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/5.2/sys/vm/vm_fault.c 122383 2003-11-10 00:44:00Z mini $");
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 #define PFBAK 4
101 #define PFFOR 4
102 #define PAGEORDER_SIZE (PFBAK+PFFOR)
103
104 static int prefault_pageorder[] = {
105 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
106 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
107 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
108 -4 * PAGE_SIZE, 4 * PAGE_SIZE
109 };
110
111 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
112 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
113
114 #define VM_FAULT_READ_AHEAD 8
115 #define VM_FAULT_READ_BEHIND 7
116 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
117
118 struct faultstate {
119 vm_page_t m;
120 vm_object_t object;
121 vm_pindex_t pindex;
122 vm_page_t first_m;
123 vm_object_t first_object;
124 vm_pindex_t first_pindex;
125 vm_map_t map;
126 vm_map_entry_t entry;
127 int lookup_still_valid;
128 struct vnode *vp;
129 };
130
131 static __inline void
132 release_page(struct faultstate *fs)
133 {
134 vm_page_lock_queues();
135 vm_page_wakeup(fs->m);
136 vm_page_deactivate(fs->m);
137 vm_page_unlock_queues();
138 fs->m = NULL;
139 }
140
141 static __inline void
142 unlock_map(struct faultstate *fs)
143 {
144 if (fs->lookup_still_valid) {
145 vm_map_lookup_done(fs->map, fs->entry);
146 fs->lookup_still_valid = FALSE;
147 }
148 }
149
150 static void
151 _unlock_things(struct faultstate *fs, int dealloc)
152 {
153
154 vm_object_pip_wakeup(fs->object);
155 VM_OBJECT_UNLOCK(fs->object);
156 if (fs->object != fs->first_object) {
157 VM_OBJECT_LOCK(fs->first_object);
158 vm_page_lock_queues();
159 vm_page_free(fs->first_m);
160 vm_page_unlock_queues();
161 vm_object_pip_wakeup(fs->first_object);
162 VM_OBJECT_UNLOCK(fs->first_object);
163 fs->first_m = NULL;
164 }
165 if (dealloc) {
166 vm_object_deallocate(fs->first_object);
167 }
168 unlock_map(fs);
169 if (fs->vp != NULL) {
170 vput(fs->vp);
171 fs->vp = NULL;
172 }
173 }
174
175 #define unlock_things(fs) _unlock_things(fs, 0)
176 #define unlock_and_deallocate(fs) _unlock_things(fs, 1)
177
178 /*
179 * TRYPAGER - used by vm_fault to calculate whether the pager for the
180 * current object *might* contain the page.
181 *
182 * default objects are zero-fill, there is no real pager.
183 */
184 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
185 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
186
187 /*
188 * vm_fault:
189 *
190 * Handle a page fault occurring at the given address,
191 * requiring the given permissions, in the map specified.
192 * If successful, the page is inserted into the
193 * associated physical map.
194 *
195 * NOTE: the given address should be truncated to the
196 * proper page address.
197 *
198 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
199 * a standard error specifying why the fault is fatal is returned.
200 *
201 *
202 * The map in question must be referenced, and remains so.
203 * Caller may hold no locks.
204 */
205 int
206 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
207 int fault_flags)
208 {
209 vm_prot_t prot;
210 int is_first_object_locked, result;
211 boolean_t growstack, wired;
212 int map_generation;
213 vm_object_t next_object;
214 vm_page_t marray[VM_FAULT_READ];
215 int hardfault;
216 int faultcount;
217 struct faultstate fs;
218
219 hardfault = 0;
220 growstack = TRUE;
221 atomic_add_int(&cnt.v_vm_faults, 1);
222
223 mtx_lock(&Giant);
224 RetryFault:;
225
226 /*
227 * Find the backing store object and offset into it to begin the
228 * search.
229 */
230 fs.map = map;
231 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
232 &fs.first_object, &fs.first_pindex, &prot, &wired);
233 if (result != KERN_SUCCESS) {
234 if (result != KERN_PROTECTION_FAILURE ||
235 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
236 if (growstack && result == KERN_INVALID_ADDRESS &&
237 map != kernel_map && curproc != NULL) {
238 result = vm_map_growstack(curproc, vaddr);
239 if (result != KERN_SUCCESS) {
240 mtx_unlock(&Giant);
241 return (KERN_FAILURE);
242 }
243 growstack = FALSE;
244 goto RetryFault;
245 }
246 mtx_unlock(&Giant);
247 return (result);
248 }
249
250 /*
251 * If we are user-wiring a r/w segment, and it is COW, then
252 * we need to do the COW operation. Note that we don't COW
253 * currently RO sections now, because it is NOT desirable
254 * to COW .text. We simply keep .text from ever being COW'ed
255 * and take the heat that one cannot debug wired .text sections.
256 */
257 result = vm_map_lookup(&fs.map, vaddr,
258 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
259 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
260 if (result != KERN_SUCCESS) {
261 mtx_unlock(&Giant);
262 return (result);
263 }
264
265 /*
266 * If we don't COW now, on a user wire, the user will never
267 * be able to write to the mapping. If we don't make this
268 * restriction, the bookkeeping would be nearly impossible.
269 *
270 * XXX The following assignment modifies the map without
271 * holding a write lock on it.
272 */
273 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
274 fs.entry->max_protection &= ~VM_PROT_WRITE;
275 }
276
277 map_generation = fs.map->timestamp;
278
279 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
280 panic("vm_fault: fault on nofault entry, addr: %lx",
281 (u_long)vaddr);
282 }
283
284 /*
285 * Make a reference to this object to prevent its disposal while we
286 * are messing with it. Once we have the reference, the map is free
287 * to be diddled. Since objects reference their shadows (and copies),
288 * they will stay around as well.
289 *
290 * Bump the paging-in-progress count to prevent size changes (e.g.
291 * truncation operations) during I/O. This must be done after
292 * obtaining the vnode lock in order to avoid possible deadlocks.
293 *
294 * XXX vnode_pager_lock() can block without releasing the map lock.
295 */
296 vm_object_reference(fs.first_object);
297 VM_OBJECT_LOCK(fs.first_object);
298 fs.vp = vnode_pager_lock(fs.first_object);
299 vm_object_pip_add(fs.first_object, 1);
300
301 fs.lookup_still_valid = TRUE;
302
303 if (wired)
304 fault_type = prot;
305
306 fs.first_m = NULL;
307
308 /*
309 * Search for the page at object/offset.
310 */
311 fs.object = fs.first_object;
312 fs.pindex = fs.first_pindex;
313 while (TRUE) {
314 /*
315 * If the object is dead, we stop here
316 */
317 if (fs.object->flags & OBJ_DEAD) {
318 unlock_and_deallocate(&fs);
319 mtx_unlock(&Giant);
320 return (KERN_PROTECTION_FAILURE);
321 }
322
323 /*
324 * See if page is resident
325 */
326 fs.m = vm_page_lookup(fs.object, fs.pindex);
327 if (fs.m != NULL) {
328 int queue, s;
329
330 /*
331 * check for page-based copy on write.
332 * We check fs.object == fs.first_object so
333 * as to ensure the legacy COW mechanism is
334 * used when the page in question is part of
335 * a shadow object. Otherwise, vm_page_cowfault()
336 * removes the page from the backing object,
337 * which is not what we want.
338 */
339 vm_page_lock_queues();
340 if ((fs.m->cow) &&
341 (fault_type & VM_PROT_WRITE) &&
342 (fs.object == fs.first_object)) {
343 s = splvm();
344 vm_page_cowfault(fs.m);
345 splx(s);
346 vm_page_unlock_queues();
347 unlock_and_deallocate(&fs);
348 goto RetryFault;
349 }
350
351 /*
352 * Wait/Retry if the page is busy. We have to do this
353 * if the page is busy via either PG_BUSY or
354 * vm_page_t->busy because the vm_pager may be using
355 * vm_page_t->busy for pageouts ( and even pageins if
356 * it is the vnode pager ), and we could end up trying
357 * to pagein and pageout the same page simultaneously.
358 *
359 * We can theoretically allow the busy case on a read
360 * fault if the page is marked valid, but since such
361 * pages are typically already pmap'd, putting that
362 * special case in might be more effort then it is
363 * worth. We cannot under any circumstances mess
364 * around with a vm_page_t->busy page except, perhaps,
365 * to pmap it.
366 */
367 if ((fs.m->flags & PG_BUSY) || fs.m->busy) {
368 vm_page_unlock_queues();
369 unlock_things(&fs);
370 vm_page_lock_queues();
371 if (!vm_page_sleep_if_busy(fs.m, TRUE, "vmpfw"))
372 vm_page_unlock_queues();
373 cnt.v_intrans++;
374 vm_object_deallocate(fs.first_object);
375 goto RetryFault;
376 }
377 queue = fs.m->queue;
378
379 s = splvm();
380 vm_pageq_remove_nowakeup(fs.m);
381 splx(s);
382
383 if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) {
384 vm_page_activate(fs.m);
385 vm_page_unlock_queues();
386 unlock_and_deallocate(&fs);
387 VM_WAITPFAULT;
388 goto RetryFault;
389 }
390
391 /*
392 * Mark page busy for other processes, and the
393 * pagedaemon. If it still isn't completely valid
394 * (readable), jump to readrest, else break-out ( we
395 * found the page ).
396 */
397 vm_page_busy(fs.m);
398 vm_page_unlock_queues();
399 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
400 fs.m->object != kernel_object && fs.m->object != kmem_object) {
401 goto readrest;
402 }
403
404 break;
405 }
406
407 /*
408 * Page is not resident, If this is the search termination
409 * or the pager might contain the page, allocate a new page.
410 */
411 if (TRYPAGER || fs.object == fs.first_object) {
412 if (fs.pindex >= fs.object->size) {
413 unlock_and_deallocate(&fs);
414 mtx_unlock(&Giant);
415 return (KERN_PROTECTION_FAILURE);
416 }
417
418 /*
419 * Allocate a new page for this object/offset pair.
420 */
421 fs.m = NULL;
422 if (!vm_page_count_severe()) {
423 fs.m = vm_page_alloc(fs.object, fs.pindex,
424 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
425 }
426 if (fs.m == NULL) {
427 unlock_and_deallocate(&fs);
428 VM_WAITPFAULT;
429 goto RetryFault;
430 }
431 }
432
433 readrest:
434 /*
435 * We have found a valid page or we have allocated a new page.
436 * The page thus may not be valid or may not be entirely
437 * valid.
438 *
439 * Attempt to fault-in the page if there is a chance that the
440 * pager has it, and potentially fault in additional pages
441 * at the same time.
442 */
443 if (TRYPAGER) {
444 int rv;
445 int reqpage;
446 int ahead, behind;
447 u_char behavior = vm_map_entry_behavior(fs.entry);
448
449 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
450 ahead = 0;
451 behind = 0;
452 } else {
453 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
454 if (behind > VM_FAULT_READ_BEHIND)
455 behind = VM_FAULT_READ_BEHIND;
456
457 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
458 if (ahead > VM_FAULT_READ_AHEAD)
459 ahead = VM_FAULT_READ_AHEAD;
460 }
461 is_first_object_locked = FALSE;
462 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
463 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
464 fs.pindex >= fs.entry->lastr &&
465 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
466 (fs.first_object == fs.object ||
467 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
468 fs.first_object->type != OBJT_DEVICE) {
469 vm_pindex_t firstpindex, tmppindex;
470
471 if (fs.first_pindex < 2 * VM_FAULT_READ)
472 firstpindex = 0;
473 else
474 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
475
476 vm_page_lock_queues();
477 /*
478 * note: partially valid pages cannot be
479 * included in the lookahead - NFS piecemeal
480 * writes will barf on it badly.
481 */
482 for (tmppindex = fs.first_pindex - 1;
483 tmppindex >= firstpindex;
484 --tmppindex) {
485 vm_page_t mt;
486
487 mt = vm_page_lookup(fs.first_object, tmppindex);
488 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
489 break;
490 if (mt->busy ||
491 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
492 mt->hold_count ||
493 mt->wire_count)
494 continue;
495 if (mt->dirty == 0)
496 vm_page_test_dirty(mt);
497 if (mt->dirty) {
498 pmap_remove_all(mt);
499 vm_page_deactivate(mt);
500 } else {
501 vm_page_cache(mt);
502 }
503 }
504 vm_page_unlock_queues();
505 ahead += behind;
506 behind = 0;
507 }
508 if (is_first_object_locked)
509 VM_OBJECT_UNLOCK(fs.first_object);
510 /*
511 * now we find out if any other pages should be paged
512 * in at this time this routine checks to see if the
513 * pages surrounding this fault reside in the same
514 * object as the page for this fault. If they do,
515 * then they are faulted in also into the object. The
516 * array "marray" returned contains an array of
517 * vm_page_t structs where one of them is the
518 * vm_page_t passed to the routine. The reqpage
519 * return value is the index into the marray for the
520 * vm_page_t passed to the routine.
521 *
522 * fs.m plus the additional pages are PG_BUSY'd.
523 *
524 * XXX vm_fault_additional_pages() can block
525 * without releasing the map lock.
526 */
527 faultcount = vm_fault_additional_pages(
528 fs.m, behind, ahead, marray, &reqpage);
529
530 /*
531 * update lastr imperfectly (we do not know how much
532 * getpages will actually read), but good enough.
533 *
534 * XXX The following assignment modifies the map
535 * without holding a write lock on it.
536 */
537 fs.entry->lastr = fs.pindex + faultcount - behind;
538
539 /*
540 * Call the pager to retrieve the data, if any, after
541 * releasing the lock on the map. We hold a ref on
542 * fs.object and the pages are PG_BUSY'd.
543 */
544 unlock_map(&fs);
545
546 rv = faultcount ?
547 vm_pager_get_pages(fs.object, marray, faultcount,
548 reqpage) : VM_PAGER_FAIL;
549
550 if (rv == VM_PAGER_OK) {
551 /*
552 * Found the page. Leave it busy while we play
553 * with it.
554 */
555
556 /*
557 * Relookup in case pager changed page. Pager
558 * is responsible for disposition of old page
559 * if moved.
560 */
561 fs.m = vm_page_lookup(fs.object, fs.pindex);
562 if (!fs.m) {
563 unlock_and_deallocate(&fs);
564 goto RetryFault;
565 }
566
567 hardfault++;
568 break; /* break to PAGE HAS BEEN FOUND */
569 }
570 /*
571 * Remove the bogus page (which does not exist at this
572 * object/offset); before doing so, we must get back
573 * our object lock to preserve our invariant.
574 *
575 * Also wake up any other process that may want to bring
576 * in this page.
577 *
578 * If this is the top-level object, we must leave the
579 * busy page to prevent another process from rushing
580 * past us, and inserting the page in that object at
581 * the same time that we are.
582 */
583 if (rv == VM_PAGER_ERROR)
584 printf("vm_fault: pager read error, pid %d (%s)\n",
585 curproc->p_pid, curproc->p_comm);
586 /*
587 * Data outside the range of the pager or an I/O error
588 */
589 /*
590 * XXX - the check for kernel_map is a kludge to work
591 * around having the machine panic on a kernel space
592 * fault w/ I/O error.
593 */
594 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
595 (rv == VM_PAGER_BAD)) {
596 vm_page_lock_queues();
597 vm_page_free(fs.m);
598 vm_page_unlock_queues();
599 fs.m = NULL;
600 unlock_and_deallocate(&fs);
601 mtx_unlock(&Giant);
602 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
603 }
604 if (fs.object != fs.first_object) {
605 vm_page_lock_queues();
606 vm_page_free(fs.m);
607 vm_page_unlock_queues();
608 fs.m = NULL;
609 /*
610 * XXX - we cannot just fall out at this
611 * point, m has been freed and is invalid!
612 */
613 }
614 }
615
616 /*
617 * We get here if the object has default pager (or unwiring)
618 * or the pager doesn't have the page.
619 */
620 if (fs.object == fs.first_object)
621 fs.first_m = fs.m;
622
623 /*
624 * Move on to the next object. Lock the next object before
625 * unlocking the current one.
626 */
627 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
628 next_object = fs.object->backing_object;
629 if (next_object == NULL) {
630 /*
631 * If there's no object left, fill the page in the top
632 * object with zeros.
633 */
634 if (fs.object != fs.first_object) {
635 vm_object_pip_wakeup(fs.object);
636 VM_OBJECT_UNLOCK(fs.object);
637
638 fs.object = fs.first_object;
639 fs.pindex = fs.first_pindex;
640 fs.m = fs.first_m;
641 VM_OBJECT_LOCK(fs.object);
642 }
643 fs.first_m = NULL;
644
645 /*
646 * Zero the page if necessary and mark it valid.
647 */
648 if ((fs.m->flags & PG_ZERO) == 0) {
649 pmap_zero_page(fs.m);
650 } else {
651 cnt.v_ozfod++;
652 }
653 cnt.v_zfod++;
654 fs.m->valid = VM_PAGE_BITS_ALL;
655 break; /* break to PAGE HAS BEEN FOUND */
656 } else {
657 KASSERT(fs.object != next_object,
658 ("object loop %p", next_object));
659 VM_OBJECT_LOCK(next_object);
660 vm_object_pip_add(next_object, 1);
661 if (fs.object != fs.first_object)
662 vm_object_pip_wakeup(fs.object);
663 VM_OBJECT_UNLOCK(fs.object);
664 fs.object = next_object;
665 }
666 }
667
668 KASSERT((fs.m->flags & PG_BUSY) != 0,
669 ("vm_fault: not busy after main loop"));
670
671 /*
672 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
673 * is held.]
674 */
675
676 /*
677 * If the page is being written, but isn't already owned by the
678 * top-level object, we have to copy it into a new page owned by the
679 * top-level object.
680 */
681 if (fs.object != fs.first_object) {
682 /*
683 * We only really need to copy if we want to write it.
684 */
685 if (fault_type & VM_PROT_WRITE) {
686 /*
687 * This allows pages to be virtually copied from a
688 * backing_object into the first_object, where the
689 * backing object has no other refs to it, and cannot
690 * gain any more refs. Instead of a bcopy, we just
691 * move the page from the backing object to the
692 * first object. Note that we must mark the page
693 * dirty in the first object so that it will go out
694 * to swap when needed.
695 */
696 is_first_object_locked = FALSE;
697 if (
698 /*
699 * Only one shadow object
700 */
701 (fs.object->shadow_count == 1) &&
702 /*
703 * No COW refs, except us
704 */
705 (fs.object->ref_count == 1) &&
706 /*
707 * No one else can look this object up
708 */
709 (fs.object->handle == NULL) &&
710 /*
711 * No other ways to look the object up
712 */
713 ((fs.object->type == OBJT_DEFAULT) ||
714 (fs.object->type == OBJT_SWAP)) &&
715 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
716 /*
717 * We don't chase down the shadow chain
718 */
719 fs.object == fs.first_object->backing_object) {
720 vm_page_lock_queues();
721 /*
722 * get rid of the unnecessary page
723 */
724 pmap_remove_all(fs.first_m);
725 vm_page_free(fs.first_m);
726 /*
727 * grab the page and put it into the
728 * process'es object. The page is
729 * automatically made dirty.
730 */
731 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
732 vm_page_busy(fs.m);
733 vm_page_unlock_queues();
734 fs.first_m = fs.m;
735 fs.m = NULL;
736 cnt.v_cow_optim++;
737 } else {
738 /*
739 * Oh, well, lets copy it.
740 */
741 pmap_copy_page(fs.m, fs.first_m);
742 fs.first_m->valid = VM_PAGE_BITS_ALL;
743 }
744 if (fs.m) {
745 /*
746 * We no longer need the old page or object.
747 */
748 release_page(&fs);
749 }
750 /*
751 * fs.object != fs.first_object due to above
752 * conditional
753 */
754 vm_object_pip_wakeup(fs.object);
755 VM_OBJECT_UNLOCK(fs.object);
756 /*
757 * Only use the new page below...
758 */
759 fs.object = fs.first_object;
760 fs.pindex = fs.first_pindex;
761 fs.m = fs.first_m;
762 if (!is_first_object_locked)
763 VM_OBJECT_LOCK(fs.object);
764 cnt.v_cow_faults++;
765 } else {
766 prot &= ~VM_PROT_WRITE;
767 }
768 }
769
770 /*
771 * We must verify that the maps have not changed since our last
772 * lookup.
773 */
774 if (!fs.lookup_still_valid &&
775 (fs.map->timestamp != map_generation)) {
776 vm_object_t retry_object;
777 vm_pindex_t retry_pindex;
778 vm_prot_t retry_prot;
779
780 /*
781 * Since map entries may be pageable, make sure we can take a
782 * page fault on them.
783 */
784
785 /*
786 * Unlock vnode before the lookup to avoid deadlock. E.G.
787 * avoid a deadlock between the inode and exec_map that can
788 * occur due to locks being obtained in different orders.
789 */
790 if (fs.vp != NULL) {
791 vput(fs.vp);
792 fs.vp = NULL;
793 }
794
795 if (fs.map->infork) {
796 release_page(&fs);
797 unlock_and_deallocate(&fs);
798 goto RetryFault;
799 }
800 VM_OBJECT_UNLOCK(fs.object);
801
802 /*
803 * To avoid trying to write_lock the map while another process
804 * has it read_locked (in vm_map_wire), we do not try for
805 * write permission. If the page is still writable, we will
806 * get write permission. If it is not, or has been marked
807 * needs_copy, we enter the mapping without write permission,
808 * and will merely take another fault.
809 */
810 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE,
811 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
812 map_generation = fs.map->timestamp;
813
814 VM_OBJECT_LOCK(fs.object);
815 /*
816 * If we don't need the page any longer, put it on the active
817 * list (the easiest thing to do here). If no one needs it,
818 * pageout will grab it eventually.
819 */
820 if (result != KERN_SUCCESS) {
821 release_page(&fs);
822 unlock_and_deallocate(&fs);
823 mtx_unlock(&Giant);
824 return (result);
825 }
826 fs.lookup_still_valid = TRUE;
827
828 if ((retry_object != fs.first_object) ||
829 (retry_pindex != fs.first_pindex)) {
830 release_page(&fs);
831 unlock_and_deallocate(&fs);
832 goto RetryFault;
833 }
834 /*
835 * Check whether the protection has changed or the object has
836 * been copied while we left the map unlocked. Changing from
837 * read to write permission is OK - we leave the page
838 * write-protected, and catch the write fault. Changing from
839 * write to read permission means that we can't mark the page
840 * write-enabled after all.
841 */
842 prot &= retry_prot;
843 }
844
845 /*
846 * Put this page into the physical map. We had to do the unlock above
847 * because pmap_enter may cause other faults. We don't put the page
848 * back on the active queue until later so that the page-out daemon
849 * won't find us (yet).
850 */
851
852 if (prot & VM_PROT_WRITE) {
853 vm_page_lock_queues();
854 vm_page_flag_set(fs.m, PG_WRITEABLE);
855 vm_object_set_writeable_dirty(fs.m->object);
856
857 /*
858 * If the fault is a write, we know that this page is being
859 * written NOW so dirty it explicitly to save on
860 * pmap_is_modified() calls later.
861 *
862 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
863 * if the page is already dirty to prevent data written with
864 * the expectation of being synced from not being synced.
865 * Likewise if this entry does not request NOSYNC then make
866 * sure the page isn't marked NOSYNC. Applications sharing
867 * data should use the same flags to avoid ping ponging.
868 *
869 * Also tell the backing pager, if any, that it should remove
870 * any swap backing since the page is now dirty.
871 */
872 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
873 if (fs.m->dirty == 0)
874 vm_page_flag_set(fs.m, PG_NOSYNC);
875 } else {
876 vm_page_flag_clear(fs.m, PG_NOSYNC);
877 }
878 vm_page_unlock_queues();
879 if (fault_flags & VM_FAULT_DIRTY) {
880 int s;
881 vm_page_dirty(fs.m);
882 s = splvm();
883 vm_pager_page_unswapped(fs.m);
884 splx(s);
885 }
886 }
887
888 /*
889 * Page had better still be busy
890 */
891 KASSERT(fs.m->flags & PG_BUSY,
892 ("vm_fault: page %p not busy!", fs.m));
893 /*
894 * Sanity check: page must be completely valid or it is not fit to
895 * map into user space. vm_pager_get_pages() ensures this.
896 */
897 if (fs.m->valid != VM_PAGE_BITS_ALL) {
898 vm_page_zero_invalid(fs.m, TRUE);
899 printf("Warning: page %p partially invalid on fault\n", fs.m);
900 }
901 unlock_things(&fs);
902
903 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired);
904 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
905 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
906 }
907 vm_page_lock_queues();
908 vm_page_flag_clear(fs.m, PG_ZERO);
909 vm_page_flag_set(fs.m, PG_REFERENCED);
910
911 /*
912 * If the page is not wired down, then put it where the pageout daemon
913 * can find it.
914 */
915 if (fault_flags & VM_FAULT_WIRE_MASK) {
916 if (wired)
917 vm_page_wire(fs.m);
918 else
919 vm_page_unwire(fs.m, 1);
920 } else {
921 vm_page_activate(fs.m);
922 }
923 vm_page_wakeup(fs.m);
924 vm_page_unlock_queues();
925
926 PROC_LOCK(curproc);
927 if ((curproc->p_sflag & PS_INMEM) && curproc->p_stats) {
928 if (hardfault) {
929 curproc->p_stats->p_ru.ru_majflt++;
930 } else {
931 curproc->p_stats->p_ru.ru_minflt++;
932 }
933 }
934 PROC_UNLOCK(curproc);
935
936 /*
937 * Unlock everything, and return
938 */
939 vm_object_deallocate(fs.first_object);
940 mtx_unlock(&Giant);
941 return (KERN_SUCCESS);
942 }
943
944 /*
945 * vm_fault_prefault provides a quick way of clustering
946 * pagefaults into a processes address space. It is a "cousin"
947 * of vm_map_pmap_enter, except it runs at page fault time instead
948 * of mmap time.
949 */
950 static void
951 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
952 {
953 int i;
954 vm_offset_t addr, starta;
955 vm_pindex_t pindex;
956 vm_page_t m, mpte;
957 vm_object_t object;
958
959 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)))
960 return;
961
962 object = entry->object.vm_object;
963
964 starta = addra - PFBAK * PAGE_SIZE;
965 if (starta < entry->start) {
966 starta = entry->start;
967 } else if (starta > addra) {
968 starta = 0;
969 }
970
971 mpte = NULL;
972 for (i = 0; i < PAGEORDER_SIZE; i++) {
973 vm_object_t backing_object, lobject;
974
975 addr = addra + prefault_pageorder[i];
976 if (addr > addra + (PFFOR * PAGE_SIZE))
977 addr = 0;
978
979 if (addr < starta || addr >= entry->end)
980 continue;
981
982 if (!pmap_is_prefaultable(pmap, addr))
983 continue;
984
985 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
986 lobject = object;
987 VM_OBJECT_LOCK(lobject);
988 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
989 lobject->type == OBJT_DEFAULT &&
990 (backing_object = lobject->backing_object) != NULL) {
991 if (lobject->backing_object_offset & PAGE_MASK)
992 break;
993 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
994 VM_OBJECT_LOCK(backing_object);
995 VM_OBJECT_UNLOCK(lobject);
996 lobject = backing_object;
997 }
998 /*
999 * give-up when a page is not in memory
1000 */
1001 if (m == NULL) {
1002 VM_OBJECT_UNLOCK(lobject);
1003 break;
1004 }
1005 vm_page_lock_queues();
1006 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1007 (m->busy == 0) &&
1008 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
1009
1010 if ((m->queue - m->pc) == PQ_CACHE) {
1011 vm_page_deactivate(m);
1012 }
1013 vm_page_busy(m);
1014 vm_page_unlock_queues();
1015 VM_OBJECT_UNLOCK(lobject);
1016 mpte = pmap_enter_quick(pmap, addr, m, mpte);
1017 VM_OBJECT_LOCK(lobject);
1018 vm_page_lock_queues();
1019 vm_page_wakeup(m);
1020 }
1021 vm_page_unlock_queues();
1022 VM_OBJECT_UNLOCK(lobject);
1023 }
1024 }
1025
1026 /*
1027 * vm_fault_quick:
1028 *
1029 * Ensure that the requested virtual address, which may be in userland,
1030 * is valid. Fault-in the page if necessary. Return -1 on failure.
1031 */
1032 int
1033 vm_fault_quick(caddr_t v, int prot)
1034 {
1035 int r;
1036
1037 if (prot & VM_PROT_WRITE)
1038 r = subyte(v, fubyte(v));
1039 else
1040 r = fubyte(v);
1041 return(r);
1042 }
1043
1044 /*
1045 * vm_fault_wire:
1046 *
1047 * Wire down a range of virtual addresses in a map.
1048 */
1049 int
1050 vm_fault_wire(map, start, end, user_wire)
1051 vm_map_t map;
1052 vm_offset_t start, end;
1053 boolean_t user_wire;
1054 {
1055 vm_offset_t va;
1056 int rv;
1057
1058 /*
1059 * We simulate a fault to get the page and enter it in the physical
1060 * map. For user wiring, we only ask for read access on currently
1061 * read-only sections.
1062 */
1063 for (va = start; va < end; va += PAGE_SIZE) {
1064 rv = vm_fault(map, va,
1065 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1066 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1067 if (rv) {
1068 if (va != start)
1069 vm_fault_unwire(map, start, va);
1070 return (rv);
1071 }
1072 }
1073 return (KERN_SUCCESS);
1074 }
1075
1076 /*
1077 * vm_fault_unwire:
1078 *
1079 * Unwire a range of virtual addresses in a map.
1080 */
1081 void
1082 vm_fault_unwire(map, start, end)
1083 vm_map_t map;
1084 vm_offset_t start, end;
1085 {
1086 vm_paddr_t pa;
1087 vm_offset_t va;
1088 pmap_t pmap;
1089
1090 pmap = vm_map_pmap(map);
1091
1092 mtx_lock(&Giant);
1093 /*
1094 * Since the pages are wired down, we must be able to get their
1095 * mappings from the physical map system.
1096 */
1097 for (va = start; va < end; va += PAGE_SIZE) {
1098 pa = pmap_extract(pmap, va);
1099 if (pa != 0) {
1100 pmap_change_wiring(pmap, va, FALSE);
1101 vm_page_lock_queues();
1102 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1103 vm_page_unlock_queues();
1104 }
1105 }
1106 mtx_unlock(&Giant);
1107 }
1108
1109 /*
1110 * Routine:
1111 * vm_fault_copy_entry
1112 * Function:
1113 * Copy all of the pages from a wired-down map entry to another.
1114 *
1115 * In/out conditions:
1116 * The source and destination maps must be locked for write.
1117 * The source map entry must be wired down (or be a sharing map
1118 * entry corresponding to a main map entry that is wired down).
1119 */
1120 void
1121 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1122 vm_map_t dst_map;
1123 vm_map_t src_map;
1124 vm_map_entry_t dst_entry;
1125 vm_map_entry_t src_entry;
1126 {
1127 vm_object_t backing_object, dst_object, object;
1128 vm_object_t src_object;
1129 vm_ooffset_t dst_offset;
1130 vm_ooffset_t src_offset;
1131 vm_pindex_t pindex;
1132 vm_prot_t prot;
1133 vm_offset_t vaddr;
1134 vm_page_t dst_m;
1135 vm_page_t src_m;
1136
1137 #ifdef lint
1138 src_map++;
1139 #endif /* lint */
1140
1141 src_object = src_entry->object.vm_object;
1142 src_offset = src_entry->offset;
1143
1144 /*
1145 * Create the top-level object for the destination entry. (Doesn't
1146 * actually shadow anything - we copy the pages directly.)
1147 */
1148 dst_object = vm_object_allocate(OBJT_DEFAULT,
1149 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start));
1150
1151 VM_OBJECT_LOCK(dst_object);
1152 dst_entry->object.vm_object = dst_object;
1153 dst_entry->offset = 0;
1154
1155 prot = dst_entry->max_protection;
1156
1157 /*
1158 * Loop through all of the pages in the entry's range, copying each
1159 * one from the source object (it should be there) to the destination
1160 * object.
1161 */
1162 for (vaddr = dst_entry->start, dst_offset = 0;
1163 vaddr < dst_entry->end;
1164 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1165
1166 /*
1167 * Allocate a page in the destination object
1168 */
1169 do {
1170 dst_m = vm_page_alloc(dst_object,
1171 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1172 if (dst_m == NULL) {
1173 VM_OBJECT_UNLOCK(dst_object);
1174 VM_WAIT;
1175 VM_OBJECT_LOCK(dst_object);
1176 }
1177 } while (dst_m == NULL);
1178
1179 /*
1180 * Find the page in the source object, and copy it in.
1181 * (Because the source is wired down, the page will be in
1182 * memory.)
1183 */
1184 VM_OBJECT_LOCK(src_object);
1185 object = src_object;
1186 pindex = 0;
1187 while ((src_m = vm_page_lookup(object, pindex +
1188 OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1189 (src_entry->protection & VM_PROT_WRITE) == 0 &&
1190 (backing_object = object->backing_object) != NULL) {
1191 /*
1192 * Allow fallback to backing objects if we are reading.
1193 */
1194 VM_OBJECT_LOCK(backing_object);
1195 pindex += OFF_TO_IDX(object->backing_object_offset);
1196 VM_OBJECT_UNLOCK(object);
1197 object = backing_object;
1198 }
1199 if (src_m == NULL)
1200 panic("vm_fault_copy_wired: page missing");
1201 pmap_copy_page(src_m, dst_m);
1202 VM_OBJECT_UNLOCK(object);
1203 dst_m->valid = VM_PAGE_BITS_ALL;
1204 VM_OBJECT_UNLOCK(dst_object);
1205
1206 /*
1207 * Enter it in the pmap...
1208 */
1209 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
1210 VM_OBJECT_LOCK(dst_object);
1211 vm_page_lock_queues();
1212 if ((prot & VM_PROT_WRITE) != 0)
1213 vm_page_flag_set(dst_m, PG_WRITEABLE);
1214
1215 /*
1216 * Mark it no longer busy, and put it on the active list.
1217 */
1218 vm_page_activate(dst_m);
1219 vm_page_wakeup(dst_m);
1220 vm_page_unlock_queues();
1221 }
1222 VM_OBJECT_UNLOCK(dst_object);
1223 }
1224
1225
1226 /*
1227 * This routine checks around the requested page for other pages that
1228 * might be able to be faulted in. This routine brackets the viable
1229 * pages for the pages to be paged in.
1230 *
1231 * Inputs:
1232 * m, rbehind, rahead
1233 *
1234 * Outputs:
1235 * marray (array of vm_page_t), reqpage (index of requested page)
1236 *
1237 * Return value:
1238 * number of pages in marray
1239 *
1240 * This routine can't block.
1241 */
1242 static int
1243 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1244 vm_page_t m;
1245 int rbehind;
1246 int rahead;
1247 vm_page_t *marray;
1248 int *reqpage;
1249 {
1250 int i,j;
1251 vm_object_t object;
1252 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1253 vm_page_t rtm;
1254 int cbehind, cahead;
1255
1256 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1257
1258 object = m->object;
1259 pindex = m->pindex;
1260
1261 /*
1262 * we don't fault-ahead for device pager
1263 */
1264 if (object->type == OBJT_DEVICE) {
1265 *reqpage = 0;
1266 marray[0] = m;
1267 return 1;
1268 }
1269
1270 /*
1271 * if the requested page is not available, then give up now
1272 */
1273 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1274 return 0;
1275 }
1276
1277 if ((cbehind == 0) && (cahead == 0)) {
1278 *reqpage = 0;
1279 marray[0] = m;
1280 return 1;
1281 }
1282
1283 if (rahead > cahead) {
1284 rahead = cahead;
1285 }
1286
1287 if (rbehind > cbehind) {
1288 rbehind = cbehind;
1289 }
1290
1291 /*
1292 * try to do any readahead that we might have free pages for.
1293 */
1294 if ((rahead + rbehind) >
1295 ((cnt.v_free_count + cnt.v_cache_count) - cnt.v_free_reserved)) {
1296 pagedaemon_wakeup();
1297 marray[0] = m;
1298 *reqpage = 0;
1299 return 1;
1300 }
1301
1302 /*
1303 * scan backward for the read behind pages -- in memory
1304 */
1305 if (pindex > 0) {
1306 if (rbehind > pindex) {
1307 rbehind = pindex;
1308 startpindex = 0;
1309 } else {
1310 startpindex = pindex - rbehind;
1311 }
1312
1313 for (tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) {
1314 if (vm_page_lookup(object, tpindex)) {
1315 startpindex = tpindex + 1;
1316 break;
1317 }
1318 if (tpindex == 0)
1319 break;
1320 }
1321
1322 for (i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) {
1323
1324 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1325 if (rtm == NULL) {
1326 vm_page_lock_queues();
1327 for (j = 0; j < i; j++) {
1328 vm_page_free(marray[j]);
1329 }
1330 vm_page_unlock_queues();
1331 marray[0] = m;
1332 *reqpage = 0;
1333 return 1;
1334 }
1335
1336 marray[i] = rtm;
1337 }
1338 } else {
1339 startpindex = 0;
1340 i = 0;
1341 }
1342
1343 marray[i] = m;
1344 /* page offset of the required page */
1345 *reqpage = i;
1346
1347 tpindex = pindex + 1;
1348 i++;
1349
1350 /*
1351 * scan forward for the read ahead pages
1352 */
1353 endpindex = tpindex + rahead;
1354 if (endpindex > object->size)
1355 endpindex = object->size;
1356
1357 for (; tpindex < endpindex; i++, tpindex++) {
1358
1359 if (vm_page_lookup(object, tpindex)) {
1360 break;
1361 }
1362
1363 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL);
1364 if (rtm == NULL) {
1365 break;
1366 }
1367
1368 marray[i] = rtm;
1369 }
1370
1371 /* return number of bytes of pages */
1372 return i;
1373 }
Cache object: d7c801561803b9f322f3538c7e775a5c
|