FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_fault.c
1 /*-
2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 *
9 *
10 * This code is derived from software contributed to Berkeley by
11 * The Mach Operating System project at Carnegie-Mellon University.
12 *
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
15 * are met:
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.
28 *
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
39 * SUCH DAMAGE.
40 *
41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
42 *
43 *
44 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45 * All rights reserved.
46 *
47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
48 *
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.
54 *
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.
58 *
59 * Carnegie Mellon requests users of this software to return to
60 *
61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
62 * School of Computer Science
63 * Carnegie Mellon University
64 * Pittsburgh PA 15213-3890
65 *
66 * any improvements or extensions that they make and grant Carnegie the
67 * rights to redistribute these changes.
68 */
69
70 /*
71 * Page fault handling module.
72 */
73
74 #include <sys/cdefs.h>
75 __FBSDID("$FreeBSD: releng/6.4/sys/vm/vm_fault.c 173367 2007-11-05 16:18:00Z ups $");
76
77 #include <sys/param.h>
78 #include <sys/systm.h>
79 #include <sys/kernel.h>
80 #include <sys/lock.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/resourcevar.h>
84 #include <sys/sysctl.h>
85 #include <sys/vmmeter.h>
86 #include <sys/vnode.h>
87
88 #include <vm/vm.h>
89 #include <vm/vm_param.h>
90 #include <vm/pmap.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>
99
100 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */
101
102 #define PFBAK 4
103 #define PFFOR 4
104 #define PAGEORDER_SIZE (PFBAK+PFFOR)
105
106 static int prefault_pageorder[] = {
107 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
108 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
109 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
110 -4 * PAGE_SIZE, 4 * PAGE_SIZE
111 };
112
113 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
114 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
115
116 #define VM_FAULT_READ_AHEAD 8
117 #define VM_FAULT_READ_BEHIND 7
118 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
119
120 struct faultstate {
121 vm_page_t m;
122 vm_object_t object;
123 vm_pindex_t pindex;
124 vm_page_t first_m;
125 vm_object_t first_object;
126 vm_pindex_t first_pindex;
127 vm_map_t map;
128 vm_map_entry_t entry;
129 int lookup_still_valid;
130 struct vnode *vp;
131 };
132
133 static __inline void
134 release_page(struct faultstate *fs)
135 {
136 vm_page_lock_queues();
137 vm_page_wakeup(fs->m);
138 vm_page_deactivate(fs->m);
139 vm_page_unlock_queues();
140 fs->m = NULL;
141 }
142
143 static __inline void
144 unlock_map(struct faultstate *fs)
145 {
146 if (fs->lookup_still_valid) {
147 vm_map_lookup_done(fs->map, fs->entry);
148 fs->lookup_still_valid = FALSE;
149 }
150 }
151
152 static void
153 unlock_and_deallocate(struct faultstate *fs)
154 {
155 boolean_t firstobjneedgiant;
156
157 vm_object_pip_wakeup(fs->object);
158 VM_OBJECT_UNLOCK(fs->object);
159 if (fs->object != fs->first_object) {
160 VM_OBJECT_LOCK(fs->first_object);
161 vm_page_lock_queues();
162 vm_page_free(fs->first_m);
163 vm_page_unlock_queues();
164 vm_object_pip_wakeup(fs->first_object);
165 VM_OBJECT_UNLOCK(fs->first_object);
166 fs->first_m = NULL;
167 }
168 firstobjneedgiant = (fs->first_object->flags & OBJ_NEEDGIANT) != 0;
169 vm_object_deallocate(fs->first_object);
170 unlock_map(fs);
171 if (fs->vp != NULL) {
172 int vfslocked;
173
174 vfslocked = VFS_LOCK_GIANT(fs->vp->v_mount);
175 vput(fs->vp);
176 fs->vp = NULL;
177 VFS_UNLOCK_GIANT(vfslocked);
178 }
179 if (firstobjneedgiant)
180 VM_UNLOCK_GIANT();
181 }
182
183 /*
184 * TRYPAGER - used by vm_fault to calculate whether the pager for the
185 * current object *might* contain the page.
186 *
187 * default objects are zero-fill, there is no real pager.
188 */
189 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
190 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
191
192 /*
193 * vm_fault:
194 *
195 * Handle a page fault occurring at the given address,
196 * requiring the given permissions, in the map specified.
197 * If successful, the page is inserted into the
198 * associated physical map.
199 *
200 * NOTE: the given address should be truncated to the
201 * proper page address.
202 *
203 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
204 * a standard error specifying why the fault is fatal is returned.
205 *
206 *
207 * The map in question must be referenced, and remains so.
208 * Caller may hold no locks.
209 */
210 int
211 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
212 int fault_flags)
213 {
214 vm_prot_t prot;
215 int is_first_object_locked, result;
216 boolean_t growstack, wired;
217 int map_generation;
218 vm_object_t next_object;
219 vm_page_t marray[VM_FAULT_READ];
220 int hardfault;
221 int faultcount;
222 struct faultstate fs;
223
224 hardfault = 0;
225 growstack = TRUE;
226 atomic_add_int(&cnt.v_vm_faults, 1);
227
228 RetryFault:;
229
230 /*
231 * Find the backing store object and offset into it to begin the
232 * search.
233 */
234 fs.map = map;
235 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
236 &fs.first_object, &fs.first_pindex, &prot, &wired);
237 if (result != KERN_SUCCESS) {
238 if (result != KERN_PROTECTION_FAILURE ||
239 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
240 if (growstack && result == KERN_INVALID_ADDRESS &&
241 map != kernel_map && curproc != NULL) {
242 result = vm_map_growstack(curproc, vaddr);
243 if (result != KERN_SUCCESS)
244 return (KERN_FAILURE);
245 growstack = FALSE;
246 goto RetryFault;
247 }
248 return (result);
249 }
250
251 /*
252 * If we are user-wiring a r/w segment, and it is COW, then
253 * we need to do the COW operation. Note that we don't COW
254 * currently RO sections now, because it is NOT desirable
255 * to COW .text. We simply keep .text from ever being COW'ed
256 * and take the heat that one cannot debug wired .text sections.
257 */
258 result = vm_map_lookup(&fs.map, vaddr,
259 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
260 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
261 if (result != KERN_SUCCESS)
262 return (result);
263
264 /*
265 * If we don't COW now, on a user wire, the user will never
266 * be able to write to the mapping. If we don't make this
267 * restriction, the bookkeeping would be nearly impossible.
268 *
269 * XXX The following assignment modifies the map without
270 * holding a write lock on it.
271 */
272 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
273 fs.entry->max_protection &= ~VM_PROT_WRITE;
274 }
275
276 map_generation = fs.map->timestamp;
277
278 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
279 panic("vm_fault: fault on nofault entry, addr: %lx",
280 (u_long)vaddr);
281 }
282
283 /*
284 * Make a reference to this object to prevent its disposal while we
285 * are messing with it. Once we have the reference, the map is free
286 * to be diddled. Since objects reference their shadows (and copies),
287 * they will stay around as well.
288 *
289 * Bump the paging-in-progress count to prevent size changes (e.g.
290 * truncation operations) during I/O. This must be done after
291 * obtaining the vnode lock in order to avoid possible deadlocks.
292 *
293 * XXX vnode_pager_lock() can block without releasing the map lock.
294 */
295 if (fs.first_object->flags & OBJ_NEEDGIANT)
296 mtx_lock(&Giant);
297 VM_OBJECT_LOCK(fs.first_object);
298 vm_object_reference_locked(fs.first_object);
299 fs.vp = vnode_pager_lock(fs.first_object);
300 KASSERT(fs.vp == NULL || !fs.map->system_map,
301 ("vm_fault: vnode-backed object mapped by system map"));
302 KASSERT((fs.first_object->flags & OBJ_NEEDGIANT) == 0 ||
303 !fs.map->system_map,
304 ("vm_fault: Object requiring giant mapped by system map"));
305 if (fs.first_object->flags & OBJ_NEEDGIANT && debug_mpsafevm)
306 mtx_unlock(&Giant);
307 vm_object_pip_add(fs.first_object, 1);
308
309 fs.lookup_still_valid = TRUE;
310
311 if (wired)
312 fault_type = prot;
313
314 fs.first_m = NULL;
315
316 /*
317 * Search for the page at object/offset.
318 */
319 fs.object = fs.first_object;
320 fs.pindex = fs.first_pindex;
321 while (TRUE) {
322 /*
323 * If the object is dead, we stop here
324 */
325 if (fs.object->flags & OBJ_DEAD) {
326 unlock_and_deallocate(&fs);
327 return (KERN_PROTECTION_FAILURE);
328 }
329
330 /*
331 * See if page is resident
332 */
333 fs.m = vm_page_lookup(fs.object, fs.pindex);
334 if (fs.m != NULL) {
335 int queue;
336
337 /*
338 * check for page-based copy on write.
339 * We check fs.object == fs.first_object so
340 * as to ensure the legacy COW mechanism is
341 * used when the page in question is part of
342 * a shadow object. Otherwise, vm_page_cowfault()
343 * removes the page from the backing object,
344 * which is not what we want.
345 */
346 vm_page_lock_queues();
347 if ((fs.m->cow) &&
348 (fault_type & VM_PROT_WRITE) &&
349 (fs.object == fs.first_object)) {
350 vm_page_cowfault(fs.m);
351 vm_page_unlock_queues();
352 unlock_and_deallocate(&fs);
353 goto RetryFault;
354 }
355
356 /*
357 * Wait/Retry if the page is busy. We have to do this
358 * if the page is busy via either PG_BUSY or
359 * vm_page_t->busy because the vm_pager may be using
360 * vm_page_t->busy for pageouts ( and even pageins if
361 * it is the vnode pager ), and we could end up trying
362 * to pagein and pageout the same page simultaneously.
363 *
364 * We can theoretically allow the busy case on a read
365 * fault if the page is marked valid, but since such
366 * pages are typically already pmap'd, putting that
367 * special case in might be more effort then it is
368 * worth. We cannot under any circumstances mess
369 * around with a vm_page_t->busy page except, perhaps,
370 * to pmap it.
371 */
372 if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
373 vm_page_unlock_queues();
374 VM_OBJECT_UNLOCK(fs.object);
375 if (fs.object != fs.first_object) {
376 VM_OBJECT_LOCK(fs.first_object);
377 vm_page_lock_queues();
378 vm_page_free(fs.first_m);
379 vm_page_unlock_queues();
380 vm_object_pip_wakeup(fs.first_object);
381 VM_OBJECT_UNLOCK(fs.first_object);
382 fs.first_m = NULL;
383 }
384 unlock_map(&fs);
385 if (fs.vp != NULL) {
386 int vfslck;
387
388 vfslck = VFS_LOCK_GIANT(fs.vp->v_mount);
389 vput(fs.vp);
390 fs.vp = NULL;
391 VFS_UNLOCK_GIANT(vfslck);
392 }
393 VM_OBJECT_LOCK(fs.object);
394 if (fs.m == vm_page_lookup(fs.object,
395 fs.pindex)) {
396 vm_page_lock_queues();
397 if (!vm_page_sleep_if_busy(fs.m, TRUE,
398 "vmpfw"))
399 vm_page_unlock_queues();
400 }
401 vm_object_pip_wakeup(fs.object);
402 VM_OBJECT_UNLOCK(fs.object);
403 atomic_add_int(&cnt.v_intrans, 1);
404 if (fs.first_object->flags & OBJ_NEEDGIANT)
405 VM_UNLOCK_GIANT();
406 vm_object_deallocate(fs.first_object);
407 goto RetryFault;
408 }
409 queue = fs.m->queue;
410
411 vm_pageq_remove_nowakeup(fs.m);
412
413 if ((queue - fs.m->pc) == PQ_CACHE) {
414 cnt.v_reactivated++;
415 if (vm_page_count_severe()) {
416 vm_page_activate(fs.m);
417 vm_page_unlock_queues();
418 unlock_and_deallocate(&fs);
419 VM_WAITPFAULT;
420 goto RetryFault;
421 }
422 }
423
424 /*
425 * Mark page busy for other processes, and the
426 * pagedaemon. If it still isn't completely valid
427 * (readable), jump to readrest, else break-out ( we
428 * found the page ).
429 */
430 vm_page_busy(fs.m);
431 vm_page_unlock_queues();
432 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
433 fs.m->object != kernel_object && fs.m->object != kmem_object) {
434 goto readrest;
435 }
436
437 break;
438 }
439
440 /*
441 * Page is not resident, If this is the search termination
442 * or the pager might contain the page, allocate a new page.
443 */
444 if (TRYPAGER || fs.object == fs.first_object) {
445 if (fs.pindex >= fs.object->size) {
446 unlock_and_deallocate(&fs);
447 return (KERN_PROTECTION_FAILURE);
448 }
449
450 /*
451 * Allocate a new page for this object/offset pair.
452 */
453 fs.m = NULL;
454 if (!vm_page_count_severe()) {
455 fs.m = vm_page_alloc(fs.object, fs.pindex,
456 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
457 }
458 if (fs.m == NULL) {
459 unlock_and_deallocate(&fs);
460 VM_WAITPFAULT;
461 goto RetryFault;
462 }
463 }
464
465 readrest:
466 /*
467 * We have found a valid page or we have allocated a new page.
468 * The page thus may not be valid or may not be entirely
469 * valid.
470 *
471 * Attempt to fault-in the page if there is a chance that the
472 * pager has it, and potentially fault in additional pages
473 * at the same time.
474 */
475 if (TRYPAGER) {
476 int rv;
477 int reqpage;
478 int ahead, behind;
479 u_char behavior = vm_map_entry_behavior(fs.entry);
480
481 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
482 ahead = 0;
483 behind = 0;
484 } else {
485 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
486 if (behind > VM_FAULT_READ_BEHIND)
487 behind = VM_FAULT_READ_BEHIND;
488
489 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
490 if (ahead > VM_FAULT_READ_AHEAD)
491 ahead = VM_FAULT_READ_AHEAD;
492 }
493 is_first_object_locked = FALSE;
494 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
495 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
496 fs.pindex >= fs.entry->lastr &&
497 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
498 (fs.first_object == fs.object ||
499 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
500 fs.first_object->type != OBJT_DEVICE) {
501 vm_pindex_t firstpindex, tmppindex;
502
503 if (fs.first_pindex < 2 * VM_FAULT_READ)
504 firstpindex = 0;
505 else
506 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
507
508 vm_page_lock_queues();
509 /*
510 * note: partially valid pages cannot be
511 * included in the lookahead - NFS piecemeal
512 * writes will barf on it badly.
513 */
514 for (tmppindex = fs.first_pindex - 1;
515 tmppindex >= firstpindex;
516 --tmppindex) {
517 vm_page_t mt;
518
519 mt = vm_page_lookup(fs.first_object, tmppindex);
520 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
521 break;
522 if (mt->busy ||
523 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
524 mt->hold_count ||
525 mt->wire_count)
526 continue;
527 pmap_remove_all(mt);
528 if (mt->dirty) {
529 vm_page_deactivate(mt);
530 } else {
531 vm_page_cache(mt);
532 }
533 }
534 vm_page_unlock_queues();
535 ahead += behind;
536 behind = 0;
537 }
538 if (is_first_object_locked)
539 VM_OBJECT_UNLOCK(fs.first_object);
540 /*
541 * now we find out if any other pages should be paged
542 * in at this time this routine checks to see if the
543 * pages surrounding this fault reside in the same
544 * object as the page for this fault. If they do,
545 * then they are faulted in also into the object. The
546 * array "marray" returned contains an array of
547 * vm_page_t structs where one of them is the
548 * vm_page_t passed to the routine. The reqpage
549 * return value is the index into the marray for the
550 * vm_page_t passed to the routine.
551 *
552 * fs.m plus the additional pages are PG_BUSY'd.
553 *
554 * XXX vm_fault_additional_pages() can block
555 * without releasing the map lock.
556 */
557 faultcount = vm_fault_additional_pages(
558 fs.m, behind, ahead, marray, &reqpage);
559
560 /*
561 * update lastr imperfectly (we do not know how much
562 * getpages will actually read), but good enough.
563 *
564 * XXX The following assignment modifies the map
565 * without holding a write lock on it.
566 */
567 fs.entry->lastr = fs.pindex + faultcount - behind;
568
569 /*
570 * Call the pager to retrieve the data, if any, after
571 * releasing the lock on the map. We hold a ref on
572 * fs.object and the pages are PG_BUSY'd.
573 */
574 unlock_map(&fs);
575
576 rv = faultcount ?
577 vm_pager_get_pages(fs.object, marray, faultcount,
578 reqpage) : VM_PAGER_FAIL;
579
580 if (rv == VM_PAGER_OK) {
581 /*
582 * Found the page. Leave it busy while we play
583 * with it.
584 */
585
586 /*
587 * Relookup in case pager changed page. Pager
588 * is responsible for disposition of old page
589 * if moved.
590 */
591 fs.m = vm_page_lookup(fs.object, fs.pindex);
592 if (!fs.m) {
593 unlock_and_deallocate(&fs);
594 goto RetryFault;
595 }
596
597 hardfault++;
598 break; /* break to PAGE HAS BEEN FOUND */
599 }
600 /*
601 * Remove the bogus page (which does not exist at this
602 * object/offset); before doing so, we must get back
603 * our object lock to preserve our invariant.
604 *
605 * Also wake up any other process that may want to bring
606 * in this page.
607 *
608 * If this is the top-level object, we must leave the
609 * busy page to prevent another process from rushing
610 * past us, and inserting the page in that object at
611 * the same time that we are.
612 */
613 if (rv == VM_PAGER_ERROR)
614 printf("vm_fault: pager read error, pid %d (%s)\n",
615 curproc->p_pid, curproc->p_comm);
616 /*
617 * Data outside the range of the pager or an I/O error
618 */
619 /*
620 * XXX - the check for kernel_map is a kludge to work
621 * around having the machine panic on a kernel space
622 * fault w/ I/O error.
623 */
624 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
625 (rv == VM_PAGER_BAD)) {
626 vm_page_lock_queues();
627 vm_page_free(fs.m);
628 vm_page_unlock_queues();
629 fs.m = NULL;
630 unlock_and_deallocate(&fs);
631 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
632 }
633 if (fs.object != fs.first_object) {
634 vm_page_lock_queues();
635 vm_page_free(fs.m);
636 vm_page_unlock_queues();
637 fs.m = NULL;
638 /*
639 * XXX - we cannot just fall out at this
640 * point, m has been freed and is invalid!
641 */
642 }
643 }
644
645 /*
646 * We get here if the object has default pager (or unwiring)
647 * or the pager doesn't have the page.
648 */
649 if (fs.object == fs.first_object)
650 fs.first_m = fs.m;
651
652 /*
653 * Move on to the next object. Lock the next object before
654 * unlocking the current one.
655 */
656 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
657 next_object = fs.object->backing_object;
658 if (next_object == NULL) {
659 /*
660 * If there's no object left, fill the page in the top
661 * object with zeros.
662 */
663 if (fs.object != fs.first_object) {
664 vm_object_pip_wakeup(fs.object);
665 VM_OBJECT_UNLOCK(fs.object);
666
667 fs.object = fs.first_object;
668 fs.pindex = fs.first_pindex;
669 fs.m = fs.first_m;
670 VM_OBJECT_LOCK(fs.object);
671 }
672 fs.first_m = NULL;
673
674 /*
675 * Zero the page if necessary and mark it valid.
676 */
677 if ((fs.m->flags & PG_ZERO) == 0) {
678 pmap_zero_page(fs.m);
679 } else {
680 atomic_add_int(&cnt.v_ozfod, 1);
681 }
682 atomic_add_int(&cnt.v_zfod, 1);
683 fs.m->valid = VM_PAGE_BITS_ALL;
684 break; /* break to PAGE HAS BEEN FOUND */
685 } else {
686 KASSERT(fs.object != next_object,
687 ("object loop %p", next_object));
688 VM_OBJECT_LOCK(next_object);
689 vm_object_pip_add(next_object, 1);
690 if (fs.object != fs.first_object)
691 vm_object_pip_wakeup(fs.object);
692 VM_OBJECT_UNLOCK(fs.object);
693 fs.object = next_object;
694 }
695 }
696
697 KASSERT((fs.m->flags & PG_BUSY) != 0,
698 ("vm_fault: not busy after main loop"));
699
700 /*
701 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
702 * is held.]
703 */
704
705 /*
706 * If the page is being written, but isn't already owned by the
707 * top-level object, we have to copy it into a new page owned by the
708 * top-level object.
709 */
710 if (fs.object != fs.first_object) {
711 /*
712 * We only really need to copy if we want to write it.
713 */
714 if (fault_type & VM_PROT_WRITE) {
715 /*
716 * This allows pages to be virtually copied from a
717 * backing_object into the first_object, where the
718 * backing object has no other refs to it, and cannot
719 * gain any more refs. Instead of a bcopy, we just
720 * move the page from the backing object to the
721 * first object. Note that we must mark the page
722 * dirty in the first object so that it will go out
723 * to swap when needed.
724 */
725 is_first_object_locked = FALSE;
726 if (
727 /*
728 * Only one shadow object
729 */
730 (fs.object->shadow_count == 1) &&
731 /*
732 * No COW refs, except us
733 */
734 (fs.object->ref_count == 1) &&
735 /*
736 * No one else can look this object up
737 */
738 (fs.object->handle == NULL) &&
739 /*
740 * No other ways to look the object up
741 */
742 ((fs.object->type == OBJT_DEFAULT) ||
743 (fs.object->type == OBJT_SWAP)) &&
744 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
745 /*
746 * We don't chase down the shadow chain
747 */
748 fs.object == fs.first_object->backing_object) {
749 vm_page_lock_queues();
750 /*
751 * get rid of the unnecessary page
752 */
753 pmap_remove_all(fs.first_m);
754 vm_page_free(fs.first_m);
755 /*
756 * grab the page and put it into the
757 * process'es object. The page is
758 * automatically made dirty.
759 */
760 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
761 vm_page_busy(fs.m);
762 vm_page_unlock_queues();
763 fs.first_m = fs.m;
764 fs.m = NULL;
765 atomic_add_int(&cnt.v_cow_optim, 1);
766 } else {
767 /*
768 * Oh, well, lets copy it.
769 */
770 pmap_copy_page(fs.m, fs.first_m);
771 fs.first_m->valid = VM_PAGE_BITS_ALL;
772 }
773 if (fs.m) {
774 /*
775 * We no longer need the old page or object.
776 */
777 release_page(&fs);
778 }
779 /*
780 * fs.object != fs.first_object due to above
781 * conditional
782 */
783 vm_object_pip_wakeup(fs.object);
784 VM_OBJECT_UNLOCK(fs.object);
785 /*
786 * Only use the new page below...
787 */
788 fs.object = fs.first_object;
789 fs.pindex = fs.first_pindex;
790 fs.m = fs.first_m;
791 if (!is_first_object_locked)
792 VM_OBJECT_LOCK(fs.object);
793 atomic_add_int(&cnt.v_cow_faults, 1);
794 } else {
795 prot &= ~VM_PROT_WRITE;
796 }
797 }
798
799 /*
800 * We must verify that the maps have not changed since our last
801 * lookup.
802 */
803 if (!fs.lookup_still_valid) {
804 vm_object_t retry_object;
805 vm_pindex_t retry_pindex;
806 vm_prot_t retry_prot;
807
808 if (!vm_map_trylock_read(fs.map)) {
809 release_page(&fs);
810 unlock_and_deallocate(&fs);
811 goto RetryFault;
812 }
813 fs.lookup_still_valid = TRUE;
814 if (fs.map->timestamp != map_generation) {
815 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
816 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
817
818 /*
819 * If we don't need the page any longer, put it on the inactive
820 * list (the easiest thing to do here). If no one needs it,
821 * pageout will grab it eventually.
822 */
823 if (result != KERN_SUCCESS) {
824 release_page(&fs);
825 unlock_and_deallocate(&fs);
826
827 /*
828 * If retry of map lookup would have blocked then
829 * retry fault from start.
830 */
831 if (result == KERN_FAILURE)
832 goto RetryFault;
833 return (result);
834 }
835 if ((retry_object != fs.first_object) ||
836 (retry_pindex != fs.first_pindex)) {
837 release_page(&fs);
838 unlock_and_deallocate(&fs);
839 goto RetryFault;
840 }
841
842 /*
843 * Check whether the protection has changed or the object has
844 * been copied while we left the map unlocked. Changing from
845 * read to write permission is OK - we leave the page
846 * write-protected, and catch the write fault. Changing from
847 * write to read permission means that we can't mark the page
848 * write-enabled after all.
849 */
850 prot &= retry_prot;
851 }
852 }
853 if (prot & VM_PROT_WRITE) {
854 vm_page_lock_queues();
855 vm_page_flag_set(fs.m, PG_WRITEABLE);
856 vm_object_set_writeable_dirty(fs.m->object);
857
858 /*
859 * If the fault is a write, we know that this page is being
860 * written NOW so dirty it explicitly to save on
861 * pmap_is_modified() calls later.
862 *
863 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
864 * if the page is already dirty to prevent data written with
865 * the expectation of being synced from not being synced.
866 * Likewise if this entry does not request NOSYNC then make
867 * sure the page isn't marked NOSYNC. Applications sharing
868 * data should use the same flags to avoid ping ponging.
869 *
870 * Also tell the backing pager, if any, that it should remove
871 * any swap backing since the page is now dirty.
872 */
873 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
874 if (fs.m->dirty == 0)
875 vm_page_flag_set(fs.m, PG_NOSYNC);
876 } else {
877 vm_page_flag_clear(fs.m, PG_NOSYNC);
878 }
879 vm_page_unlock_queues();
880 if (fault_flags & VM_FAULT_DIRTY) {
881 vm_page_dirty(fs.m);
882 vm_pager_page_unswapped(fs.m);
883 }
884 }
885
886 /*
887 * Page had better still be busy
888 */
889 KASSERT(fs.m->flags & PG_BUSY,
890 ("vm_fault: page %p not busy!", fs.m));
891 /*
892 * Sanity check: page must be completely valid or it is not fit to
893 * map into user space. vm_pager_get_pages() ensures this.
894 */
895 if (fs.m->valid != VM_PAGE_BITS_ALL) {
896 vm_page_zero_invalid(fs.m, TRUE);
897 printf("Warning: page %p partially invalid on fault\n", fs.m);
898 }
899 VM_OBJECT_UNLOCK(fs.object);
900
901 /*
902 * Put this page into the physical map. We had to do the unlock above
903 * because pmap_enter() may sleep. We don't put the page
904 * back on the active queue until later so that the pageout daemon
905 * won't find it (yet).
906 */
907 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
908 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
909 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
910 }
911 VM_OBJECT_LOCK(fs.object);
912 vm_page_lock_queues();
913 vm_page_flag_set(fs.m, PG_REFERENCED);
914
915 /*
916 * If the page is not wired down, then put it where the pageout daemon
917 * can find it.
918 */
919 if (fault_flags & VM_FAULT_WIRE_MASK) {
920 if (wired)
921 vm_page_wire(fs.m);
922 else
923 vm_page_unwire(fs.m, 1);
924 } else {
925 vm_page_activate(fs.m);
926 }
927 vm_page_wakeup(fs.m);
928 vm_page_unlock_queues();
929
930 /*
931 * Unlock everything, and return
932 */
933 unlock_and_deallocate(&fs);
934 PROC_LOCK(curproc);
935 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
936 if (hardfault) {
937 curproc->p_stats->p_ru.ru_majflt++;
938 } else {
939 curproc->p_stats->p_ru.ru_minflt++;
940 }
941 }
942 PROC_UNLOCK(curproc);
943
944 return (KERN_SUCCESS);
945 }
946
947 /*
948 * vm_fault_prefault provides a quick way of clustering
949 * pagefaults into a processes address space. It is a "cousin"
950 * of vm_map_pmap_enter, except it runs at page fault time instead
951 * of mmap time.
952 */
953 static void
954 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
955 {
956 int i;
957 vm_offset_t addr, starta;
958 vm_pindex_t pindex;
959 vm_page_t m;
960 vm_object_t object;
961
962 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
963 return;
964
965 object = entry->object.vm_object;
966
967 starta = addra - PFBAK * PAGE_SIZE;
968 if (starta < entry->start) {
969 starta = entry->start;
970 } else if (starta > addra) {
971 starta = 0;
972 }
973
974 for (i = 0; i < PAGEORDER_SIZE; i++) {
975 vm_object_t backing_object, lobject;
976
977 addr = addra + prefault_pageorder[i];
978 if (addr > addra + (PFFOR * PAGE_SIZE))
979 addr = 0;
980
981 if (addr < starta || addr >= entry->end)
982 continue;
983
984 if (!pmap_is_prefaultable(pmap, addr))
985 continue;
986
987 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
988 lobject = object;
989 VM_OBJECT_LOCK(lobject);
990 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
991 lobject->type == OBJT_DEFAULT &&
992 (backing_object = lobject->backing_object) != NULL) {
993 if (lobject->backing_object_offset & PAGE_MASK)
994 break;
995 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
996 VM_OBJECT_LOCK(backing_object);
997 VM_OBJECT_UNLOCK(lobject);
998 lobject = backing_object;
999 }
1000 /*
1001 * give-up when a page is not in memory
1002 */
1003 if (m == NULL) {
1004 VM_OBJECT_UNLOCK(lobject);
1005 break;
1006 }
1007 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1008 (m->busy == 0) &&
1009 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
1010
1011 vm_page_lock_queues();
1012 if ((m->queue - m->pc) == PQ_CACHE)
1013 vm_page_deactivate(m);
1014 pmap_enter_quick(pmap, addr, m, entry->protection);
1015 vm_page_unlock_queues();
1016 }
1017 VM_OBJECT_UNLOCK(lobject);
1018 }
1019 }
1020
1021 /*
1022 * vm_fault_quick:
1023 *
1024 * Ensure that the requested virtual address, which may be in userland,
1025 * is valid. Fault-in the page if necessary. Return -1 on failure.
1026 */
1027 int
1028 vm_fault_quick(caddr_t v, int prot)
1029 {
1030 int r;
1031
1032 if (prot & VM_PROT_WRITE)
1033 r = subyte(v, fubyte(v));
1034 else
1035 r = fubyte(v);
1036 return(r);
1037 }
1038
1039 /*
1040 * vm_fault_wire:
1041 *
1042 * Wire down a range of virtual addresses in a map.
1043 */
1044 int
1045 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1046 boolean_t user_wire, boolean_t fictitious)
1047 {
1048 vm_offset_t va;
1049 int rv;
1050
1051 /*
1052 * We simulate a fault to get the page and enter it in the physical
1053 * map. For user wiring, we only ask for read access on currently
1054 * read-only sections.
1055 */
1056 for (va = start; va < end; va += PAGE_SIZE) {
1057 rv = vm_fault(map, va,
1058 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1059 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1060 if (rv) {
1061 if (va != start)
1062 vm_fault_unwire(map, start, va, fictitious);
1063 return (rv);
1064 }
1065 }
1066 return (KERN_SUCCESS);
1067 }
1068
1069 /*
1070 * vm_fault_unwire:
1071 *
1072 * Unwire a range of virtual addresses in a map.
1073 */
1074 void
1075 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1076 boolean_t fictitious)
1077 {
1078 vm_paddr_t pa;
1079 vm_offset_t va;
1080 pmap_t pmap;
1081
1082 pmap = vm_map_pmap(map);
1083
1084 /*
1085 * Since the pages are wired down, we must be able to get their
1086 * mappings from the physical map system.
1087 */
1088 for (va = start; va < end; va += PAGE_SIZE) {
1089 pa = pmap_extract(pmap, va);
1090 if (pa != 0) {
1091 pmap_change_wiring(pmap, va, FALSE);
1092 if (!fictitious) {
1093 vm_page_lock_queues();
1094 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1095 vm_page_unlock_queues();
1096 }
1097 }
1098 }
1099 }
1100
1101 /*
1102 * Routine:
1103 * vm_fault_copy_entry
1104 * Function:
1105 * Copy all of the pages from a wired-down map entry to another.
1106 *
1107 * In/out conditions:
1108 * The source and destination maps must be locked for write.
1109 * The source map entry must be wired down (or be a sharing map
1110 * entry corresponding to a main map entry that is wired down).
1111 */
1112 void
1113 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1114 vm_map_t dst_map;
1115 vm_map_t src_map;
1116 vm_map_entry_t dst_entry;
1117 vm_map_entry_t src_entry;
1118 {
1119 vm_object_t backing_object, dst_object, object;
1120 vm_object_t src_object;
1121 vm_ooffset_t dst_offset;
1122 vm_ooffset_t src_offset;
1123 vm_pindex_t pindex;
1124 vm_prot_t prot;
1125 vm_offset_t vaddr;
1126 vm_page_t dst_m;
1127 vm_page_t src_m;
1128
1129 #ifdef lint
1130 src_map++;
1131 #endif /* lint */
1132
1133 src_object = src_entry->object.vm_object;
1134 src_offset = src_entry->offset;
1135
1136 /*
1137 * Create the top-level object for the destination entry. (Doesn't
1138 * actually shadow anything - we copy the pages directly.)
1139 */
1140 dst_object = vm_object_allocate(OBJT_DEFAULT,
1141 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1142
1143 VM_OBJECT_LOCK(dst_object);
1144 dst_entry->object.vm_object = dst_object;
1145 dst_entry->offset = 0;
1146
1147 prot = dst_entry->max_protection;
1148
1149 /*
1150 * Loop through all of the pages in the entry's range, copying each
1151 * one from the source object (it should be there) to the destination
1152 * object.
1153 */
1154 for (vaddr = dst_entry->start, dst_offset = 0;
1155 vaddr < dst_entry->end;
1156 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1157
1158 /*
1159 * Allocate a page in the destination object
1160 */
1161 do {
1162 dst_m = vm_page_alloc(dst_object,
1163 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1164 if (dst_m == NULL) {
1165 VM_OBJECT_UNLOCK(dst_object);
1166 VM_WAIT;
1167 VM_OBJECT_LOCK(dst_object);
1168 }
1169 } while (dst_m == NULL);
1170
1171 /*
1172 * Find the page in the source object, and copy it in.
1173 * (Because the source is wired down, the page will be in
1174 * memory.)
1175 */
1176 VM_OBJECT_LOCK(src_object);
1177 object = src_object;
1178 pindex = 0;
1179 while ((src_m = vm_page_lookup(object, pindex +
1180 OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1181 (src_entry->protection & VM_PROT_WRITE) == 0 &&
1182 (backing_object = object->backing_object) != NULL) {
1183 /*
1184 * Allow fallback to backing objects if we are reading.
1185 */
1186 VM_OBJECT_LOCK(backing_object);
1187 pindex += OFF_TO_IDX(object->backing_object_offset);
1188 VM_OBJECT_UNLOCK(object);
1189 object = backing_object;
1190 }
1191 if (src_m == NULL)
1192 panic("vm_fault_copy_wired: page missing");
1193 pmap_copy_page(src_m, dst_m);
1194 VM_OBJECT_UNLOCK(object);
1195 dst_m->valid = VM_PAGE_BITS_ALL;
1196 VM_OBJECT_UNLOCK(dst_object);
1197
1198 /*
1199 * Enter it in the pmap...
1200 */
1201 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1202 VM_OBJECT_LOCK(dst_object);
1203 vm_page_lock_queues();
1204 if ((prot & VM_PROT_WRITE) != 0)
1205 vm_page_flag_set(dst_m, PG_WRITEABLE);
1206
1207 /*
1208 * Mark it no longer busy, and put it on the active list.
1209 */
1210 vm_page_activate(dst_m);
1211 vm_page_wakeup(dst_m);
1212 vm_page_unlock_queues();
1213 }
1214 VM_OBJECT_UNLOCK(dst_object);
1215 }
1216
1217
1218 /*
1219 * This routine checks around the requested page for other pages that
1220 * might be able to be faulted in. This routine brackets the viable
1221 * pages for the pages to be paged in.
1222 *
1223 * Inputs:
1224 * m, rbehind, rahead
1225 *
1226 * Outputs:
1227 * marray (array of vm_page_t), reqpage (index of requested page)
1228 *
1229 * Return value:
1230 * number of pages in marray
1231 *
1232 * This routine can't block.
1233 */
1234 static int
1235 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1236 vm_page_t m;
1237 int rbehind;
1238 int rahead;
1239 vm_page_t *marray;
1240 int *reqpage;
1241 {
1242 int i,j;
1243 vm_object_t object;
1244 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1245 vm_page_t rtm;
1246 int cbehind, cahead;
1247
1248 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1249
1250 object = m->object;
1251 pindex = m->pindex;
1252
1253 /*
1254 * if the requested page is not available, then give up now
1255 */
1256 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1257 return 0;
1258 }
1259
1260 if ((cbehind == 0) && (cahead == 0)) {
1261 *reqpage = 0;
1262 marray[0] = m;
1263 return 1;
1264 }
1265
1266 if (rahead > cahead) {
1267 rahead = cahead;
1268 }
1269
1270 if (rbehind > cbehind) {
1271 rbehind = cbehind;
1272 }
1273
1274 /*
1275 * scan backward for the read behind pages -- in memory
1276 */
1277 if (pindex > 0) {
1278 if (rbehind > pindex) {
1279 rbehind = pindex;
1280 startpindex = 0;
1281 } else {
1282 startpindex = pindex - rbehind;
1283 }
1284
1285 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1286 if (vm_page_lookup(object, tpindex)) {
1287 startpindex = tpindex + 1;
1288 break;
1289 }
1290 if (tpindex == 0)
1291 break;
1292 }
1293
1294 /* tpindex is unsigned; beware of numeric underflow. */
1295 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1296 tpindex < pindex; i++, tpindex--) {
1297
1298 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1299 if (rtm == NULL) {
1300 /*
1301 * Shift the allocated pages to the
1302 * beginning of the array.
1303 */
1304 for (j = 0; j < i; j++) {
1305 marray[j] = marray[j + tpindex + 1 -
1306 startpindex];
1307 }
1308 break;
1309 }
1310
1311 marray[tpindex - startpindex] = rtm;
1312 }
1313 } else {
1314 startpindex = 0;
1315 i = 0;
1316 }
1317
1318 marray[i] = m;
1319 /* page offset of the required page */
1320 *reqpage = i;
1321
1322 tpindex = pindex + 1;
1323 i++;
1324
1325 /*
1326 * scan forward for the read ahead pages
1327 */
1328 endpindex = tpindex + rahead;
1329 if (endpindex > object->size)
1330 endpindex = object->size;
1331
1332 for (; tpindex < endpindex; i++, tpindex++) {
1333
1334 if (vm_page_lookup(object, tpindex)) {
1335 break;
1336 }
1337
1338 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1339 if (rtm == NULL) {
1340 break;
1341 }
1342
1343 marray[i] = rtm;
1344 }
1345
1346 /* return number of pages */
1347 return i;
1348 }
Cache object: 3f82cbdc06424686ece5a933b91aae50
|