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)
118 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX)
120 #define VM_FAULT_DONTNEED_MIN 1048576
129 vm_object_t first_object;
130 vm_pindex_t first_pindex;
132 vm_map_entry_t entry;
134 bool lookup_still_valid;
138 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
140 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
141 int backward, int forward, bool obj_locked);
143 static int vm_pfault_oom_attempts = 3;
144 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
145 &vm_pfault_oom_attempts, 0,
146 "Number of page allocation attempts in page fault handler before it "
147 "triggers OOM handling");
149 static int vm_pfault_oom_wait = 10;
150 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
151 &vm_pfault_oom_wait, 0,
152 "Number of seconds to wait for free pages before retrying "
153 "the page fault handler");
156 fault_page_release(vm_page_t *mp)
163 * We are likely to loop around again and attempt to busy
164 * this page. Deactivating it leaves it available for
165 * pageout while optimizing fault restarts.
167 vm_page_deactivate(m);
174 fault_page_free(vm_page_t *mp)
180 VM_OBJECT_ASSERT_WLOCKED(m->object);
181 if (!vm_page_wired(m))
190 unlock_map(struct faultstate *fs)
193 if (fs->lookup_still_valid) {
194 vm_map_lookup_done(fs->map, fs->entry);
195 fs->lookup_still_valid = false;
200 unlock_vp(struct faultstate *fs)
203 if (fs->vp != NULL) {
210 fault_deallocate(struct faultstate *fs)
213 fault_page_release(&fs->m_cow);
214 fault_page_release(&fs->m);
215 vm_object_pip_wakeup(fs->object);
216 if (fs->object != fs->first_object) {
217 VM_OBJECT_WLOCK(fs->first_object);
218 fault_page_free(&fs->first_m);
219 VM_OBJECT_WUNLOCK(fs->first_object);
220 vm_object_pip_wakeup(fs->first_object);
222 vm_object_deallocate(fs->first_object);
228 unlock_and_deallocate(struct faultstate *fs)
231 VM_OBJECT_WUNLOCK(fs->object);
232 fault_deallocate(fs);
236 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
237 vm_prot_t fault_type, int fault_flags)
241 if (((prot & VM_PROT_WRITE) == 0 &&
242 (fault_flags & VM_FAULT_DIRTY) == 0) ||
243 (m->oflags & VPO_UNMANAGED) != 0)
246 VM_PAGE_OBJECT_BUSY_ASSERT(m);
248 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
249 (fault_flags & VM_FAULT_WIRE) == 0) ||
250 (fault_flags & VM_FAULT_DIRTY) != 0;
252 vm_object_set_writeable_dirty(m->object);
255 * If the fault is a write, we know that this page is being
256 * written NOW so dirty it explicitly to save on
257 * pmap_is_modified() calls later.
259 * Also, since the page is now dirty, we can possibly tell
260 * the pager to release any swap backing the page.
262 if (need_dirty && vm_page_set_dirty(m) == 0) {
264 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
265 * if the page is already dirty to prevent data written with
266 * the expectation of being synced from not being synced.
267 * Likewise if this entry does not request NOSYNC then make
268 * sure the page isn't marked NOSYNC. Applications sharing
269 * data should use the same flags to avoid ping ponging.
271 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0)
272 vm_page_aflag_set(m, PGA_NOSYNC);
274 vm_page_aflag_clear(m, PGA_NOSYNC);
280 * Unlocks fs.first_object and fs.map on success.
283 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
284 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
287 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
288 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
295 MPASS(fs->vp == NULL);
296 vm_object_busy(fs->first_object);
297 m = vm_page_lookup(fs->first_object, fs->first_pindex);
298 /* A busy page can be mapped for read|execute access. */
299 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
300 vm_page_busied(m)) || !vm_page_all_valid(m)) {
306 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
307 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
309 if ((m->flags & PG_FICTITIOUS) == 0 &&
310 (m_super = vm_reserv_to_superpage(m)) != NULL &&
311 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
312 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
313 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
314 (pagesizes[m_super->psind] - 1)) && !wired &&
315 pmap_ps_enabled(fs->map->pmap)) {
316 flags = PS_ALL_VALID;
317 if ((prot & VM_PROT_WRITE) != 0) {
319 * Create a superpage mapping allowing write access
320 * only if none of the constituent pages are busy and
321 * all of them are already dirty (except possibly for
322 * the page that was faulted on).
324 flags |= PS_NONE_BUSY;
325 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
326 flags |= PS_ALL_DIRTY;
328 if (vm_page_ps_test(m_super, flags, m)) {
330 psind = m_super->psind;
331 vaddr = rounddown2(vaddr, pagesizes[psind]);
332 /* Preset the modified bit for dirty superpages. */
333 if ((flags & PS_ALL_DIRTY) != 0)
334 fault_type |= VM_PROT_WRITE;
338 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
339 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
340 if (rv != KERN_SUCCESS)
342 if (m_hold != NULL) {
346 if (psind == 0 && !wired)
347 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
348 VM_OBJECT_RUNLOCK(fs->first_object);
349 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags);
350 vm_map_lookup_done(fs->map, fs->entry);
351 curthread->td_ru.ru_minflt++;
354 vm_object_unbusy(fs->first_object);
359 vm_fault_restore_map_lock(struct faultstate *fs)
362 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
363 MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
365 if (!vm_map_trylock_read(fs->map)) {
366 VM_OBJECT_WUNLOCK(fs->first_object);
367 vm_map_lock_read(fs->map);
368 VM_OBJECT_WLOCK(fs->first_object);
370 fs->lookup_still_valid = true;
374 vm_fault_populate_check_page(vm_page_t m)
378 * Check each page to ensure that the pager is obeying the
379 * interface: the page must be installed in the object, fully
380 * valid, and exclusively busied.
383 MPASS(vm_page_all_valid(m));
384 MPASS(vm_page_xbusied(m));
388 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
394 VM_OBJECT_ASSERT_WLOCKED(object);
395 MPASS(first <= last);
396 for (pidx = first, m = vm_page_lookup(object, pidx);
397 pidx <= last; pidx++, m = vm_page_next(m)) {
398 vm_fault_populate_check_page(m);
399 vm_page_deactivate(m);
405 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
406 int fault_flags, boolean_t wired, vm_page_t *m_hold)
410 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
411 int i, npages, psind, rv;
413 MPASS(fs->object == fs->first_object);
414 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
415 MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
416 MPASS(fs->first_object->backing_object == NULL);
417 MPASS(fs->lookup_still_valid);
419 pager_first = OFF_TO_IDX(fs->entry->offset);
420 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
425 * Call the pager (driver) populate() method.
427 * There is no guarantee that the method will be called again
428 * if the current fault is for read, and a future fault is
429 * for write. Report the entry's maximum allowed protection
432 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
433 fault_type, fs->entry->max_protection, &pager_first, &pager_last);
435 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
436 if (rv == VM_PAGER_BAD) {
438 * VM_PAGER_BAD is the backdoor for a pager to request
439 * normal fault handling.
441 vm_fault_restore_map_lock(fs);
442 if (fs->map->timestamp != fs->map_generation)
443 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
444 return (KERN_NOT_RECEIVER);
446 if (rv != VM_PAGER_OK)
447 return (KERN_FAILURE); /* AKA SIGSEGV */
449 /* Ensure that the driver is obeying the interface. */
450 MPASS(pager_first <= pager_last);
451 MPASS(fs->first_pindex <= pager_last);
452 MPASS(fs->first_pindex >= pager_first);
453 MPASS(pager_last < fs->first_object->size);
455 vm_fault_restore_map_lock(fs);
456 if (fs->map->timestamp != fs->map_generation) {
457 vm_fault_populate_cleanup(fs->first_object, pager_first,
459 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
463 * The map is unchanged after our last unlock. Process the fault.
465 * The range [pager_first, pager_last] that is given to the
466 * pager is only a hint. The pager may populate any range
467 * within the object that includes the requested page index.
468 * In case the pager expanded the range, clip it to fit into
471 map_first = OFF_TO_IDX(fs->entry->offset);
472 if (map_first > pager_first) {
473 vm_fault_populate_cleanup(fs->first_object, pager_first,
475 pager_first = map_first;
477 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
478 if (map_last < pager_last) {
479 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
481 pager_last = map_last;
483 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
485 pidx += npages, m = vm_page_next(&m[npages - 1])) {
486 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
487 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
488 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
490 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
491 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
492 !pmap_ps_enabled(fs->map->pmap) || wired))
497 npages = atop(pagesizes[psind]);
498 for (i = 0; i < npages; i++) {
499 vm_fault_populate_check_page(&m[i]);
500 vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
503 VM_OBJECT_WUNLOCK(fs->first_object);
504 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
505 (wired ? PMAP_ENTER_WIRED : 0), psind);
506 #if defined(__amd64__)
507 if (psind > 0 && rv == KERN_FAILURE) {
508 for (i = 0; i < npages; i++) {
509 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
510 &m[i], prot, fault_type |
511 (wired ? PMAP_ENTER_WIRED : 0), 0);
512 MPASS(rv == KERN_SUCCESS);
516 MPASS(rv == KERN_SUCCESS);
518 VM_OBJECT_WLOCK(fs->first_object);
519 for (i = 0; i < npages; i++) {
520 if ((fault_flags & VM_FAULT_WIRE) != 0)
523 vm_page_activate(&m[i]);
524 if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
528 vm_page_xunbusy(&m[i]);
531 curthread->td_ru.ru_majflt++;
532 return (KERN_SUCCESS);
535 static int prot_fault_translation;
536 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
537 &prot_fault_translation, 0,
538 "Control signal to deliver on protection fault");
540 /* compat definition to keep common code for signal translation */
541 #define UCODE_PAGEFLT 12
543 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
549 * Handle a page fault occurring at the given address,
550 * requiring the given permissions, in the map specified.
551 * If successful, the page is inserted into the
552 * associated physical map.
554 * NOTE: the given address should be truncated to the
555 * proper page address.
557 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
558 * a standard error specifying why the fault is fatal is returned.
560 * The map in question must be referenced, and remains so.
561 * Caller may hold no locks.
564 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
565 int fault_flags, int *signo, int *ucode)
569 MPASS(signo == NULL || ucode != NULL);
571 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
572 ktrfault(vaddr, fault_type);
574 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
576 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
577 result == KERN_INVALID_ADDRESS ||
578 result == KERN_RESOURCE_SHORTAGE ||
579 result == KERN_PROTECTION_FAILURE ||
580 result == KERN_OUT_OF_BOUNDS,
581 ("Unexpected Mach error %d from vm_fault()", result));
583 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
586 if (result != KERN_SUCCESS && signo != NULL) {
589 case KERN_INVALID_ADDRESS:
591 *ucode = SEGV_MAPERR;
593 case KERN_RESOURCE_SHORTAGE:
597 case KERN_OUT_OF_BOUNDS:
601 case KERN_PROTECTION_FAILURE:
602 if (prot_fault_translation == 0) {
604 * Autodetect. This check also covers
605 * the images without the ABI-tag ELF
608 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
609 curproc->p_osrel >= P_OSREL_SIGSEGV) {
611 *ucode = SEGV_ACCERR;
614 *ucode = UCODE_PAGEFLT;
616 } else if (prot_fault_translation == 1) {
617 /* Always compat mode. */
619 *ucode = UCODE_PAGEFLT;
621 /* Always SIGSEGV mode. */
623 *ucode = SEGV_ACCERR;
627 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
636 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
641 if (fs->object->type != OBJT_VNODE)
642 return (KERN_SUCCESS);
643 vp = fs->object->handle;
645 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
646 return (KERN_SUCCESS);
650 * Perform an unlock in case the desired vnode changed while
651 * the map was unlocked during a retry.
655 locked = VOP_ISLOCKED(vp);
656 if (locked != LK_EXCLUSIVE)
660 * We must not sleep acquiring the vnode lock while we have
661 * the page exclusive busied or the object's
662 * paging-in-progress count incremented. Otherwise, we could
665 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread);
668 return (KERN_SUCCESS);
673 unlock_and_deallocate(fs);
675 fault_deallocate(fs);
676 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread);
679 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
680 return (KERN_RESOURCE_SHORTAGE);
684 * Calculate the desired readahead. Handle drop-behind.
686 * Returns the number of readahead blocks to pass to the pager.
689 vm_fault_readahead(struct faultstate *fs)
694 KASSERT(fs->lookup_still_valid, ("map unlocked"));
695 era = fs->entry->read_ahead;
696 behavior = vm_map_entry_behavior(fs->entry);
697 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
699 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
700 nera = VM_FAULT_READ_AHEAD_MAX;
701 if (fs->vaddr == fs->entry->next_read)
702 vm_fault_dontneed(fs, fs->vaddr, nera);
703 } else if (fs->vaddr == fs->entry->next_read) {
705 * This is a sequential fault. Arithmetically
706 * increase the requested number of pages in
707 * the read-ahead window. The requested
708 * number of pages is "# of sequential faults
709 * x (read ahead min + 1) + read ahead min"
711 nera = VM_FAULT_READ_AHEAD_MIN;
714 if (nera > VM_FAULT_READ_AHEAD_MAX)
715 nera = VM_FAULT_READ_AHEAD_MAX;
717 if (era == VM_FAULT_READ_AHEAD_MAX)
718 vm_fault_dontneed(fs, fs->vaddr, nera);
721 * This is a non-sequential fault.
727 * A read lock on the map suffices to update
728 * the read ahead count safely.
730 fs->entry->read_ahead = nera;
737 * Wait/Retry if the page is busy. We have to do this if the page is
738 * either exclusive or shared busy because the vm_pager may be using
739 * read busy for pageouts (and even pageins if it is the vnode pager),
740 * and we could end up trying to pagein and pageout the same page
743 * We can theoretically allow the busy case on a read fault if the page
744 * is marked valid, but since such pages are typically already pmap'd,
745 * putting that special case in might be more effort then it is worth.
746 * We cannot under any circumstances mess around with a shared busied
747 * page except, perhaps, to pmap it.
750 vm_fault_busy_sleep(struct faultstate *fs)
753 * Reference the page before unlocking and
754 * sleeping so that the page daemon is less
755 * likely to reclaim it.
757 vm_page_aflag_set(fs->m, PGA_REFERENCED);
758 if (fs->object != fs->first_object) {
759 fault_page_release(&fs->first_m);
760 vm_object_pip_wakeup(fs->first_object);
762 vm_object_pip_wakeup(fs->object);
764 if (fs->m == vm_page_lookup(fs->object, fs->pindex))
765 vm_page_busy_sleep(fs->m, "vmpfw", false);
767 VM_OBJECT_WUNLOCK(fs->object);
768 VM_CNT_INC(v_intrans);
769 vm_object_deallocate(fs->first_object);
773 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
774 int fault_flags, vm_page_t *m_hold)
776 struct faultstate fs;
777 struct domainset *dset;
778 vm_object_t next_object, retry_object;
779 vm_offset_t e_end, e_start;
780 vm_pindex_t retry_pindex;
781 vm_prot_t prot, retry_prot;
782 int ahead, alloc_req, behind, cluster_offset, faultcount;
783 int nera, oom, result, rv;
785 boolean_t wired; /* Passed by reference. */
786 bool dead, hardfault, is_first_object_locked;
788 VM_CNT_INC(v_vm_faults);
790 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
791 return (KERN_PROTECTION_FAILURE);
804 * Find the backing store object and offset into it to begin the
808 result = vm_map_lookup(&fs.map, vaddr, fault_type |
809 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
810 &fs.first_pindex, &prot, &wired);
811 if (result != KERN_SUCCESS) {
816 fs.map_generation = fs.map->timestamp;
818 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
819 panic("%s: fault on nofault entry, addr: %#lx",
820 __func__, (u_long)vaddr);
823 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
824 fs.entry->wiring_thread != curthread) {
825 vm_map_unlock_read(fs.map);
827 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
828 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
830 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
831 vm_map_unlock_and_wait(fs.map, 0);
833 vm_map_unlock(fs.map);
837 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
840 fault_type = prot | (fault_type & VM_PROT_COPY);
842 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
843 ("!wired && VM_FAULT_WIRE"));
846 * Try to avoid lock contention on the top-level object through
847 * special-case handling of some types of page faults, specifically,
848 * those that are mapping an existing page from the top-level object.
849 * Under this condition, a read lock on the object suffices, allowing
850 * multiple page faults of a similar type to run in parallel.
852 if (fs.vp == NULL /* avoid locked vnode leak */ &&
853 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
854 VM_OBJECT_RLOCK(fs.first_object);
855 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
856 fault_flags, wired, m_hold);
857 if (rv == KERN_SUCCESS)
859 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
860 VM_OBJECT_RUNLOCK(fs.first_object);
861 VM_OBJECT_WLOCK(fs.first_object);
864 VM_OBJECT_WLOCK(fs.first_object);
868 * Make a reference to this object to prevent its disposal while we
869 * are messing with it. Once we have the reference, the map is free
870 * to be diddled. Since objects reference their shadows (and copies),
871 * they will stay around as well.
873 * Bump the paging-in-progress count to prevent size changes (e.g.
874 * truncation operations) during I/O.
876 vm_object_reference_locked(fs.first_object);
877 vm_object_pip_add(fs.first_object, 1);
879 fs.lookup_still_valid = true;
881 fs.m_cow = fs.m = fs.first_m = NULL;
884 * Search for the page at object/offset.
886 fs.object = fs.first_object;
887 fs.pindex = fs.first_pindex;
889 KASSERT(fs.m == NULL,
890 ("page still set %p at loop start", fs.m));
892 * If the object is marked for imminent termination,
893 * we retry here, since the collapse pass has raced
894 * with us. Otherwise, if we see terminally dead
895 * object, return fail.
897 if ((fs.object->flags & OBJ_DEAD) != 0) {
898 dead = fs.object->type == OBJT_DEAD;
899 unlock_and_deallocate(&fs);
901 return (KERN_PROTECTION_FAILURE);
907 * See if page is resident
909 fs.m = vm_page_lookup(fs.object, fs.pindex);
911 if (vm_page_tryxbusy(fs.m) == 0) {
912 vm_fault_busy_sleep(&fs);
917 * The page is marked busy for other processes and the
918 * pagedaemon. If it still isn't completely valid
919 * (readable), jump to readrest, else break-out ( we
922 if (!vm_page_all_valid(fs.m))
924 VM_OBJECT_WUNLOCK(fs.object);
925 break; /* break to PAGE HAS BEEN FOUND. */
927 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
928 VM_OBJECT_ASSERT_WLOCKED(fs.object);
931 * Page is not resident. If the pager might contain the page
932 * or this is the beginning of the search, allocate a new
933 * page. (Default objects are zero-fill, so there is no real
936 if (fs.object->type != OBJT_DEFAULT ||
937 fs.object == fs.first_object) {
938 if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) {
939 rv = vm_fault_lock_vnode(&fs, true);
940 MPASS(rv == KERN_SUCCESS ||
941 rv == KERN_RESOURCE_SHORTAGE);
942 if (rv == KERN_RESOURCE_SHORTAGE)
945 if (fs.pindex >= fs.object->size) {
946 unlock_and_deallocate(&fs);
947 return (KERN_OUT_OF_BOUNDS);
950 if (fs.object == fs.first_object &&
951 (fs.first_object->flags & OBJ_POPULATE) != 0 &&
952 fs.first_object->shadow_count == 0) {
953 rv = vm_fault_populate(&fs, prot, fault_type,
954 fault_flags, wired, m_hold);
958 unlock_and_deallocate(&fs);
960 case KERN_RESOURCE_SHORTAGE:
961 unlock_and_deallocate(&fs);
963 case KERN_NOT_RECEIVER:
965 * Pager's populate() method
966 * returned VM_PAGER_BAD.
970 panic("inconsistent return codes");
975 * Allocate a new page for this object/offset pair.
977 * Unlocked read of the p_flag is harmless. At
978 * worst, the P_KILLED might be not observed
979 * there, and allocation can fail, causing
980 * restart and new reading of the p_flag.
982 dset = fs.object->domain.dr_policy;
984 dset = curthread->td_domain.dr_policy;
985 if (!vm_page_count_severe_set(&dset->ds_mask) ||
987 #if VM_NRESERVLEVEL > 0
988 vm_object_color(fs.object, atop(vaddr) -
991 alloc_req = P_KILLED(curproc) ?
992 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
993 if (fs.object->type != OBJT_VNODE &&
994 fs.object->backing_object == NULL)
995 alloc_req |= VM_ALLOC_ZERO;
996 fs.m = vm_page_alloc(fs.object, fs.pindex,
1000 unlock_and_deallocate(&fs);
1001 if (vm_pfault_oom_attempts < 0 ||
1002 oom < vm_pfault_oom_attempts) {
1005 vm_pfault_oom_wait * hz);
1006 goto RetryFault_oom;
1010 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1011 curproc->p_pid, curproc->p_comm);
1012 vm_pageout_oom(VM_OOM_MEM_PF);
1019 * Default objects have no pager so no exclusive busy exists
1020 * to protect this page in the chain. Skip to the next
1021 * object without dropping the lock to preserve atomicity of
1024 if (fs.object->type == OBJT_DEFAULT)
1028 * At this point, we have either allocated a new page or found
1029 * an existing page that is only partially valid.
1031 * We hold a reference on the current object and the page is
1032 * exclusive busied. The exclusive busy prevents simultaneous
1033 * faults and collapses while the object lock is dropped.
1035 VM_OBJECT_WUNLOCK(fs.object);
1038 * If the pager for the current object might have the page,
1039 * then determine the number of additional pages to read and
1040 * potentially reprioritize previously read pages for earlier
1041 * reclamation. These operations should only be performed
1042 * once per page fault. Even if the current pager doesn't
1043 * have the page, the number of additional pages to read will
1044 * apply to subsequent objects in the shadow chain.
1046 if (nera == -1 && !P_KILLED(curproc)) {
1047 nera = vm_fault_readahead(&fs);
1049 * Prepare for unlocking the map. Save the map
1050 * entry's start and end addresses, which are used to
1051 * optimize the size of the pager operation below.
1052 * Even if the map entry's addresses change after
1053 * unlocking the map, using the saved addresses is
1056 e_start = fs.entry->start;
1057 e_end = fs.entry->end;
1058 behavior = vm_map_entry_behavior(fs.entry);
1062 * Call the pager to retrieve the page if there is a chance
1063 * that the pager has it, and potentially retrieve additional
1064 * pages at the same time.
1066 if (fs.object->type != OBJT_DEFAULT) {
1068 * Release the map lock before locking the vnode or
1069 * sleeping in the pager. (If the current object has
1070 * a shadow, then an earlier iteration of this loop
1071 * may have already unlocked the map.)
1075 rv = vm_fault_lock_vnode(&fs, false);
1076 MPASS(rv == KERN_SUCCESS ||
1077 rv == KERN_RESOURCE_SHORTAGE);
1078 if (rv == KERN_RESOURCE_SHORTAGE)
1080 KASSERT(fs.vp == NULL || !fs.map->system_map,
1081 ("vm_fault: vnode-backed object mapped by system map"));
1084 * Page in the requested page and hint the pager,
1085 * that it may bring up surrounding pages.
1087 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1088 P_KILLED(curproc)) {
1092 /* Is this a sequential fault? */
1098 * Request a cluster of pages that is
1099 * aligned to a VM_FAULT_READ_DEFAULT
1100 * page offset boundary within the
1101 * object. Alignment to a page offset
1102 * boundary is more likely to coincide
1103 * with the underlying file system
1104 * block than alignment to a virtual
1107 cluster_offset = fs.pindex %
1108 VM_FAULT_READ_DEFAULT;
1109 behind = ulmin(cluster_offset,
1110 atop(vaddr - e_start));
1111 ahead = VM_FAULT_READ_DEFAULT - 1 -
1114 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1116 rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1118 if (rv == VM_PAGER_OK) {
1119 faultcount = behind + 1 + ahead;
1121 break; /* break to PAGE HAS BEEN FOUND. */
1123 VM_OBJECT_WLOCK(fs.object);
1124 if (rv == VM_PAGER_ERROR)
1125 printf("vm_fault: pager read error, pid %d (%s)\n",
1126 curproc->p_pid, curproc->p_comm);
1129 * If an I/O error occurred or the requested page was
1130 * outside the range of the pager, clean up and return
1133 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1134 fault_page_free(&fs.m);
1135 unlock_and_deallocate(&fs);
1136 return (KERN_OUT_OF_BOUNDS);
1143 * The requested page does not exist at this object/
1144 * offset. Remove the invalid page from the object,
1145 * waking up anyone waiting for it, and continue on to
1146 * the next object. However, if this is the top-level
1147 * object, we must leave the busy page in place to
1148 * prevent another process from rushing past us, and
1149 * inserting the page in that object at the same time
1152 if (fs.object == fs.first_object) {
1156 fault_page_free(&fs.m);
1159 * Move on to the next object. Lock the next object before
1160 * unlocking the current one.
1162 VM_OBJECT_ASSERT_WLOCKED(fs.object);
1163 next_object = fs.object->backing_object;
1164 if (next_object == NULL) {
1166 * If there's no object left, fill the page in the top
1167 * object with zeros.
1169 VM_OBJECT_WUNLOCK(fs.object);
1170 if (fs.object != fs.first_object) {
1171 vm_object_pip_wakeup(fs.object);
1172 fs.object = fs.first_object;
1173 fs.pindex = fs.first_pindex;
1175 MPASS(fs.first_m != NULL);
1176 MPASS(fs.m == NULL);
1181 * Zero the page if necessary and mark it valid.
1183 if ((fs.m->flags & PG_ZERO) == 0) {
1184 pmap_zero_page(fs.m);
1186 VM_CNT_INC(v_ozfod);
1189 vm_page_valid(fs.m);
1190 /* Don't try to prefault neighboring pages. */
1192 break; /* break to PAGE HAS BEEN FOUND. */
1194 MPASS(fs.first_m != NULL);
1195 KASSERT(fs.object != next_object,
1196 ("object loop %p", next_object));
1197 VM_OBJECT_WLOCK(next_object);
1198 vm_object_pip_add(next_object, 1);
1199 if (fs.object != fs.first_object)
1200 vm_object_pip_wakeup(fs.object);
1202 OFF_TO_IDX(fs.object->backing_object_offset);
1203 VM_OBJECT_WUNLOCK(fs.object);
1204 fs.object = next_object;
1209 * PAGE HAS BEEN FOUND. A valid page has been found and exclusively
1210 * busied. The object lock must no longer be held.
1212 vm_page_assert_xbusied(fs.m);
1213 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1216 * If the page is being written, but isn't already owned by the
1217 * top-level object, we have to copy it into a new page owned by the
1220 if (fs.object != fs.first_object) {
1222 * We only really need to copy if we want to write it.
1224 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1226 * This allows pages to be virtually copied from a
1227 * backing_object into the first_object, where the
1228 * backing object has no other refs to it, and cannot
1229 * gain any more refs. Instead of a bcopy, we just
1230 * move the page from the backing object to the
1231 * first object. Note that we must mark the page
1232 * dirty in the first object so that it will go out
1233 * to swap when needed.
1235 is_first_object_locked = false;
1238 * Only one shadow object
1240 fs.object->shadow_count == 1 &&
1242 * No COW refs, except us
1244 fs.object->ref_count == 1 &&
1246 * No one else can look this object up
1248 fs.object->handle == NULL &&
1250 * No other ways to look the object up
1252 (fs.object->flags & OBJ_ANON) != 0 &&
1253 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1255 * We don't chase down the shadow chain
1257 fs.object == fs.first_object->backing_object &&
1258 VM_OBJECT_TRYWLOCK(fs.object)) {
1261 * Remove but keep xbusy for replace. fs.m is
1262 * moved into fs.first_object and left busy
1263 * while fs.first_m is conditionally freed.
1265 vm_page_remove_xbusy(fs.m);
1266 vm_page_replace(fs.m, fs.first_object,
1267 fs.first_pindex, fs.first_m);
1268 vm_page_dirty(fs.m);
1269 #if VM_NRESERVLEVEL > 0
1271 * Rename the reservation.
1273 vm_reserv_rename(fs.m, fs.first_object,
1274 fs.object, OFF_TO_IDX(
1275 fs.first_object->backing_object_offset));
1277 VM_OBJECT_WUNLOCK(fs.object);
1278 VM_OBJECT_WUNLOCK(fs.first_object);
1281 VM_CNT_INC(v_cow_optim);
1283 if (is_first_object_locked)
1284 VM_OBJECT_WUNLOCK(fs.first_object);
1286 * Oh, well, lets copy it.
1288 pmap_copy_page(fs.m, fs.first_m);
1289 vm_page_valid(fs.first_m);
1290 if (wired && (fault_flags &
1291 VM_FAULT_WIRE) == 0) {
1292 vm_page_wire(fs.first_m);
1293 vm_page_unwire(fs.m, PQ_INACTIVE);
1296 * Save the cow page to be released after
1297 * pmap_enter is complete.
1303 * fs.object != fs.first_object due to above
1306 vm_object_pip_wakeup(fs.object);
1309 * We only try to prefault read-only mappings to the
1310 * neighboring pages when this copy-on-write fault is
1311 * a hard fault. In other cases, trying to prefault
1312 * is typically wasted effort.
1314 if (faultcount == 0)
1318 * Only use the new page below...
1320 fs.object = fs.first_object;
1321 fs.pindex = fs.first_pindex;
1323 VM_CNT_INC(v_cow_faults);
1324 curthread->td_cow++;
1326 prot &= ~VM_PROT_WRITE;
1331 * We must verify that the maps have not changed since our last
1334 if (!fs.lookup_still_valid) {
1335 if (!vm_map_trylock_read(fs.map)) {
1336 fault_deallocate(&fs);
1339 fs.lookup_still_valid = true;
1340 if (fs.map->timestamp != fs.map_generation) {
1341 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1342 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1345 * If we don't need the page any longer, put it on the inactive
1346 * list (the easiest thing to do here). If no one needs it,
1347 * pageout will grab it eventually.
1349 if (result != KERN_SUCCESS) {
1350 fault_deallocate(&fs);
1353 * If retry of map lookup would have blocked then
1354 * retry fault from start.
1356 if (result == KERN_FAILURE)
1360 if ((retry_object != fs.first_object) ||
1361 (retry_pindex != fs.first_pindex)) {
1362 fault_deallocate(&fs);
1367 * Check whether the protection has changed or the object has
1368 * been copied while we left the map unlocked. Changing from
1369 * read to write permission is OK - we leave the page
1370 * write-protected, and catch the write fault. Changing from
1371 * write to read permission means that we can't mark the page
1372 * write-enabled after all.
1375 fault_type &= retry_prot;
1377 fault_deallocate(&fs);
1381 /* Reassert because wired may have changed. */
1382 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1383 ("!wired && VM_FAULT_WIRE"));
1386 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1389 * If the page was filled by a pager, save the virtual address that
1390 * should be faulted on next under a sequential access pattern to the
1391 * map entry. A read lock on the map suffices to update this address
1395 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1398 * Page must be completely valid or it is not fit to
1399 * map into user space. vm_pager_get_pages() ensures this.
1401 vm_page_assert_xbusied(fs.m);
1402 KASSERT(vm_page_all_valid(fs.m),
1403 ("vm_fault: page %p partially invalid", fs.m));
1405 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags);
1408 * Put this page into the physical map. We had to do the unlock above
1409 * because pmap_enter() may sleep. We don't put the page
1410 * back on the active queue until later so that the pageout daemon
1411 * won't find it (yet).
1413 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1414 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1415 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1417 vm_fault_prefault(&fs, vaddr,
1418 faultcount > 0 ? behind : PFBAK,
1419 faultcount > 0 ? ahead : PFFOR, false);
1422 * If the page is not wired down, then put it where the pageout daemon
1425 if ((fault_flags & VM_FAULT_WIRE) != 0)
1428 vm_page_activate(fs.m);
1429 if (m_hold != NULL) {
1433 vm_page_xunbusy(fs.m);
1437 * Unlock everything, and return
1439 fault_deallocate(&fs);
1441 VM_CNT_INC(v_io_faults);
1442 curthread->td_ru.ru_majflt++;
1444 if (racct_enable && fs.object->type == OBJT_VNODE) {
1446 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1447 racct_add_force(curproc, RACCT_WRITEBPS,
1448 PAGE_SIZE + behind * PAGE_SIZE);
1449 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1451 racct_add_force(curproc, RACCT_READBPS,
1452 PAGE_SIZE + ahead * PAGE_SIZE);
1453 racct_add_force(curproc, RACCT_READIOPS, 1);
1455 PROC_UNLOCK(curproc);
1459 curthread->td_ru.ru_minflt++;
1461 return (KERN_SUCCESS);
1465 * Speed up the reclamation of pages that precede the faulting pindex within
1466 * the first object of the shadow chain. Essentially, perform the equivalent
1467 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1468 * the faulting pindex by the cluster size when the pages read by vm_fault()
1469 * cross a cluster-size boundary. The cluster size is the greater of the
1470 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1472 * When "fs->first_object" is a shadow object, the pages in the backing object
1473 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1474 * function must only be concerned with pages in the first object.
1477 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1479 vm_map_entry_t entry;
1480 vm_object_t first_object, object;
1481 vm_offset_t end, start;
1482 vm_page_t m, m_next;
1483 vm_pindex_t pend, pstart;
1486 object = fs->object;
1487 VM_OBJECT_ASSERT_UNLOCKED(object);
1488 first_object = fs->first_object;
1489 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1490 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1491 VM_OBJECT_RLOCK(first_object);
1492 size = VM_FAULT_DONTNEED_MIN;
1493 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1494 size = pagesizes[1];
1495 end = rounddown2(vaddr, size);
1496 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1497 (entry = fs->entry)->start < end) {
1498 if (end - entry->start < size)
1499 start = entry->start;
1502 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1503 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1505 m_next = vm_page_find_least(first_object, pstart);
1506 pend = OFF_TO_IDX(entry->offset) + atop(end -
1508 while ((m = m_next) != NULL && m->pindex < pend) {
1509 m_next = TAILQ_NEXT(m, listq);
1510 if (!vm_page_all_valid(m) ||
1515 * Don't clear PGA_REFERENCED, since it would
1516 * likely represent a reference by a different
1519 * Typically, at this point, prefetched pages
1520 * are still in the inactive queue. Only
1521 * pages that triggered page faults are in the
1522 * active queue. The test for whether the page
1523 * is in the inactive queue is racy; in the
1524 * worst case we will requeue the page
1527 if (!vm_page_inactive(m))
1528 vm_page_deactivate(m);
1531 VM_OBJECT_RUNLOCK(first_object);
1536 * vm_fault_prefault provides a quick way of clustering
1537 * pagefaults into a processes address space. It is a "cousin"
1538 * of vm_map_pmap_enter, except it runs at page fault time instead
1542 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1543 int backward, int forward, bool obj_locked)
1546 vm_map_entry_t entry;
1547 vm_object_t backing_object, lobject;
1548 vm_offset_t addr, starta;
1553 pmap = fs->map->pmap;
1554 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1559 if (addra < backward * PAGE_SIZE) {
1560 starta = entry->start;
1562 starta = addra - backward * PAGE_SIZE;
1563 if (starta < entry->start)
1564 starta = entry->start;
1568 * Generate the sequence of virtual addresses that are candidates for
1569 * prefaulting in an outward spiral from the faulting virtual address,
1570 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1571 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1572 * If the candidate address doesn't have a backing physical page, then
1573 * the loop immediately terminates.
1575 for (i = 0; i < 2 * imax(backward, forward); i++) {
1576 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1578 if (addr > addra + forward * PAGE_SIZE)
1581 if (addr < starta || addr >= entry->end)
1584 if (!pmap_is_prefaultable(pmap, addr))
1587 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1588 lobject = entry->object.vm_object;
1590 VM_OBJECT_RLOCK(lobject);
1591 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1592 lobject->type == OBJT_DEFAULT &&
1593 (backing_object = lobject->backing_object) != NULL) {
1594 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1595 0, ("vm_fault_prefault: unaligned object offset"));
1596 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1597 VM_OBJECT_RLOCK(backing_object);
1598 if (!obj_locked || lobject != entry->object.vm_object)
1599 VM_OBJECT_RUNLOCK(lobject);
1600 lobject = backing_object;
1603 if (!obj_locked || lobject != entry->object.vm_object)
1604 VM_OBJECT_RUNLOCK(lobject);
1607 if (vm_page_all_valid(m) &&
1608 (m->flags & PG_FICTITIOUS) == 0)
1609 pmap_enter_quick(pmap, addr, m, entry->protection);
1610 if (!obj_locked || lobject != entry->object.vm_object)
1611 VM_OBJECT_RUNLOCK(lobject);
1616 * Hold each of the physical pages that are mapped by the specified range of
1617 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1618 * and allow the specified types of access, "prot". If all of the implied
1619 * pages are successfully held, then the number of held pages is returned
1620 * together with pointers to those pages in the array "ma". However, if any
1621 * of the pages cannot be held, -1 is returned.
1624 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1625 vm_prot_t prot, vm_page_t *ma, int max_count)
1627 vm_offset_t end, va;
1630 boolean_t pmap_failed;
1634 end = round_page(addr + len);
1635 addr = trunc_page(addr);
1638 * Check for illegal addresses.
1640 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1643 if (atop(end - addr) > max_count)
1644 panic("vm_fault_quick_hold_pages: count > max_count");
1645 count = atop(end - addr);
1648 * Most likely, the physical pages are resident in the pmap, so it is
1649 * faster to try pmap_extract_and_hold() first.
1651 pmap_failed = FALSE;
1652 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1653 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1656 else if ((prot & VM_PROT_WRITE) != 0 &&
1657 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1659 * Explicitly dirty the physical page. Otherwise, the
1660 * caller's changes may go unnoticed because they are
1661 * performed through an unmanaged mapping or by a DMA
1664 * The object lock is not held here.
1665 * See vm_page_clear_dirty_mask().
1672 * One or more pages could not be held by the pmap. Either no
1673 * page was mapped at the specified virtual address or that
1674 * mapping had insufficient permissions. Attempt to fault in
1675 * and hold these pages.
1677 * If vm_fault_disable_pagefaults() was called,
1678 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1679 * acquire MD VM locks, which means we must not call
1680 * vm_fault(). Some (out of tree) callers mark
1681 * too wide a code area with vm_fault_disable_pagefaults()
1682 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1683 * the proper behaviour explicitly.
1685 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1686 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1688 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1689 if (*mp == NULL && vm_fault(map, va, prot,
1690 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1695 for (mp = ma; mp < ma + count; mp++)
1697 vm_page_unwire(*mp, PQ_INACTIVE);
1703 * vm_fault_copy_entry
1705 * Create new shadow object backing dst_entry with private copy of
1706 * all underlying pages. When src_entry is equal to dst_entry,
1707 * function implements COW for wired-down map entry. Otherwise,
1708 * it forks wired entry into dst_map.
1710 * In/out conditions:
1711 * The source and destination maps must be locked for write.
1712 * The source map entry must be wired down (or be a sharing map
1713 * entry corresponding to a main map entry that is wired down).
1716 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1717 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1718 vm_ooffset_t *fork_charge)
1720 vm_object_t backing_object, dst_object, object, src_object;
1721 vm_pindex_t dst_pindex, pindex, src_pindex;
1722 vm_prot_t access, prot;
1732 upgrade = src_entry == dst_entry;
1733 access = prot = dst_entry->protection;
1735 src_object = src_entry->object.vm_object;
1736 src_pindex = OFF_TO_IDX(src_entry->offset);
1738 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1739 dst_object = src_object;
1740 vm_object_reference(dst_object);
1743 * Create the top-level object for the destination entry.
1744 * Doesn't actually shadow anything - we copy the pages
1747 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1748 dst_entry->start), NULL, NULL, 0);
1749 #if VM_NRESERVLEVEL > 0
1750 dst_object->flags |= OBJ_COLORED;
1751 dst_object->pg_color = atop(dst_entry->start);
1753 dst_object->domain = src_object->domain;
1754 dst_object->charge = dst_entry->end - dst_entry->start;
1757 VM_OBJECT_WLOCK(dst_object);
1758 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1759 ("vm_fault_copy_entry: vm_object not NULL"));
1760 if (src_object != dst_object) {
1761 dst_entry->object.vm_object = dst_object;
1762 dst_entry->offset = 0;
1763 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1765 if (fork_charge != NULL) {
1766 KASSERT(dst_entry->cred == NULL,
1767 ("vm_fault_copy_entry: leaked swp charge"));
1768 dst_object->cred = curthread->td_ucred;
1769 crhold(dst_object->cred);
1770 *fork_charge += dst_object->charge;
1771 } else if ((dst_object->type == OBJT_DEFAULT ||
1772 dst_object->type == OBJT_SWAP) &&
1773 dst_object->cred == NULL) {
1774 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1776 dst_object->cred = dst_entry->cred;
1777 dst_entry->cred = NULL;
1781 * If not an upgrade, then enter the mappings in the pmap as
1782 * read and/or execute accesses. Otherwise, enter them as
1785 * A writeable large page mapping is only created if all of
1786 * the constituent small page mappings are modified. Marking
1787 * PTEs as modified on inception allows promotion to happen
1788 * without taking potentially large number of soft faults.
1791 access &= ~VM_PROT_WRITE;
1794 * Loop through all of the virtual pages within the entry's
1795 * range, copying each page from the source object to the
1796 * destination object. Since the source is wired, those pages
1797 * must exist. In contrast, the destination is pageable.
1798 * Since the destination object doesn't share any backing storage
1799 * with the source object, all of its pages must be dirtied,
1800 * regardless of whether they can be written.
1802 for (vaddr = dst_entry->start, dst_pindex = 0;
1803 vaddr < dst_entry->end;
1804 vaddr += PAGE_SIZE, dst_pindex++) {
1807 * Find the page in the source object, and copy it in.
1808 * Because the source is wired down, the page will be
1811 if (src_object != dst_object)
1812 VM_OBJECT_RLOCK(src_object);
1813 object = src_object;
1814 pindex = src_pindex + dst_pindex;
1815 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1816 (backing_object = object->backing_object) != NULL) {
1818 * Unless the source mapping is read-only or
1819 * it is presently being upgraded from
1820 * read-only, the first object in the shadow
1821 * chain should provide all of the pages. In
1822 * other words, this loop body should never be
1823 * executed when the source mapping is already
1826 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1828 ("vm_fault_copy_entry: main object missing page"));
1830 VM_OBJECT_RLOCK(backing_object);
1831 pindex += OFF_TO_IDX(object->backing_object_offset);
1832 if (object != dst_object)
1833 VM_OBJECT_RUNLOCK(object);
1834 object = backing_object;
1836 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1838 if (object != dst_object) {
1840 * Allocate a page in the destination object.
1842 dst_m = vm_page_alloc(dst_object, (src_object ==
1843 dst_object ? src_pindex : 0) + dst_pindex,
1845 if (dst_m == NULL) {
1846 VM_OBJECT_WUNLOCK(dst_object);
1847 VM_OBJECT_RUNLOCK(object);
1848 vm_wait(dst_object);
1849 VM_OBJECT_WLOCK(dst_object);
1852 pmap_copy_page(src_m, dst_m);
1853 VM_OBJECT_RUNLOCK(object);
1854 dst_m->dirty = dst_m->valid = src_m->valid;
1857 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1859 if (dst_m->pindex >= dst_object->size) {
1861 * We are upgrading. Index can occur
1862 * out of bounds if the object type is
1863 * vnode and the file was truncated.
1865 vm_page_xunbusy(dst_m);
1869 VM_OBJECT_WUNLOCK(dst_object);
1872 * Enter it in the pmap. If a wired, copy-on-write
1873 * mapping is being replaced by a write-enabled
1874 * mapping, then wire that new mapping.
1876 * The page can be invalid if the user called
1877 * msync(MS_INVALIDATE) or truncated the backing vnode
1878 * or shared memory object. In this case, do not
1879 * insert it into pmap, but still do the copy so that
1880 * all copies of the wired map entry have similar
1883 if (vm_page_all_valid(dst_m)) {
1884 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1885 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1889 * Mark it no longer busy, and put it on the active list.
1891 VM_OBJECT_WLOCK(dst_object);
1894 if (src_m != dst_m) {
1895 vm_page_unwire(src_m, PQ_INACTIVE);
1896 vm_page_wire(dst_m);
1898 KASSERT(vm_page_wired(dst_m),
1899 ("dst_m %p is not wired", dst_m));
1902 vm_page_activate(dst_m);
1904 vm_page_xunbusy(dst_m);
1906 VM_OBJECT_WUNLOCK(dst_object);
1908 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1909 vm_object_deallocate(src_object);
1914 * Block entry into the machine-independent layer's page fault handler by
1915 * the calling thread. Subsequent calls to vm_fault() by that thread will
1916 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1917 * spurious page faults.
1920 vm_fault_disable_pagefaults(void)
1923 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1927 vm_fault_enable_pagefaults(int save)
1930 curthread_pflags_restore(save);