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
123 /* Fault parameters. */
126 vm_prot_t fault_type;
132 /* Page reference for cow. */
135 /* Current object. */
140 /* Top-level map object. */
141 vm_object_t first_object;
142 vm_pindex_t first_pindex;
147 vm_map_entry_t entry;
149 bool lookup_still_valid;
151 /* Vnode if locked. */
155 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
157 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
158 int backward, int forward, bool obj_locked);
160 static int vm_pfault_oom_attempts = 3;
161 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
162 &vm_pfault_oom_attempts, 0,
163 "Number of page allocation attempts in page fault handler before it "
164 "triggers OOM handling");
166 static int vm_pfault_oom_wait = 10;
167 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
168 &vm_pfault_oom_wait, 0,
169 "Number of seconds to wait for free pages before retrying "
170 "the page fault handler");
173 fault_page_release(vm_page_t *mp)
180 * We are likely to loop around again and attempt to busy
181 * this page. Deactivating it leaves it available for
182 * pageout while optimizing fault restarts.
184 vm_page_deactivate(m);
191 fault_page_free(vm_page_t *mp)
197 VM_OBJECT_ASSERT_WLOCKED(m->object);
198 if (!vm_page_wired(m))
207 unlock_map(struct faultstate *fs)
210 if (fs->lookup_still_valid) {
211 vm_map_lookup_done(fs->map, fs->entry);
212 fs->lookup_still_valid = false;
217 unlock_vp(struct faultstate *fs)
220 if (fs->vp != NULL) {
227 fault_deallocate(struct faultstate *fs)
230 fault_page_release(&fs->m_cow);
231 fault_page_release(&fs->m);
232 vm_object_pip_wakeup(fs->object);
233 if (fs->object != fs->first_object) {
234 VM_OBJECT_WLOCK(fs->first_object);
235 fault_page_free(&fs->first_m);
236 VM_OBJECT_WUNLOCK(fs->first_object);
237 vm_object_pip_wakeup(fs->first_object);
239 vm_object_deallocate(fs->first_object);
245 unlock_and_deallocate(struct faultstate *fs)
248 VM_OBJECT_WUNLOCK(fs->object);
249 fault_deallocate(fs);
253 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
257 if (((fs->prot & VM_PROT_WRITE) == 0 &&
258 (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
259 (m->oflags & VPO_UNMANAGED) != 0)
262 VM_PAGE_OBJECT_BUSY_ASSERT(m);
264 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
265 (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
266 (fs->fault_flags & VM_FAULT_DIRTY) != 0;
268 vm_object_set_writeable_dirty(m->object);
271 * If the fault is a write, we know that this page is being
272 * written NOW so dirty it explicitly to save on
273 * pmap_is_modified() calls later.
275 * Also, since the page is now dirty, we can possibly tell
276 * the pager to release any swap backing the page.
278 if (need_dirty && vm_page_set_dirty(m) == 0) {
280 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
281 * if the page is already dirty to prevent data written with
282 * the expectation of being synced from not being synced.
283 * Likewise if this entry does not request NOSYNC then make
284 * sure the page isn't marked NOSYNC. Applications sharing
285 * data should use the same flags to avoid ping ponging.
287 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
288 vm_page_aflag_set(m, PGA_NOSYNC);
290 vm_page_aflag_clear(m, PGA_NOSYNC);
296 * Unlocks fs.first_object and fs.map on success.
299 vm_fault_soft_fast(struct faultstate *fs)
302 #if VM_NRESERVLEVEL > 0
309 MPASS(fs->vp == NULL);
311 vm_object_busy(fs->first_object);
312 m = vm_page_lookup(fs->first_object, fs->first_pindex);
313 /* A busy page can be mapped for read|execute access. */
314 if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 &&
315 vm_page_busied(m)) || !vm_page_all_valid(m)) {
321 #if VM_NRESERVLEVEL > 0
322 if ((m->flags & PG_FICTITIOUS) == 0 &&
323 (m_super = vm_reserv_to_superpage(m)) != NULL &&
324 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
325 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
326 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
327 (pagesizes[m_super->psind] - 1)) && !fs->wired &&
328 pmap_ps_enabled(fs->map->pmap)) {
329 flags = PS_ALL_VALID;
330 if ((fs->prot & VM_PROT_WRITE) != 0) {
332 * Create a superpage mapping allowing write access
333 * only if none of the constituent pages are busy and
334 * all of them are already dirty (except possibly for
335 * the page that was faulted on).
337 flags |= PS_NONE_BUSY;
338 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
339 flags |= PS_ALL_DIRTY;
341 if (vm_page_ps_test(m_super, flags, m)) {
343 psind = m_super->psind;
344 vaddr = rounddown2(vaddr, pagesizes[psind]);
345 /* Preset the modified bit for dirty superpages. */
346 if ((flags & PS_ALL_DIRTY) != 0)
347 fs->fault_type |= VM_PROT_WRITE;
351 rv = pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
352 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
353 if (rv != KERN_SUCCESS)
355 if (fs->m_hold != NULL) {
359 if (psind == 0 && !fs->wired)
360 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
361 VM_OBJECT_RUNLOCK(fs->first_object);
362 vm_fault_dirty(fs, m);
363 vm_map_lookup_done(fs->map, fs->entry);
364 curthread->td_ru.ru_minflt++;
367 vm_object_unbusy(fs->first_object);
372 vm_fault_restore_map_lock(struct faultstate *fs)
375 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
376 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
378 if (!vm_map_trylock_read(fs->map)) {
379 VM_OBJECT_WUNLOCK(fs->first_object);
380 vm_map_lock_read(fs->map);
381 VM_OBJECT_WLOCK(fs->first_object);
383 fs->lookup_still_valid = true;
387 vm_fault_populate_check_page(vm_page_t m)
391 * Check each page to ensure that the pager is obeying the
392 * interface: the page must be installed in the object, fully
393 * valid, and exclusively busied.
396 MPASS(vm_page_all_valid(m));
397 MPASS(vm_page_xbusied(m));
401 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
407 VM_OBJECT_ASSERT_WLOCKED(object);
408 MPASS(first <= last);
409 for (pidx = first, m = vm_page_lookup(object, pidx);
410 pidx <= last; pidx++, m = vm_page_next(m)) {
411 vm_fault_populate_check_page(m);
412 vm_page_deactivate(m);
418 vm_fault_populate(struct faultstate *fs)
422 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
423 int bdry_idx, i, npages, psind, rv;
425 MPASS(fs->object == fs->first_object);
426 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
427 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
428 MPASS(fs->first_object->backing_object == NULL);
429 MPASS(fs->lookup_still_valid);
431 pager_first = OFF_TO_IDX(fs->entry->offset);
432 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
437 * Call the pager (driver) populate() method.
439 * There is no guarantee that the method will be called again
440 * if the current fault is for read, and a future fault is
441 * for write. Report the entry's maximum allowed protection
444 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
445 fs->fault_type, fs->entry->max_protection, &pager_first,
448 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
449 if (rv == VM_PAGER_BAD) {
451 * VM_PAGER_BAD is the backdoor for a pager to request
452 * normal fault handling.
454 vm_fault_restore_map_lock(fs);
455 if (fs->map->timestamp != fs->map_generation)
456 return (KERN_RESTART);
457 return (KERN_NOT_RECEIVER);
459 if (rv != VM_PAGER_OK)
460 return (KERN_FAILURE); /* AKA SIGSEGV */
462 /* Ensure that the driver is obeying the interface. */
463 MPASS(pager_first <= pager_last);
464 MPASS(fs->first_pindex <= pager_last);
465 MPASS(fs->first_pindex >= pager_first);
466 MPASS(pager_last < fs->first_object->size);
468 vm_fault_restore_map_lock(fs);
469 bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >>
470 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT;
471 if (fs->map->timestamp != fs->map_generation) {
473 vm_fault_populate_cleanup(fs->first_object, pager_first,
476 m = vm_page_lookup(fs->first_object, pager_first);
480 return (KERN_RESTART);
484 * The map is unchanged after our last unlock. Process the fault.
486 * First, the special case of largepage mappings, where
487 * populate only busies the first page in superpage run.
490 m = vm_page_lookup(fs->first_object, pager_first);
491 vm_fault_populate_check_page(m);
492 VM_OBJECT_WUNLOCK(fs->first_object);
493 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
495 /* assert alignment for entry */
496 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
497 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
498 (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
499 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
500 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
501 ("unaligned superpage m %p %#jx", m,
502 (uintmax_t)VM_PAGE_TO_PHYS(m)));
503 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
504 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
505 PMAP_ENTER_LARGEPAGE, bdry_idx);
506 VM_OBJECT_WLOCK(fs->first_object);
508 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
509 for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
512 if (fs->m_hold != NULL) {
513 *fs->m_hold = m + (fs->first_pindex - pager_first);
514 vm_page_wire(*fs->m_hold);
520 * The range [pager_first, pager_last] that is given to the
521 * pager is only a hint. The pager may populate any range
522 * within the object that includes the requested page index.
523 * In case the pager expanded the range, clip it to fit into
526 map_first = OFF_TO_IDX(fs->entry->offset);
527 if (map_first > pager_first) {
528 vm_fault_populate_cleanup(fs->first_object, pager_first,
530 pager_first = map_first;
532 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
533 if (map_last < pager_last) {
534 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
536 pager_last = map_last;
538 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
540 pidx += npages, m = vm_page_next(&m[npages - 1])) {
541 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
542 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \
543 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)
545 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
546 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
547 !pmap_ps_enabled(fs->map->pmap) || fs->wired))
552 npages = atop(pagesizes[psind]);
553 for (i = 0; i < npages; i++) {
554 vm_fault_populate_check_page(&m[i]);
555 vm_fault_dirty(fs, &m[i]);
557 VM_OBJECT_WUNLOCK(fs->first_object);
558 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
559 (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
560 #if defined(__amd64__)
561 if (psind > 0 && rv == KERN_FAILURE) {
562 for (i = 0; i < npages; i++) {
563 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
564 &m[i], fs->prot, fs->fault_type |
565 (fs->wired ? PMAP_ENTER_WIRED : 0), 0);
566 MPASS(rv == KERN_SUCCESS);
570 MPASS(rv == KERN_SUCCESS);
572 VM_OBJECT_WLOCK(fs->first_object);
573 for (i = 0; i < npages; i++) {
574 if ((fs->fault_flags & VM_FAULT_WIRE) != 0)
577 vm_page_activate(&m[i]);
578 if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) {
579 (*fs->m_hold) = &m[i];
582 vm_page_xunbusy(&m[i]);
586 curthread->td_ru.ru_majflt++;
587 return (KERN_SUCCESS);
590 static int prot_fault_translation;
591 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
592 &prot_fault_translation, 0,
593 "Control signal to deliver on protection fault");
595 /* compat definition to keep common code for signal translation */
596 #define UCODE_PAGEFLT 12
598 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
604 * Handle a page fault occurring at the given address,
605 * requiring the given permissions, in the map specified.
606 * If successful, the page is inserted into the
607 * associated physical map.
609 * NOTE: the given address should be truncated to the
610 * proper page address.
612 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
613 * a standard error specifying why the fault is fatal is returned.
615 * The map in question must be referenced, and remains so.
616 * Caller may hold no locks.
619 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
620 int fault_flags, int *signo, int *ucode)
624 MPASS(signo == NULL || ucode != NULL);
626 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
627 ktrfault(vaddr, fault_type);
629 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
631 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
632 result == KERN_INVALID_ADDRESS ||
633 result == KERN_RESOURCE_SHORTAGE ||
634 result == KERN_PROTECTION_FAILURE ||
635 result == KERN_OUT_OF_BOUNDS,
636 ("Unexpected Mach error %d from vm_fault()", result));
638 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
641 if (result != KERN_SUCCESS && signo != NULL) {
644 case KERN_INVALID_ADDRESS:
646 *ucode = SEGV_MAPERR;
648 case KERN_RESOURCE_SHORTAGE:
652 case KERN_OUT_OF_BOUNDS:
656 case KERN_PROTECTION_FAILURE:
657 if (prot_fault_translation == 0) {
659 * Autodetect. This check also covers
660 * the images without the ABI-tag ELF
663 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
664 curproc->p_osrel >= P_OSREL_SIGSEGV) {
666 *ucode = SEGV_ACCERR;
669 *ucode = UCODE_PAGEFLT;
671 } else if (prot_fault_translation == 1) {
672 /* Always compat mode. */
674 *ucode = UCODE_PAGEFLT;
676 /* Always SIGSEGV mode. */
678 *ucode = SEGV_ACCERR;
682 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
691 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
696 if (fs->object->type != OBJT_VNODE)
697 return (KERN_SUCCESS);
698 vp = fs->object->handle;
700 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
701 return (KERN_SUCCESS);
705 * Perform an unlock in case the desired vnode changed while
706 * the map was unlocked during a retry.
710 locked = VOP_ISLOCKED(vp);
711 if (locked != LK_EXCLUSIVE)
715 * We must not sleep acquiring the vnode lock while we have
716 * the page exclusive busied or the object's
717 * paging-in-progress count incremented. Otherwise, we could
720 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
723 return (KERN_SUCCESS);
728 unlock_and_deallocate(fs);
730 fault_deallocate(fs);
731 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
734 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
735 return (KERN_RESOURCE_SHORTAGE);
739 * Calculate the desired readahead. Handle drop-behind.
741 * Returns the number of readahead blocks to pass to the pager.
744 vm_fault_readahead(struct faultstate *fs)
749 KASSERT(fs->lookup_still_valid, ("map unlocked"));
750 era = fs->entry->read_ahead;
751 behavior = vm_map_entry_behavior(fs->entry);
752 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
754 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
755 nera = VM_FAULT_READ_AHEAD_MAX;
756 if (fs->vaddr == fs->entry->next_read)
757 vm_fault_dontneed(fs, fs->vaddr, nera);
758 } else if (fs->vaddr == fs->entry->next_read) {
760 * This is a sequential fault. Arithmetically
761 * increase the requested number of pages in
762 * the read-ahead window. The requested
763 * number of pages is "# of sequential faults
764 * x (read ahead min + 1) + read ahead min"
766 nera = VM_FAULT_READ_AHEAD_MIN;
769 if (nera > VM_FAULT_READ_AHEAD_MAX)
770 nera = VM_FAULT_READ_AHEAD_MAX;
772 if (era == VM_FAULT_READ_AHEAD_MAX)
773 vm_fault_dontneed(fs, fs->vaddr, nera);
776 * This is a non-sequential fault.
782 * A read lock on the map suffices to update
783 * the read ahead count safely.
785 fs->entry->read_ahead = nera;
792 vm_fault_lookup(struct faultstate *fs)
796 KASSERT(!fs->lookup_still_valid,
797 ("vm_fault_lookup: Map already locked."));
798 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
799 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
800 &fs->first_pindex, &fs->prot, &fs->wired);
801 if (result != KERN_SUCCESS) {
806 fs->map_generation = fs->map->timestamp;
808 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
809 panic("%s: fault on nofault entry, addr: %#lx",
810 __func__, (u_long)fs->vaddr);
813 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
814 fs->entry->wiring_thread != curthread) {
815 vm_map_unlock_read(fs->map);
816 vm_map_lock(fs->map);
817 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
818 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
820 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
821 vm_map_unlock_and_wait(fs->map, 0);
823 vm_map_unlock(fs->map);
824 return (KERN_RESOURCE_SHORTAGE);
827 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
830 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
832 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
833 ("!fs->wired && VM_FAULT_WIRE"));
834 fs->lookup_still_valid = true;
836 return (KERN_SUCCESS);
840 vm_fault_relookup(struct faultstate *fs)
842 vm_object_t retry_object;
843 vm_pindex_t retry_pindex;
844 vm_prot_t retry_prot;
847 if (!vm_map_trylock_read(fs->map))
848 return (KERN_RESTART);
850 fs->lookup_still_valid = true;
851 if (fs->map->timestamp == fs->map_generation)
852 return (KERN_SUCCESS);
854 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
855 &fs->entry, &retry_object, &retry_pindex, &retry_prot,
857 if (result != KERN_SUCCESS) {
859 * If retry of map lookup would have blocked then
860 * retry fault from start.
862 if (result == KERN_FAILURE)
863 return (KERN_RESTART);
866 if (retry_object != fs->first_object ||
867 retry_pindex != fs->first_pindex)
868 return (KERN_RESTART);
871 * Check whether the protection has changed or the object has
872 * been copied while we left the map unlocked. Changing from
873 * read to write permission is OK - we leave the page
874 * write-protected, and catch the write fault. Changing from
875 * write to read permission means that we can't mark the page
876 * write-enabled after all.
878 fs->prot &= retry_prot;
879 fs->fault_type &= retry_prot;
881 return (KERN_RESTART);
883 /* Reassert because wired may have changed. */
884 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
885 ("!wired && VM_FAULT_WIRE"));
887 return (KERN_SUCCESS);
891 vm_fault_cow(struct faultstate *fs)
893 bool is_first_object_locked;
896 * This allows pages to be virtually copied from a backing_object
897 * into the first_object, where the backing object has no other
898 * refs to it, and cannot gain any more refs. Instead of a bcopy,
899 * we just move the page from the backing object to the first
900 * object. Note that we must mark the page dirty in the first
901 * object so that it will go out to swap when needed.
903 is_first_object_locked = false;
906 * Only one shadow object and no other refs.
908 fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
910 * No other ways to look the object up
912 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
914 * We don't chase down the shadow chain and we can acquire locks.
916 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
917 fs->object == fs->first_object->backing_object &&
918 VM_OBJECT_TRYWLOCK(fs->object)) {
920 * Remove but keep xbusy for replace. fs->m is moved into
921 * fs->first_object and left busy while fs->first_m is
922 * conditionally freed.
924 vm_page_remove_xbusy(fs->m);
925 vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
927 vm_page_dirty(fs->m);
928 #if VM_NRESERVLEVEL > 0
930 * Rename the reservation.
932 vm_reserv_rename(fs->m, fs->first_object, fs->object,
933 OFF_TO_IDX(fs->first_object->backing_object_offset));
935 VM_OBJECT_WUNLOCK(fs->object);
936 VM_OBJECT_WUNLOCK(fs->first_object);
939 VM_CNT_INC(v_cow_optim);
941 if (is_first_object_locked)
942 VM_OBJECT_WUNLOCK(fs->first_object);
944 * Oh, well, lets copy it.
946 pmap_copy_page(fs->m, fs->first_m);
947 vm_page_valid(fs->first_m);
948 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
949 vm_page_wire(fs->first_m);
950 vm_page_unwire(fs->m, PQ_INACTIVE);
953 * Save the cow page to be released after
954 * pmap_enter is complete.
960 * fs->object != fs->first_object due to above
963 vm_object_pip_wakeup(fs->object);
966 * Only use the new page below...
968 fs->object = fs->first_object;
969 fs->pindex = fs->first_pindex;
971 VM_CNT_INC(v_cow_faults);
976 vm_fault_next(struct faultstate *fs)
978 vm_object_t next_object;
981 * The requested page does not exist at this object/
982 * offset. Remove the invalid page from the object,
983 * waking up anyone waiting for it, and continue on to
984 * the next object. However, if this is the top-level
985 * object, we must leave the busy page in place to
986 * prevent another process from rushing past us, and
987 * inserting the page in that object at the same time
990 if (fs->object == fs->first_object) {
994 fault_page_free(&fs->m);
997 * Move on to the next object. Lock the next object before
998 * unlocking the current one.
1000 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1001 next_object = fs->object->backing_object;
1002 if (next_object == NULL)
1004 MPASS(fs->first_m != NULL);
1005 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1006 VM_OBJECT_WLOCK(next_object);
1007 vm_object_pip_add(next_object, 1);
1008 if (fs->object != fs->first_object)
1009 vm_object_pip_wakeup(fs->object);
1010 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1011 VM_OBJECT_WUNLOCK(fs->object);
1012 fs->object = next_object;
1018 vm_fault_zerofill(struct faultstate *fs)
1022 * If there's no object left, fill the page in the top
1023 * object with zeros.
1025 if (fs->object != fs->first_object) {
1026 vm_object_pip_wakeup(fs->object);
1027 fs->object = fs->first_object;
1028 fs->pindex = fs->first_pindex;
1030 MPASS(fs->first_m != NULL);
1031 MPASS(fs->m == NULL);
1032 fs->m = fs->first_m;
1036 * Zero the page if necessary and mark it valid.
1038 if ((fs->m->flags & PG_ZERO) == 0) {
1039 pmap_zero_page(fs->m);
1041 VM_CNT_INC(v_ozfod);
1044 vm_page_valid(fs->m);
1048 * Allocate a page directly or via the object populate method.
1051 vm_fault_allocate(struct faultstate *fs)
1053 struct domainset *dset;
1057 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1058 rv = vm_fault_lock_vnode(fs, true);
1059 MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1060 if (rv == KERN_RESOURCE_SHORTAGE)
1064 if (fs->pindex >= fs->object->size)
1065 return (KERN_OUT_OF_BOUNDS);
1067 if (fs->object == fs->first_object &&
1068 (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1069 fs->first_object->shadow_count == 0) {
1070 rv = vm_fault_populate(fs);
1076 case KERN_NOT_RECEIVER:
1078 * Pager's populate() method
1079 * returned VM_PAGER_BAD.
1083 panic("inconsistent return codes");
1088 * Allocate a new page for this object/offset pair.
1090 * Unlocked read of the p_flag is harmless. At worst, the P_KILLED
1091 * might be not observed there, and allocation can fail, causing
1092 * restart and new reading of the p_flag.
1094 dset = fs->object->domain.dr_policy;
1096 dset = curthread->td_domain.dr_policy;
1097 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1098 #if VM_NRESERVLEVEL > 0
1099 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1101 alloc_req = P_KILLED(curproc) ?
1102 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL;
1103 if (fs->object->type != OBJT_VNODE &&
1104 fs->object->backing_object == NULL)
1105 alloc_req |= VM_ALLOC_ZERO;
1106 fs->m = vm_page_alloc(fs->object, fs->pindex, alloc_req);
1108 if (fs->m == NULL) {
1109 unlock_and_deallocate(fs);
1110 if (vm_pfault_oom_attempts < 0 ||
1111 fs->oom < vm_pfault_oom_attempts) {
1113 vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1117 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1118 curproc->p_pid, curproc->p_comm);
1119 vm_pageout_oom(VM_OOM_MEM_PF);
1122 return (KERN_RESOURCE_SHORTAGE);
1126 return (KERN_NOT_RECEIVER);
1130 * Call the pager to retrieve the page if there is a chance
1131 * that the pager has it, and potentially retrieve additional
1132 * pages at the same time.
1135 vm_fault_getpages(struct faultstate *fs, int nera, int *behindp, int *aheadp)
1137 vm_offset_t e_end, e_start;
1138 int ahead, behind, cluster_offset, rv;
1142 * Prepare for unlocking the map. Save the map
1143 * entry's start and end addresses, which are used to
1144 * optimize the size of the pager operation below.
1145 * Even if the map entry's addresses change after
1146 * unlocking the map, using the saved addresses is
1149 e_start = fs->entry->start;
1150 e_end = fs->entry->end;
1151 behavior = vm_map_entry_behavior(fs->entry);
1154 * Release the map lock before locking the vnode or
1155 * sleeping in the pager. (If the current object has
1156 * a shadow, then an earlier iteration of this loop
1157 * may have already unlocked the map.)
1161 rv = vm_fault_lock_vnode(fs, false);
1162 MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE);
1163 if (rv == KERN_RESOURCE_SHORTAGE)
1165 KASSERT(fs->vp == NULL || !fs->map->system_map,
1166 ("vm_fault: vnode-backed object mapped by system map"));
1169 * Page in the requested page and hint the pager,
1170 * that it may bring up surrounding pages.
1172 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1173 P_KILLED(curproc)) {
1177 /* Is this a sequential fault? */
1183 * Request a cluster of pages that is
1184 * aligned to a VM_FAULT_READ_DEFAULT
1185 * page offset boundary within the
1186 * object. Alignment to a page offset
1187 * boundary is more likely to coincide
1188 * with the underlying file system
1189 * block than alignment to a virtual
1192 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1193 behind = ulmin(cluster_offset,
1194 atop(fs->vaddr - e_start));
1195 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1197 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1201 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1202 if (rv == VM_PAGER_OK)
1203 return (KERN_SUCCESS);
1204 if (rv == VM_PAGER_ERROR)
1205 printf("vm_fault: pager read error, pid %d (%s)\n",
1206 curproc->p_pid, curproc->p_comm);
1208 * If an I/O error occurred or the requested page was
1209 * outside the range of the pager, clean up and return
1212 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD)
1213 return (KERN_OUT_OF_BOUNDS);
1214 return (KERN_NOT_RECEIVER);
1218 * Wait/Retry if the page is busy. We have to do this if the page is
1219 * either exclusive or shared busy because the vm_pager may be using
1220 * read busy for pageouts (and even pageins if it is the vnode pager),
1221 * and we could end up trying to pagein and pageout the same page
1224 * We can theoretically allow the busy case on a read fault if the page
1225 * is marked valid, but since such pages are typically already pmap'd,
1226 * putting that special case in might be more effort then it is worth.
1227 * We cannot under any circumstances mess around with a shared busied
1228 * page except, perhaps, to pmap it.
1231 vm_fault_busy_sleep(struct faultstate *fs)
1234 * Reference the page before unlocking and
1235 * sleeping so that the page daemon is less
1236 * likely to reclaim it.
1238 vm_page_aflag_set(fs->m, PGA_REFERENCED);
1239 if (fs->object != fs->first_object) {
1240 fault_page_release(&fs->first_m);
1241 vm_object_pip_wakeup(fs->first_object);
1243 vm_object_pip_wakeup(fs->object);
1245 if (fs->m == vm_page_lookup(fs->object, fs->pindex))
1246 vm_page_busy_sleep(fs->m, "vmpfw", false);
1248 VM_OBJECT_WUNLOCK(fs->object);
1249 VM_CNT_INC(v_intrans);
1250 vm_object_deallocate(fs->first_object);
1254 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1255 int fault_flags, vm_page_t *m_hold)
1257 struct faultstate fs;
1258 int ahead, behind, faultcount;
1259 int nera, result, rv;
1260 bool dead, hardfault;
1262 VM_CNT_INC(v_vm_faults);
1264 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1265 return (KERN_PROTECTION_FAILURE);
1270 fs.fault_flags = fault_flags;
1272 fs.lookup_still_valid = false;
1279 fs.fault_type = fault_type;
1282 * Find the backing store object and offset into it to begin the
1285 result = vm_fault_lookup(&fs);
1286 if (result != KERN_SUCCESS) {
1287 if (result == KERN_RESOURCE_SHORTAGE)
1293 * Try to avoid lock contention on the top-level object through
1294 * special-case handling of some types of page faults, specifically,
1295 * those that are mapping an existing page from the top-level object.
1296 * Under this condition, a read lock on the object suffices, allowing
1297 * multiple page faults of a similar type to run in parallel.
1299 if (fs.vp == NULL /* avoid locked vnode leak */ &&
1300 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1301 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1302 VM_OBJECT_RLOCK(fs.first_object);
1303 rv = vm_fault_soft_fast(&fs);
1304 if (rv == KERN_SUCCESS)
1306 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) {
1307 VM_OBJECT_RUNLOCK(fs.first_object);
1308 VM_OBJECT_WLOCK(fs.first_object);
1311 VM_OBJECT_WLOCK(fs.first_object);
1315 * Make a reference to this object to prevent its disposal while we
1316 * are messing with it. Once we have the reference, the map is free
1317 * to be diddled. Since objects reference their shadows (and copies),
1318 * they will stay around as well.
1320 * Bump the paging-in-progress count to prevent size changes (e.g.
1321 * truncation operations) during I/O.
1323 vm_object_reference_locked(fs.first_object);
1324 vm_object_pip_add(fs.first_object, 1);
1326 fs.m_cow = fs.m = fs.first_m = NULL;
1329 * Search for the page at object/offset.
1331 fs.object = fs.first_object;
1332 fs.pindex = fs.first_pindex;
1334 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1335 rv = vm_fault_allocate(&fs);
1338 unlock_and_deallocate(&fs);
1340 case KERN_RESOURCE_SHORTAGE:
1344 case KERN_OUT_OF_BOUNDS:
1345 unlock_and_deallocate(&fs);
1347 case KERN_NOT_RECEIVER:
1350 panic("vm_fault: Unhandled rv %d", rv);
1355 KASSERT(fs.m == NULL,
1356 ("page still set %p at loop start", fs.m));
1358 * If the object is marked for imminent termination,
1359 * we retry here, since the collapse pass has raced
1360 * with us. Otherwise, if we see terminally dead
1361 * object, return fail.
1363 if ((fs.object->flags & OBJ_DEAD) != 0) {
1364 dead = fs.object->type == OBJT_DEAD;
1365 unlock_and_deallocate(&fs);
1367 return (KERN_PROTECTION_FAILURE);
1373 * See if page is resident
1375 fs.m = vm_page_lookup(fs.object, fs.pindex);
1377 if (vm_page_tryxbusy(fs.m) == 0) {
1378 vm_fault_busy_sleep(&fs);
1383 * The page is marked busy for other processes and the
1384 * pagedaemon. If it still is completely valid we
1387 if (vm_page_all_valid(fs.m)) {
1388 VM_OBJECT_WUNLOCK(fs.object);
1389 break; /* break to PAGE HAS BEEN FOUND. */
1392 VM_OBJECT_ASSERT_WLOCKED(fs.object);
1395 * Page is not resident. If the pager might contain the page
1396 * or this is the beginning of the search, allocate a new
1397 * page. (Default objects are zero-fill, so there is no real
1400 if (fs.m == NULL && (fs.object->type != OBJT_DEFAULT ||
1401 fs.object == fs.first_object)) {
1402 rv = vm_fault_allocate(&fs);
1405 unlock_and_deallocate(&fs);
1407 case KERN_RESOURCE_SHORTAGE:
1411 case KERN_OUT_OF_BOUNDS:
1412 unlock_and_deallocate(&fs);
1414 case KERN_NOT_RECEIVER:
1417 panic("vm_fault: Unhandled rv %d", rv);
1422 * Default objects have no pager so no exclusive busy exists
1423 * to protect this page in the chain. Skip to the next
1424 * object without dropping the lock to preserve atomicity of
1427 if (fs.object->type != OBJT_DEFAULT) {
1429 * At this point, we have either allocated a new page
1430 * or found an existing page that is only partially
1433 * We hold a reference on the current object and the
1434 * page is exclusive busied. The exclusive busy
1435 * prevents simultaneous faults and collapses while
1436 * the object lock is dropped.
1438 VM_OBJECT_WUNLOCK(fs.object);
1441 * If the pager for the current object might have
1442 * the page, then determine the number of additional
1443 * pages to read and potentially reprioritize
1444 * previously read pages for earlier reclamation.
1445 * These operations should only be performed once per
1446 * page fault. Even if the current pager doesn't
1447 * have the page, the number of additional pages to
1448 * read will apply to subsequent objects in the
1451 if (nera == -1 && !P_KILLED(curproc))
1452 nera = vm_fault_readahead(&fs);
1454 rv = vm_fault_getpages(&fs, nera, &behind, &ahead);
1455 if (rv == KERN_SUCCESS) {
1456 faultcount = behind + 1 + ahead;
1458 break; /* break to PAGE HAS BEEN FOUND. */
1460 if (rv == KERN_RESOURCE_SHORTAGE)
1462 VM_OBJECT_WLOCK(fs.object);
1463 if (rv == KERN_OUT_OF_BOUNDS) {
1464 fault_page_free(&fs.m);
1465 unlock_and_deallocate(&fs);
1471 * The page was not found in the current object. Try to
1472 * traverse into a backing object or zero fill if none is
1475 if (vm_fault_next(&fs))
1477 VM_OBJECT_WUNLOCK(fs.object);
1478 vm_fault_zerofill(&fs);
1479 /* Don't try to prefault neighboring pages. */
1481 break; /* break to PAGE HAS BEEN FOUND. */
1485 * PAGE HAS BEEN FOUND. A valid page has been found and exclusively
1486 * busied. The object lock must no longer be held.
1488 vm_page_assert_xbusied(fs.m);
1489 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1492 * If the page is being written, but isn't already owned by the
1493 * top-level object, we have to copy it into a new page owned by the
1496 if (fs.object != fs.first_object) {
1498 * We only really need to copy if we want to write it.
1500 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1503 * We only try to prefault read-only mappings to the
1504 * neighboring pages when this copy-on-write fault is
1505 * a hard fault. In other cases, trying to prefault
1506 * is typically wasted effort.
1508 if (faultcount == 0)
1512 fs.prot &= ~VM_PROT_WRITE;
1517 * We must verify that the maps have not changed since our last
1520 if (!fs.lookup_still_valid) {
1521 result = vm_fault_relookup(&fs);
1522 if (result != KERN_SUCCESS) {
1523 fault_deallocate(&fs);
1524 if (result == KERN_RESTART)
1529 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1532 * If the page was filled by a pager, save the virtual address that
1533 * should be faulted on next under a sequential access pattern to the
1534 * map entry. A read lock on the map suffices to update this address
1538 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1541 * Page must be completely valid or it is not fit to
1542 * map into user space. vm_pager_get_pages() ensures this.
1544 vm_page_assert_xbusied(fs.m);
1545 KASSERT(vm_page_all_valid(fs.m),
1546 ("vm_fault: page %p partially invalid", fs.m));
1548 vm_fault_dirty(&fs, fs.m);
1551 * Put this page into the physical map. We had to do the unlock above
1552 * because pmap_enter() may sleep. We don't put the page
1553 * back on the active queue until later so that the pageout daemon
1554 * won't find it (yet).
1556 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1557 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1558 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1560 vm_fault_prefault(&fs, vaddr,
1561 faultcount > 0 ? behind : PFBAK,
1562 faultcount > 0 ? ahead : PFFOR, false);
1565 * If the page is not wired down, then put it where the pageout daemon
1568 if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1571 vm_page_activate(fs.m);
1572 if (fs.m_hold != NULL) {
1573 (*fs.m_hold) = fs.m;
1576 vm_page_xunbusy(fs.m);
1580 * Unlock everything, and return
1582 fault_deallocate(&fs);
1584 VM_CNT_INC(v_io_faults);
1585 curthread->td_ru.ru_majflt++;
1587 if (racct_enable && fs.object->type == OBJT_VNODE) {
1589 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1590 racct_add_force(curproc, RACCT_WRITEBPS,
1591 PAGE_SIZE + behind * PAGE_SIZE);
1592 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1594 racct_add_force(curproc, RACCT_READBPS,
1595 PAGE_SIZE + ahead * PAGE_SIZE);
1596 racct_add_force(curproc, RACCT_READIOPS, 1);
1598 PROC_UNLOCK(curproc);
1602 curthread->td_ru.ru_minflt++;
1604 return (KERN_SUCCESS);
1608 * Speed up the reclamation of pages that precede the faulting pindex within
1609 * the first object of the shadow chain. Essentially, perform the equivalent
1610 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1611 * the faulting pindex by the cluster size when the pages read by vm_fault()
1612 * cross a cluster-size boundary. The cluster size is the greater of the
1613 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1615 * When "fs->first_object" is a shadow object, the pages in the backing object
1616 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1617 * function must only be concerned with pages in the first object.
1620 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1622 vm_map_entry_t entry;
1623 vm_object_t first_object, object;
1624 vm_offset_t end, start;
1625 vm_page_t m, m_next;
1626 vm_pindex_t pend, pstart;
1629 object = fs->object;
1630 VM_OBJECT_ASSERT_UNLOCKED(object);
1631 first_object = fs->first_object;
1632 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1633 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1634 VM_OBJECT_RLOCK(first_object);
1635 size = VM_FAULT_DONTNEED_MIN;
1636 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1637 size = pagesizes[1];
1638 end = rounddown2(vaddr, size);
1639 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1640 (entry = fs->entry)->start < end) {
1641 if (end - entry->start < size)
1642 start = entry->start;
1645 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1646 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1648 m_next = vm_page_find_least(first_object, pstart);
1649 pend = OFF_TO_IDX(entry->offset) + atop(end -
1651 while ((m = m_next) != NULL && m->pindex < pend) {
1652 m_next = TAILQ_NEXT(m, listq);
1653 if (!vm_page_all_valid(m) ||
1658 * Don't clear PGA_REFERENCED, since it would
1659 * likely represent a reference by a different
1662 * Typically, at this point, prefetched pages
1663 * are still in the inactive queue. Only
1664 * pages that triggered page faults are in the
1665 * active queue. The test for whether the page
1666 * is in the inactive queue is racy; in the
1667 * worst case we will requeue the page
1670 if (!vm_page_inactive(m))
1671 vm_page_deactivate(m);
1674 VM_OBJECT_RUNLOCK(first_object);
1679 * vm_fault_prefault provides a quick way of clustering
1680 * pagefaults into a processes address space. It is a "cousin"
1681 * of vm_map_pmap_enter, except it runs at page fault time instead
1685 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1686 int backward, int forward, bool obj_locked)
1689 vm_map_entry_t entry;
1690 vm_object_t backing_object, lobject;
1691 vm_offset_t addr, starta;
1696 pmap = fs->map->pmap;
1697 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1702 if (addra < backward * PAGE_SIZE) {
1703 starta = entry->start;
1705 starta = addra - backward * PAGE_SIZE;
1706 if (starta < entry->start)
1707 starta = entry->start;
1711 * Generate the sequence of virtual addresses that are candidates for
1712 * prefaulting in an outward spiral from the faulting virtual address,
1713 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1714 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1715 * If the candidate address doesn't have a backing physical page, then
1716 * the loop immediately terminates.
1718 for (i = 0; i < 2 * imax(backward, forward); i++) {
1719 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1721 if (addr > addra + forward * PAGE_SIZE)
1724 if (addr < starta || addr >= entry->end)
1727 if (!pmap_is_prefaultable(pmap, addr))
1730 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1731 lobject = entry->object.vm_object;
1733 VM_OBJECT_RLOCK(lobject);
1734 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1735 lobject->type == OBJT_DEFAULT &&
1736 (backing_object = lobject->backing_object) != NULL) {
1737 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1738 0, ("vm_fault_prefault: unaligned object offset"));
1739 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1740 VM_OBJECT_RLOCK(backing_object);
1741 if (!obj_locked || lobject != entry->object.vm_object)
1742 VM_OBJECT_RUNLOCK(lobject);
1743 lobject = backing_object;
1746 if (!obj_locked || lobject != entry->object.vm_object)
1747 VM_OBJECT_RUNLOCK(lobject);
1750 if (vm_page_all_valid(m) &&
1751 (m->flags & PG_FICTITIOUS) == 0)
1752 pmap_enter_quick(pmap, addr, m, entry->protection);
1753 if (!obj_locked || lobject != entry->object.vm_object)
1754 VM_OBJECT_RUNLOCK(lobject);
1759 * Hold each of the physical pages that are mapped by the specified range of
1760 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1761 * and allow the specified types of access, "prot". If all of the implied
1762 * pages are successfully held, then the number of held pages is returned
1763 * together with pointers to those pages in the array "ma". However, if any
1764 * of the pages cannot be held, -1 is returned.
1767 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1768 vm_prot_t prot, vm_page_t *ma, int max_count)
1770 vm_offset_t end, va;
1773 boolean_t pmap_failed;
1777 end = round_page(addr + len);
1778 addr = trunc_page(addr);
1780 if (!vm_map_range_valid(map, addr, end))
1783 if (atop(end - addr) > max_count)
1784 panic("vm_fault_quick_hold_pages: count > max_count");
1785 count = atop(end - addr);
1788 * Most likely, the physical pages are resident in the pmap, so it is
1789 * faster to try pmap_extract_and_hold() first.
1791 pmap_failed = FALSE;
1792 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1793 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1796 else if ((prot & VM_PROT_WRITE) != 0 &&
1797 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1799 * Explicitly dirty the physical page. Otherwise, the
1800 * caller's changes may go unnoticed because they are
1801 * performed through an unmanaged mapping or by a DMA
1804 * The object lock is not held here.
1805 * See vm_page_clear_dirty_mask().
1812 * One or more pages could not be held by the pmap. Either no
1813 * page was mapped at the specified virtual address or that
1814 * mapping had insufficient permissions. Attempt to fault in
1815 * and hold these pages.
1817 * If vm_fault_disable_pagefaults() was called,
1818 * i.e., TDP_NOFAULTING is set, we must not sleep nor
1819 * acquire MD VM locks, which means we must not call
1820 * vm_fault(). Some (out of tree) callers mark
1821 * too wide a code area with vm_fault_disable_pagefaults()
1822 * already, use the VM_PROT_QUICK_NOFAULT flag to request
1823 * the proper behaviour explicitly.
1825 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
1826 (curthread->td_pflags & TDP_NOFAULTING) != 0)
1828 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
1829 if (*mp == NULL && vm_fault(map, va, prot,
1830 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
1835 for (mp = ma; mp < ma + count; mp++)
1837 vm_page_unwire(*mp, PQ_INACTIVE);
1843 * vm_fault_copy_entry
1845 * Create new shadow object backing dst_entry with private copy of
1846 * all underlying pages. When src_entry is equal to dst_entry,
1847 * function implements COW for wired-down map entry. Otherwise,
1848 * it forks wired entry into dst_map.
1850 * In/out conditions:
1851 * The source and destination maps must be locked for write.
1852 * The source map entry must be wired down (or be a sharing map
1853 * entry corresponding to a main map entry that is wired down).
1856 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1857 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
1858 vm_ooffset_t *fork_charge)
1860 vm_object_t backing_object, dst_object, object, src_object;
1861 vm_pindex_t dst_pindex, pindex, src_pindex;
1862 vm_prot_t access, prot;
1872 upgrade = src_entry == dst_entry;
1873 access = prot = dst_entry->protection;
1875 src_object = src_entry->object.vm_object;
1876 src_pindex = OFF_TO_IDX(src_entry->offset);
1878 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
1879 dst_object = src_object;
1880 vm_object_reference(dst_object);
1883 * Create the top-level object for the destination entry.
1884 * Doesn't actually shadow anything - we copy the pages
1887 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
1888 dst_entry->start), NULL, NULL, 0);
1889 #if VM_NRESERVLEVEL > 0
1890 dst_object->flags |= OBJ_COLORED;
1891 dst_object->pg_color = atop(dst_entry->start);
1893 dst_object->domain = src_object->domain;
1894 dst_object->charge = dst_entry->end - dst_entry->start;
1897 VM_OBJECT_WLOCK(dst_object);
1898 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
1899 ("vm_fault_copy_entry: vm_object not NULL"));
1900 if (src_object != dst_object) {
1901 dst_entry->object.vm_object = dst_object;
1902 dst_entry->offset = 0;
1903 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
1905 if (fork_charge != NULL) {
1906 KASSERT(dst_entry->cred == NULL,
1907 ("vm_fault_copy_entry: leaked swp charge"));
1908 dst_object->cred = curthread->td_ucred;
1909 crhold(dst_object->cred);
1910 *fork_charge += dst_object->charge;
1911 } else if ((dst_object->type == OBJT_DEFAULT ||
1912 dst_object->type == OBJT_SWAP) &&
1913 dst_object->cred == NULL) {
1914 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
1916 dst_object->cred = dst_entry->cred;
1917 dst_entry->cred = NULL;
1921 * If not an upgrade, then enter the mappings in the pmap as
1922 * read and/or execute accesses. Otherwise, enter them as
1925 * A writeable large page mapping is only created if all of
1926 * the constituent small page mappings are modified. Marking
1927 * PTEs as modified on inception allows promotion to happen
1928 * without taking potentially large number of soft faults.
1931 access &= ~VM_PROT_WRITE;
1934 * Loop through all of the virtual pages within the entry's
1935 * range, copying each page from the source object to the
1936 * destination object. Since the source is wired, those pages
1937 * must exist. In contrast, the destination is pageable.
1938 * Since the destination object doesn't share any backing storage
1939 * with the source object, all of its pages must be dirtied,
1940 * regardless of whether they can be written.
1942 for (vaddr = dst_entry->start, dst_pindex = 0;
1943 vaddr < dst_entry->end;
1944 vaddr += PAGE_SIZE, dst_pindex++) {
1947 * Find the page in the source object, and copy it in.
1948 * Because the source is wired down, the page will be
1951 if (src_object != dst_object)
1952 VM_OBJECT_RLOCK(src_object);
1953 object = src_object;
1954 pindex = src_pindex + dst_pindex;
1955 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
1956 (backing_object = object->backing_object) != NULL) {
1958 * Unless the source mapping is read-only or
1959 * it is presently being upgraded from
1960 * read-only, the first object in the shadow
1961 * chain should provide all of the pages. In
1962 * other words, this loop body should never be
1963 * executed when the source mapping is already
1966 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
1968 ("vm_fault_copy_entry: main object missing page"));
1970 VM_OBJECT_RLOCK(backing_object);
1971 pindex += OFF_TO_IDX(object->backing_object_offset);
1972 if (object != dst_object)
1973 VM_OBJECT_RUNLOCK(object);
1974 object = backing_object;
1976 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
1978 if (object != dst_object) {
1980 * Allocate a page in the destination object.
1982 dst_m = vm_page_alloc(dst_object, (src_object ==
1983 dst_object ? src_pindex : 0) + dst_pindex,
1985 if (dst_m == NULL) {
1986 VM_OBJECT_WUNLOCK(dst_object);
1987 VM_OBJECT_RUNLOCK(object);
1988 vm_wait(dst_object);
1989 VM_OBJECT_WLOCK(dst_object);
1992 pmap_copy_page(src_m, dst_m);
1993 VM_OBJECT_RUNLOCK(object);
1994 dst_m->dirty = dst_m->valid = src_m->valid;
1997 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
1999 if (dst_m->pindex >= dst_object->size) {
2001 * We are upgrading. Index can occur
2002 * out of bounds if the object type is
2003 * vnode and the file was truncated.
2005 vm_page_xunbusy(dst_m);
2009 VM_OBJECT_WUNLOCK(dst_object);
2012 * Enter it in the pmap. If a wired, copy-on-write
2013 * mapping is being replaced by a write-enabled
2014 * mapping, then wire that new mapping.
2016 * The page can be invalid if the user called
2017 * msync(MS_INVALIDATE) or truncated the backing vnode
2018 * or shared memory object. In this case, do not
2019 * insert it into pmap, but still do the copy so that
2020 * all copies of the wired map entry have similar
2023 if (vm_page_all_valid(dst_m)) {
2024 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2025 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2029 * Mark it no longer busy, and put it on the active list.
2031 VM_OBJECT_WLOCK(dst_object);
2034 if (src_m != dst_m) {
2035 vm_page_unwire(src_m, PQ_INACTIVE);
2036 vm_page_wire(dst_m);
2038 KASSERT(vm_page_wired(dst_m),
2039 ("dst_m %p is not wired", dst_m));
2042 vm_page_activate(dst_m);
2044 vm_page_xunbusy(dst_m);
2046 VM_OBJECT_WUNLOCK(dst_object);
2048 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2049 vm_object_deallocate(src_object);
2054 * Block entry into the machine-independent layer's page fault handler by
2055 * the calling thread. Subsequent calls to vm_fault() by that thread will
2056 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
2057 * spurious page faults.
2060 vm_fault_disable_pagefaults(void)
2063 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2067 vm_fault_enable_pagefaults(int save)
2070 curthread_pflags_restore(save);