2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
10 * This code is derived from software contributed to Berkeley by
11 * The Mach Operating System project at Carnegie-Mellon University.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
44 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45 * All rights reserved.
47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
49 * Permission to use, copy, modify and distribute this software and
50 * its documentation is hereby granted, provided that both the copyright
51 * notice and this permission notice appear in all copies of the
52 * software, derivative works or modified versions, and any portions
53 * thereof, and that both notices appear in supporting documentation.
55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
59 * Carnegie Mellon requests users of this software to return to
61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
62 * School of Computer Science
63 * Carnegie Mellon University
64 * Pittsburgh PA 15213-3890
66 * any improvements or extensions that they make and grant Carnegie the
67 * rights to redistribute these changes.
69 * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
70 * $DragonFly: src/sys/vm/vm_fault.c,v 1.39 2007/01/11 10:15:21 dillon Exp $
74 * Page fault handling module.
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
81 #include <sys/vnode.h>
82 #include <sys/resourcevar.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vkernel.h>
85 #include <sys/sfbuf.h>
89 #include <vm/vm_param.h>
91 #include <vm/vm_map.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_pageout.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vnode_pager.h>
98 #include <vm/vm_extern.h>
100 #include <sys/thread2.h>
101 #include <vm/vm_page2.h>
103 #define VM_FAULT_READ_AHEAD 8
104 #define VM_FAULT_READ_BEHIND 7
105 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
113 vm_object_t first_object;
114 vm_prot_t first_prot;
116 vm_map_entry_t entry;
117 int lookup_still_valid;
126 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
127 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
128 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
129 static int vm_fault_ratelimit(struct vmspace *);
132 release_page(struct faultstate *fs)
134 vm_page_wakeup(fs->m);
135 vm_page_deactivate(fs->m);
140 unlock_map(struct faultstate *fs)
142 if (fs->lookup_still_valid) {
143 vm_map_lookup_done(fs->map, fs->entry, 0);
144 fs->lookup_still_valid = FALSE;
149 * Clean up after a successful call to vm_fault_object() so another call
150 * to vm_fault_object() can be made.
153 _cleanup_successful_fault(struct faultstate *fs, int relock)
155 if (fs->object != fs->first_object) {
156 vm_page_free(fs->first_m);
157 vm_object_pip_wakeup(fs->object);
160 fs->object = fs->first_object;
161 if (relock && fs->lookup_still_valid == FALSE) {
162 vm_map_lock_read(fs->map);
163 fs->lookup_still_valid = TRUE;
168 _unlock_things(struct faultstate *fs, int dealloc)
170 vm_object_pip_wakeup(fs->first_object);
171 _cleanup_successful_fault(fs, 0);
173 vm_object_deallocate(fs->first_object);
176 if (fs->vp != NULL) {
182 #define unlock_things(fs) _unlock_things(fs, 0)
183 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
184 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
189 * Determine if the pager for the current object *might* contain the page.
191 * We only need to try the pager if this is not a default object (default
192 * objects are zero-fill and have no real pager), and if we are not taking
193 * a wiring fault or if the FS entry is wired.
195 #define TRYPAGER(fs) \
196 (fs->object->type != OBJT_DEFAULT && \
197 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
202 * Handle a page fault occuring at the given address, requiring the given
203 * permissions, in the map specified. If successful, the page is inserted
204 * into the associated physical map.
206 * NOTE: The given address should be truncated to the proper page address.
208 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
209 * a standard error specifying why the fault is fatal is returned.
211 * The map in question must be referenced, and remains so.
212 * The caller may hold no locks.
215 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
218 vm_pindex_t first_pindex;
219 struct faultstate fs;
221 mycpu->gd_cnt.v_vm_faults++;
225 fs.fault_flags = fault_flags;
229 * Find the vm_map_entry representing the backing store and resolve
230 * the top level object and page index. This may have the side
231 * effect of executing a copy-on-write on the map entry and/or
232 * creating a shadow object, but will not COW any actual VM pages.
234 * On success fs.map is left read-locked and various other fields
235 * are initialized but not otherwise referenced or locked.
237 * NOTE! vm_map_lookup will try to upgrade the fault_type to
238 * VM_FAULT_WRITE if the map entry is a virtual page table and also
239 * writable, so we can set the 'A'accessed bit in the virtual page
243 result = vm_map_lookup(&fs.map, vaddr, fault_type,
244 &fs.entry, &fs.first_object,
245 &first_pindex, &fs.first_prot, &fs.wired);
248 * If the lookup failed or the map protections are incompatible,
249 * the fault generally fails. However, if the caller is trying
250 * to do a user wiring we have more work to do.
252 if (result != KERN_SUCCESS) {
253 if (result != KERN_PROTECTION_FAILURE)
255 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
259 * If we are user-wiring a r/w segment, and it is COW, then
260 * we need to do the COW operation. Note that we don't
261 * currently COW RO sections now, because it is NOT desirable
262 * to COW .text. We simply keep .text from ever being COW'ed
263 * and take the heat that one cannot debug wired .text sections.
265 result = vm_map_lookup(&fs.map, vaddr,
266 VM_PROT_READ|VM_PROT_WRITE|
267 VM_PROT_OVERRIDE_WRITE,
268 &fs.entry, &fs.first_object,
269 &first_pindex, &fs.first_prot,
271 if (result != KERN_SUCCESS)
275 * If we don't COW now, on a user wire, the user will never
276 * be able to write to the mapping. If we don't make this
277 * restriction, the bookkeeping would be nearly impossible.
279 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
280 fs.entry->max_protection &= ~VM_PROT_WRITE;
284 * fs.map is read-locked
286 * Misc checks. Save the map generation number to detect races.
288 fs.map_generation = fs.map->timestamp;
290 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
291 panic("vm_fault: fault on nofault entry, addr: %lx",
296 * A system map entry may return a NULL object. No object means
297 * no pager means an unrecoverable kernel fault.
299 if (fs.first_object == NULL) {
300 panic("vm_fault: unrecoverable fault at %p in entry %p",
301 (void *)vaddr, fs.entry);
305 * Make a reference to this object to prevent its disposal while we
306 * are messing with it. Once we have the reference, the map is free
307 * to be diddled. Since objects reference their shadows (and copies),
308 * they will stay around as well.
310 * Bump the paging-in-progress count to prevent size changes (e.g.
311 * truncation operations) during I/O. This must be done after
312 * obtaining the vnode lock in order to avoid possible deadlocks.
314 vm_object_reference(fs.first_object);
315 fs.vp = vnode_pager_lock(fs.first_object);
316 vm_object_pip_add(fs.first_object, 1);
318 fs.lookup_still_valid = TRUE;
320 fs.object = fs.first_object; /* so unlock_and_deallocate works */
323 * If the entry is wired we cannot change the page protection.
326 fault_type = fs.first_prot;
329 * The page we want is at (first_object, first_pindex), but if the
330 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
331 * page table to figure out the actual pindex.
333 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
336 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
337 result = vm_fault_vpagetable(&fs, &first_pindex,
338 fs.entry->aux.master_pde,
340 if (result == KERN_TRY_AGAIN)
342 if (result != KERN_SUCCESS)
347 * Now we have the actual (object, pindex), fault in the page. If
348 * vm_fault_object() fails it will unlock and deallocate the FS
349 * data. If it succeeds everything remains locked and fs->object
350 * will have an additinal PIP count if it is not equal to
353 * vm_fault_object will set fs->prot for the pmap operation. It is
354 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
355 * page can be safely written. However, it will force a read-only
356 * mapping for a read fault if the memory is managed by a virtual
359 result = vm_fault_object(&fs, first_pindex, fault_type);
361 if (result == KERN_TRY_AGAIN)
363 if (result != KERN_SUCCESS)
367 * On success vm_fault_object() does not unlock or deallocate, and fs.m
368 * will contain a busied page.
370 * Enter the page into the pmap and do pmap-related adjustments.
373 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
375 if (((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0) && (fs.wired == 0)) {
376 pmap_prefault(fs.map->pmap, vaddr, fs.entry);
379 vm_page_flag_clear(fs.m, PG_ZERO);
380 vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED);
383 * If the page is not wired down, then put it where the pageout daemon
386 if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
390 vm_page_unwire(fs.m, 1);
392 vm_page_activate(fs.m);
395 if (curthread->td_lwp) {
397 curthread->td_lwp->lwp_ru.ru_majflt++;
399 curthread->td_lwp->lwp_ru.ru_minflt++;
404 * Unlock everything, and return
406 vm_page_wakeup(fs.m);
407 vm_object_deallocate(fs.first_object);
409 return (KERN_SUCCESS);
413 * Fault in the specified virtual address in the current process map,
414 * returning a held VM page or NULL. See vm_fault_page() for more
418 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
422 m = vm_fault_page(&curproc->p_vmspace->vm_map, va,
423 fault_type, VM_FAULT_NORMAL, errorp);
428 * Fault in the specified virtual address in the specified map, doing all
429 * necessary manipulation of the object store and all necessary I/O. Return
430 * a held VM page or NULL, and set *errorp. The related pmap is not
433 * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
434 * and marked PG_REFERENCED as well.
437 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
438 int fault_flags, int *errorp)
441 vm_pindex_t first_pindex;
442 struct faultstate fs;
444 mycpu->gd_cnt.v_vm_faults++;
448 fs.fault_flags = fault_flags;
449 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
453 * Find the vm_map_entry representing the backing store and resolve
454 * the top level object and page index. This may have the side
455 * effect of executing a copy-on-write on the map entry and/or
456 * creating a shadow object, but will not COW any actual VM pages.
458 * On success fs.map is left read-locked and various other fields
459 * are initialized but not otherwise referenced or locked.
461 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
462 * if the map entry is a virtual page table and also writable,
463 * so we can set the 'A'accessed bit in the virtual page table entry.
466 result = vm_map_lookup(&fs.map, vaddr, fault_type,
467 &fs.entry, &fs.first_object,
468 &first_pindex, &fs.first_prot, &fs.wired);
470 if (result != KERN_SUCCESS) {
476 * fs.map is read-locked
478 * Misc checks. Save the map generation number to detect races.
480 fs.map_generation = fs.map->timestamp;
482 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
483 panic("vm_fault: fault on nofault entry, addr: %lx",
488 * A system map entry may return a NULL object. No object means
489 * no pager means an unrecoverable kernel fault.
491 if (fs.first_object == NULL) {
492 panic("vm_fault: unrecoverable fault at %p in entry %p",
493 (void *)vaddr, fs.entry);
497 * Make a reference to this object to prevent its disposal while we
498 * are messing with it. Once we have the reference, the map is free
499 * to be diddled. Since objects reference their shadows (and copies),
500 * they will stay around as well.
502 * Bump the paging-in-progress count to prevent size changes (e.g.
503 * truncation operations) during I/O. This must be done after
504 * obtaining the vnode lock in order to avoid possible deadlocks.
506 vm_object_reference(fs.first_object);
507 fs.vp = vnode_pager_lock(fs.first_object);
508 vm_object_pip_add(fs.first_object, 1);
510 fs.lookup_still_valid = TRUE;
512 fs.object = fs.first_object; /* so unlock_and_deallocate works */
515 * If the entry is wired we cannot change the page protection.
518 fault_type = fs.first_prot;
521 * The page we want is at (first_object, first_pindex), but if the
522 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
523 * page table to figure out the actual pindex.
525 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
528 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
529 result = vm_fault_vpagetable(&fs, &first_pindex,
530 fs.entry->aux.master_pde,
532 if (result == KERN_TRY_AGAIN)
534 if (result != KERN_SUCCESS) {
541 * Now we have the actual (object, pindex), fault in the page. If
542 * vm_fault_object() fails it will unlock and deallocate the FS
543 * data. If it succeeds everything remains locked and fs->object
544 * will have an additinal PIP count if it is not equal to
547 result = vm_fault_object(&fs, first_pindex, fault_type);
549 if (result == KERN_TRY_AGAIN)
551 if (result != KERN_SUCCESS) {
557 * On success vm_fault_object() does not unlock or deallocate, and fs.m
558 * will contain a busied page.
563 * Return a held page. We are not doing any pmap manipulation so do
564 * not set PG_MAPPED. However, adjust the page flags according to
565 * the fault type because the caller may not use a managed pmapping
566 * (so we don't want to lose the fact that the page will be dirtied
567 * if a write fault was specified).
570 vm_page_flag_clear(fs.m, PG_ZERO);
571 if (fault_type & VM_PROT_WRITE)
575 * Update the pmap. We really only have to do this if a COW
576 * occured to replace the read-only page with the new page. For
577 * now just do it unconditionally. XXX
579 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
580 vm_page_flag_set(fs.m, PG_REFERENCED|PG_MAPPED);
583 * Unbusy the page by activating it. It remains held and will not
586 vm_page_activate(fs.m);
588 if (curthread->td_lwp) {
590 curthread->td_lwp->lwp_ru.ru_majflt++;
592 curthread->td_lwp->lwp_ru.ru_minflt++;
597 * Unlock everything, and return the held page.
599 vm_page_wakeup(fs.m);
600 vm_object_deallocate(fs.first_object);
607 * Translate the virtual page number (first_pindex) that is relative
608 * to the address space into a logical page number that is relative to the
609 * backing object. Use the virtual page table pointed to by (vpte).
611 * This implements an N-level page table. Any level can terminate the
612 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
613 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
617 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
618 vpte_t vpte, int fault_type)
621 int vshift = 32 - PAGE_SHIFT; /* page index bits remaining */
622 int result = KERN_SUCCESS;
627 * We cannot proceed if the vpte is not valid, not readable
628 * for a read fault, or not writable for a write fault.
630 if ((vpte & VPTE_V) == 0) {
631 unlock_and_deallocate(fs);
632 return (KERN_FAILURE);
634 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
635 unlock_and_deallocate(fs);
636 return (KERN_FAILURE);
638 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
639 unlock_and_deallocate(fs);
640 return (KERN_FAILURE);
642 if ((vpte & VPTE_PS) || vshift == 0)
644 KKASSERT(vshift >= VPTE_PAGE_BITS);
647 * Get the page table page. Nominally we only read the page
648 * table, but since we are actively setting VPTE_M and VPTE_A,
649 * tell vm_fault_object() that we are writing it.
651 * There is currently no real need to optimize this.
653 result = vm_fault_object(fs, vpte >> PAGE_SHIFT,
654 VM_PROT_READ|VM_PROT_WRITE);
655 if (result != KERN_SUCCESS)
659 * Process the returned fs.m and look up the page table
660 * entry in the page table page.
662 vshift -= VPTE_PAGE_BITS;
663 sf = sf_buf_alloc(fs->m, SFB_CPUPRIVATE);
664 ptep = ((vpte_t *)sf_buf_kva(sf) +
665 ((*pindex >> vshift) & VPTE_PAGE_MASK));
669 * Page table write-back. If the vpte is valid for the
670 * requested operation, do a write-back to the page table.
672 * XXX VPTE_M is not set properly for page directory pages.
673 * It doesn't get set in the page directory if the page table
674 * is modified during a read access.
676 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
678 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
679 atomic_set_int(ptep, VPTE_M|VPTE_A);
680 vm_page_dirty(fs->m);
683 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
685 if ((vpte & VPTE_A) == 0) {
686 atomic_set_int(ptep, VPTE_A);
687 vm_page_dirty(fs->m);
691 vm_page_flag_set(fs->m, PG_REFERENCED);
692 vm_page_activate(fs->m);
693 vm_page_wakeup(fs->m);
694 cleanup_successful_fault(fs);
697 * Combine remaining address bits with the vpte.
699 *pindex = (vpte >> PAGE_SHIFT) +
700 (*pindex & ((1 << vshift) - 1));
701 return (KERN_SUCCESS);
706 * Do all operations required to fault-in (fs.first_object, pindex). Run
707 * through the shadow chain as necessary and do required COW or virtual
708 * copy operations. The caller has already fully resolved the vm_map_entry
709 * and, if appropriate, has created a copy-on-write layer. All we need to
710 * do is iterate the object chain.
712 * On failure (fs) is unlocked and deallocated and the caller may return or
713 * retry depending on the failure code. On success (fs) is NOT unlocked or
714 * deallocated, fs.m will contained a resolved, busied page, and fs.object
715 * will have an additional PIP count if it is not equal to fs.first_object.
719 vm_fault_object(struct faultstate *fs,
720 vm_pindex_t first_pindex, vm_prot_t fault_type)
722 vm_object_t next_object;
723 vm_page_t marray[VM_FAULT_READ];
727 fs->prot = fs->first_prot;
728 fs->object = fs->first_object;
729 pindex = first_pindex;
732 * If a read fault occurs we try to make the page writable if
733 * possible. There are three cases where we cannot make the
734 * page mapping writable:
736 * (1) The mapping is read-only or the VM object is read-only,
737 * fs->prot above will simply not have VM_PROT_WRITE set.
739 * (2) If the mapping is a virtual page table we need to be able
740 * to detect writes so we can set VPTE_M.
742 * (3) If the VM page is read-only or copy-on-write, upgrading would
743 * just result in an unnecessary COW fault.
745 * VM_PROT_VPAGED is set if faulting via a virtual page table and
746 * causes adjustments to the 'M'odify bit to also turn off write
747 * access to force a re-fault.
749 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
750 if ((fault_type & VM_PROT_WRITE) == 0)
751 fs->prot &= ~VM_PROT_WRITE;
756 * If the object is dead, we stop here
758 if (fs->object->flags & OBJ_DEAD) {
759 unlock_and_deallocate(fs);
760 return (KERN_PROTECTION_FAILURE);
764 * See if page is resident. spl protection is required
765 * to avoid an interrupt unbusy/free race against our
766 * lookup. We must hold the protection through a page
767 * allocation or busy.
770 fs->m = vm_page_lookup(fs->object, pindex);
774 * Wait/Retry if the page is busy. We have to do this
775 * if the page is busy via either PG_BUSY or
776 * vm_page_t->busy because the vm_pager may be using
777 * vm_page_t->busy for pageouts ( and even pageins if
778 * it is the vnode pager ), and we could end up trying
779 * to pagein and pageout the same page simultaneously.
781 * We can theoretically allow the busy case on a read
782 * fault if the page is marked valid, but since such
783 * pages are typically already pmap'd, putting that
784 * special case in might be more effort then it is
785 * worth. We cannot under any circumstances mess
786 * around with a vm_page_t->busy page except, perhaps,
789 if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
791 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
792 mycpu->gd_cnt.v_intrans++;
793 vm_object_deallocate(fs->first_object);
795 return (KERN_TRY_AGAIN);
799 * If reactivating a page from PQ_CACHE we may have
802 queue = fs->m->queue;
803 vm_page_unqueue_nowakeup(fs->m);
805 if ((queue - fs->m->pc) == PQ_CACHE &&
806 vm_page_count_severe()) {
807 vm_page_activate(fs->m);
808 unlock_and_deallocate(fs);
811 return (KERN_TRY_AGAIN);
815 * Mark page busy for other processes, and the
816 * pagedaemon. If it still isn't completely valid
817 * (readable), jump to readrest, else we found the
818 * page and can return.
820 * We can release the spl once we have marked the
826 if (((fs->m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
827 fs->m->object != &kernel_object) {
830 break; /* break to PAGE HAS BEEN FOUND */
834 * Page is not resident, If this is the search termination
835 * or the pager might contain the page, allocate a new page.
837 * NOTE: We are still in a critical section.
839 if (TRYPAGER(fs) || fs->object == fs->first_object) {
841 * If the page is beyond the object size we fail
843 if (pindex >= fs->object->size) {
845 unlock_and_deallocate(fs);
846 return (KERN_PROTECTION_FAILURE);
852 if (fs->didlimit == 0 && curproc != NULL) {
855 limticks = vm_fault_ratelimit(curproc->p_vmspace);
858 unlock_and_deallocate(fs);
859 tsleep(curproc, 0, "vmrate", limticks);
861 return (KERN_TRY_AGAIN);
866 * Allocate a new page for this object/offset pair.
869 if (!vm_page_count_severe()) {
870 fs->m = vm_page_alloc(fs->object, pindex,
871 (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
875 unlock_and_deallocate(fs);
877 return (KERN_TRY_AGAIN);
884 * We have found a valid page or we have allocated a new page.
885 * The page thus may not be valid or may not be entirely
888 * Attempt to fault-in the page if there is a chance that the
889 * pager has it, and potentially fault in additional pages
892 * We are NOT in splvm here and if TRYPAGER is true then
893 * fs.m will be non-NULL and will be PG_BUSY for us.
900 u_char behavior = vm_map_entry_behavior(fs->entry);
902 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
907 if (behind > VM_FAULT_READ_BEHIND)
908 behind = VM_FAULT_READ_BEHIND;
910 ahead = fs->object->size - pindex;
913 if (ahead > VM_FAULT_READ_AHEAD)
914 ahead = VM_FAULT_READ_AHEAD;
917 if ((fs->first_object->type != OBJT_DEVICE) &&
918 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
919 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
920 pindex >= fs->entry->lastr &&
921 pindex < fs->entry->lastr + VM_FAULT_READ))
923 vm_pindex_t firstpindex, tmppindex;
925 if (first_pindex < 2 * VM_FAULT_READ)
928 firstpindex = first_pindex - 2 * VM_FAULT_READ;
931 * note: partially valid pages cannot be
932 * included in the lookahead - NFS piecemeal
933 * writes will barf on it badly.
935 * spl protection is required to avoid races
936 * between the lookup and an interrupt
937 * unbusy/free sequence occuring prior to
941 for (tmppindex = first_pindex - 1;
942 tmppindex >= firstpindex;
947 mt = vm_page_lookup(fs->first_object, tmppindex);
948 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
951 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
956 vm_page_test_dirty(mt);
958 vm_page_protect(mt, VM_PROT_NONE);
959 vm_page_deactivate(mt);
971 * now we find out if any other pages should be paged
972 * in at this time this routine checks to see if the
973 * pages surrounding this fault reside in the same
974 * object as the page for this fault. If they do,
975 * then they are faulted in also into the object. The
976 * array "marray" returned contains an array of
977 * vm_page_t structs where one of them is the
978 * vm_page_t passed to the routine. The reqpage
979 * return value is the index into the marray for the
980 * vm_page_t passed to the routine.
982 * fs.m plus the additional pages are PG_BUSY'd.
984 faultcount = vm_fault_additional_pages(
985 fs->m, behind, ahead, marray, &reqpage);
988 * update lastr imperfectly (we do not know how much
989 * getpages will actually read), but good enough.
991 fs->entry->lastr = pindex + faultcount - behind;
994 * Call the pager to retrieve the data, if any, after
995 * releasing the lock on the map. We hold a ref on
996 * fs.object and the pages are PG_BUSY'd.
1001 rv = vm_pager_get_pages(fs->object, marray,
1002 faultcount, reqpage);
1007 if (rv == VM_PAGER_OK) {
1009 * Found the page. Leave it busy while we play
1014 * Relookup in case pager changed page. Pager
1015 * is responsible for disposition of old page
1018 * XXX other code segments do relookups too.
1019 * It's a bad abstraction that needs to be
1022 fs->m = vm_page_lookup(fs->object, pindex);
1023 if (fs->m == NULL) {
1024 unlock_and_deallocate(fs);
1025 return (KERN_TRY_AGAIN);
1029 break; /* break to PAGE HAS BEEN FOUND */
1033 * Remove the bogus page (which does not exist at this
1034 * object/offset); before doing so, we must get back
1035 * our object lock to preserve our invariant.
1037 * Also wake up any other process that may want to bring
1040 * If this is the top-level object, we must leave the
1041 * busy page to prevent another process from rushing
1042 * past us, and inserting the page in that object at
1043 * the same time that we are.
1045 if (rv == VM_PAGER_ERROR) {
1047 kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
1049 kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
1052 * Data outside the range of the pager or an I/O error
1055 * XXX - the check for kernel_map is a kludge to work
1056 * around having the machine panic on a kernel space
1057 * fault w/ I/O error.
1059 if (((fs->map != &kernel_map) && (rv == VM_PAGER_ERROR)) ||
1060 (rv == VM_PAGER_BAD)) {
1061 vm_page_free(fs->m);
1063 unlock_and_deallocate(fs);
1064 if (rv == VM_PAGER_ERROR)
1065 return (KERN_FAILURE);
1067 return (KERN_PROTECTION_FAILURE);
1070 if (fs->object != fs->first_object) {
1071 vm_page_free(fs->m);
1074 * XXX - we cannot just fall out at this
1075 * point, m has been freed and is invalid!
1081 * We get here if the object has a default pager (or unwiring)
1082 * or the pager doesn't have the page.
1084 if (fs->object == fs->first_object)
1085 fs->first_m = fs->m;
1088 * Move on to the next object. Lock the next object before
1089 * unlocking the current one.
1091 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1092 next_object = fs->object->backing_object;
1093 if (next_object == NULL) {
1095 * If there's no object left, fill the page in the top
1096 * object with zeros.
1098 if (fs->object != fs->first_object) {
1099 vm_object_pip_wakeup(fs->object);
1101 fs->object = fs->first_object;
1102 pindex = first_pindex;
1103 fs->m = fs->first_m;
1108 * Zero the page if necessary and mark it valid.
1110 if ((fs->m->flags & PG_ZERO) == 0) {
1111 vm_page_zero_fill(fs->m);
1113 mycpu->gd_cnt.v_ozfod++;
1115 mycpu->gd_cnt.v_zfod++;
1116 fs->m->valid = VM_PAGE_BITS_ALL;
1117 break; /* break to PAGE HAS BEEN FOUND */
1119 if (fs->object != fs->first_object) {
1120 vm_object_pip_wakeup(fs->object);
1122 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1123 fs->object = next_object;
1124 vm_object_pip_add(fs->object, 1);
1128 KASSERT((fs->m->flags & PG_BUSY) != 0,
1129 ("vm_fault: not busy after main loop"));
1132 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1137 * If the page is being written, but isn't already owned by the
1138 * top-level object, we have to copy it into a new page owned by the
1141 if (fs->object != fs->first_object) {
1143 * We only really need to copy if we want to write it.
1145 if (fault_type & VM_PROT_WRITE) {
1147 * This allows pages to be virtually copied from a
1148 * backing_object into the first_object, where the
1149 * backing object has no other refs to it, and cannot
1150 * gain any more refs. Instead of a bcopy, we just
1151 * move the page from the backing object to the
1152 * first object. Note that we must mark the page
1153 * dirty in the first object so that it will go out
1154 * to swap when needed.
1156 if (fs->map_generation == fs->map->timestamp &&
1158 * Only one shadow object
1160 (fs->object->shadow_count == 1) &&
1162 * No COW refs, except us
1164 (fs->object->ref_count == 1) &&
1166 * No one else can look this object up
1168 (fs->object->handle == NULL) &&
1170 * No other ways to look the object up
1172 ((fs->object->type == OBJT_DEFAULT) ||
1173 (fs->object->type == OBJT_SWAP)) &&
1175 * We don't chase down the shadow chain
1177 (fs->object == fs->first_object->backing_object) &&
1180 * grab the lock if we need to
1182 (fs->lookup_still_valid ||
1183 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
1186 fs->lookup_still_valid = 1;
1188 * get rid of the unnecessary page
1190 vm_page_protect(fs->first_m, VM_PROT_NONE);
1191 vm_page_free(fs->first_m);
1195 * grab the page and put it into the
1196 * process'es object. The page is
1197 * automatically made dirty.
1199 vm_page_rename(fs->m, fs->first_object, first_pindex);
1200 fs->first_m = fs->m;
1201 vm_page_busy(fs->first_m);
1203 mycpu->gd_cnt.v_cow_optim++;
1206 * Oh, well, lets copy it.
1208 vm_page_copy(fs->m, fs->first_m);
1213 * We no longer need the old page or object.
1219 * fs->object != fs->first_object due to above
1222 vm_object_pip_wakeup(fs->object);
1225 * Only use the new page below...
1228 mycpu->gd_cnt.v_cow_faults++;
1229 fs->m = fs->first_m;
1230 fs->object = fs->first_object;
1231 pindex = first_pindex;
1234 * If it wasn't a write fault avoid having to copy
1235 * the page by mapping it read-only.
1237 fs->prot &= ~VM_PROT_WRITE;
1242 * We may have had to unlock a map to do I/O. If we did then
1243 * lookup_still_valid will be FALSE. If the map generation count
1244 * also changed then all sorts of things could have happened while
1245 * we were doing the I/O and we need to retry.
1248 if (!fs->lookup_still_valid &&
1249 (fs->map->timestamp != fs->map_generation)) {
1251 unlock_and_deallocate(fs);
1252 return (KERN_TRY_AGAIN);
1256 * Put this page into the physical map. We had to do the unlock above
1257 * because pmap_enter may cause other faults. We don't put the page
1258 * back on the active queue until later so that the page-out daemon
1259 * won't find us (yet).
1261 if (fs->prot & VM_PROT_WRITE) {
1262 vm_page_flag_set(fs->m, PG_WRITEABLE);
1263 vm_object_set_writeable_dirty(fs->m->object);
1266 * If the fault is a write, we know that this page is being
1267 * written NOW so dirty it explicitly to save on
1268 * pmap_is_modified() calls later.
1270 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1271 * if the page is already dirty to prevent data written with
1272 * the expectation of being synced from not being synced.
1273 * Likewise if this entry does not request NOSYNC then make
1274 * sure the page isn't marked NOSYNC. Applications sharing
1275 * data should use the same flags to avoid ping ponging.
1277 * Also tell the backing pager, if any, that it should remove
1278 * any swap backing since the page is now dirty.
1280 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) {
1281 if (fs->m->dirty == 0)
1282 vm_page_flag_set(fs->m, PG_NOSYNC);
1284 vm_page_flag_clear(fs->m, PG_NOSYNC);
1286 if (fs->fault_flags & VM_FAULT_DIRTY) {
1288 vm_page_dirty(fs->m);
1289 vm_pager_page_unswapped(fs->m);
1295 * Page had better still be busy. We are still locked up and
1296 * fs->object will have another PIP reference if it is not equal
1297 * to fs->first_object.
1299 KASSERT(fs->m->flags & PG_BUSY,
1300 ("vm_fault: page %p not busy!", fs->m));
1303 * Sanity check: page must be completely valid or it is not fit to
1304 * map into user space. vm_pager_get_pages() ensures this.
1306 if (fs->m->valid != VM_PAGE_BITS_ALL) {
1307 vm_page_zero_invalid(fs->m, TRUE);
1308 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
1311 return (KERN_SUCCESS);
1315 * quick version of vm_fault
1318 vm_fault_quick(caddr_t v, int prot)
1322 if (prot & VM_PROT_WRITE)
1323 r = subyte(v, fubyte(v));
1330 * Wire down a range of virtual addresses in a map. The entry in question
1331 * should be marked in-transition and the map must be locked. We must
1332 * release the map temporarily while faulting-in the page to avoid a
1333 * deadlock. Note that the entry may be clipped while we are blocked but
1334 * will never be freed.
1337 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
1339 boolean_t fictitious;
1347 pmap = vm_map_pmap(map);
1348 start = entry->start;
1350 fictitious = entry->object.vm_object &&
1351 (entry->object.vm_object->type == OBJT_DEVICE);
1357 * We simulate a fault to get the page and enter it in the physical
1360 for (va = start; va < end; va += PAGE_SIZE) {
1362 rv = vm_fault(map, va, VM_PROT_READ,
1363 VM_FAULT_USER_WIRE);
1365 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1366 VM_FAULT_CHANGE_WIRING);
1369 while (va > start) {
1371 if ((pa = pmap_extract(pmap, va)) == 0)
1373 pmap_change_wiring(pmap, va, FALSE);
1375 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1382 return (KERN_SUCCESS);
1386 * Unwire a range of virtual addresses in a map. The map should be
1390 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
1392 boolean_t fictitious;
1399 pmap = vm_map_pmap(map);
1400 start = entry->start;
1402 fictitious = entry->object.vm_object &&
1403 (entry->object.vm_object->type == OBJT_DEVICE);
1406 * Since the pages are wired down, we must be able to get their
1407 * mappings from the physical map system.
1409 for (va = start; va < end; va += PAGE_SIZE) {
1410 pa = pmap_extract(pmap, va);
1412 pmap_change_wiring(pmap, va, FALSE);
1414 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1420 * Reduce the rate at which memory is allocated to a process based
1421 * on the perceived load on the VM system. As the load increases
1422 * the allocation burst rate goes down and the delay increases.
1424 * Rate limiting does not apply when faulting active or inactive
1425 * pages. When faulting 'cache' pages, rate limiting only applies
1426 * if the system currently has a severe page deficit.
1428 * XXX vm_pagesupply should be increased when a page is freed.
1430 * We sleep up to 1/10 of a second.
1433 vm_fault_ratelimit(struct vmspace *vmspace)
1435 if (vm_load_enable == 0)
1437 if (vmspace->vm_pagesupply > 0) {
1438 --vmspace->vm_pagesupply;
1442 if (vm_load_debug) {
1443 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
1445 (1000 - vm_load ) / 10, vm_load * hz / 10000,
1446 curproc->p_pid, curproc->p_comm);
1449 vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1450 return(vm_load * hz / 10000);
1455 * vm_fault_copy_entry
1457 * Copy all of the pages from a wired-down map entry to another.
1459 * In/out conditions:
1460 * The source and destination maps must be locked for write.
1461 * The source map entry must be wired down (or be a sharing map
1462 * entry corresponding to a main map entry that is wired down).
1466 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1467 vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
1469 vm_object_t dst_object;
1470 vm_object_t src_object;
1471 vm_ooffset_t dst_offset;
1472 vm_ooffset_t src_offset;
1482 src_object = src_entry->object.vm_object;
1483 src_offset = src_entry->offset;
1486 * Create the top-level object for the destination entry. (Doesn't
1487 * actually shadow anything - we copy the pages directly.)
1489 vm_map_entry_allocate_object(dst_entry);
1490 dst_object = dst_entry->object.vm_object;
1492 prot = dst_entry->max_protection;
1495 * Loop through all of the pages in the entry's range, copying each
1496 * one from the source object (it should be there) to the destination
1499 for (vaddr = dst_entry->start, dst_offset = 0;
1500 vaddr < dst_entry->end;
1501 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1504 * Allocate a page in the destination object
1507 dst_m = vm_page_alloc(dst_object,
1508 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1509 if (dst_m == NULL) {
1512 } while (dst_m == NULL);
1515 * Find the page in the source object, and copy it in.
1516 * (Because the source is wired down, the page will be in
1519 src_m = vm_page_lookup(src_object,
1520 OFF_TO_IDX(dst_offset + src_offset));
1522 panic("vm_fault_copy_wired: page missing");
1524 vm_page_copy(src_m, dst_m);
1527 * Enter it in the pmap...
1530 vm_page_flag_clear(dst_m, PG_ZERO);
1531 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1532 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED);
1535 * Mark it no longer busy, and put it on the active list.
1537 vm_page_activate(dst_m);
1538 vm_page_wakeup(dst_m);
1544 * This routine checks around the requested page for other pages that
1545 * might be able to be faulted in. This routine brackets the viable
1546 * pages for the pages to be paged in.
1549 * m, rbehind, rahead
1552 * marray (array of vm_page_t), reqpage (index of requested page)
1555 * number of pages in marray
1558 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
1559 vm_page_t *marray, int *reqpage)
1563 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1565 int cbehind, cahead;
1571 * we don't fault-ahead for device pager
1573 if (object->type == OBJT_DEVICE) {
1580 * if the requested page is not available, then give up now
1583 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1587 if ((cbehind == 0) && (cahead == 0)) {
1593 if (rahead > cahead) {
1597 if (rbehind > cbehind) {
1602 * try to do any readahead that we might have free pages for.
1604 if ((rahead + rbehind) >
1605 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) {
1606 pagedaemon_wakeup();
1613 * scan backward for the read behind pages -- in memory
1615 * Assume that if the page is not found an interrupt will not
1616 * create it. Theoretically interrupts can only remove (busy)
1617 * pages, not create new associations.
1620 if (rbehind > pindex) {
1624 startpindex = pindex - rbehind;
1628 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1629 if (vm_page_lookup( object, tpindex)) {
1630 startpindex = tpindex + 1;
1637 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1639 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1642 for (j = 0; j < i; j++) {
1643 vm_page_free(marray[j]);
1659 /* page offset of the required page */
1662 tpindex = pindex + 1;
1666 * scan forward for the read ahead pages
1668 endpindex = tpindex + rahead;
1669 if (endpindex > object->size)
1670 endpindex = object->size;
1673 for( ; tpindex < endpindex; i++, tpindex++) {
1675 if (vm_page_lookup(object, tpindex)) {
1679 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1688 /* return number of bytes of pages */