL,
1611 ("page still set %p at loop start", fs.m));
1613 res = vm_fault_object(&fs, &behind, &ahead);
1618 faultcount = behind + 1 + ahead;
1624 return (KERN_SUCCESS);
1626 return (KERN_FAILURE);
1627 case FAULT_OUT_OF_BOUNDS:
1628 return (KERN_OUT_OF_BOUNDS);
1629 case FAULT_PROTECTION_FAILURE:
1630 return (KERN_PROTECTION_FAILURE);
1631 case FAULT_CONTINUE:
1634 panic("vm_fault: Unhandled status %d", res);
1638 * The page was not found in the current object. Try to
1639 * traverse into a backing object or zero fill if none is
1642 res_next = vm_fault_next(&fs);
1643 if (res_next == FAULT_NEXT_RESTART)
1645 else if (res_next == FAULT_NEXT_GOTOBJ)
1647 MPASS(res_next == FAULT_NEXT_NOOBJ);
1648 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1649 if (fs.first_object == fs.object)
1650 vm_fault_page_free(&fs.first_m);
1651 vm_fault_unlock_and_deallocate(&fs);
1652 return (KERN_OUT_OF_BOUNDS);
1654 VM_OBJECT_UNLOCK(fs.object);
1655 vm_fault_zerofill(&fs);
1656 /* Don't try to prefault neighboring pages. */
1663 * A valid page has been found and exclusively busied. The
1664 * object lock must no longer be held.
1666 vm_page_assert_xbusied(fs.m);
1667 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1670 * If the page is being written, but isn't already owned by the
1671 * top-level object, we have to copy it into a new page owned by the
1674 if (fs.object != fs.first_object) {
1676 * We only really need to copy if we want to write it.
1678 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1681 * We only try to prefault read-only mappings to the
1682 * neighboring pages when this copy-on-write fault is
1683 * a hard fault. In other cases, trying to prefault
1684 * is typically wasted effort.
1686 if (faultcount == 0)
1690 fs.prot &= ~VM_PROT_WRITE;
1695 * We must verify that the maps have not changed since our last
1698 if (!fs.lookup_still_valid) {
1699 rv = vm_fault_relookup(&fs);
1700 if (rv != KERN_SUCCESS) {
1701 vm_fault_deallocate(&fs);
1702 if (rv == KERN_RESTART)
1707 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1710 * If the page was filled by a pager, save the virtual address that
1711 * should be faulted on next under a sequential access pattern to the
1712 * map entry. A read lock on the map suffices to update this address
1716 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1719 * Page must be completely valid or it is not fit to
1720 * map into user space. vm_pager_get_pages() ensures this.
1722 vm_page_assert_xbusied(fs.m);
1723 KASSERT(vm_page_all_valid(fs.m),
1724 ("vm_fault: page %p partially invalid", fs.m));
1726 vm_fault_dirty(&fs, fs.m);
1729 * Put this page into the physical map. We had to do the unlock above
1730 * because pmap_enter() may sleep. We don't put the page
1731 * back on the active queue until later so that the pageout daemon
1732 * won't find it (yet).
1734 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1735 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1736 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1738 vm_fault_prefault(&fs, vaddr,
1739 faultcount > 0 ? behind : PFBAK,
1740 faultcount > 0 ? ahead : PFFOR, false);
1743 * If the page is not wired down, then put it where the pageout daemon
1746 if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1749 vm_page_activate(fs.m);
1750 if (fs.m_hold != NULL) {
1751 (*fs.m_hold) = fs.m;
1754 vm_page_xunbusy(fs.m);
1758 * Unlock everything, and return
1760 vm_fault_deallocate(&fs);
1762 VM_CNT_INC(v_io_faults);
1763 curthread->td_ru.ru_majflt++;
1765 if (racct_enable && fs.object->type == OBJT_VNODE) {
1767 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1768 racct_add_force(curproc, RACCT_WRITEBPS,
1769 PAGE_SIZE + behind * PAGE_SIZE);
1770 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1772 racct_add_force(curproc, RACCT_READBPS,
1773 PAGE_SIZE + ahead * PAGE_SIZE);
1774 racct_add_force(curproc, RACCT_READIOPS, 1);
1776 PROC_UNLOCK(curproc);
1780 curthread->td_ru.ru_minflt++;
1782 return (KERN_SUCCESS);
1786 * Speed up the reclamation of pages that precede the faulting pindex within
1787 * the first object of the shadow chain. Essentially, perform the equivalent
1788 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1789 * the faulting pindex by the cluster size when the pages read by vm_fault()
1790 * cross a cluster-size boundary. The cluster size is the greater of the
1791 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1793 * When "fs->first_object" is a shadow object, the pages in the backing object
1794 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1795 * function must only be concerned with pages in the first object.
1798 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1800 vm_map_entry_t entry;
1801 vm_object_t first_object;
1802 vm_offset_t end, start;
1803 vm_page_t m, m_next;
1804 vm_pindex_t pend, pstart;
1807 VM_OBJECT_ASSERT_UNLOCKED(fs->object);
1808 first_object = fs->first_object;
1809 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1810 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1811 VM_OBJECT_RLOCK(first_object);
1812 size = VM_FAULT_DONTNEED_MIN;
1813 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1814 size = pagesizes[1];
1815 end = rounddown2(vaddr, size);
1816 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1817 (entry = fs->entry)->start < end) {
1818 if (end - entry->start < size)
1819 start = entry->start;
1822 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1823 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1825 m_next = vm_page_find_least(first_object, pstart);
1826 pend = OFF_TO_IDX(entry->offset) + atop(end -
1828 while ((m = m_next) != NULL && m->pindex < pend) {
1829 m_next = TAILQ_NEXT(m, listq);
1830 if (!vm_page_all_valid(m) ||
1835 * Don't clear PGA_REFERENCED, since it would
1836 * likely represent a reference by a different
1839 * Typically, at this point, prefetched pages
1840 * are still in the inactive queue. Only
1841 * pages that triggered page faults are in the
1842 * active queue. The test for whether the page
1843 * is in the inactive queue is racy; in the
1844 * worst case we will requeue the page
1847 if (!vm_page_inactive(m))
1848 vm_page_deactivate(m);
1851 VM_OBJECT_RUNLOCK(first_object);
1856 * vm_fault_prefault provides a quick way of clustering
1857 * pagefaults into a processes address space. It is a "cousin"
1858 * of vm_map_pmap_enter, except it runs at page fault time instead
1862 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1863 int backward, int forward, bool obj_locked)
1866 vm_map_entry_t entry;
1867 vm_object_t backing_object, lobject;
1868 vm_offset_t addr, starta;
1873 pmap = fs->map->pmap;
1874 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1879 if (addra < backward * PAGE_SIZE) {
1880 starta = entry->start;
1882 starta = addra - backward * PAGE_SIZE;
1883 if (starta < entry->start)
1884 starta = entry->start;
1888 * Generate the sequence of virtual addresses that are candidates for
1889 * prefaulting in an outward spiral from the faulting virtual address,
1890 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1891 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1892 * If the candidate address doesn't have a backing physical page, then
1893 * the loop immediately terminates.
1895 for (i = 0; i < 2 * imax(backward, forward); i++) {
1896 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1898 if (addr > addra + forward * PAGE_SIZE)
1901 if (addr < starta || addr >= entry->end)
1904 if (!pmap_is_prefaultable(pmap, addr))
1907 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1908 lobject = entry->object.vm_object;
1910 VM_OBJECT_RLOCK(lobject);
1911 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1912 !vm_fault_object_needs_getpages(lobject) &&
1913 (backing_object = lobject->backing_object) != NULL) {
1914 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1915 0, ("vm_fault_prefault: unaligned object offset"));
1916 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1917 VM_OBJECT_RLOCK(backing_object);
1918 if (!obj_locked || lobject != entry->object.vm_object)
1919 VM_OBJECT_RUNLOCK(lobject);
1920 lobject = backing_object;
1923 if (!obj_locked || lobject != entry->object.vm_object)
1924 VM_OBJECT_RUNLOCK(lobject);
1927 if (vm_page_all_valid(m) &&
1928 (m->flags & PG_FICTITIOUS) == 0)
1929 pmap_enter_quick(pmap, addr, m, entry->protection);
1930 if (!obj_locked || lobject != entry->object.vm_object)
1931 VM_OBJECT_RUNLOCK(lobject);
1936 * Hold each of the physical pages that are mapped by the specified range of
1937 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1938 * and allow the specified types of access, "prot". If all of the implied
1939 * pages are successfully held, then the number of held pages is returned
1940 * together with pointers to those pages in the array "ma". However, if any
1941 * of the pages cannot be held, -1 is returned.
1944 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1945 vm_prot_t prot, vm_page_t *ma, int max_count)
1947 vm_offset_t end, va;
1950 boolean_t pmap_failed;
1954 end = round_page(addr + len);
1955 addr = trunc_page(addr);
1957 if (!vm_map_range_valid(map, addr, end))
1960 if (atop(end - addr) > max_count)
1961 panic("vm_fault_quick_hold_pages: count > max_count");
1962 count = atop(end - addr);
1965 * Most likely, the physical pages are resident in the pmap, so it is
1966 * faster to try pmap_extract_and_hold() first.
1968 pmap_failed = FALSE;
1969 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1970 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1973 else if ((prot & VM_PROT_WRITE) != 0 &&
1974 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1976 * Explicitly dirty the physical page. Otherwise, the
1977 * caller's changes may go unnoticed because they are
1978 * performed through an unmanaged mapping or by a DMA
1981 * The object lock is not held here.
1982 * See vm_page_clear_dirty_mask().
1989 * One or more pages could not be held by the pmap. Either no
1990 * page was mapped at the specified virtual address or that
1991 * mapping had insufficient permissions. Attempt to fault in
1992 * and hold these pages.
1994 * If vm_fault_disable_pagefaults() was called,
1995 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1996 * acquire MD VM locks, which means we must not call
1997 * vm_fault(). Some (out of tree) callers mark
1998 * too wide a code area with vm_fault_disable_pagefaults()
1999 * already, use the VM_PROT_QUICK_NOFAULT flag to request
2000 * the proper behaviour explicitly.
2002 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
2003 (curthread->td_pflags & TDP_NOFAULTING) != 0)
2005 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
2006 if (*mp == NULL && vm_fault(map, va, prot,
2007 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
2012 for (mp = ma; mp < ma + count; mp++)
2014 vm_page_unwire(*mp, PQ_INACTIVE);
2020 * vm_fault_copy_entry
2022 * Create new object backing dst_entry with private copy of all
2023 * underlying pages. When src_entry is equal to dst_entry, function
2024 * implements COW for wired-down map entry. Otherwise, it forks
2025 * wired entry into dst_map.
2027 * In/out conditions:
2028 * The source and destination maps must be locked for write.
2029 * The source map entry must be wired down (or be a sharing map
2030 * entry corresponding to a main map entry that is wired down).
2033 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map __unused,
2034 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
2035 vm_ooffset_t *fork_charge)
2037 vm_object_t backing_object, dst_object, object, src_object;
2038 vm_pindex_t dst_pindex, pindex, src_pindex;
2039 vm_prot_t access, prot;
2045 upgrade = src_entry == dst_entry;
2046 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
2047 ("vm_fault_copy_entry: vm_object not NULL"));
2050 * If not an upgrade, then enter the mappings in the pmap as
2051 * read and/or execute accesses. Otherwise, enter them as
2054 * A writeable large page mapping is only created if all of
2055 * the constituent small page mappings are modified. Marking
2056 * PTEs as modified on inception allows promotion to happen
2057 * without taking potentially large number of soft faults.
2059 access = prot = dst_entry->protection;
2061 access &= ~VM_PROT_WRITE;
2063 src_object = src_entry->object.vm_object;
2064 src_pindex = OFF_TO_IDX(src_entry->offset);
2066 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2067 dst_object = src_object;
2068 vm_object_reference(dst_object);
2071 * Create the top-level object for the destination entry.
2072 * Doesn't actually shadow anything - we copy the pages
2075 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
2076 dst_entry->start), NULL, NULL, 0);
2077 #if VM_NRESERVLEVEL > 0
2078 dst_object->flags |= OBJ_COLORED;
2079 dst_object->pg_color = atop(dst_entry->start);
2081 dst_object->domain = src_object->domain;
2082 dst_object->charge = dst_entry->end - dst_entry->start;
2084 dst_entry->object.vm_object = dst_object;
2085 dst_entry->offset = 0;
2086 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
2089 VM_OBJECT_WLOCK(dst_object);
2090 if (fork_charge != NULL) {
2091 KASSERT(dst_entry->cred == NULL,
2092 ("vm_fault_copy_entry: leaked swp charge"));
2093 dst_object->cred = curthread->td_ucred;
2094 crhold(dst_object->cred);
2095 *fork_charge += dst_object->charge;
2096 } else if ((dst_object->flags & OBJ_SWAP) != 0 &&
2097 dst_object->cred == NULL) {
2098 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
2100 dst_object->cred = dst_entry->cred;
2101 dst_entry->cred = NULL;
2105 * Loop through all of the virtual pages within the entry's
2106 * range, copying each page from the source object to the
2107 * destination object. Since the source is wired, those pages
2108 * must exist. In contrast, the destination is pageable.
2109 * Since the destination object doesn't share any backing storage
2110 * with the source object, all of its pages must be dirtied,
2111 * regardless of whether they can be written.
2113 for (vaddr = dst_entry->start, dst_pindex = 0;
2114 vaddr < dst_entry->end;
2115 vaddr += PAGE_SIZE, dst_pindex++) {
2118 * Find the page in the source object, and copy it in.
2119 * Because the source is wired down, the page will be
2122 if (src_object != dst_object)
2123 VM_OBJECT_RLOCK(src_object);
2124 object = src_object;
2125 pindex = src_pindex + dst_pindex;
2126 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
2127 (backing_object = object->backing_object) != NULL) {
2129 * Unless the source mapping is read-only or
2130 * it is presently being upgraded from
2131 * read-only, the first object in the shadow
2132 * chain should provide all of the pages. In
2133 * other words, this loop body should never be
2134 * executed when the source mapping is already
2137 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
2139 ("vm_fault_copy_entry: main object missing page"));
2141 VM_OBJECT_RLOCK(backing_object);
2142 pindex += OFF_TO_IDX(object->backing_object_offset);
2143 if (object != dst_object)
2144 VM_OBJECT_RUNLOCK(object);
2145 object = backing_object;
2147 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2149 if (object != dst_object) {
2151 * Allocate a page in the destination object.
2153 dst_m = vm_page_alloc(dst_object, (src_object ==
2154 dst_object ? src_pindex : 0) + dst_pindex,
2156 if (dst_m == NULL) {
2157 VM_OBJECT_WUNLOCK(dst_object);
2158 VM_OBJECT_RUNLOCK(object);
2159 vm_wait(dst_object);
2160 VM_OBJECT_WLOCK(dst_object);
2165 * See the comment in vm_fault_cow().
2167 if (src_object == dst_object &&
2168 (object->flags & OBJ_ONEMAPPING) == 0)
2169 pmap_remove_all(src_m);
2170 pmap_copy_page(src_m, dst_m);
2173 * The object lock does not guarantee that "src_m" will
2174 * transition from invalid to valid, but it does ensure
2175 * that "src_m" will not transition from valid to
2178 dst_m->dirty = dst_m->valid = src_m->valid;
2179 VM_OBJECT_RUNLOCK(object);
2182 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2184 if (dst_m->pindex >= dst_object->size) {
2186 * We are upgrading. Index can occur
2187 * out of bounds if the object type is
2188 * vnode and the file was truncated.
2190 vm_page_xunbusy(dst_m);
2196 * Enter it in the pmap. If a wired, copy-on-write
2197 * mapping is being replaced by a write-enabled
2198 * mapping, then wire that new mapping.
2200 * The page can be invalid if the user called
2201 * msync(MS_INVALIDATE) or truncated the backing vnode
2202 * or shared memory object. In this case, do not
2203 * insert it into pmap, but still do the copy so that
2204 * all copies of the wired map entry have similar
2207 if (vm_page_all_valid(dst_m)) {
2208 VM_OBJECT_WUNLOCK(dst_object);
2209 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2210 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2211 VM_OBJECT_WLOCK(dst_object);
2215 * Mark it no longer busy, and put it on the active list.
2218 if (src_m != dst_m) {
2219 vm_page_unwire(src_m, PQ_INACTIVE);
2220 vm_page_wire(dst_m);
2222 KASSERT(vm_page_wired(dst_m),
2223 ("dst_m %p is not wired", dst_m));
2226 vm_page_activate(dst_m);
2228 vm_page_xunbusy(dst_m);
2230 VM_OBJECT_WUNLOCK(dst_object);
2232 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2233 vm_object_deallocate(src_object);
2238 * Block entry into the machine-independent layer's page fault handler by
2239 * the calling thread. Subsequent calls to vm_fault() by that thread will
2240 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
2241 * spurious page faults.
2244 vm_fault_disable_pagefaults(void)
2247 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2251 vm_fault_enable_pagefaults(int save)
2254 curthread_pflags_restore(save);