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
127 vm_object_t first_object;
128 vm_pindex_t first_pindex;
130 vm_map_entry_t entry;
132 bool lookup_still_valid;
136 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
138 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
139 int backward, int forward, bool obj_locked);
141 static int vm_pfault_oom_attempts = 3;
142 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
143 &vm_pfault_oom_attempts, 0,
144 "Number of page allocation attempts in page fault handler before it "
145 "triggers OOM handling");
147 static int vm_pfault_oom_wait = 10;
148 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
149 &vm_pfault_oom_wait, 0,
150 "Number of seconds to wait for free pages before retrying "
151 "the page fault handler");
154 release_page(struct faultstate *fs)
158 vm_page_xunbusy(fs->m);
160 vm_page_deactivate(fs->m);
161 vm_page_unlock(fs->m);
167 unlock_map(struct faultstate *fs)
170 if (fs->lookup_still_valid) {
171 vm_map_lookup_done(fs->map, fs->entry);
172 fs->lookup_still_valid = false;
177 unlock_vp(struct faultstate *fs)
180 if (fs->vp != NULL) {
187 fault_deallocate(struct faultstate *fs)
190 vm_object_pip_wakeup(fs->object);
191 if (fs->object != fs->first_object) {
192 VM_OBJECT_WLOCK(fs->first_object);
193 vm_page_free(fs->first_m);
194 vm_object_pip_wakeup(fs->first_object);
195 VM_OBJECT_WUNLOCK(fs->first_object);
198 vm_object_deallocate(fs->first_object);
204 unlock_and_deallocate(struct faultstate *fs)
207 VM_OBJECT_WUNLOCK(fs->object);
208 fault_deallocate(fs);
212 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
213 vm_prot_t fault_type, int fault_flags, bool excl)
217 if (((prot & VM_PROT_WRITE) == 0 &&
218 (fault_flags & VM_FAULT_DIRTY) == 0) ||
219 (m->oflags & VPO_UNMANAGED) != 0)
222 VM_OBJECT_ASSERT_LOCKED(m->object);
223 VM_PAGE_OBJECT_BUSY_ASSERT(m);
225 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
226 (fault_flags & VM_FAULT_WIRE) == 0) ||
227 (fault_flags & VM_FAULT_DIRTY) != 0;
229 vm_object_set_writeable_dirty(m->object);
233 * If two callers of vm_fault_dirty() with excl ==
234 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
235 * flag set, other with flag clear, race, it is
236 * possible for the no-NOSYNC thread to see m->dirty
237 * != 0 and not clear PGA_NOSYNC. Take vm_page lock
238 * around manipulation of PGA_NOSYNC and
239 * vm_page_dirty() call to avoid the race.
244 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
245 * if the page is already dirty to prevent data written with
246 * the expectation of being synced from not being synced.
247 * Likewise if this entry does not request NOSYNC then make
248 * sure the page isn't marked NOSYNC. Applications sharing
249 * data should use the same flags to avoid ping ponging.
251 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
253 vm_page_aflag_set(m, PGA_NOSYNC);
256 vm_page_aflag_clear(m, PGA_NOSYNC);
260 * If the fault is a write, we know that this page is being
261 * written NOW so dirty it explicitly to save on
262 * pmap_is_modified() calls later.
264 * Also, since the page is now dirty, we can possibly tell
265 * the pager to release any swap backing the page. Calling
266 * the pager requires a write lock on the object.
273 vm_pager_page_unswapped(m);
277 * Unlocks fs.first_object and fs.map on success.
280 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
281 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
284 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
285 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
292 MPASS(fs->vp == NULL);
293 vm_object_busy(fs->first_object);
294 m = vm_page_lookup(fs->first_object, fs->first_pindex);
295 /* A busy page can be mapped for read|execute access. */
296 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
297 vm_page_busied(m)) || !vm_page_all_valid(m)) {
303 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
304 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
306 if ((m->flags & PG_FICTITIOUS) == 0 &&
307 (m_super = vm_reserv_to_superpage(m)) != NULL &&
308 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
309 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
310 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
311 (pagesizes[m_super->psind] - 1)) && !wired &&
312 pmap_ps_enabled(fs->map->pmap)) {
313 flags = PS_ALL_VALID;
314 if ((prot & VM_PROT_WRITE) != 0) {
316 * Create a superpage mapping allowing write access
317 * only if none of the constituent pages are busy and
318 * all of them are already dirty (except possibly for
319 * the page that was faulted on).
321 flags |= PS_NONE_BUSY;
322 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
323 flags |= PS_ALL_DIRTY;
325 if (vm_page_ps_test(m_super, flags, m)) {
327 psind = m_super->psind;
328 vaddr = rounddown2(vaddr, pagesizes[psind]);
329 /* Preset the modified bit for dirty superpages. */
330 if ((flags & PS_ALL_DIRTY) != 0)
331 fault_type |= VM_PROT_WRITE;
335 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
336 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
337 if (rv != KERN_SUCCESS)
339 if (m_hold != NULL) {
343 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false);
344 if (psind == 0 && !wired)
345 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
346 VM_OBJECT_RUNLOCK(fs->first_object);
347 vm_map_lookup_done(fs->map, fs->entry);
348 curthread->td_ru.ru_minflt++;
351 vm_object_unbusy(fs->first_object);
356 vm_fault_restore_map_lock(struct faultstate *fs)
359 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
360 MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0);
362 if (!vm_map_trylock_read(fs->map)) {
363 VM_OBJECT_WUNLOCK(fs->first_object);
364 vm_map_lock_read(fs->map);
365 VM_OBJECT_WLOCK(fs->first_object);
367 fs->lookup_still_valid = true;
371 vm_fault_populate_check_page(vm_page_t m)
375 * Check each page to ensure that the pager is obeying the
376 * interface: the page must be installed in the object, fully
377 * valid, and exclusively busied.
380 MPASS(vm_page_all_valid(m));
381 MPASS(vm_page_xbusied(m));
385 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
391 VM_OBJECT_ASSERT_WLOCKED(object);
392 MPASS(first <= last);
393 for (pidx = first, m = vm_page_lookup(object, pidx);
394 pidx <= last; pidx++, m = vm_page_next(m)) {
395 vm_fault_populate_check_page(m);
397 vm_page_deactivate(m);
404 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
405 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);
520 for (i = 0; i < npages; i++) {
521 if ((fault_flags & VM_FAULT_WIRE) != 0) {
524 vm_page_change_lock(&m[i], &m_mtx);
525 vm_page_activate(&m[i]);
527 if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
531 vm_page_xunbusy(&m[i]);
536 curthread->td_ru.ru_majflt++;
537 return (KERN_SUCCESS);
540 static int prot_fault_translation;
541 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
542 &prot_fault_translation, 0,
543 "Control signal to deliver on protection fault");
545 /* compat definition to keep common code for signal translation */
546 #define UCODE_PAGEFLT 12
548 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
554 * Handle a page fault occurring at the given address,
555 * requiring the given permissions, in the map specified.
556 * If successful, the page is inserted into the
557 * associated physical map.
559 * NOTE: the given address should be truncated to the
560 * proper page address.
562 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
563 * a standard error specifying why the fault is fatal is returned.
565 * The map in question must be referenced, and remains so.
566 * Caller may hold no locks.
569 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
570 int fault_flags, int *signo, int *ucode)
574 MPASS(signo == NULL || ucode != NULL);
576 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
577 ktrfault(vaddr, fault_type);
579 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
581 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
582 result == KERN_INVALID_ADDRESS ||
583 result == KERN_RESOURCE_SHORTAGE ||
584 result == KERN_PROTECTION_FAILURE ||
585 result == KERN_OUT_OF_BOUNDS,
586 ("Unexpected Mach error %d from vm_fault()", result));
588 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
591 if (result != KERN_SUCCESS && signo != NULL) {
594 case KERN_INVALID_ADDRESS:
596 *ucode = SEGV_MAPERR;
598 case KERN_RESOURCE_SHORTAGE:
602 case KERN_OUT_OF_BOUNDS:
606 case KERN_PROTECTION_FAILURE:
607 if (prot_fault_translation == 0) {
609 * Autodetect. This check also covers
610 * the images without the ABI-tag ELF
613 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
614 curproc->p_osrel >= P_OSREL_SIGSEGV) {
616 *ucode = SEGV_ACCERR;
619 *ucode = UCODE_PAGEFLT;
621 } else if (prot_fault_translation == 1) {
622 /* Always compat mode. */
624 *ucode = UCODE_PAGEFLT;
626 /* Always SIGSEGV mode. */
628 *ucode = SEGV_ACCERR;
632 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
641 vm_fault_lock_vnode(struct faultstate *fs)
646 if (fs->object->type != OBJT_VNODE)
647 return (KERN_SUCCESS);
648 vp = fs->object->handle;
650 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
651 return (KERN_SUCCESS);
655 * Perform an unlock in case the desired vnode changed while
656 * the map was unlocked during a retry.
660 locked = VOP_ISLOCKED(vp);
661 if (locked != LK_EXCLUSIVE)
665 * We must not sleep acquiring the vnode lock while we have
666 * the page exclusive busied or the object's
667 * paging-in-progress count incremented. Otherwise, we could
670 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT, curthread);
673 return (KERN_SUCCESS);
678 unlock_and_deallocate(fs);
679 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread);
682 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
683 return (KERN_RESOURCE_SHORTAGE);
687 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
688 int fault_flags, vm_page_t *m_hold)
690 struct faultstate fs;
691 struct domainset *dset;
692 vm_object_t next_object, retry_object;
693 vm_offset_t e_end, e_start;
694 vm_pindex_t retry_pindex;
695 vm_prot_t prot, retry_prot;
696 int ahead, alloc_req, behind, cluster_offset, era, faultcount;
697 int nera, oom, result, rv;
699 boolean_t wired; /* Passed by reference. */
700 bool dead, hardfault, is_first_object_locked;
702 VM_CNT_INC(v_vm_faults);
704 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
705 return (KERN_PROTECTION_FAILURE);
717 * Find the backing store object and offset into it to begin the
721 result = vm_map_lookup(&fs.map, vaddr, fault_type |
722 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
723 &fs.first_pindex, &prot, &wired);
724 if (result != KERN_SUCCESS) {
729 fs.map_generation = fs.map->timestamp;
731 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
732 panic("%s: fault on nofault entry, addr: %#lx",
733 __func__, (u_long)vaddr);
736 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
737 fs.entry->wiring_thread != curthread) {
738 vm_map_unlock_read(fs.map);
740 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
741 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
743 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
744 vm_map_unlock_and_wait(fs.map, 0);
746 vm_map_unlock(fs.map);
750 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
753 fault_type = prot | (fault_type & VM_PROT_COPY);
755 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
756 ("!wired && VM_FAULT_WIRE"));
759 * Try to avoid lock contention on the top-level object through
760 * special-case handling of some types of page faults, specifically,
761 * those that are mapping an existing page from the top-level object.
762 * Under this condition, a read lock on the object suffices, allowing
763 * multiple page faults of a similar type to run in parallel.
765 if (fs.vp == NULL /* avoid locked vnode leak */ &&
766 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
767 VM_OBJECT_RLOCK(fs.first_object);
768 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
769 fault_flags, wired, m_hold);
770 if (rv == KERN_SUCCESS)
772 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
773 VM_OBJECT_RUNLOCK(fs.first_object);
774 VM_OBJECT_WLOCK(fs.first_object);
777 VM_OBJECT_WLOCK(fs.first_object);
781 * Make a reference to this object to prevent its disposal while we
782 * are messing with it. Once we have the reference, the map is free
783 * to be diddled. Since objects reference their shadows (and copies),
784 * they will stay around as well.
786 * Bump the paging-in-progress count to prevent size changes (e.g.
787 * truncation operations) during I/O.
789 vm_object_reference_locked(fs.first_object);
790 vm_object_pip_add(fs.first_object, 1);
792 fs.lookup_still_valid = true;
797 * Search for the page at object/offset.
799 fs.object = fs.first_object;
800 fs.pindex = fs.first_pindex;
803 * If the object is marked for imminent termination,
804 * we retry here, since the collapse pass has raced
805 * with us. Otherwise, if we see terminally dead
806 * object, return fail.
808 if ((fs.object->flags & OBJ_DEAD) != 0) {
809 dead = fs.object->type == OBJT_DEAD;
810 unlock_and_deallocate(&fs);
812 return (KERN_PROTECTION_FAILURE);
818 * See if page is resident
820 fs.m = vm_page_lookup(fs.object, fs.pindex);
823 * Wait/Retry if the page is busy. We have to do this
824 * if the page is either exclusive or shared busy
825 * because the vm_pager may be using read busy for
826 * pageouts (and even pageins if it is the vnode
827 * pager), and we could end up trying to pagein and
828 * pageout the same page simultaneously.
830 * We can theoretically allow the busy case on a read
831 * fault if the page is marked valid, but since such
832 * pages are typically already pmap'd, putting that
833 * special case in might be more effort then it is
834 * worth. We cannot under any circumstances mess
835 * around with a shared busied page except, perhaps,
838 if (vm_page_tryxbusy(fs.m) == 0) {
840 * Reference the page before unlocking and
841 * sleeping so that the page daemon is less
842 * likely to reclaim it.
844 vm_page_aflag_set(fs.m, PGA_REFERENCED);
845 if (fs.object != fs.first_object) {
846 if (!VM_OBJECT_TRYWLOCK(
848 VM_OBJECT_WUNLOCK(fs.object);
849 VM_OBJECT_WLOCK(fs.first_object);
850 VM_OBJECT_WLOCK(fs.object);
852 vm_page_free(fs.first_m);
853 vm_object_pip_wakeup(fs.first_object);
854 VM_OBJECT_WUNLOCK(fs.first_object);
858 if (fs.m == vm_page_lookup(fs.object,
860 vm_page_sleep_if_busy(fs.m, "vmpfw");
862 vm_object_pip_wakeup(fs.object);
863 VM_OBJECT_WUNLOCK(fs.object);
864 VM_CNT_INC(v_intrans);
865 vm_object_deallocate(fs.first_object);
870 * The page is marked busy for other processes and the
871 * pagedaemon. If it still isn't completely valid
872 * (readable), jump to readrest, else break-out ( we
875 if (!vm_page_all_valid(fs.m))
877 break; /* break to PAGE HAS BEEN FOUND */
879 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
882 * Page is not resident. If the pager might contain the page
883 * or this is the beginning of the search, allocate a new
884 * page. (Default objects are zero-fill, so there is no real
887 if (fs.object->type != OBJT_DEFAULT ||
888 fs.object == fs.first_object) {
889 if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) {
890 rv = vm_fault_lock_vnode(&fs);
891 MPASS(rv == KERN_SUCCESS ||
892 rv == KERN_RESOURCE_SHORTAGE);
893 if (rv == KERN_RESOURCE_SHORTAGE)
896 if (fs.pindex >= fs.object->size) {
897 unlock_and_deallocate(&fs);
898 return (KERN_OUT_OF_BOUNDS);
901 if (fs.object == fs.first_object &&
902 (fs.first_object->flags & OBJ_POPULATE) != 0 &&
903 fs.first_object->shadow_count == 0) {
904 rv = vm_fault_populate(&fs, prot, fault_type,
905 fault_flags, wired, m_hold);
909 unlock_and_deallocate(&fs);
911 case KERN_RESOURCE_SHORTAGE:
912 unlock_and_deallocate(&fs);
914 case KERN_NOT_RECEIVER:
916 * Pager's populate() method
917 * returned VM_PAGER_BAD.
921 panic("inconsistent return codes");
926 * Allocate a new page for this object/offset pair.
928 * Unlocked read of the p_flag is harmless. At
929 * worst, the P_KILLED might be not observed
930 * there, and allocation can fail, causing
931 * restart and new reading of the p_flag.
933 dset = fs.object->domain.dr_policy;
935 dset = curthread->td_domain.dr_policy;
936 if (!vm_page_count_severe_set(&dset->ds_mask) ||
938 #if VM_NRESERVLEVEL > 0
939 vm_object_color(fs.object, atop(vaddr) -
942 alloc_req = P_KILLED(curproc) ?
943 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
944 if (fs.object->type != OBJT_VNODE &&
945 fs.object->backing_object == NULL)
946 alloc_req |= VM_ALLOC_ZERO;
947 fs.m = vm_page_alloc(fs.object, fs.pindex,
951 unlock_and_deallocate(&fs);
952 if (vm_pfault_oom_attempts < 0 ||
953 oom < vm_pfault_oom_attempts) {
956 vm_pfault_oom_wait * hz);
961 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
962 curproc->p_pid, curproc->p_comm);
963 vm_pageout_oom(VM_OOM_MEM_PF);
970 * At this point, we have either allocated a new page or found
971 * an existing page that is only partially valid.
973 * We hold a reference on the current object and the page is
978 * If the pager for the current object might have the page,
979 * then determine the number of additional pages to read and
980 * potentially reprioritize previously read pages for earlier
981 * reclamation. These operations should only be performed
982 * once per page fault. Even if the current pager doesn't
983 * have the page, the number of additional pages to read will
984 * apply to subsequent objects in the shadow chain.
986 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
987 !P_KILLED(curproc)) {
988 KASSERT(fs.lookup_still_valid, ("map unlocked"));
989 era = fs.entry->read_ahead;
990 behavior = vm_map_entry_behavior(fs.entry);
991 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
993 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
994 nera = VM_FAULT_READ_AHEAD_MAX;
995 if (vaddr == fs.entry->next_read)
996 vm_fault_dontneed(&fs, vaddr, nera);
997 } else if (vaddr == fs.entry->next_read) {
999 * This is a sequential fault. Arithmetically
1000 * increase the requested number of pages in
1001 * the read-ahead window. The requested
1002 * number of pages is "# of sequential faults
1003 * x (read ahead min + 1) + read ahead min"
1005 nera = VM_FAULT_READ_AHEAD_MIN;
1008 if (nera > VM_FAULT_READ_AHEAD_MAX)
1009 nera = VM_FAULT_READ_AHEAD_MAX;
1011 if (era == VM_FAULT_READ_AHEAD_MAX)
1012 vm_fault_dontneed(&fs, vaddr, nera);
1015 * This is a non-sequential fault.
1021 * A read lock on the map suffices to update
1022 * the read ahead count safely.
1024 fs.entry->read_ahead = nera;
1028 * Prepare for unlocking the map. Save the map
1029 * entry's start and end addresses, which are used to
1030 * optimize the size of the pager operation below.
1031 * Even if the map entry's addresses change after
1032 * unlocking the map, using the saved addresses is
1035 e_start = fs.entry->start;
1036 e_end = fs.entry->end;
1040 * Call the pager to retrieve the page if there is a chance
1041 * that the pager has it, and potentially retrieve additional
1042 * pages at the same time.
1044 if (fs.object->type != OBJT_DEFAULT) {
1046 * Release the map lock before locking the vnode or
1047 * sleeping in the pager. (If the current object has
1048 * a shadow, then an earlier iteration of this loop
1049 * may have already unlocked the map.)
1053 rv = vm_fault_lock_vnode(&fs);
1054 MPASS(rv == KERN_SUCCESS ||
1055 rv == KERN_RESOURCE_SHORTAGE);
1056 if (rv == KERN_RESOURCE_SHORTAGE)
1058 KASSERT(fs.vp == NULL || !fs.map->system_map,
1059 ("vm_fault: vnode-backed object mapped by system map"));
1062 * Page in the requested page and hint the pager,
1063 * that it may bring up surrounding pages.
1065 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1066 P_KILLED(curproc)) {
1070 /* Is this a sequential fault? */
1076 * Request a cluster of pages that is
1077 * aligned to a VM_FAULT_READ_DEFAULT
1078 * page offset boundary within the
1079 * object. Alignment to a page offset
1080 * boundary is more likely to coincide
1081 * with the underlying file system
1082 * block than alignment to a virtual
1085 cluster_offset = fs.pindex %
1086 VM_FAULT_READ_DEFAULT;
1087 behind = ulmin(cluster_offset,
1088 atop(vaddr - e_start));
1089 ahead = VM_FAULT_READ_DEFAULT - 1 -
1092 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1094 rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1096 if (rv == VM_PAGER_OK) {
1097 faultcount = behind + 1 + ahead;
1099 break; /* break to PAGE HAS BEEN FOUND */
1101 if (rv == VM_PAGER_ERROR)
1102 printf("vm_fault: pager read error, pid %d (%s)\n",
1103 curproc->p_pid, curproc->p_comm);
1106 * If an I/O error occurred or the requested page was
1107 * outside the range of the pager, clean up and return
1110 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1111 if (!vm_page_wired(fs.m))
1114 vm_page_xunbusy(fs.m);
1116 unlock_and_deallocate(&fs);
1117 return (KERN_OUT_OF_BOUNDS);
1121 * The requested page does not exist at this object/
1122 * offset. Remove the invalid page from the object,
1123 * waking up anyone waiting for it, and continue on to
1124 * the next object. However, if this is the top-level
1125 * object, we must leave the busy page in place to
1126 * prevent another process from rushing past us, and
1127 * inserting the page in that object at the same time
1130 if (fs.object != fs.first_object) {
1131 if (!vm_page_wired(fs.m))
1134 vm_page_xunbusy(fs.m);
1140 * We get here if the object has default pager (or unwiring)
1141 * or the pager doesn't have the page.
1143 if (fs.object == fs.first_object)
1147 * Move on to the next object. Lock the next object before
1148 * unlocking the current one.
1150 next_object = fs.object->backing_object;
1151 if (next_object == NULL) {
1153 * If there's no object left, fill the page in the top
1154 * object with zeros.
1156 if (fs.object != fs.first_object) {
1157 vm_object_pip_wakeup(fs.object);
1158 VM_OBJECT_WUNLOCK(fs.object);
1160 fs.object = fs.first_object;
1161 fs.pindex = fs.first_pindex;
1163 VM_OBJECT_WLOCK(fs.object);
1168 * Zero the page if necessary and mark it valid.
1170 if ((fs.m->flags & PG_ZERO) == 0) {
1171 pmap_zero_page(fs.m);
1173 VM_CNT_INC(v_ozfod);
1176 vm_page_valid(fs.m);
1177 /* Don't try to prefault neighboring pages. */
1179 break; /* break to PAGE HAS BEEN FOUND */
1181 KASSERT(fs.object != next_object,
1182 ("object loop %p", next_object));
1183 VM_OBJECT_WLOCK(next_object);
1184 vm_object_pip_add(next_object, 1);
1185 if (fs.object != fs.first_object)
1186 vm_object_pip_wakeup(fs.object);
1188 OFF_TO_IDX(fs.object->backing_object_offset);
1189 VM_OBJECT_WUNLOCK(fs.object);
1190 fs.object = next_object;
1194 vm_page_assert_xbusied(fs.m);
1197 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1202 * If the page is being written, but isn't already owned by the
1203 * top-level object, we have to copy it into a new page owned by the
1206 if (fs.object != fs.first_object) {
1208 * We only really need to copy if we want to write it.
1210 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1212 * This allows pages to be virtually copied from a
1213 * backing_object into the first_object, where the
1214 * backing object has no other refs to it, and cannot
1215 * gain any more refs. Instead of a bcopy, we just
1216 * move the page from the backing object to the
1217 * first object. Note that we must mark the page
1218 * dirty in the first object so that it will go out
1219 * to swap when needed.
1221 is_first_object_locked = false;
1224 * Only one shadow object
1226 (fs.object->shadow_count == 1) &&
1228 * No COW refs, except us
1230 (fs.object->ref_count == 1) &&
1232 * No one else can look this object up
1234 (fs.object->handle == NULL) &&
1236 * No other ways to look the object up
1238 ((fs.object->type == OBJT_DEFAULT) ||
1239 (fs.object->type == OBJT_SWAP)) &&
1240 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1242 * We don't chase down the shadow chain
1244 fs.object == fs.first_object->backing_object) {
1246 (void)vm_page_remove(fs.m);
1247 vm_page_replace_checked(fs.m, fs.first_object,
1248 fs.first_pindex, fs.first_m);
1249 vm_page_free(fs.first_m);
1250 vm_page_dirty(fs.m);
1251 #if VM_NRESERVLEVEL > 0
1253 * Rename the reservation.
1255 vm_reserv_rename(fs.m, fs.first_object,
1256 fs.object, OFF_TO_IDX(
1257 fs.first_object->backing_object_offset));
1259 VM_OBJECT_WUNLOCK(fs.object);
1262 VM_CNT_INC(v_cow_optim);
1264 VM_OBJECT_WUNLOCK(fs.object);
1266 * Oh, well, lets copy it.
1268 pmap_copy_page(fs.m, fs.first_m);
1269 vm_page_valid(fs.first_m);
1270 if (wired && (fault_flags &
1271 VM_FAULT_WIRE) == 0) {
1272 vm_page_wire(fs.first_m);
1273 vm_page_unwire(fs.m, PQ_INACTIVE);
1276 * We no longer need the old page or object.
1281 * fs.object != fs.first_object due to above
1284 vm_object_pip_wakeup(fs.object);
1287 * We only try to prefault read-only mappings to the
1288 * neighboring pages when this copy-on-write fault is
1289 * a hard fault. In other cases, trying to prefault
1290 * is typically wasted effort.
1292 if (faultcount == 0)
1296 * Only use the new page below...
1298 fs.object = fs.first_object;
1299 fs.pindex = fs.first_pindex;
1301 if (!is_first_object_locked)
1302 VM_OBJECT_WLOCK(fs.object);
1303 VM_CNT_INC(v_cow_faults);
1304 curthread->td_cow++;
1306 prot &= ~VM_PROT_WRITE;
1311 * We must verify that the maps have not changed since our last
1314 if (!fs.lookup_still_valid) {
1315 if (!vm_map_trylock_read(fs.map)) {
1317 unlock_and_deallocate(&fs);
1320 fs.lookup_still_valid = true;
1321 if (fs.map->timestamp != fs.map_generation) {
1322 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1323 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1326 * If we don't need the page any longer, put it on the inactive
1327 * list (the easiest thing to do here). If no one needs it,
1328 * pageout will grab it eventually.
1330 if (result != KERN_SUCCESS) {
1332 unlock_and_deallocate(&fs);
1335 * If retry of map lookup would have blocked then
1336 * retry fault from start.
1338 if (result == KERN_FAILURE)
1342 if ((retry_object != fs.first_object) ||
1343 (retry_pindex != fs.first_pindex)) {
1345 unlock_and_deallocate(&fs);
1350 * Check whether the protection has changed or the object has
1351 * been copied while we left the map unlocked. Changing from
1352 * read to write permission is OK - we leave the page
1353 * write-protected, and catch the write fault. Changing from
1354 * write to read permission means that we can't mark the page
1355 * write-enabled after all.
1358 fault_type &= retry_prot;
1361 unlock_and_deallocate(&fs);
1365 /* Reassert because wired may have changed. */
1366 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1367 ("!wired && VM_FAULT_WIRE"));
1372 * If the page was filled by a pager, save the virtual address that
1373 * should be faulted on next under a sequential access pattern to the
1374 * map entry. A read lock on the map suffices to update this address
1378 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1380 vm_page_assert_xbusied(fs.m);
1381 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1384 * Page must be completely valid or it is not fit to
1385 * map into user space. vm_pager_get_pages() ensures this.
1387 KASSERT(vm_page_all_valid(fs.m),
1388 ("vm_fault: page %p partially invalid", fs.m));
1389 VM_OBJECT_WUNLOCK(fs.object);
1392 * Put this page into the physical map. We had to do the unlock above
1393 * because pmap_enter() may sleep. We don't put the page
1394 * back on the active queue until later so that the pageout daemon
1395 * won't find it (yet).
1397 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1398 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1399 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1401 vm_fault_prefault(&fs, vaddr,
1402 faultcount > 0 ? behind : PFBAK,
1403 faultcount > 0 ? ahead : PFFOR, false);
1406 * If the page is not wired down, then put it where the pageout daemon
1409 if ((fault_flags & VM_FAULT_WIRE) != 0) {
1413 vm_page_activate(fs.m);
1414 vm_page_unlock(fs.m);
1416 if (m_hold != NULL) {
1420 vm_page_xunbusy(fs.m);
1423 * Unlock everything, and return
1425 fault_deallocate(&fs);
1427 VM_CNT_INC(v_io_faults);
1428 curthread->td_ru.ru_majflt++;
1430 if (racct_enable && fs.object->type == OBJT_VNODE) {
1432 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1433 racct_add_force(curproc, RACCT_WRITEBPS,
1434 PAGE_SIZE + behind * PAGE_SIZE);
1435 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1437 racct_add_force(curproc, RACCT_READBPS,
1438 PAGE_SIZE + ahead * PAGE_SIZE);
1439 racct_add_force(curproc, RACCT_READIOPS, 1);
1441 PROC_UNLOCK(curproc);
1445 curthread->td_ru.ru_minflt++;
1447 return (KERN_SUCCESS);
1451 * Speed up the reclamation of pages that precede the faulting pindex within
1452 * the first object of the shadow chain. Essentially, perform the equivalent
1453 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1454 * the faulting pindex by the cluster size when the pages read by vm_fault()
1455 * cross a cluster-size boundary. The cluster size is the greater of the
1456 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1458 * When "fs->first_object" is a shadow object, the pages in the backing object
1459 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1460 * function must only be concerned with pages in the first object.
1463 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1465 vm_map_entry_t entry;
1466 vm_object_t first_object, object;
1467 vm_offset_t end, start;
1468 vm_page_t m, m_next;
1469 vm_pindex_t pend, pstart;
1472 object = fs->object;
1473 VM_OBJECT_ASSERT_WLOCKED(object);
1474 first_object = fs->first_object;
1475 if (first_object != object) {
1476 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1477 VM_OBJECT_WUNLOCK(object);
1478 VM_OBJECT_WLOCK(first_object);
1479 VM_OBJECT_WLOCK(object);
1482 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1483 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1484 size = VM_FAULT_DONTNEED_MIN;
1485 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1486 size = pagesizes[1];
1487 end = rounddown2(vaddr, size);
1488 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1489 (entry = fs->entry)->start < end) {
1490 if (end - entry->start < size)
1491 start = entry->start;
1494 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1495 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1497 m_next = vm_page_find_least(first_object, pstart);
1498 pend = OFF_TO_IDX(entry->offset) + atop(end -
1500 while ((m = m_next) != NULL && m->pindex < pend) {
1501 m_next = TAILQ_NEXT(m, listq);
1502 if (!vm_page_all_valid(m) ||
1507 * Don't clear PGA_REFERENCED, since it would
1508 * likely represent a reference by a different
1511 * Typically, at this point, prefetched pages
1512 * are still in the inactive queue. Only
1513 * pages that triggered page faults are in the
1517 if (!vm_page_inactive(m))
1518 vm_page_deactivate(m);
1523 if (first_object != object)
1524 VM_OBJECT_WUNLOCK(first_object);
1528 * vm_fault_prefault provides a quick way of clustering
1529 * pagefaults into a processes address space. It is a "cousin"
1530 * of vm_map_pmap_enter, except it runs at page fault time instead
1534 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1535 int backward, int forward, bool obj_locked)
1538 vm_map_entry_t entry;
1539 vm_object_t backing_object, lobject;
1540 vm_offset_t addr, starta;
1545 pmap = fs->map->pmap;
1546 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1551 if (addra < backward * PAGE_SIZE) {
1552 starta = entry->start;
1554 starta = addra - backward * PAGE_SIZE;
1555 if (starta < entry->start)
1556 starta = entry->start;
1560 * Generate the sequence of virtual addresses that are candidates for
1561 * prefaulting in an outward spiral from the faulting virtual address,
1562 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1563 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1564 * If the candidate address doesn't have a backing physical page, then
1565 * the loop immediately terminates.
1567 for (i = 0; i < 2 * imax(backward, forward); i++) {
1568 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1570 if (addr > addra + forward * PAGE_SIZE)
1573 if (addr < starta || addr >= entry->end)
1576 if (!pmap_is_prefaultable(pmap, addr))
1579 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1580 lobject = entry->object.vm_object;
1582 VM_OBJECT_RLOCK(lobject);
1583 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1584 lobject->type == OBJT_DEFAULT &&
1585 (backing_object = lobject->backing_object) != NULL) {
1586 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1587 0, ("vm_fault_prefault: unaligned object offset"));
1588 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1589 VM_OBJECT_RLOCK(backing_object);
1590 if (!obj_locked || lobject != entry->object.vm_object)
1591 VM_OBJECT_RUNLOCK(lobject);
1592 lobject = backing_object;
1595 if (!obj_locked || lobject != entry->object.vm_object)
1596 VM_OBJECT_RUNLOCK(lobject);
1599 if (vm_page_all_valid(m) &&
1600 (m->flags & PG_FICTITIOUS) == 0)
1601 pmap_enter_quick(pmap, addr, m, entry->protection);
1602 if (!obj_locked || lobject != entry->object.vm_object)
1603 VM_OBJECT_RUNLOCK(lobject);
1608 * Hold each of the physical pages that are mapped by the specified range of
1609 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1610 * and allow the specified types of access, "prot". If all of the implied
1611 * pages are successfully held, then the number of held pages is returned
1612 * together with pointers to those pages in the array "ma". However, if any
1613 * of the pages cannot be held, -1 is returned.
1616 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1617 vm_prot_t prot, vm_page_t *ma, int max_count)
1619 vm_offset_t end, va;
1622 boolean_t pmap_failed;
1626 end = round_page(addr + len);
1627 addr = trunc_page(addr);
1630 * Check for illegal addresses.
1632 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map))
1635 if (atop(end - addr) > max_count)
1636 panic("vm_fault_quick_hold_pages: count > max_count");
1637 count = atop(end - addr);
1640 * Most likely, the physical pages are resident in the pmap, so it is
1641 * faster to try pmap_extract_and_hold() first.
1643 pmap_failed = FALSE;
1644 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1645 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1648 else if ((prot & VM_PROT_WRITE) != 0 &&
1649 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1651 * Explicitly dirty the physical page. Otherwise, the
1652 * caller's changes may go unnoticed because they are
1653 * performed through an unmanaged mapping or by a DMA
1656 * The object lock is not held here.
1657 * See vm_page_clear_dirty_mask().
1664 * One or more pages could not be held by the pmap. Either no
1665 * page was mapped at the specified virtual address or that
1666 * mapping had insufficient permissions. Attempt to fault in
1667 * and hold these pages.
1669 * If vm_fault_disable_pagefaults() was called,
1670 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1671 * acquire MD VM locks, which means we must not call
1672 * vm_fault(). Some (out of tree) callers mark
1673 * too wide a code area with vm_fault_disable_pagefaults()
1674 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1675 * the proper behaviour explicitly.
1677 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1678 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1680 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1681 if (*mp == NULL && vm_fault(map, va, prot,
1682 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1687 for (mp = ma; mp < ma + count; mp++)
1689 vm_page_unwire(*mp, PQ_INACTIVE);
1695 * vm_fault_copy_entry
1697 * Create new shadow object backing dst_entry with private copy of
1698 * all underlying pages. When src_entry is equal to dst_entry,
1699 * function implements COW for wired-down map entry. Otherwise,
1700 * it forks wired entry into dst_map.
1702 * In/out conditions:
1703 * The source and destination maps must be locked for write.
1704 * The source map entry must be wired down (or be a sharing map
1705 * entry corresponding to a main map entry that is wired down).
1708 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1709 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1710 vm_ooffset_t *fork_charge)
1712 vm_object_t backing_object, dst_object, object, src_object;
1713 vm_pindex_t dst_pindex, pindex, src_pindex;
1714 vm_prot_t access, prot;
1724 upgrade = src_entry == dst_entry;
1725 access = prot = dst_entry->protection;
1727 src_object = src_entry->object.vm_object;
1728 src_pindex = OFF_TO_IDX(src_entry->offset);
1730 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1731 dst_object = src_object;
1732 vm_object_reference(dst_object);
1735 * Create the top-level object for the destination entry. (Doesn't
1736 * actually shadow anything - we copy the pages directly.)
1738 dst_object = vm_object_allocate(OBJT_DEFAULT,
1739 atop(dst_entry->end - dst_entry->start));
1740 #if VM_NRESERVLEVEL > 0
1741 dst_object->flags |= OBJ_COLORED;
1742 dst_object->pg_color = atop(dst_entry->start);
1744 dst_object->domain = src_object->domain;
1745 dst_object->charge = dst_entry->end - dst_entry->start;
1748 VM_OBJECT_WLOCK(dst_object);
1749 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1750 ("vm_fault_copy_entry: vm_object not NULL"));
1751 if (src_object != dst_object) {
1752 dst_entry->object.vm_object = dst_object;
1753 dst_entry->offset = 0;
1754 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1756 if (fork_charge != NULL) {
1757 KASSERT(dst_entry->cred == NULL,
1758 ("vm_fault_copy_entry: leaked swp charge"));
1759 dst_object->cred = curthread->td_ucred;
1760 crhold(dst_object->cred);
1761 *fork_charge += dst_object->charge;
1762 } else if ((dst_object->type == OBJT_DEFAULT ||
1763 dst_object->type == OBJT_SWAP) &&
1764 dst_object->cred == NULL) {
1765 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1767 dst_object->cred = dst_entry->cred;
1768 dst_entry->cred = NULL;
1772 * If not an upgrade, then enter the mappings in the pmap as
1773 * read and/or execute accesses. Otherwise, enter them as
1776 * A writeable large page mapping is only created if all of
1777 * the constituent small page mappings are modified. Marking
1778 * PTEs as modified on inception allows promotion to happen
1779 * without taking potentially large number of soft faults.
1782 access &= ~VM_PROT_WRITE;
1785 * Loop through all of the virtual pages within the entry's
1786 * range, copying each page from the source object to the
1787 * destination object. Since the source is wired, those pages
1788 * must exist. In contrast, the destination is pageable.
1789 * Since the destination object doesn't share any backing storage
1790 * with the source object, all of its pages must be dirtied,
1791 * regardless of whether they can be written.
1793 for (vaddr = dst_entry->start, dst_pindex = 0;
1794 vaddr < dst_entry->end;
1795 vaddr += PAGE_SIZE, dst_pindex++) {
1798 * Find the page in the source object, and copy it in.
1799 * Because the source is wired down, the page will be
1802 if (src_object != dst_object)
1803 VM_OBJECT_RLOCK(src_object);
1804 object = src_object;
1805 pindex = src_pindex + dst_pindex;
1806 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1807 (backing_object = object->backing_object) != NULL) {
1809 * Unless the source mapping is read-only or
1810 * it is presently being upgraded from
1811 * read-only, the first object in the shadow
1812 * chain should provide all of the pages. In
1813 * other words, this loop body should never be
1814 * executed when the source mapping is already
1817 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1819 ("vm_fault_copy_entry: main object missing page"));
1821 VM_OBJECT_RLOCK(backing_object);
1822 pindex += OFF_TO_IDX(object->backing_object_offset);
1823 if (object != dst_object)
1824 VM_OBJECT_RUNLOCK(object);
1825 object = backing_object;
1827 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1829 if (object != dst_object) {
1831 * Allocate a page in the destination object.
1833 dst_m = vm_page_alloc(dst_object, (src_object ==
1834 dst_object ? src_pindex : 0) + dst_pindex,
1836 if (dst_m == NULL) {
1837 VM_OBJECT_WUNLOCK(dst_object);
1838 VM_OBJECT_RUNLOCK(object);
1839 vm_wait(dst_object);
1840 VM_OBJECT_WLOCK(dst_object);
1843 pmap_copy_page(src_m, dst_m);
1844 VM_OBJECT_RUNLOCK(object);
1845 dst_m->dirty = dst_m->valid = src_m->valid;
1848 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1850 if (dst_m->pindex >= dst_object->size) {
1852 * We are upgrading. Index can occur
1853 * out of bounds if the object type is
1854 * vnode and the file was truncated.
1856 vm_page_xunbusy(dst_m);
1860 VM_OBJECT_WUNLOCK(dst_object);
1863 * Enter it in the pmap. If a wired, copy-on-write
1864 * mapping is being replaced by a write-enabled
1865 * mapping, then wire that new mapping.
1867 * The page can be invalid if the user called
1868 * msync(MS_INVALIDATE) or truncated the backing vnode
1869 * or shared memory object. In this case, do not
1870 * insert it into pmap, but still do the copy so that
1871 * all copies of the wired map entry have similar
1874 if (vm_page_all_valid(dst_m)) {
1875 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1876 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1880 * Mark it no longer busy, and put it on the active list.
1882 VM_OBJECT_WLOCK(dst_object);
1885 if (src_m != dst_m) {
1886 vm_page_unwire(src_m, PQ_INACTIVE);
1887 vm_page_wire(dst_m);
1889 KASSERT(vm_page_wired(dst_m),
1890 ("dst_m %p is not wired", dst_m));
1893 vm_page_lock(dst_m);
1894 vm_page_activate(dst_m);
1895 vm_page_unlock(dst_m);
1897 vm_page_xunbusy(dst_m);
1899 VM_OBJECT_WUNLOCK(dst_object);
1901 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1902 vm_object_deallocate(src_object);
1907 * Block entry into the machine-independent layer's page fault handler by
1908 * the calling thread. Subsequent calls to vm_fault() by that thread will
1909 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1910 * spurious page faults.
1913 vm_fault_disable_pagefaults(void)
1916 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1920 vm_fault_enable_pagefaults(int save)
1923 curthread_pflags_restore(save);