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>
88 #include <sys/racct.h>
89 #include <sys/resourcevar.h>
90 #include <sys/rwlock.h>
91 #include <sys/signalvar.h>
92 #include <sys/sysctl.h>
93 #include <sys/sysent.h>
94 #include <sys/vmmeter.h>
95 #include <sys/vnode.h>
97 #include <sys/ktrace.h>
101 #include <vm/vm_param.h>
103 #include <vm/vm_map.h>
104 #include <vm/vm_object.h>
105 #include <vm/vm_page.h>
106 #include <vm/vm_pageout.h>
107 #include <vm/vm_kern.h>
108 #include <vm/vm_pager.h>
109 #include <vm/vm_extern.h>
110 #include <vm/vm_reserv.h>
115 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
116 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX)
118 #define VM_FAULT_DONTNEED_MIN 1048576
125 vm_object_t first_object;
126 vm_pindex_t first_pindex;
128 vm_map_entry_t entry;
130 bool lookup_still_valid;
134 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
136 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
137 int backward, int forward, bool obj_locked);
139 static int vm_pfault_oom_attempts = 3;
140 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
141 &vm_pfault_oom_attempts, 0,
142 "Number of page allocation attempts in page fault handler before it "
143 "triggers OOM handling");
145 static int vm_pfault_oom_wait = 10;
146 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
147 &vm_pfault_oom_wait, 0,
148 "Number of seconds to wait for free pages before retrying "
149 "the page fault handler");
152 release_page(struct faultstate *fs)
156 vm_page_xunbusy(fs->m);
158 vm_page_deactivate(fs->m);
159 vm_page_unlock(fs->m);
165 unlock_map(struct faultstate *fs)
168 if (fs->lookup_still_valid) {
169 vm_map_lookup_done(fs->map, fs->entry);
170 fs->lookup_still_valid = false;
175 unlock_vp(struct faultstate *fs)
178 if (fs->vp != NULL) {
185 unlock_and_deallocate(struct faultstate *fs)
188 vm_object_pip_wakeup(fs->object);
189 VM_OBJECT_WUNLOCK(fs->object);
190 if (fs->object != fs->first_object) {
191 VM_OBJECT_WLOCK(fs->first_object);
192 vm_page_lock(fs->first_m);
193 vm_page_free(fs->first_m);
194 vm_page_unlock(fs->first_m);
195 vm_object_pip_wakeup(fs->first_object);
196 VM_OBJECT_WUNLOCK(fs->first_object);
199 vm_object_deallocate(fs->first_object);
205 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot,
206 vm_prot_t fault_type, int fault_flags, bool set_wd)
210 if (((prot & VM_PROT_WRITE) == 0 &&
211 (fault_flags & VM_FAULT_DIRTY) == 0) ||
212 (m->oflags & VPO_UNMANAGED) != 0)
215 VM_OBJECT_ASSERT_LOCKED(m->object);
217 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 &&
218 (fault_flags & VM_FAULT_WIRE) == 0) ||
219 (fault_flags & VM_FAULT_DIRTY) != 0;
222 vm_object_set_writeable_dirty(m->object);
225 * If two callers of vm_fault_dirty() with set_wd ==
226 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC
227 * flag set, other with flag clear, race, it is
228 * possible for the no-NOSYNC thread to see m->dirty
229 * != 0 and not clear VPO_NOSYNC. Take vm_page lock
230 * around manipulation of VPO_NOSYNC and
231 * vm_page_dirty() call, to avoid the race and keep
232 * m->oflags consistent.
237 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
238 * if the page is already dirty to prevent data written with
239 * the expectation of being synced from not being synced.
240 * Likewise if this entry does not request NOSYNC then make
241 * sure the page isn't marked NOSYNC. Applications sharing
242 * data should use the same flags to avoid ping ponging.
244 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) {
246 m->oflags |= VPO_NOSYNC;
249 m->oflags &= ~VPO_NOSYNC;
253 * If the fault is a write, we know that this page is being
254 * written NOW so dirty it explicitly to save on
255 * pmap_is_modified() calls later.
257 * Also, since the page is now dirty, we can possibly tell
258 * the pager to release any swap backing the page. Calling
259 * the pager requires a write lock on the object.
266 vm_pager_page_unswapped(m);
270 vm_fault_fill_hold(vm_page_t *m_hold, vm_page_t m)
273 if (m_hold != NULL) {
282 * Unlocks fs.first_object and fs.map on success.
285 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot,
286 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold)
289 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
290 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
297 MPASS(fs->vp == NULL);
298 m = vm_page_lookup(fs->first_object, fs->first_pindex);
299 /* A busy page can be mapped for read|execute access. */
300 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 &&
301 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL)
302 return (KERN_FAILURE);
305 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
306 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \
308 if ((m->flags & PG_FICTITIOUS) == 0 &&
309 (m_super = vm_reserv_to_superpage(m)) != NULL &&
310 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
311 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
312 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
313 (pagesizes[m_super->psind] - 1)) &&
314 pmap_ps_enabled(fs->map->pmap)) {
315 flags = PS_ALL_VALID;
316 if ((prot & VM_PROT_WRITE) != 0) {
318 * Create a superpage mapping allowing write access
319 * only if none of the constituent pages are busy and
320 * all of them are already dirty (except possibly for
321 * the page that was faulted on).
323 flags |= PS_NONE_BUSY;
324 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
325 flags |= PS_ALL_DIRTY;
327 if (vm_page_ps_test(m_super, flags, m)) {
329 psind = m_super->psind;
330 vaddr = rounddown2(vaddr, pagesizes[psind]);
331 /* Preset the modified bit for dirty superpages. */
332 if ((flags & PS_ALL_DIRTY) != 0)
333 fault_type |= VM_PROT_WRITE;
337 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type |
338 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind);
339 if (rv != KERN_SUCCESS)
341 vm_fault_fill_hold(m_hold, m);
342 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false);
343 if (psind == 0 && !wired)
344 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
345 VM_OBJECT_RUNLOCK(fs->first_object);
346 vm_map_lookup_done(fs->map, fs->entry);
347 curthread->td_ru.ru_minflt++;
348 return (KERN_SUCCESS);
352 vm_fault_restore_map_lock(struct faultstate *fs)
355 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
356 MPASS(fs->first_object->paging_in_progress > 0);
358 if (!vm_map_trylock_read(fs->map)) {
359 VM_OBJECT_WUNLOCK(fs->first_object);
360 vm_map_lock_read(fs->map);
361 VM_OBJECT_WLOCK(fs->first_object);
363 fs->lookup_still_valid = true;
367 vm_fault_populate_check_page(vm_page_t m)
371 * Check each page to ensure that the pager is obeying the
372 * interface: the page must be installed in the object, fully
373 * valid, and exclusively busied.
376 MPASS(m->valid == VM_PAGE_BITS_ALL);
377 MPASS(vm_page_xbusied(m));
381 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
387 VM_OBJECT_ASSERT_WLOCKED(object);
388 MPASS(first <= last);
389 for (pidx = first, m = vm_page_lookup(object, pidx);
390 pidx <= last; pidx++, m = vm_page_next(m)) {
391 vm_fault_populate_check_page(m);
393 vm_page_deactivate(m);
400 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type,
401 int fault_flags, boolean_t wired, vm_page_t *m_hold)
406 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
407 int i, npages, psind, rv;
409 MPASS(fs->object == fs->first_object);
410 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
411 MPASS(fs->first_object->paging_in_progress > 0);
412 MPASS(fs->first_object->backing_object == NULL);
413 MPASS(fs->lookup_still_valid);
415 pager_first = OFF_TO_IDX(fs->entry->offset);
416 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
421 * Call the pager (driver) populate() method.
423 * There is no guarantee that the method will be called again
424 * if the current fault is for read, and a future fault is
425 * for write. Report the entry's maximum allowed protection
428 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
429 fault_type, fs->entry->max_protection, &pager_first, &pager_last);
431 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
432 if (rv == VM_PAGER_BAD) {
434 * VM_PAGER_BAD is the backdoor for a pager to request
435 * normal fault handling.
437 vm_fault_restore_map_lock(fs);
438 if (fs->map->timestamp != fs->map_generation)
439 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
440 return (KERN_NOT_RECEIVER);
442 if (rv != VM_PAGER_OK)
443 return (KERN_FAILURE); /* AKA SIGSEGV */
445 /* Ensure that the driver is obeying the interface. */
446 MPASS(pager_first <= pager_last);
447 MPASS(fs->first_pindex <= pager_last);
448 MPASS(fs->first_pindex >= pager_first);
449 MPASS(pager_last < fs->first_object->size);
451 vm_fault_restore_map_lock(fs);
452 if (fs->map->timestamp != fs->map_generation) {
453 vm_fault_populate_cleanup(fs->first_object, pager_first,
455 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */
459 * The map is unchanged after our last unlock. Process the fault.
461 * The range [pager_first, pager_last] that is given to the
462 * pager is only a hint. The pager may populate any range
463 * within the object that includes the requested page index.
464 * In case the pager expanded the range, clip it to fit into
467 map_first = OFF_TO_IDX(fs->entry->offset);
468 if (map_first > pager_first) {
469 vm_fault_populate_cleanup(fs->first_object, pager_first,
471 pager_first = map_first;
473 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
474 if (map_last < pager_last) {
475 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
477 pager_last = map_last;
479 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
481 pidx += npages, m = vm_page_next(&m[npages - 1])) {
482 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
483 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
484 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
486 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
487 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
488 !pmap_ps_enabled(fs->map->pmap)))
493 npages = atop(pagesizes[psind]);
494 for (i = 0; i < npages; i++) {
495 vm_fault_populate_check_page(&m[i]);
496 vm_fault_dirty(fs->entry, &m[i], prot, fault_type,
499 VM_OBJECT_WUNLOCK(fs->first_object);
500 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type |
501 (wired ? PMAP_ENTER_WIRED : 0), psind);
502 #if defined(__amd64__)
503 if (psind > 0 && rv == KERN_FAILURE) {
504 for (i = 0; i < npages; i++) {
505 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
506 &m[i], prot, fault_type |
507 (wired ? PMAP_ENTER_WIRED : 0), 0);
508 MPASS(rv == KERN_SUCCESS);
512 MPASS(rv == KERN_SUCCESS);
514 VM_OBJECT_WLOCK(fs->first_object);
516 for (i = 0; i < npages; i++) {
517 vm_page_change_lock(&m[i], &m_mtx);
518 if ((fault_flags & VM_FAULT_WIRE) != 0)
521 vm_page_activate(&m[i]);
522 if (m_hold != NULL && m[i].pindex == fs->first_pindex) {
526 vm_page_xunbusy_maybelocked(&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)
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);
674 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE, curthread);
677 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
678 return (KERN_RESOURCE_SHORTAGE);
682 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
683 int fault_flags, vm_page_t *m_hold)
685 struct faultstate fs;
686 struct domainset *dset;
687 vm_object_t next_object, retry_object;
688 vm_offset_t e_end, e_start;
689 vm_pindex_t retry_pindex;
690 vm_prot_t prot, retry_prot;
691 int ahead, alloc_req, behind, cluster_offset, era, faultcount;
692 int nera, oom, result, rv;
694 boolean_t wired; /* Passed by reference. */
695 bool dead, hardfault, is_first_object_locked;
697 VM_CNT_INC(v_vm_faults);
699 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
700 return (KERN_PROTECTION_FAILURE);
712 * Find the backing store object and offset into it to begin the
716 result = vm_map_lookup(&fs.map, vaddr, fault_type |
717 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object,
718 &fs.first_pindex, &prot, &wired);
719 if (result != KERN_SUCCESS) {
724 fs.map_generation = fs.map->timestamp;
726 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
727 panic("%s: fault on nofault entry, addr: %#lx",
728 __func__, (u_long)vaddr);
731 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION &&
732 fs.entry->wiring_thread != curthread) {
733 vm_map_unlock_read(fs.map);
735 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) &&
736 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
738 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
739 vm_map_unlock_and_wait(fs.map, 0);
741 vm_map_unlock(fs.map);
745 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0);
748 fault_type = prot | (fault_type & VM_PROT_COPY);
750 KASSERT((fault_flags & VM_FAULT_WIRE) == 0,
751 ("!wired && VM_FAULT_WIRE"));
754 * Try to avoid lock contention on the top-level object through
755 * special-case handling of some types of page faults, specifically,
756 * those that are both (1) mapping an existing page from the top-
757 * level object and (2) not having to mark that object as containing
758 * dirty pages. Under these conditions, a read lock on the top-level
759 * object suffices, allowing multiple page faults of a similar type to
760 * run in parallel on the same top-level object.
762 if (fs.vp == NULL /* avoid locked vnode leak */ &&
763 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 &&
764 /* avoid calling vm_object_set_writeable_dirty() */
765 ((prot & VM_PROT_WRITE) == 0 ||
766 (fs.first_object->type != OBJT_VNODE &&
767 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
768 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) {
769 VM_OBJECT_RLOCK(fs.first_object);
770 if ((prot & VM_PROT_WRITE) == 0 ||
771 (fs.first_object->type != OBJT_VNODE &&
772 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) ||
773 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) {
774 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type,
775 fault_flags, wired, m_hold);
776 if (rv == KERN_SUCCESS)
779 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
780 VM_OBJECT_RUNLOCK(fs.first_object);
781 VM_OBJECT_WLOCK(fs.first_object);
784 VM_OBJECT_WLOCK(fs.first_object);
788 * Make a reference to this object to prevent its disposal while we
789 * are messing with it. Once we have the reference, the map is free
790 * to be diddled. Since objects reference their shadows (and copies),
791 * they will stay around as well.
793 * Bump the paging-in-progress count to prevent size changes (e.g.
794 * truncation operations) during I/O.
796 vm_object_reference_locked(fs.first_object);
797 vm_object_pip_add(fs.first_object, 1);
799 fs.lookup_still_valid = true;
804 * Search for the page at object/offset.
806 fs.object = fs.first_object;
807 fs.pindex = fs.first_pindex;
810 * If the object is marked for imminent termination,
811 * we retry here, since the collapse pass has raced
812 * with us. Otherwise, if we see terminally dead
813 * object, return fail.
815 if ((fs.object->flags & OBJ_DEAD) != 0) {
816 dead = fs.object->type == OBJT_DEAD;
817 unlock_and_deallocate(&fs);
819 return (KERN_PROTECTION_FAILURE);
825 * See if page is resident
827 fs.m = vm_page_lookup(fs.object, fs.pindex);
830 * Wait/Retry if the page is busy. We have to do this
831 * if the page is either exclusive or shared busy
832 * because the vm_pager may be using read busy for
833 * pageouts (and even pageins if it is the vnode
834 * pager), and we could end up trying to pagein and
835 * pageout the same page simultaneously.
837 * We can theoretically allow the busy case on a read
838 * fault if the page is marked valid, but since such
839 * pages are typically already pmap'd, putting that
840 * special case in might be more effort then it is
841 * worth. We cannot under any circumstances mess
842 * around with a shared busied page except, perhaps,
845 if (vm_page_busied(fs.m)) {
847 * Reference the page before unlocking and
848 * sleeping so that the page daemon is less
849 * likely to reclaim it.
851 vm_page_aflag_set(fs.m, PGA_REFERENCED);
852 if (fs.object != fs.first_object) {
853 if (!VM_OBJECT_TRYWLOCK(
855 VM_OBJECT_WUNLOCK(fs.object);
856 VM_OBJECT_WLOCK(fs.first_object);
857 VM_OBJECT_WLOCK(fs.object);
859 vm_page_lock(fs.first_m);
860 vm_page_free(fs.first_m);
861 vm_page_unlock(fs.first_m);
862 vm_object_pip_wakeup(fs.first_object);
863 VM_OBJECT_WUNLOCK(fs.first_object);
867 if (fs.m == vm_page_lookup(fs.object,
869 vm_page_sleep_if_busy(fs.m, "vmpfw");
871 vm_object_pip_wakeup(fs.object);
872 VM_OBJECT_WUNLOCK(fs.object);
873 VM_CNT_INC(v_intrans);
874 vm_object_deallocate(fs.first_object);
879 * Mark page busy for other processes, and the
880 * pagedaemon. If it still isn't completely valid
881 * (readable), jump to readrest, else break-out ( we
885 if (fs.m->valid != VM_PAGE_BITS_ALL)
887 break; /* break to PAGE HAS BEEN FOUND */
889 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m));
892 * Page is not resident. If the pager might contain the page
893 * or this is the beginning of the search, allocate a new
894 * page. (Default objects are zero-fill, so there is no real
897 if (fs.object->type != OBJT_DEFAULT ||
898 fs.object == fs.first_object) {
899 if ((fs.object->flags & OBJ_SIZEVNLOCK) != 0) {
900 rv = vm_fault_lock_vnode(&fs);
901 MPASS(rv == KERN_SUCCESS ||
902 rv == KERN_RESOURCE_SHORTAGE);
903 if (rv == KERN_RESOURCE_SHORTAGE)
906 if (fs.pindex >= fs.object->size) {
907 unlock_and_deallocate(&fs);
908 return (KERN_OUT_OF_BOUNDS);
911 if (fs.object == fs.first_object &&
912 (fs.first_object->flags & OBJ_POPULATE) != 0 &&
913 fs.first_object->shadow_count == 0) {
914 rv = vm_fault_populate(&fs, prot, fault_type,
915 fault_flags, wired, m_hold);
919 unlock_and_deallocate(&fs);
921 case KERN_RESOURCE_SHORTAGE:
922 unlock_and_deallocate(&fs);
924 case KERN_NOT_RECEIVER:
926 * Pager's populate() method
927 * returned VM_PAGER_BAD.
931 panic("inconsistent return codes");
936 * Allocate a new page for this object/offset pair.
938 * Unlocked read of the p_flag is harmless. At
939 * worst, the P_KILLED might be not observed
940 * there, and allocation can fail, causing
941 * restart and new reading of the p_flag.
943 dset = fs.object->domain.dr_policy;
945 dset = curthread->td_domain.dr_policy;
946 if (!vm_page_count_severe_set(&dset->ds_mask) ||
948 #if VM_NRESERVLEVEL > 0
949 vm_object_color(fs.object, atop(vaddr) -
952 alloc_req = P_KILLED(curproc) ?
953 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
954 if (fs.object->type != OBJT_VNODE &&
955 fs.object->backing_object == NULL)
956 alloc_req |= VM_ALLOC_ZERO;
957 fs.m = vm_page_alloc(fs.object, fs.pindex,
961 unlock_and_deallocate(&fs);
962 if (vm_pfault_oom_attempts < 0 ||
963 oom < vm_pfault_oom_attempts) {
966 vm_pfault_oom_wait * hz);
971 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
972 curproc->p_pid, curproc->p_comm);
973 vm_pageout_oom(VM_OOM_MEM_PF);
980 * At this point, we have either allocated a new page or found
981 * an existing page that is only partially valid.
983 * We hold a reference on the current object and the page is
988 * If the pager for the current object might have the page,
989 * then determine the number of additional pages to read and
990 * potentially reprioritize previously read pages for earlier
991 * reclamation. These operations should only be performed
992 * once per page fault. Even if the current pager doesn't
993 * have the page, the number of additional pages to read will
994 * apply to subsequent objects in the shadow chain.
996 if (fs.object->type != OBJT_DEFAULT && nera == -1 &&
997 !P_KILLED(curproc)) {
998 KASSERT(fs.lookup_still_valid, ("map unlocked"));
999 era = fs.entry->read_ahead;
1000 behavior = vm_map_entry_behavior(fs.entry);
1001 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
1003 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
1004 nera = VM_FAULT_READ_AHEAD_MAX;
1005 if (vaddr == fs.entry->next_read)
1006 vm_fault_dontneed(&fs, vaddr, nera);
1007 } else if (vaddr == fs.entry->next_read) {
1009 * This is a sequential fault. Arithmetically
1010 * increase the requested number of pages in
1011 * the read-ahead window. The requested
1012 * number of pages is "# of sequential faults
1013 * x (read ahead min + 1) + read ahead min"
1015 nera = VM_FAULT_READ_AHEAD_MIN;
1018 if (nera > VM_FAULT_READ_AHEAD_MAX)
1019 nera = VM_FAULT_READ_AHEAD_MAX;
1021 if (era == VM_FAULT_READ_AHEAD_MAX)
1022 vm_fault_dontneed(&fs, vaddr, nera);
1025 * This is a non-sequential fault.
1031 * A read lock on the map suffices to update
1032 * the read ahead count safely.
1034 fs.entry->read_ahead = nera;
1038 * Prepare for unlocking the map. Save the map
1039 * entry's start and end addresses, which are used to
1040 * optimize the size of the pager operation below.
1041 * Even if the map entry's addresses change after
1042 * unlocking the map, using the saved addresses is
1045 e_start = fs.entry->start;
1046 e_end = fs.entry->end;
1050 * Call the pager to retrieve the page if there is a chance
1051 * that the pager has it, and potentially retrieve additional
1052 * pages at the same time.
1054 if (fs.object->type != OBJT_DEFAULT) {
1056 * Release the map lock before locking the vnode or
1057 * sleeping in the pager. (If the current object has
1058 * a shadow, then an earlier iteration of this loop
1059 * may have already unlocked the map.)
1063 rv = vm_fault_lock_vnode(&fs);
1064 MPASS(rv == KERN_SUCCESS ||
1065 rv == KERN_RESOURCE_SHORTAGE);
1066 if (rv == KERN_RESOURCE_SHORTAGE)
1068 KASSERT(fs.vp == NULL || !fs.map->system_map,
1069 ("vm_fault: vnode-backed object mapped by system map"));
1072 * Page in the requested page and hint the pager,
1073 * that it may bring up surrounding pages.
1075 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1076 P_KILLED(curproc)) {
1080 /* Is this a sequential fault? */
1086 * Request a cluster of pages that is
1087 * aligned to a VM_FAULT_READ_DEFAULT
1088 * page offset boundary within the
1089 * object. Alignment to a page offset
1090 * boundary is more likely to coincide
1091 * with the underlying file system
1092 * block than alignment to a virtual
1095 cluster_offset = fs.pindex %
1096 VM_FAULT_READ_DEFAULT;
1097 behind = ulmin(cluster_offset,
1098 atop(vaddr - e_start));
1099 ahead = VM_FAULT_READ_DEFAULT - 1 -
1102 ahead = ulmin(ahead, atop(e_end - vaddr) - 1);
1104 rv = vm_pager_get_pages(fs.object, &fs.m, 1,
1106 if (rv == VM_PAGER_OK) {
1107 faultcount = behind + 1 + ahead;
1109 break; /* break to PAGE HAS BEEN FOUND */
1111 if (rv == VM_PAGER_ERROR)
1112 printf("vm_fault: pager read error, pid %d (%s)\n",
1113 curproc->p_pid, curproc->p_comm);
1116 * If an I/O error occurred or the requested page was
1117 * outside the range of the pager, clean up and return
1120 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1122 if (!vm_page_wired(fs.m))
1125 vm_page_xunbusy_maybelocked(fs.m);
1126 vm_page_unlock(fs.m);
1128 unlock_and_deallocate(&fs);
1129 return (KERN_OUT_OF_BOUNDS);
1133 * The requested page does not exist at this object/
1134 * offset. Remove the invalid page from the object,
1135 * waking up anyone waiting for it, and continue on to
1136 * the next object. However, if this is the top-level
1137 * object, we must leave the busy page in place to
1138 * prevent another process from rushing past us, and
1139 * inserting the page in that object at the same time
1142 if (fs.object != fs.first_object) {
1144 if (!vm_page_wired(fs.m))
1147 vm_page_xunbusy_maybelocked(fs.m);
1148 vm_page_unlock(fs.m);
1154 * We get here if the object has default pager (or unwiring)
1155 * or the pager doesn't have the page.
1157 if (fs.object == fs.first_object)
1161 * Move on to the next object. Lock the next object before
1162 * unlocking the current one.
1164 next_object = fs.object->backing_object;
1165 if (next_object == NULL) {
1167 * If there's no object left, fill the page in the top
1168 * object with zeros.
1170 if (fs.object != fs.first_object) {
1171 vm_object_pip_wakeup(fs.object);
1172 VM_OBJECT_WUNLOCK(fs.object);
1174 fs.object = fs.first_object;
1175 fs.pindex = fs.first_pindex;
1177 VM_OBJECT_WLOCK(fs.object);
1182 * Zero the page if necessary and mark it valid.
1184 if ((fs.m->flags & PG_ZERO) == 0) {
1185 pmap_zero_page(fs.m);
1187 VM_CNT_INC(v_ozfod);
1190 fs.m->valid = VM_PAGE_BITS_ALL;
1191 /* Don't try to prefault neighboring pages. */
1193 break; /* break to PAGE HAS BEEN FOUND */
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;
1208 vm_page_assert_xbusied(fs.m);
1211 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
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->type == OBJT_DEFAULT) ||
1253 (fs.object->type == OBJT_SWAP)) &&
1254 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) &&
1256 * We don't chase down the shadow chain
1258 fs.object == fs.first_object->backing_object) {
1260 vm_page_dequeue(fs.m);
1261 (void)vm_page_remove(fs.m);
1262 vm_page_unlock(fs.m);
1263 vm_page_lock(fs.first_m);
1264 vm_page_replace_checked(fs.m, fs.first_object,
1265 fs.first_pindex, fs.first_m);
1266 vm_page_free(fs.first_m);
1267 vm_page_unlock(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));
1278 * Removing the page from the backing object
1281 vm_page_xbusy(fs.m);
1284 VM_CNT_INC(v_cow_optim);
1287 * Oh, well, lets copy it.
1289 pmap_copy_page(fs.m, fs.first_m);
1290 fs.first_m->valid = VM_PAGE_BITS_ALL;
1291 if (wired && (fault_flags &
1292 VM_FAULT_WIRE) == 0) {
1293 vm_page_lock(fs.first_m);
1294 vm_page_wire(fs.first_m);
1295 vm_page_unlock(fs.first_m);
1298 vm_page_unwire(fs.m, PQ_INACTIVE);
1299 vm_page_unlock(fs.m);
1303 * Typically, the shadow object is either
1304 * private to this address space
1305 * (OBJ_ONEMAPPING) or its pages are read only.
1306 * In the highly unusual case where the pages of
1307 * a shadow object are read/write shared between
1308 * this and other address spaces, we need to
1309 * ensure that any pmap-level mappings to the
1310 * original, copy-on-write page from the backing
1311 * object are removed from those other address
1314 * The flag check is racy, but this is
1315 * tolerable: if OBJ_ONEMAPPING is cleared after
1316 * the check, the busy state ensures that new
1317 * mappings of fs.m can't be created.
1318 * pmap_enter() will replace an existing mapping
1319 * in the current address space. If
1320 * OBJ_ONEMAPPING is set after the check,
1321 * removing mappings will at worse trigger some
1322 * unnecessary page faults.
1324 vm_page_assert_xbusied(fs.m);
1325 if ((fs.first_object->flags & OBJ_ONEMAPPING) == 0)
1326 pmap_remove_all(fs.m);
1329 * We no longer need the old page or object.
1334 * fs.object != fs.first_object due to above
1337 vm_object_pip_wakeup(fs.object);
1338 VM_OBJECT_WUNLOCK(fs.object);
1340 * Only use the new page below...
1342 fs.object = fs.first_object;
1343 fs.pindex = fs.first_pindex;
1345 if (!is_first_object_locked)
1346 VM_OBJECT_WLOCK(fs.object);
1347 VM_CNT_INC(v_cow_faults);
1348 curthread->td_cow++;
1350 prot &= ~VM_PROT_WRITE;
1355 * We must verify that the maps have not changed since our last
1358 if (!fs.lookup_still_valid) {
1359 if (!vm_map_trylock_read(fs.map)) {
1361 unlock_and_deallocate(&fs);
1364 fs.lookup_still_valid = true;
1365 if (fs.map->timestamp != fs.map_generation) {
1366 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
1367 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
1370 * If we don't need the page any longer, put it on the inactive
1371 * list (the easiest thing to do here). If no one needs it,
1372 * pageout will grab it eventually.
1374 if (result != KERN_SUCCESS) {
1376 unlock_and_deallocate(&fs);
1379 * If retry of map lookup would have blocked then
1380 * retry fault from start.
1382 if (result == KERN_FAILURE)
1386 if ((retry_object != fs.first_object) ||
1387 (retry_pindex != fs.first_pindex)) {
1389 unlock_and_deallocate(&fs);
1394 * Check whether the protection has changed or the object has
1395 * been copied while we left the map unlocked. Changing from
1396 * read to write permission is OK - we leave the page
1397 * write-protected, and catch the write fault. Changing from
1398 * write to read permission means that we can't mark the page
1399 * write-enabled after all.
1402 fault_type &= retry_prot;
1405 unlock_and_deallocate(&fs);
1409 /* Reassert because wired may have changed. */
1410 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0,
1411 ("!wired && VM_FAULT_WIRE"));
1416 * If the page was filled by a pager, save the virtual address that
1417 * should be faulted on next under a sequential access pattern to the
1418 * map entry. A read lock on the map suffices to update this address
1422 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1424 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true);
1425 vm_page_assert_xbusied(fs.m);
1428 * Page must be completely valid or it is not fit to
1429 * map into user space. vm_pager_get_pages() ensures this.
1431 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL,
1432 ("vm_fault: page %p partially invalid", fs.m));
1433 VM_OBJECT_WUNLOCK(fs.object);
1436 * Put this page into the physical map. We had to do the unlock above
1437 * because pmap_enter() may sleep. We don't put the page
1438 * back on the active queue until later so that the pageout daemon
1439 * won't find it (yet).
1441 pmap_enter(fs.map->pmap, vaddr, fs.m, prot,
1442 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0);
1443 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 &&
1445 vm_fault_prefault(&fs, vaddr,
1446 faultcount > 0 ? behind : PFBAK,
1447 faultcount > 0 ? ahead : PFFOR, false);
1448 VM_OBJECT_WLOCK(fs.object);
1452 * If the page is not wired down, then put it where the pageout daemon
1455 if ((fault_flags & VM_FAULT_WIRE) != 0)
1458 vm_page_activate(fs.m);
1459 if (m_hold != NULL) {
1463 vm_page_unlock(fs.m);
1464 vm_page_xunbusy(fs.m);
1467 * Unlock everything, and return
1469 unlock_and_deallocate(&fs);
1471 VM_CNT_INC(v_io_faults);
1472 curthread->td_ru.ru_majflt++;
1474 if (racct_enable && fs.object->type == OBJT_VNODE) {
1476 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1477 racct_add_force(curproc, RACCT_WRITEBPS,
1478 PAGE_SIZE + behind * PAGE_SIZE);
1479 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1481 racct_add_force(curproc, RACCT_READBPS,
1482 PAGE_SIZE + ahead * PAGE_SIZE);
1483 racct_add_force(curproc, RACCT_READIOPS, 1);
1485 PROC_UNLOCK(curproc);
1489 curthread->td_ru.ru_minflt++;
1491 return (KERN_SUCCESS);
1495 * Speed up the reclamation of pages that precede the faulting pindex within
1496 * the first object of the shadow chain. Essentially, perform the equivalent
1497 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1498 * the faulting pindex by the cluster size when the pages read by vm_fault()
1499 * cross a cluster-size boundary. The cluster size is the greater of the
1500 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1502 * When "fs->first_object" is a shadow object, the pages in the backing object
1503 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1504 * function must only be concerned with pages in the first object.
1507 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1509 vm_map_entry_t entry;
1510 vm_object_t first_object, object;
1511 vm_offset_t end, start;
1512 vm_page_t m, m_next;
1513 vm_pindex_t pend, pstart;
1516 object = fs->object;
1517 VM_OBJECT_ASSERT_WLOCKED(object);
1518 first_object = fs->first_object;
1519 if (first_object != object) {
1520 if (!VM_OBJECT_TRYWLOCK(first_object)) {
1521 VM_OBJECT_WUNLOCK(object);
1522 VM_OBJECT_WLOCK(first_object);
1523 VM_OBJECT_WLOCK(object);
1526 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1527 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1528 size = VM_FAULT_DONTNEED_MIN;
1529 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1530 size = pagesizes[1];
1531 end = rounddown2(vaddr, size);
1532 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1533 (entry = fs->entry)->start < end) {
1534 if (end - entry->start < size)
1535 start = entry->start;
1538 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1539 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1541 m_next = vm_page_find_least(first_object, pstart);
1542 pend = OFF_TO_IDX(entry->offset) + atop(end -
1544 while ((m = m_next) != NULL && m->pindex < pend) {
1545 m_next = TAILQ_NEXT(m, listq);
1546 if (m->valid != VM_PAGE_BITS_ALL ||
1551 * Don't clear PGA_REFERENCED, since it would
1552 * likely represent a reference by a different
1555 * Typically, at this point, prefetched pages
1556 * are still in the inactive queue. Only
1557 * pages that triggered page faults are in the
1561 if (!vm_page_inactive(m))
1562 vm_page_deactivate(m);
1567 if (first_object != object)
1568 VM_OBJECT_WUNLOCK(first_object);
1572 * vm_fault_prefault provides a quick way of clustering
1573 * pagefaults into a processes address space. It is a "cousin"
1574 * of vm_map_pmap_enter, except it runs at page fault time instead
1578 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1579 int backward, int forward, bool obj_locked)
1582 vm_map_entry_t entry;
1583 vm_object_t backing_object, lobject;
1584 vm_offset_t addr, starta;
1589 pmap = fs->map->pmap;
1590 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1595 if (addra < backward * PAGE_SIZE) {
1596 starta = entry->start;
1598 starta = addra - backward * PAGE_SIZE;
1599 if (starta < entry->start)
1600 starta = entry->start;
1604 * Generate the sequence of virtual addresses that are candidates for
1605 * prefaulting in an outward spiral from the faulting virtual address,
1606 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1607 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1608 * If the candidate address doesn't have a backing physical page, then
1609 * the loop immediately terminates.
1611 for (i = 0; i < 2 * imax(backward, forward); i++) {
1612 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1614 if (addr > addra + forward * PAGE_SIZE)
1617 if (addr < starta || addr >= entry->end)
1620 if (!pmap_is_prefaultable(pmap, addr))
1623 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1624 lobject = entry->object.vm_object;
1626 VM_OBJECT_RLOCK(lobject);
1627 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1628 lobject->type == OBJT_DEFAULT &&
1629 (backing_object = lobject->backing_object) != NULL) {
1630 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1631 0, ("vm_fault_prefault: unaligned object offset"));
1632 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1633 VM_OBJECT_RLOCK(backing_object);
1634 if (!obj_locked || lobject != entry->object.vm_object)
1635 VM_OBJECT_RUNLOCK(lobject);
1636 lobject = backing_object;
1639 if (!obj_locked || lobject != entry->object.vm_object)
1640 VM_OBJECT_RUNLOCK(lobject);
1643 if (m->valid == VM_PAGE_BITS_ALL &&
1644 (m->flags & PG_FICTITIOUS) == 0)
1645 pmap_enter_quick(pmap, addr, m, entry->protection);
1646 if (!obj_locked || lobject != entry->object.vm_object)
1647 VM_OBJECT_RUNLOCK(lobject);
1652 * Hold each of the physical pages that are mapped by the specified range of
1653 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1654 * and allow the specified types of access, "prot". If all of the implied
1655 * pages are successfully held, then the number of held pages is returned
1656 * together with pointers to those pages in the array "ma". However, if any
1657 * of the pages cannot be held, -1 is returned.
1660 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1661 vm_prot_t prot, vm_page_t *ma, int max_count)
1663 vm_offset_t end, va;
1666 boolean_t pmap_failed;
1670 end = round_page(addr + len);
1671 addr = trunc_page(addr);
1673 if (!vm_map_range_valid(map, addr, end))
1676 if (atop(end - addr) > max_count)
1677 panic("vm_fault_quick_hold_pages: count > max_count");
1678 count = atop(end - addr);
1681 * Most likely, the physical pages are resident in the pmap, so it is
1682 * faster to try pmap_extract_and_hold() first.
1684 pmap_failed = FALSE;
1685 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1686 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1689 else if ((prot & VM_PROT_WRITE) != 0 &&
1690 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1692 * Explicitly dirty the physical page. Otherwise, the
1693 * caller's changes may go unnoticed because they are
1694 * performed through an unmanaged mapping or by a DMA
1697 * The object lock is not held here.
1698 * See vm_page_clear_dirty_mask().
1705 * One or more pages could not be held by the pmap. Either no
1706 * page was mapped at the specified virtual address or that
1707 * mapping had insufficient permissions. Attempt to fault in
1708 * and hold these pages.
1710 * If vm_fault_disable_pagefaults() was called,
1711 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1712 * acquire MD VM locks, which means we must not call
1713 * vm_fault(). Some (out of tree) callers mark
1714 * too wide a code area with vm_fault_disable_pagefaults()
1715 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1716 * the proper behaviour explicitly.
1718 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1719 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1721 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1722 if (*mp == NULL && vm_fault(map, va, prot,
1723 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1728 for (mp = ma; mp < ma + count; mp++)
1731 vm_page_unhold(*mp);
1732 vm_page_unlock(*mp);
1739 * vm_fault_copy_entry
1741 * Create new shadow object backing dst_entry with private copy of
1742 * all underlying pages. When src_entry is equal to dst_entry,
1743 * function implements COW for wired-down map entry. Otherwise,
1744 * it forks wired entry into dst_map.
1746 * In/out conditions:
1747 * The source and destination maps must be locked for write.
1748 * The source map entry must be wired down (or be a sharing map
1749 * entry corresponding to a main map entry that is wired down).
1752 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1753 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1754 vm_ooffset_t *fork_charge)
1756 vm_object_t backing_object, dst_object, object, src_object;
1757 vm_pindex_t dst_pindex, pindex, src_pindex;
1758 vm_prot_t access, prot;
1768 upgrade = src_entry == dst_entry;
1769 access = prot = dst_entry->protection;
1771 src_object = src_entry->object.vm_object;
1772 src_pindex = OFF_TO_IDX(src_entry->offset);
1774 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1775 dst_object = src_object;
1776 vm_object_reference(dst_object);
1779 * Create the top-level object for the destination entry. (Doesn't
1780 * actually shadow anything - we copy the pages directly.)
1782 dst_object = vm_object_allocate(OBJT_DEFAULT,
1783 atop(dst_entry->end - dst_entry->start));
1784 #if VM_NRESERVLEVEL > 0
1785 dst_object->flags |= OBJ_COLORED;
1786 dst_object->pg_color = atop(dst_entry->start);
1788 dst_object->domain = src_object->domain;
1789 dst_object->charge = dst_entry->end - dst_entry->start;
1792 VM_OBJECT_WLOCK(dst_object);
1793 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1794 ("vm_fault_copy_entry: vm_object not NULL"));
1795 if (src_object != dst_object) {
1796 dst_entry->object.vm_object = dst_object;
1797 dst_entry->offset = 0;
1798 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1800 if (fork_charge != NULL) {
1801 KASSERT(dst_entry->cred == NULL,
1802 ("vm_fault_copy_entry: leaked swp charge"));
1803 dst_object->cred = curthread->td_ucred;
1804 crhold(dst_object->cred);
1805 *fork_charge += dst_object->charge;
1806 } else if ((dst_object->type == OBJT_DEFAULT ||
1807 dst_object->type == OBJT_SWAP) &&
1808 dst_object->cred == NULL) {
1809 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1811 dst_object->cred = dst_entry->cred;
1812 dst_entry->cred = NULL;
1816 * If not an upgrade, then enter the mappings in the pmap as
1817 * read and/or execute accesses. Otherwise, enter them as
1820 * A writeable large page mapping is only created if all of
1821 * the constituent small page mappings are modified. Marking
1822 * PTEs as modified on inception allows promotion to happen
1823 * without taking potentially large number of soft faults.
1826 access &= ~VM_PROT_WRITE;
1829 * Loop through all of the virtual pages within the entry's
1830 * range, copying each page from the source object to the
1831 * destination object. Since the source is wired, those pages
1832 * must exist. In contrast, the destination is pageable.
1833 * Since the destination object doesn't share any backing storage
1834 * with the source object, all of its pages must be dirtied,
1835 * regardless of whether they can be written.
1837 for (vaddr = dst_entry->start, dst_pindex = 0;
1838 vaddr < dst_entry->end;
1839 vaddr += PAGE_SIZE, dst_pindex++) {
1842 * Find the page in the source object, and copy it in.
1843 * Because the source is wired down, the page will be
1846 if (src_object != dst_object)
1847 VM_OBJECT_RLOCK(src_object);
1848 object = src_object;
1849 pindex = src_pindex + dst_pindex;
1850 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1851 (backing_object = object->backing_object) != NULL) {
1853 * Unless the source mapping is read-only or
1854 * it is presently being upgraded from
1855 * read-only, the first object in the shadow
1856 * chain should provide all of the pages. In
1857 * other words, this loop body should never be
1858 * executed when the source mapping is already
1861 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1863 ("vm_fault_copy_entry: main object missing page"));
1865 VM_OBJECT_RLOCK(backing_object);
1866 pindex += OFF_TO_IDX(object->backing_object_offset);
1867 if (object != dst_object)
1868 VM_OBJECT_RUNLOCK(object);
1869 object = backing_object;
1871 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1873 if (object != dst_object) {
1875 * Allocate a page in the destination object.
1877 dst_m = vm_page_alloc(dst_object, (src_object ==
1878 dst_object ? src_pindex : 0) + dst_pindex,
1880 if (dst_m == NULL) {
1881 VM_OBJECT_WUNLOCK(dst_object);
1882 VM_OBJECT_RUNLOCK(object);
1883 vm_wait(dst_object);
1884 VM_OBJECT_WLOCK(dst_object);
1887 pmap_copy_page(src_m, dst_m);
1888 VM_OBJECT_RUNLOCK(object);
1889 dst_m->dirty = dst_m->valid = src_m->valid;
1892 if (vm_page_sleep_if_busy(dst_m, "fltupg"))
1894 if (dst_m->pindex >= dst_object->size)
1896 * We are upgrading. Index can occur
1897 * out of bounds if the object type is
1898 * vnode and the file was truncated.
1901 vm_page_xbusy(dst_m);
1903 VM_OBJECT_WUNLOCK(dst_object);
1906 * Enter it in the pmap. If a wired, copy-on-write
1907 * mapping is being replaced by a write-enabled
1908 * mapping, then wire that new mapping.
1910 * The page can be invalid if the user called
1911 * msync(MS_INVALIDATE) or truncated the backing vnode
1912 * or shared memory object. In this case, do not
1913 * insert it into pmap, but still do the copy so that
1914 * all copies of the wired map entry have similar
1917 if (dst_m->valid == VM_PAGE_BITS_ALL) {
1918 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
1919 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
1923 * Mark it no longer busy, and put it on the active list.
1925 VM_OBJECT_WLOCK(dst_object);
1928 if (src_m != dst_m) {
1929 vm_page_lock(src_m);
1930 vm_page_unwire(src_m, PQ_INACTIVE);
1931 vm_page_unlock(src_m);
1932 vm_page_lock(dst_m);
1933 vm_page_wire(dst_m);
1934 vm_page_unlock(dst_m);
1936 KASSERT(vm_page_wired(dst_m),
1937 ("dst_m %p is not wired", dst_m));
1940 vm_page_lock(dst_m);
1941 vm_page_activate(dst_m);
1942 vm_page_unlock(dst_m);
1944 vm_page_xunbusy(dst_m);
1946 VM_OBJECT_WUNLOCK(dst_object);
1948 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
1949 vm_object_deallocate(src_object);
1954 * Block entry into the machine-independent layer's page fault handler by
1955 * the calling thread. Subsequent calls to vm_fault() by that thread will
1956 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
1957 * spurious page faults.
1960 vm_fault_disable_pagefaults(void)
1963 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
1967 vm_fault_enable_pagefaults(int save)
1970 curthread_pflags_restore(save);