2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 * Copyright (c) 1994 John S. Dyson
8 * Copyright (c) 1994 David Greenman
12 * This code is derived from software contributed to Berkeley by
13 * The Mach Operating System project at Carnegie-Mellon University.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
46 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
47 * All rights reserved.
49 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
51 * Permission to use, copy, modify and distribute this software and
52 * its documentation is hereby granted, provided that both the copyright
53 * notice and this permission notice appear in all copies of the
54 * software, derivative works or modified versions, and any portions
55 * thereof, and that both notices appear in supporting documentation.
57 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
58 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
59 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
61 * Carnegie Mellon requests users of this software to return to
63 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
64 * School of Computer Science
65 * Carnegie Mellon University
66 * Pittsburgh PA 15213-3890
68 * any improvements or extensions that they make and grant Carnegie the
69 * rights to redistribute these changes.
73 * Page fault handling module.
76 #include <sys/cdefs.h>
77 __FBSDID("$FreeBSD$");
79 #include "opt_ktrace.h"
82 #include <sys/param.h>
83 #include <sys/systm.h>
84 #include <sys/kernel.h>
87 #include <sys/mutex.h>
89 #include <sys/racct.h>
90 #include <sys/refcount.h>
91 #include <sys/resourcevar.h>
92 #include <sys/rwlock.h>
93 #include <sys/signalvar.h>
94 #include <sys/sysctl.h>
95 #include <sys/sysent.h>
96 #include <sys/vmmeter.h>
97 #include <sys/vnode.h>
99 #include <sys/ktrace.h>
103 #include <vm/vm_param.h>
105 #include <vm/vm_map.h>
106 #include <vm/vm_object.h>
107 #include <vm/vm_page.h>
108 #include <vm/vm_pageout.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_pager.h>
111 #include <vm/vm_extern.h>
112 #include <vm/vm_reserv.h>
117 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
119 #define VM_FAULT_DONTNEED_MIN 1048576
122 /* Fault parameters. */
125 vm_prot_t fault_type;
131 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 = 1, /* 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 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
174 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
175 int backward, int forward, bool obj_locked);
177 static int vm_pfault_oom_attempts = 3;
178 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
179 &vm_pfault_oom_attempts, 0,
180 "Number of page allocation attempts in page fault handler before it "
181 "triggers OOM handling");
183 static int vm_pfault_oom_wait = 10;
184 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
185 &vm_pfault_oom_wait, 0,
186 "Number of seconds to wait for free pages before retrying "
187 "the page fault handler");
190 fault_page_release(vm_page_t *mp)
197 * We are likely to loop around again and attempt to busy
198 * this page. Deactivating it leaves it available for
199 * pageout while optimizing fault restarts.
201 vm_page_deactivate(m);
208 fault_page_free(vm_page_t *mp)
214 VM_OBJECT_ASSERT_WLOCKED(m->object);
215 if (!vm_page_wired(m))
224 unlock_map(struct faultstate *fs)
227 if (fs->lookup_still_valid) {
228 vm_map_lookup_done(fs->map, fs->entry);
229 fs->lookup_still_valid = false;
234 unlock_vp(struct faultstate *fs)
237 if (fs->vp != NULL) {
244 fault_deallocate(struct faultstate *fs)
247 fault_page_release(&fs->m_cow);
248 fault_page_release(&fs->m);
249 vm_object_pip_wakeup(fs->object);
250 if (fs->object != fs->first_object) {
251 VM_OBJECT_WLOCK(fs->first_object);
252 fault_page_free(&fs->first_m);
253 VM_OBJECT_WUNLOCK(fs->first_object);
254 vm_object_pip_wakeup(fs->first_object);
256 vm_object_deallocate(fs->first_object);
262 unlock_and_deallocate(struct faultstate *fs)
265 VM_OBJECT_WUNLOCK(fs->object);
266 fault_deallocate(fs);
270 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
274 if (((fs->prot & VM_PROT_WRITE) == 0 &&
275 (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
276 (m->oflags & VPO_UNMANAGED) != 0)
279 VM_PAGE_OBJECT_BUSY_ASSERT(m);
281 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
282 (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
283 (fs->fault_flags & VM_FAULT_DIRTY) != 0;
285 vm_object_set_writeable_dirty(m->object);
288 * If the fault is a write, we know that this page is being
289 * written NOW so dirty it explicitly to save on
290 * pmap_is_modified() calls later.
292 * Also, since the page is now dirty, we can possibly tell
293 * the pager to release any swap backing the page.
295 if (need_dirty && vm_page_set_dirty(m) == 0) {
297 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
298 * if the page is already dirty to prevent data written with
299 * the expectation of being synced from not being synced.
300 * Likewise if this entry does not request NOSYNC then make
301 * sure the page isn't marked NOSYNC. Applications sharing
302 * data should use the same flags to avoid ping ponging.
304 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
305 vm_page_aflag_set(m, PGA_NOSYNC);
307 vm_page_aflag_clear(m, PGA_NOSYNC);
313 * Unlocks fs.first_object and fs.map on success.
315 static enum fault_status
316 vm_fault_soft_fast(struct faultstate *fs)
319 #if VM_NRESERVLEVEL > 0
325 enum fault_status res;
327 MPASS(fs->vp == NULL);
331 vm_object_busy(fs->first_object);
332 m = vm_page_lookup(fs->first_object, fs->first_pindex);
333 /* A busy page can be mapped for read|execute access. */
334 if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 &&
335 vm_page_busied(m)) || !vm_page_all_valid(m)) {
341 #if VM_NRESERVLEVEL > 0
342 if ((m->flags & PG_FICTITIOUS) == 0 &&
343 (m_super = vm_reserv_to_superpage(m)) != NULL &&
344 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
345 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
346 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
347 (pagesizes[m_super->psind] - 1)) && !fs->wired &&
348 pmap_ps_enabled(fs->map->pmap)) {
349 flags = PS_ALL_VALID;
350 if ((fs->prot & VM_PROT_WRITE) != 0) {
352 * Create a superpage mapping allowing write access
353 * only if none of the constituent pages are busy and
354 * all of them are already dirty (except possibly for
355 * the page that was faulted on).
357 flags |= PS_NONE_BUSY;
358 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
359 flags |= PS_ALL_DIRTY;
361 if (vm_page_ps_test(m_super, flags, m)) {
363 psind = m_super->psind;
364 vaddr = rounddown2(vaddr, pagesizes[psind]);
365 /* Preset the modified bit for dirty superpages. */
366 if ((flags & PS_ALL_DIRTY) != 0)
367 fs->fault_type |= VM_PROT_WRITE;
371 if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
372 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) !=
377 if (fs->m_hold != NULL) {
381 if (psind == 0 && !fs->wired)
382 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
383 VM_OBJECT_RUNLOCK(fs->first_object);
384 vm_fault_dirty(fs, m);
385 vm_map_lookup_done(fs->map, fs->entry);
386 curthread->td_ru.ru_minflt++;
389 vm_object_unbusy(fs->first_object);
394 vm_fault_restore_map_lock(struct faultstate *fs)
397 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
398 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
400 if (!vm_map_trylock_read(fs->map)) {
401 VM_OBJECT_WUNLOCK(fs->first_object);
402 vm_map_lock_read(fs->map);
403 VM_OBJECT_WLOCK(fs->first_object);
405 fs->lookup_still_valid = true;
409 vm_fault_populate_check_page(vm_page_t m)
413 * Check each page to ensure that the pager is obeying the
414 * interface: the page must be installed in the object, fully
415 * valid, and exclusively busied.
418 MPASS(vm_page_all_valid(m));
419 MPASS(vm_page_xbusied(m));
423 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
429 VM_OBJECT_ASSERT_WLOCKED(object);
430 MPASS(first <= last);
431 for (pidx = first, m = vm_page_lookup(object, pidx);
432 pidx <= last; pidx++, m = vm_page_next(m)) {
433 vm_fault_populate_check_page(m);
434 vm_page_deactivate(m);
439 static enum fault_status
440 vm_fault_populate(struct faultstate *fs)
444 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
445 int bdry_idx, i, npages, psind, rv;
446 enum fault_status res;
448 MPASS(fs->object == fs->first_object);
449 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
450 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
451 MPASS(fs->first_object->backing_object == NULL);
452 MPASS(fs->lookup_still_valid);
454 pager_first = OFF_TO_IDX(fs->entry->offset);
455 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
462 * Call the pager (driver) populate() method.
464 * There is no guarantee that the method will be called again
465 * if the current fault is for read, and a future fault is
466 * for write. Report the entry's maximum allowed protection
469 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
470 fs->fault_type, fs->entry->max_protection, &pager_first,
473 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
474 if (rv == VM_PAGER_BAD) {
476 * VM_PAGER_BAD is the backdoor for a pager to request
477 * normal fault handling.
479 vm_fault_restore_map_lock(fs);
480 if (fs->map->timestamp != fs->map_generation)
481 return (FAULT_RESTART);
482 return (FAULT_CONTINUE);
484 if (rv != VM_PAGER_OK)
485 return (FAULT_FAILURE); /* AKA SIGSEGV */
487 /* Ensure that the driver is obeying the interface. */
488 MPASS(pager_first <= pager_last);
489 MPASS(fs->first_pindex <= pager_last);
490 MPASS(fs->first_pindex >= pager_first);
491 MPASS(pager_last < fs->first_object->size);
493 vm_fault_restore_map_lock(fs);
494 bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
495 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
496 if (fs->map->timestamp != fs->map_generation) {
498 vm_fault_populate_cleanup(fs->first_object, pager_first,
501 m = vm_page_lookup(fs->first_object, pager_first);
505 return (FAULT_RESTART);
509 * The map is unchanged after our last unlock. Process the fault.
511 * First, the special case of largepage mappings, where
512 * populate only busies the first page in superpage run.
515 KASSERT(PMAP_HAS_LARGEPAGES,
516 ("missing pmap support for large pages"));
517 m = vm_page_lookup(fs->first_object, pager_first);
518 vm_fault_populate_check_page(m);
519 VM_OBJECT_WUNLOCK(fs->first_object);
520 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
522 /* assert alignment for entry */
523 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
524 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
525 (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
526 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
527 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
528 ("unaligned superpage m %p %#jx", m,
529 (uintmax_t)VM_PAGE_TO_PHYS(m)));
530 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
531 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
532 PMAP_ENTER_LARGEPAGE, bdry_idx);
533 VM_OBJECT_WLOCK(fs->first_object);
535 if (rv != KERN_SUCCESS) {
539 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
540 for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
543 if (fs->m_hold != NULL) {
544 *fs->m_hold = m + (fs->first_pindex - pager_first);
545 vm_page_wire(*fs->m_hold);
551 * The range [pager_first, pager_last] that is given to the
552 * pager is only a hint. The pager may populate any range
553 * within the object that includes the requested page index.
554 * In case the pager expanded the range, clip it to fit into
557 map_first = OFF_TO_IDX(fs->entry->offset);
558 if (map_first > pager_first) {
559 vm_fault_populate_cleanup(fs->first_object, pager_first,
561 pager_first = map_first;
563 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
564 if (map_last < pager_last) {
565 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
567 pager_last = map_last;
569 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
571 pidx += npages, m = vm_page_next(&m[npages - 1])) {
572 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
575 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
576 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
577 !pmap_ps_enabled(fs->map->pmap) || fs->wired))
580 npages = atop(pagesizes[psind]);
581 for (i = 0; i < npages; i++) {
582 vm_fault_populate_check_page(&m[i]);
583 vm_fault_dirty(fs, &m[i]);
585 VM_OBJECT_WUNLOCK(fs->first_object);
586 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
587 (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
590 * pmap_enter() may fail for a superpage mapping if additional
591 * protection policies prevent the full mapping.
592 * For example, this will happen on amd64 if the entire
593 * address range does not share the same userspace protection
594 * key. Revert to single-page mappings if this happens.
596 MPASS(rv == KERN_SUCCESS ||
597 (psind > 0 && rv == KERN_PROTECTION_FAILURE));
598 if (__predict_false(psind > 0 &&
599 rv == KERN_PROTECTION_FAILURE)) {
601 for (i = 0; i < npages; i++) {
602 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
603 &m[i], fs->prot, fs->fault_type, 0);
604 MPASS(rv == KERN_SUCCESS);
608 VM_OBJECT_WLOCK(fs->first_object);
609 for (i = 0; i < npages; i++) {
610 if ((fs->fault_flags & VM_FAULT_WIRE) != 0 &&
611 m[i].pindex == fs->first_pindex)
614 vm_page_activate(&m[i]);
615 if (fs->m_hold != NULL &&
616 m[i].pindex == fs->first_pindex) {
617 (*fs->m_hold) = &m[i];
620 vm_page_xunbusy(&m[i]);
624 curthread->td_ru.ru_majflt++;
628 static int prot_fault_translation;
629 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
630 &prot_fault_translation, 0,
631 "Control signal to deliver on protection fault");
633 /* compat definition to keep common code for signal translation */
634 #define UCODE_PAGEFLT 12
636 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
642 * Handle a page fault occurring at the given address,
643 * requiring the given permissions, in the map specified.
644 * If successful, the page is inserted into the
645 * associated physical map.
647 * NOTE: the given address should be truncated to the
648 * proper page address.
650 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
651 * a standard error specifying why the fault is fatal is returned.
653 * The map in question must be referenced, and remains so.
654 * Caller may hold no locks.
657 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
658 int fault_flags, int *signo, int *ucode)
662 MPASS(signo == NULL || ucode != NULL);
664 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
665 ktrfault(vaddr, fault_type);
667 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
669 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
670 result == KERN_INVALID_ADDRESS ||
671 result == KERN_RESOURCE_SHORTAGE ||
672 result == KERN_PROTECTION_FAILURE ||
673 result == KERN_OUT_OF_BOUNDS,
674 ("Unexpected Mach error %d from vm_fault()", result));
676 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
679 if (result != KERN_SUCCESS && signo != NULL) {
682 case KERN_INVALID_ADDRESS:
684 *ucode = SEGV_MAPERR;
686 case KERN_RESOURCE_SHORTAGE:
690 case KERN_OUT_OF_BOUNDS:
694 case KERN_PROTECTION_FAILURE:
695 if (prot_fault_translation == 0) {
697 * Autodetect. This check also covers
698 * the images without the ABI-tag ELF
701 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
702 curproc->p_osrel >= P_OSREL_SIGSEGV) {
704 *ucode = SEGV_ACCERR;
707 *ucode = UCODE_PAGEFLT;
709 } else if (prot_fault_translation == 1) {
710 /* Always compat mode. */
712 *ucode = UCODE_PAGEFLT;
714 /* Always SIGSEGV mode. */
716 *ucode = SEGV_ACCERR;
720 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
728 static enum fault_status
729 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
734 if (fs->object->type != OBJT_VNODE)
735 return (FAULT_CONTINUE);
736 vp = fs->object->handle;
738 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
739 return (FAULT_CONTINUE);
743 * Perform an unlock in case the desired vnode changed while
744 * the map was unlocked during a retry.
748 locked = VOP_ISLOCKED(vp);
749 if (locked != LK_EXCLUSIVE)
753 * We must not sleep acquiring the vnode lock while we have
754 * the page exclusive busied or the object's
755 * paging-in-progress count incremented. Otherwise, we could
758 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
761 return (FAULT_CONTINUE);
766 unlock_and_deallocate(fs);
768 fault_deallocate(fs);
769 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
772 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
773 return (FAULT_RESTART);
777 * Calculate the desired readahead. Handle drop-behind.
779 * Returns the number of readahead blocks to pass to the pager.
782 vm_fault_readahead(struct faultstate *fs)
787 KASSERT(fs->lookup_still_valid, ("map unlocked"));
788 era = fs->entry->read_ahead;
789 behavior = vm_map_entry_behavior(fs->entry);
790 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
792 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
793 nera = VM_FAULT_READ_AHEAD_MAX;
794 if (fs->vaddr == fs->entry->next_read)
795 vm_fault_dontneed(fs, fs->vaddr, nera);
796 } else if (fs->vaddr == fs->entry->next_read) {
798 * This is a sequential fault. Arithmetically
799 * increase the requested number of pages in
800 * the read-ahead window. The requested
801 * number of pages is "# of sequential faults
802 * x (read ahead min + 1) + read ahead min"
804 nera = VM_FAULT_READ_AHEAD_MIN;
807 if (nera > VM_FAULT_READ_AHEAD_MAX)
808 nera = VM_FAULT_READ_AHEAD_MAX;
810 if (era == VM_FAULT_READ_AHEAD_MAX)
811 vm_fault_dontneed(fs, fs->vaddr, nera);
814 * This is a non-sequential fault.
820 * A read lock on the map suffices to update
821 * the read ahead count safely.
823 fs->entry->read_ahead = nera;
830 vm_fault_lookup(struct faultstate *fs)
834 KASSERT(!fs->lookup_still_valid,
835 ("vm_fault_lookup: Map already locked."));
836 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
837 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
838 &fs->first_pindex, &fs->prot, &fs->wired);
839 if (result != KERN_SUCCESS) {
844 fs->map_generation = fs->map->timestamp;
846 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
847 panic("%s: fault on nofault entry, addr: %#lx",
848 __func__, (u_long)fs->vaddr);
851 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
852 fs->entry->wiring_thread != curthread) {
853 vm_map_unlock_read(fs->map);
854 vm_map_lock(fs->map);
855 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
856 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
858 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
859 vm_map_unlock_and_wait(fs->map, 0);
861 vm_map_unlock(fs->map);
862 return (KERN_RESOURCE_SHORTAGE);
865 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
868 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
870 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
871 ("!fs->wired && VM_FAULT_WIRE"));
872 fs->lookup_still_valid = true;
874 return (KERN_SUCCESS);
878 vm_fault_relookup(struct faultstate *fs)
880 vm_object_t retry_object;
881 vm_pindex_t retry_pindex;
882 vm_prot_t retry_prot;
885 if (!vm_map_trylock_read(fs->map))
886 return (KERN_RESTART);
888 fs->lookup_still_valid = true;
889 if (fs->map->timestamp == fs->map_generation)
890 return (KERN_SUCCESS);
892 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
893 &fs->entry, &retry_object, &retry_pindex, &retry_prot,
895 if (result != KERN_SUCCESS) {
897 * If retry of map lookup would have blocked then
898 * retry fault from start.
900 if (result == KERN_FAILURE)
901 return (KERN_RESTART);
904 if (retry_object != fs->first_object ||
905 retry_pindex != fs->first_pindex)
906 return (KERN_RESTART);
909 * Check whether the protection has changed or the object has
910 * been copied while we left the map unlocked. Changing from
911 * read to write permission is OK - we leave the page
912 * write-protected, and catch the write fault. Changing from
913 * write to read permission means that we can't mark the page
914 * write-enabled after all.
916 fs->prot &= retry_prot;
917 fs->fault_type &= retry_prot;
919 return (KERN_RESTART);
921 /* Reassert because wired may have changed. */
922 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
923 ("!wired && VM_FAULT_WIRE"));
925 return (KERN_SUCCESS);
929 vm_fault_cow(struct faultstate *fs)
931 bool is_first_object_locked;
933 KASSERT(fs->object != fs->first_object,
934 ("source and target COW objects are identical"));
937 * This allows pages to be virtually copied from a backing_object
938 * into the first_object, where the backing object has no other
939 * refs to it, and cannot gain any more refs. Instead of a bcopy,
940 * we just move the page from the backing object to the first
941 * object. Note that we must mark the page dirty in the first
942 * object so that it will go out to swap when needed.
944 is_first_object_locked = false;
947 * Only one shadow object and no other refs.
949 fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
951 * No other ways to look the object up
953 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
955 * We don't chase down the shadow chain and we can acquire locks.
957 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
958 fs->object == fs->first_object->backing_object &&
959 VM_OBJECT_TRYWLOCK(fs->object)) {
961 * Remove but keep xbusy for replace. fs->m is moved into
962 * fs->first_object and left busy while fs->first_m is
963 * conditionally freed.
965 vm_page_remove_xbusy(fs->m);
966 vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
968 vm_page_dirty(fs->m);
969 #if VM_NRESERVLEVEL > 0
971 * Rename the reservation.
973 vm_reserv_rename(fs->m, fs->first_object, fs->object,
974 OFF_TO_IDX(fs->first_object->backing_object_offset));
976 VM_OBJECT_WUNLOCK(fs->object);
977 VM_OBJECT_WUNLOCK(fs->first_object);
980 VM_CNT_INC(v_cow_optim);
982 if (is_first_object_locked)
983 VM_OBJECT_WUNLOCK(fs->first_object);
985 * Oh, well, lets copy it.
987 pmap_copy_page(fs->m, fs->first_m);
988 vm_page_valid(fs->first_m);
989 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
990 vm_page_wire(fs->first_m);
991 vm_page_unwire(fs->m, PQ_INACTIVE);
994 * Save the cow page to be released after
995 * pmap_enter is complete.
1001 * Typically, the shadow object is either private to this
1002 * address space (OBJ_ONEMAPPING) or its pages are read only.
1003 * In the highly unusual case where the pages of a shadow object
1004 * are read/write shared between this and other address spaces,
1005 * we need to ensure that any pmap-level mappings to the
1006 * original, copy-on-write page from the backing object are
1007 * removed from those other address spaces.
1009 * The flag check is racy, but this is tolerable: if
1010 * OBJ_ONEMAPPING is cleared after the check, the busy state
1011 * ensures that new mappings of m_cow can't be created.
1012 * pmap_enter() will replace an existing mapping in the current
1013 * address space. If OBJ_ONEMAPPING is set after the check,
1014 * removing mappings will at worse trigger some unnecessary page
1017 vm_page_assert_xbusied(fs->m_cow);
1018 if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
1019 pmap_remove_all(fs->m_cow);
1022 vm_object_pip_wakeup(fs->object);
1025 * Only use the new page below...
1027 fs->object = fs->first_object;
1028 fs->pindex = fs->first_pindex;
1029 fs->m = fs->first_m;
1030 VM_CNT_INC(v_cow_faults);
1031 curthread->td_cow++;
1035 vm_fault_next(struct faultstate *fs)
1037 vm_object_t next_object;
1040 * The requested page does not exist at this object/
1041 * offset. Remove the invalid page from the object,
1042 * waking up anyone waiting for it, and continue on to
1043 * the next object. However, if this is the top-level
1044 * object, we must leave the busy page in place to
1045 * prevent another process from rushing past us, and
1046 * inserting the page in that object at the same time
1049 if (fs->object == fs->first_object) {
1050 fs->first_m = fs->m;
1053 fault_page_free(&fs->m);
1056 * Move on to the next object. Lock the next object before
1057 * unlocking the current one.
1059 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1060 next_object = fs->object->backing_object;
1061 if (next_object == NULL)
1063 MPASS(fs->first_m != NULL);
1064 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1065 VM_OBJECT_WLOCK(next_object);
1066 vm_object_pip_add(next_object, 1);
1067 if (fs->object != fs->first_object)
1068 vm_object_pip_wakeup(fs->object);
1069 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1070 VM_OBJECT_WUNLOCK(fs->object);
1071 fs->object = next_object;
1077 vm_fault_zerofill(struct faultstate *fs)
1081 * If there's no object left, fill the page in the top
1082 * object with zeros.
1084 if (fs->object != fs->first_object) {
1085 vm_object_pip_wakeup(fs->object);
1086 fs->object = fs->first_object;
1087 fs->pindex = fs->first_pindex;
1089 MPASS(fs->first_m != NULL);
1090 MPASS(fs->m == NULL);
1091 fs->m = fs->first_m;
1095 * Zero the page if necessary and mark it valid.
1097 if ((fs->m->flags & PG_ZERO) == 0) {
1098 pmap_zero_page(fs->m);
1100 VM_CNT_INC(v_ozfod);
1103 vm_page_valid(fs->m);
1107 * Initiate page fault after timeout. Returns true if caller should
1108 * do vm_waitpfault() after the call.
1111 vm_fault_allocate_oom(struct faultstate *fs)
1115 unlock_and_deallocate(fs);
1116 if (vm_pfault_oom_attempts < 0)
1118 if (!fs->oom_started) {
1119 fs->oom_started = true;
1120 getmicrotime(&fs->oom_start_time);
1125 timevalsub(&now, &fs->oom_start_time);
1126 if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait)
1131 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1132 curproc->p_pid, curproc->p_comm);
1133 vm_pageout_oom(VM_OOM_MEM_PF);
1134 fs->oom_started = false;
1139 * Allocate a page directly or via the object populate method.
1141 static enum fault_status
1142 vm_fault_allocate(struct faultstate *fs)
1144 struct domainset *dset;
1145 enum fault_status res;
1147 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1148 res = vm_fault_lock_vnode(fs, true);
1149 MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART);
1150 if (res == FAULT_RESTART)
1154 if (fs->pindex >= fs->object->size) {
1155 unlock_and_deallocate(fs);
1156 return (FAULT_OUT_OF_BOUNDS);
1159 if (fs->object == fs->first_object &&
1160 (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1161 fs->first_object->shadow_count == 0) {
1162 res = vm_fault_populate(fs);
1167 unlock_and_deallocate(fs);
1169 case FAULT_CONTINUE:
1171 * Pager's populate() method
1172 * returned VM_PAGER_BAD.
1176 panic("inconsistent return codes");
1181 * Allocate a new page for this object/offset pair.
1183 * If the process has a fatal signal pending, prioritize the allocation
1184 * with the expectation that the process will exit shortly and free some
1185 * pages. In particular, the signal may have been posted by the page
1186 * daemon in an attempt to resolve an out-of-memory condition.
1188 * The unlocked read of the p_flag is harmless. At worst, the P_KILLED
1189 * might be not observed here, and allocation fails, causing a restart
1190 * and new reading of the p_flag.
1192 dset = fs->object->domain.dr_policy;
1194 dset = curthread->td_domain.dr_policy;
1195 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1196 #if VM_NRESERVLEVEL > 0
1197 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1199 fs->m = vm_page_alloc(fs->object, fs->pindex,
1200 P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0);
1202 if (fs->m == NULL) {
1203 if (vm_fault_allocate_oom(fs))
1204 vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1205 return (FAULT_RESTART);
1207 fs->oom_started = false;
1209 return (FAULT_CONTINUE);
1213 * Call the pager to retrieve the page if there is a chance
1214 * that the pager has it, and potentially retrieve additional
1215 * pages at the same time.
1217 static enum fault_status
1218 vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp)
1220 vm_offset_t e_end, e_start;
1221 int ahead, behind, cluster_offset, rv;
1222 enum fault_status status;
1226 * Prepare for unlocking the map. Save the map
1227 * entry's start and end addresses, which are used to
1228 * optimize the size of the pager operation below.
1229 * Even if the map entry's addresses change after
1230 * unlocking the map, using the saved addresses is
1233 e_start = fs->entry->start;
1234 e_end = fs->entry->end;
1235 behavior = vm_map_entry_behavior(fs->entry);
1238 * If the pager for the current object might have
1239 * the page, then determine the number of additional
1240 * pages to read and potentially reprioritize
1241 * previously read pages for earlier reclamation.
1242 * These operations should only be performed once per
1243 * page fault. Even if the current pager doesn't
1244 * have the page, the number of additional pages to
1245 * read will apply to subsequent objects in the
1248 if (fs->nera == -1 && !P_KILLED(curproc))
1249 fs->nera = vm_fault_readahead(fs);
1252 * Release the map lock before locking the vnode or
1253 * sleeping in the pager. (If the current object has
1254 * a shadow, then an earlier iteration of this loop
1255 * may have already unlocked the map.)
1259 status = vm_fault_lock_vnode(fs, false);
1260 MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART);
1261 if (status == FAULT_RESTART)
1263 KASSERT(fs->vp == NULL || !fs->map->system_map,
1264 ("vm_fault: vnode-backed object mapped by system map"));
1267 * Page in the requested page and hint the pager,
1268 * that it may bring up surrounding pages.
1270 if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1271 P_KILLED(curproc)) {
1275 /* Is this a sequential fault? */
1281 * Request a cluster of pages that is
1282 * aligned to a VM_FAULT_READ_DEFAULT
1283 * page offset boundary within the
1284 * object. Alignment to a page offset
1285 * boundary is more likely to coincide
1286 * with the underlying file system
1287 * block than alignment to a virtual
1290 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1291 behind = ulmin(cluster_offset,
1292 atop(fs->vaddr - e_start));
1293 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1295 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1299 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1300 if (rv == VM_PAGER_OK)
1301 return (FAULT_HARD);
1302 if (rv == VM_PAGER_ERROR)
1303 printf("vm_fault: pager read error, pid %d (%s)\n",
1304 curproc->p_pid, curproc->p_comm);
1306 * If an I/O error occurred or the requested page was
1307 * outside the range of the pager, clean up and return
1310 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1311 VM_OBJECT_WLOCK(fs->object);
1312 fault_page_free(&fs->m);
1313 unlock_and_deallocate(fs);
1314 return (FAULT_OUT_OF_BOUNDS);
1316 KASSERT(rv == VM_PAGER_FAIL,
1317 ("%s: unexpected pager error %d", __func__, rv));
1318 return (FAULT_CONTINUE);
1322 * Wait/Retry if the page is busy. We have to do this if the page is
1323 * either exclusive or shared busy because the vm_pager may be using
1324 * read busy for pageouts (and even pageins if it is the vnode pager),
1325 * and we could end up trying to pagein and pageout the same page
1328 * We can theoretically allow the busy case on a read fault if the page
1329 * is marked valid, but since such pages are typically already pmap'd,
1330 * putting that special case in might be more effort then it is worth.
1331 * We cannot under any circumstances mess around with a shared busied
1332 * page except, perhaps, to pmap it.
1335 vm_fault_busy_sleep(struct faultstate *fs)
1338 * Reference the page before unlocking and
1339 * sleeping so that the page daemon is less
1340 * likely to reclaim it.
1342 vm_page_aflag_set(fs->m, PGA_REFERENCED);
1343 if (fs->object != fs->first_object) {
1344 fault_page_release(&fs->first_m);
1345 vm_object_pip_wakeup(fs->first_object);
1347 vm_object_pip_wakeup(fs->object);
1349 if (fs->m != vm_page_lookup(fs->object, fs->pindex) ||
1350 !vm_page_busy_sleep(fs->m, "vmpfw", 0))
1351 VM_OBJECT_WUNLOCK(fs->object);
1352 VM_CNT_INC(v_intrans);
1353 vm_object_deallocate(fs->first_object);
1357 * Handle page lookup, populate, allocate, page-in for the current
1360 * The object is locked on entry and will remain locked with a return
1361 * code of FAULT_CONTINUE so that fault may follow the shadow chain.
1362 * Otherwise, the object will be unlocked upon return.
1364 static enum fault_status
1365 vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp)
1367 enum fault_status res;
1371 * If the object is marked for imminent termination, we retry
1372 * here, since the collapse pass has raced with us. Otherwise,
1373 * if we see terminally dead object, return fail.
1375 if ((fs->object->flags & OBJ_DEAD) != 0) {
1376 dead = fs->object->type == OBJT_DEAD;
1377 unlock_and_deallocate(fs);
1379 return (FAULT_PROTECTION_FAILURE);
1381 return (FAULT_RESTART);
1385 * See if the page is resident.
1387 fs->m = vm_page_lookup(fs->object, fs->pindex);
1388 if (fs->m != NULL) {
1389 if (!vm_page_tryxbusy(fs->m)) {
1390 vm_fault_busy_sleep(fs);
1391 return (FAULT_RESTART);
1395 * The page is marked busy for other processes and the
1396 * pagedaemon. If it is still completely valid we are
1399 if (vm_page_all_valid(fs->m)) {
1400 VM_OBJECT_WUNLOCK(fs->object);
1401 return (FAULT_SOFT);
1404 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1407 * Page is not resident. If the pager might contain the page
1408 * or this is the beginning of the search, allocate a new
1409 * page. (Default objects are zero-fill, so there is no real
1412 if (fs->m == NULL && (fs->object->type != OBJT_DEFAULT ||
1413 fs->object == fs->first_object)) {
1414 res = vm_fault_allocate(fs);
1415 if (res != FAULT_CONTINUE)
1420 * Default objects have no pager so no exclusive busy exists
1421 * to protect this page in the chain. Skip to the next
1422 * object without dropping the lock to preserve atomicity of
1425 if (fs->object->type != OBJT_DEFAULT) {
1427 * At this point, we have either allocated a new page
1428 * or found an existing page that is only partially
1431 * We hold a reference on the current object and the
1432 * page is exclusive busied. The exclusive busy
1433 * prevents simultaneous faults and collapses while
1434 * the object lock is dropped.
1436 VM_OBJECT_WUNLOCK(fs->object);
1437 res = vm_fault_getpages(fs, behindp, aheadp);
1438 if (res == FAULT_CONTINUE)
1439 VM_OBJECT_WLOCK(fs->object);
1441 res = FAULT_CONTINUE;
1447 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1448 int fault_flags, vm_page_t *m_hold)
1450 struct faultstate fs;
1451 int ahead, behind, faultcount, rv;
1452 enum fault_status res;
1455 VM_CNT_INC(v_vm_faults);
1457 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1458 return (KERN_PROTECTION_FAILURE);
1463 fs.fault_flags = fault_flags;
1465 fs.lookup_still_valid = false;
1466 fs.oom_started = false;
1472 fs.fault_type = fault_type;
1475 * Find the backing store object and offset into it to begin the
1478 rv = vm_fault_lookup(&fs);
1479 if (rv != KERN_SUCCESS) {
1480 if (rv == KERN_RESOURCE_SHORTAGE)
1486 * Try to avoid lock contention on the top-level object through
1487 * special-case handling of some types of page faults, specifically,
1488 * those that are mapping an existing page from the top-level object.
1489 * Under this condition, a read lock on the object suffices, allowing
1490 * multiple page faults of a similar type to run in parallel.
1492 if (fs.vp == NULL /* avoid locked vnode leak */ &&
1493 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1494 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1495 VM_OBJECT_RLOCK(fs.first_object);
1496 res = vm_fault_soft_fast(&fs);
1497 if (res == FAULT_SUCCESS)
1498 return (KERN_SUCCESS);
1499 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
1500 VM_OBJECT_RUNLOCK(fs.first_object);
1501 VM_OBJECT_WLOCK(fs.first_object);
1504 VM_OBJECT_WLOCK(fs.first_object);
1508 * Make a reference to this object to prevent its disposal while we
1509 * are messing with it. Once we have the reference, the map is free
1510 * to be diddled. Since objects reference their shadows (and copies),
1511 * they will stay around as well.
1513 * Bump the paging-in-progress count to prevent size changes (e.g.
1514 * truncation operations) during I/O.
1516 vm_object_reference_locked(fs.first_object);
1517 vm_object_pip_add(fs.first_object, 1);
1519 fs.m_cow = fs.m = fs.first_m = NULL;
1522 * Search for the page at object/offset.
1524 fs.object = fs.first_object;
1525 fs.pindex = fs.first_pindex;
1527 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1528 res = vm_fault_allocate(&fs);
1533 return (KERN_SUCCESS);
1535 return (KERN_FAILURE);
1536 case FAULT_OUT_OF_BOUNDS:
1537 return (KERN_OUT_OF_BOUNDS);
1538 case FAULT_CONTINUE:
1541 panic("vm_fault: Unhandled status %d", res);
1546 KASSERT(fs.m == NULL,
1547 ("page still set %p at loop start", fs.m));
1549 res = vm_fault_object(&fs, &behind, &ahead);
1554 faultcount = behind + 1 + ahead;
1560 return (KERN_SUCCESS);
1562 return (KERN_FAILURE);
1563 case FAULT_OUT_OF_BOUNDS:
1564 return (KERN_OUT_OF_BOUNDS);
1565 case FAULT_PROTECTION_FAILURE:
1566 return (KERN_PROTECTION_FAILURE);
1567 case FAULT_CONTINUE:
1570 panic("vm_fault: Unhandled status %d", res);
1574 * The page was not found in the current object. Try to
1575 * traverse into a backing object or zero fill if none is
1578 if (vm_fault_next(&fs))
1580 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1581 if (fs.first_object == fs.object)
1582 fault_page_free(&fs.first_m);
1583 unlock_and_deallocate(&fs);
1584 return (KERN_OUT_OF_BOUNDS);
1586 VM_OBJECT_WUNLOCK(fs.object);
1587 vm_fault_zerofill(&fs);
1588 /* Don't try to prefault neighboring pages. */
1595 * A valid page has been found and exclusively busied. The
1596 * object lock must no longer be held.
1598 vm_page_assert_xbusied(fs.m);
1599 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1602 * If the page is being written, but isn't already owned by the
1603 * top-level object, we have to copy it into a new page owned by the
1606 if (fs.object != fs.first_object) {
1608 * We only really need to copy if we want to write it.
1610 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1613 * We only try to prefault read-only mappings to the
1614 * neighboring pages when this copy-on-write fault is
1615 * a hard fault. In other cases, trying to prefault
1616 * is typically wasted effort.
1618 if (faultcount == 0)
1622 fs.prot &= ~VM_PROT_WRITE;
1627 * We must verify that the maps have not changed since our last
1630 if (!fs.lookup_still_valid) {
1631 rv = vm_fault_relookup(&fs);
1632 if (rv != KERN_SUCCESS) {
1633 fault_deallocate(&fs);
1634 if (rv == KERN_RESTART)
1639 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1642 * If the page was filled by a pager, save the virtual address that
1643 * should be faulted on next under a sequential access pattern to the
1644 * map entry. A read lock on the map suffices to update this address
1648 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1651 * Page must be completely valid or it is not fit to
1652 * map into user space. vm_pager_get_pages() ensures this.
1654 vm_page_assert_xbusied(fs.m);
1655 KASSERT(vm_page_all_valid(fs.m),
1656 ("vm_fault: page %p partially invalid", fs.m));
1658 vm_fault_dirty(&fs, fs.m);
1661 * Put this page into the physical map. We had to do the unlock above
1662 * because pmap_enter() may sleep. We don't put the page
1663 * back on the active queue until later so that the pageout daemon
1664 * won't find it (yet).
1666 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1667 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1668 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1670 vm_fault_prefault(&fs, vaddr,
1671 faultcount > 0 ? behind : PFBAK,
1672 faultcount > 0 ? ahead : PFFOR, false);
1675 * If the page is not wired down, then put it where the pageout daemon
1678 if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1681 vm_page_activate(fs.m);
1682 if (fs.m_hold != NULL) {
1683 (*fs.m_hold) = fs.m;
1686 vm_page_xunbusy(fs.m);
1690 * Unlock everything, and return
1692 fault_deallocate(&fs);
1694 VM_CNT_INC(v_io_faults);
1695 curthread->td_ru.ru_majflt++;
1697 if (racct_enable && fs.object->type == OBJT_VNODE) {
1699 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1700 racct_add_force(curproc, RACCT_WRITEBPS,
1701 PAGE_SIZE + behind * PAGE_SIZE);
1702 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1704 racct_add_force(curproc, RACCT_READBPS,
1705 PAGE_SIZE + ahead * PAGE_SIZE);
1706 racct_add_force(curproc, RACCT_READIOPS, 1);
1708 PROC_UNLOCK(curproc);
1712 curthread->td_ru.ru_minflt++;
1714 return (KERN_SUCCESS);
1718 * Speed up the reclamation of pages that precede the faulting pindex within
1719 * the first object of the shadow chain. Essentially, perform the equivalent
1720 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1721 * the faulting pindex by the cluster size when the pages read by vm_fault()
1722 * cross a cluster-size boundary. The cluster size is the greater of the
1723 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1725 * When "fs->first_object" is a shadow object, the pages in the backing object
1726 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1727 * function must only be concerned with pages in the first object.
1730 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1732 vm_map_entry_t entry;
1733 vm_object_t first_object, object;
1734 vm_offset_t end, start;
1735 vm_page_t m, m_next;
1736 vm_pindex_t pend, pstart;
1739 object = fs->object;
1740 VM_OBJECT_ASSERT_UNLOCKED(object);
1741 first_object = fs->first_object;
1742 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1743 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1744 VM_OBJECT_RLOCK(first_object);
1745 size = VM_FAULT_DONTNEED_MIN;
1746 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1747 size = pagesizes[1];
1748 end = rounddown2(vaddr, size);
1749 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1750 (entry = fs->entry)->start < end) {
1751 if (end - entry->start < size)
1752 start = entry->start;
1755 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1756 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1758 m_next = vm_page_find_least(first_object, pstart);
1759 pend = OFF_TO_IDX(entry->offset) + atop(end -
1761 while ((m = m_next) != NULL && m->pindex < pend) {
1762 m_next = TAILQ_NEXT(m, listq);
1763 if (!vm_page_all_valid(m) ||
1768 * Don't clear PGA_REFERENCED, since it would
1769 * likely represent a reference by a different
1772 * Typically, at this point, prefetched pages
1773 * are still in the inactive queue. Only
1774 * pages that triggered page faults are in the
1775 * active queue. The test for whether the page
1776 * is in the inactive queue is racy; in the
1777 * worst case we will requeue the page
1780 if (!vm_page_inactive(m))
1781 vm_page_deactivate(m);
1784 VM_OBJECT_RUNLOCK(first_object);
1789 * vm_fault_prefault provides a quick way of clustering
1790 * pagefaults into a processes address space. It is a "cousin"
1791 * of vm_map_pmap_enter, except it runs at page fault time instead
1795 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1796 int backward, int forward, bool obj_locked)
1799 vm_map_entry_t entry;
1800 vm_object_t backing_object, lobject;
1801 vm_offset_t addr, starta;
1806 pmap = fs->map->pmap;
1807 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1812 if (addra < backward * PAGE_SIZE) {
1813 starta = entry->start;
1815 starta = addra - backward * PAGE_SIZE;
1816 if (starta < entry->start)
1817 starta = entry->start;
1821 * Generate the sequence of virtual addresses that are candidates for
1822 * prefaulting in an outward spiral from the faulting virtual address,
1823 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1824 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1825 * If the candidate address doesn't have a backing physical page, then
1826 * the loop immediately terminates.
1828 for (i = 0; i < 2 * imax(backward, forward); i++) {
1829 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1831 if (addr > addra + forward * PAGE_SIZE)
1834 if (addr < starta || addr >= entry->end)
1837 if (!pmap_is_prefaultable(pmap, addr))
1840 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1841 lobject = entry->object.vm_object;
1843 VM_OBJECT_RLOCK(lobject);
1844 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1845 lobject->type == OBJT_DEFAULT &&
1846 (backing_object = lobject->backing_object) != NULL) {
1847 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1848 0, ("vm_fault_prefault: unaligned object offset"));
1849 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1850 VM_OBJECT_RLOCK(backing_object);
1851 if (!obj_locked || lobject != entry->object.vm_object)
1852 VM_OBJECT_RUNLOCK(lobject);
1853 lobject = backing_object;
1856 if (!obj_locked || lobject != entry->object.vm_object)
1857 VM_OBJECT_RUNLOCK(lobject);
1860 if (vm_page_all_valid(m) &&
1861 (m->flags & PG_FICTITIOUS) == 0)
1862 pmap_enter_quick(pmap, addr, m, entry->protection);
1863 if (!obj_locked || lobject != entry->object.vm_object)
1864 VM_OBJECT_RUNLOCK(lobject);
1869 * Hold each of the physical pages that are mapped by the specified range of
1870 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1871 * and allow the specified types of access, "prot". If all of the implied
1872 * pages are successfully held, then the number of held pages is returned
1873 * together with pointers to those pages in the array "ma". However, if any
1874 * of the pages cannot be held, -1 is returned.
1877 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1878 vm_prot_t prot, vm_page_t *ma, int max_count)
1880 vm_offset_t end, va;
1883 boolean_t pmap_failed;
1887 end = round_page(addr + len);
1888 addr = trunc_page(addr);
1890 if (!vm_map_range_valid(map, addr, end))
1893 if (atop(end - addr) > max_count)
1894 panic("vm_fault_quick_hold_pages: count > max_count");
1895 count = atop(end - addr);
1898 * Most likely, the physical pages are resident in the pmap, so it is
1899 * faster to try pmap_extract_and_hold() first.
1901 pmap_failed = FALSE;
1902 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1903 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1906 else if ((prot & VM_PROT_WRITE) != 0 &&
1907 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1909 * Explicitly dirty the physical page. Otherwise, the
1910 * caller's changes may go unnoticed because they are
1911 * performed through an unmanaged mapping or by a DMA
1914 * The object lock is not held here.
1915 * See vm_page_clear_dirty_mask().
1922 * One or more pages could not be held by the pmap. Either no
1923 * page was mapped at the specified virtual address or that
1924 * mapping had insufficient permissions. Attempt to fault in
1925 * and hold these pages.
1927 * If vm_fault_disable_pagefaults() was called,
1928 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1929 * acquire MD VM locks, which means we must not call
1930 * vm_fault(). Some (out of tree) callers mark
1931 * too wide a code area with vm_fault_disable_pagefaults()
1932 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1933 * the proper behaviour explicitly.
1935 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1936 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1938 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1939 if (*mp == NULL && vm_fault(map, va, prot,
1940 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1945 for (mp = ma; mp < ma + count; mp++)
1947 vm_page_unwire(*mp, PQ_INACTIVE);
1953 * vm_fault_copy_entry
1955 * Create new shadow object backing dst_entry with private copy of
1956 * all underlying pages. When src_entry is equal to dst_entry,
1957 * function implements COW for wired-down map entry. Otherwise,
1958 * it forks wired entry into dst_map.
1960 * In/out conditions:
1961 * The source and destination maps must be locked for write.
1962 * The source map entry must be wired down (or be a sharing map
1963 * entry corresponding to a main map entry that is wired down).
1966 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1967 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1968 vm_ooffset_t *fork_charge)
1970 vm_object_t backing_object, dst_object, object, src_object;
1971 vm_pindex_t dst_pindex, pindex, src_pindex;
1972 vm_prot_t access, prot;
1982 upgrade = src_entry == dst_entry;
1983 access = prot = dst_entry->protection;
1985 src_object = src_entry->object.vm_object;
1986 src_pindex = OFF_TO_IDX(src_entry->offset);
1988 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1989 dst_object = src_object;
1990 vm_object_reference(dst_object);
1993 * Create the top-level object for the destination entry.
1994 * Doesn't actually shadow anything - we copy the pages
1997 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1998 dst_entry->start), NULL, NULL, 0);
1999 #if VM_NRESERVLEVEL > 0
2000 dst_object->flags |= OBJ_COLORED;
2001 dst_object->pg_color = atop(dst_entry->start);
2003 dst_object->domain = src_object->domain;
2004 dst_object->charge = dst_entry->end - dst_entry->start;
2007 VM_OBJECT_WLOCK(dst_object);
2008 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
2009 ("vm_fault_copy_entry: vm_object not NULL"));
2010 if (src_object != dst_object) {
2011 dst_entry->object.vm_object = dst_object;
2012 dst_entry->offset = 0;
2013 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
2015 if (fork_charge != NULL) {
2016 KASSERT(dst_entry->cred == NULL,
2017 ("vm_fault_copy_entry: leaked swp charge"));
2018 dst_object->cred = curthread->td_ucred;
2019 crhold(dst_object->cred);
2020 *fork_charge += dst_object->charge;
2021 } else if ((dst_object->type == OBJT_DEFAULT ||
2022 (dst_object->flags & OBJ_SWAP) != 0) &&
2023 dst_object->cred == NULL) {
2024 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
2026 dst_object->cred = dst_entry->cred;
2027 dst_entry->cred = NULL;
2031 * If not an upgrade, then enter the mappings in the pmap as
2032 * read and/or execute accesses. Otherwise, enter them as
2035 * A writeable large page mapping is only created if all of
2036 * the constituent small page mappings are modified. Marking
2037 * PTEs as modified on inception allows promotion to happen
2038 * without taking potentially large number of soft faults.
2041 access &= ~VM_PROT_WRITE;
2044 * Loop through all of the virtual pages within the entry's
2045 * range, copying each page from the source object to the
2046 * destination object. Since the source is wired, those pages
2047 * must exist. In contrast, the destination is pageable.
2048 * Since the destination object doesn't share any backing storage
2049 * with the source object, all of its pages must be dirtied,
2050 * regardless of whether they can be written.
2052 for (vaddr = dst_entry->start, dst_pindex = 0;
2053 vaddr < dst_entry->end;
2054 vaddr += PAGE_SIZE, dst_pindex++) {
2057 * Find the page in the source object, and copy it in.
2058 * Because the source is wired down, the page will be
2061 if (src_object != dst_object)
2062 VM_OBJECT_RLOCK(src_object);
2063 object = src_object;
2064 pindex = src_pindex + dst_pindex;
2065 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
2066 (backing_object = object->backing_object) != NULL) {
2068 * Unless the source mapping is read-only or
2069 * it is presently being upgraded from
2070 * read-only, the first object in the shadow
2071 * chain should provide all of the pages. In
2072 * other words, this loop body should never be
2073 * executed when the source mapping is already
2076 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
2078 ("vm_fault_copy_entry: main object missing page"));
2080 VM_OBJECT_RLOCK(backing_object);
2081 pindex += OFF_TO_IDX(object->backing_object_offset);
2082 if (object != dst_object)
2083 VM_OBJECT_RUNLOCK(object);
2084 object = backing_object;
2086 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2088 if (object != dst_object) {
2090 * Allocate a page in the destination object.
2092 dst_m = vm_page_alloc(dst_object, (src_object ==
2093 dst_object ? src_pindex : 0) + dst_pindex,
2095 if (dst_m == NULL) {
2096 VM_OBJECT_WUNLOCK(dst_object);
2097 VM_OBJECT_RUNLOCK(object);
2098 vm_wait(dst_object);
2099 VM_OBJECT_WLOCK(dst_object);
2102 pmap_copy_page(src_m, dst_m);
2103 VM_OBJECT_RUNLOCK(object);
2104 dst_m->dirty = dst_m->valid = src_m->valid;
2107 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2109 if (dst_m->pindex >= dst_object->size) {
2111 * We are upgrading. Index can occur
2112 * out of bounds if the object type is
2113 * vnode and the file was truncated.
2115 vm_page_xunbusy(dst_m);
2119 VM_OBJECT_WUNLOCK(dst_object);
2122 * Enter it in the pmap. If a wired, copy-on-write
2123 * mapping is being replaced by a write-enabled
2124 * mapping, then wire that new mapping.
2126 * The page can be invalid if the user called
2127 * msync(MS_INVALIDATE) or truncated the backing vnode
2128 * or shared memory object. In this case, do not
2129 * insert it into pmap, but still do the copy so that
2130 * all copies of the wired map entry have similar
2133 if (vm_page_all_valid(dst_m)) {
2134 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2135 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2139 * Mark it no longer busy, and put it on the active list.
2141 VM_OBJECT_WLOCK(dst_object);
2144 if (src_m != dst_m) {
2145 vm_page_unwire(src_m, PQ_INACTIVE);
2146 vm_page_wire(dst_m);
2148 KASSERT(vm_page_wired(dst_m),
2149 ("dst_m %p is not wired", dst_m));
2152 vm_page_activate(dst_m);
2154 vm_page_xunbusy(dst_m);
2156 VM_OBJECT_WUNLOCK(dst_object);
2158 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2159 vm_object_deallocate(src_object);
2164 * Block entry into the machine-independent layer's page fault handler by
2165 * the calling thread. Subsequent calls to vm_fault() by that thread will
2166 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
2167 * spurious page faults.
2170 vm_fault_disable_pagefaults(void)
2173 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2177 vm_fault_enable_pagefaults(int save)
2180 curthread_pflags_restore(save);