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/7.4/sys/vm/vm_fault.c 215433 2010-11-17 18:21:29Z alc $");
76
77 #include "opt_vm.h"
78
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
82 #include <sys/lock.h>
83 #include <sys/mutex.h>
84 #include <sys/proc.h>
85 #include <sys/resourcevar.h>
86 #include <sys/sysctl.h>
87 #include <sys/vmmeter.h>
88 #include <sys/vnode.h>
89
90 #include <vm/vm.h>
91 #include <vm/vm_param.h>
92 #include <vm/pmap.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_page.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vnode_pager.h>
100 #include <vm/vm_extern.h>
101
102 #include <sys/mount.h> /* XXX Temporary for VFS_LOCK_GIANT() */
103
104 #define PFBAK 4
105 #define PFFOR 4
106 #define PAGEORDER_SIZE (PFBAK+PFFOR)
107
108 static int prefault_pageorder[] = {
109 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
110 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
111 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
112 -4 * PAGE_SIZE, 4 * PAGE_SIZE
113 };
114
115 static int vm_fault_additional_pages(vm_page_t, int, int, vm_page_t *, int *);
116 static void vm_fault_prefault(pmap_t, vm_offset_t, vm_map_entry_t);
117
118 #define VM_FAULT_READ_AHEAD 8
119 #define VM_FAULT_READ_BEHIND 7
120 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1)
121
122 struct faultstate {
123 vm_page_t m;
124 vm_object_t object;
125 vm_pindex_t pindex;
126 vm_page_t first_m;
127 vm_object_t first_object;
128 vm_pindex_t first_pindex;
129 vm_map_t map;
130 vm_map_entry_t entry;
131 int lookup_still_valid;
132 struct vnode *vp;
133 };
134
135 static inline void
136 release_page(struct faultstate *fs)
137 {
138 vm_page_wakeup(fs->m);
139 vm_page_lock_queues();
140 vm_page_deactivate(fs->m);
141 vm_page_unlock_queues();
142 fs->m = NULL;
143 }
144
145 static inline void
146 unlock_map(struct faultstate *fs)
147 {
148 if (fs->lookup_still_valid) {
149 vm_map_lookup_done(fs->map, fs->entry);
150 fs->lookup_still_valid = FALSE;
151 }
152 }
153
154 static void
155 unlock_and_deallocate(struct faultstate *fs)
156 {
157
158 vm_object_pip_wakeup(fs->object);
159 VM_OBJECT_UNLOCK(fs->object);
160 if (fs->object != fs->first_object) {
161 VM_OBJECT_LOCK(fs->first_object);
162 vm_page_lock_queues();
163 vm_page_free(fs->first_m);
164 vm_page_unlock_queues();
165 vm_object_pip_wakeup(fs->first_object);
166 VM_OBJECT_UNLOCK(fs->first_object);
167 fs->first_m = NULL;
168 }
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 }
180
181 /*
182 * TRYPAGER - used by vm_fault to calculate whether the pager for the
183 * current object *might* contain the page.
184 *
185 * default objects are zero-fill, there is no real pager.
186 */
187 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \
188 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired))
189
190 /*
191 * vm_fault:
192 *
193 * Handle a page fault occurring at the given address,
194 * requiring the given permissions, in the map specified.
195 * If successful, the page is inserted into the
196 * associated physical map.
197 *
198 * NOTE: the given address should be truncated to the
199 * proper page address.
200 *
201 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
202 * a standard error specifying why the fault is fatal is returned.
203 *
204 *
205 * The map in question must be referenced, and remains so.
206 * Caller may hold no locks.
207 */
208 int
209 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
210 int fault_flags)
211 {
212 vm_prot_t prot;
213 int is_first_object_locked, result;
214 boolean_t growstack, wired;
215 int map_generation;
216 vm_object_t next_object;
217 vm_page_t marray[VM_FAULT_READ], mt, mt_prev;
218 int hardfault;
219 int faultcount;
220 struct faultstate fs;
221
222 hardfault = 0;
223 growstack = TRUE;
224 PCPU_INC(cnt.v_vm_faults);
225
226 RetryFault:;
227
228 /*
229 * Find the backing store object and offset into it to begin the
230 * search.
231 */
232 fs.map = map;
233 result = vm_map_lookup(&fs.map, vaddr, fault_type, &fs.entry,
234 &fs.first_object, &fs.first_pindex, &prot, &wired);
235 if (result != KERN_SUCCESS) {
236 if (result != KERN_PROTECTION_FAILURE ||
237 (fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) {
238 if (growstack && result == KERN_INVALID_ADDRESS &&
239 map != kernel_map && curproc != NULL) {
240 result = vm_map_growstack(curproc, vaddr);
241 if (result != KERN_SUCCESS)
242 return (KERN_FAILURE);
243 growstack = FALSE;
244 goto RetryFault;
245 }
246 return (result);
247 }
248
249 /*
250 * If we are user-wiring a r/w segment, and it is COW, then
251 * we need to do the COW operation. Note that we don't COW
252 * currently RO sections now, because it is NOT desirable
253 * to COW .text. We simply keep .text from ever being COW'ed
254 * and take the heat that one cannot debug wired .text sections.
255 */
256 result = vm_map_lookup(&fs.map, vaddr,
257 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE,
258 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired);
259 if (result != KERN_SUCCESS)
260 return (result);
261
262 /*
263 * If we don't COW now, on a user wire, the user will never
264 * be able to write to the mapping. If we don't make this
265 * restriction, the bookkeeping would be nearly impossible.
266 *
267 * XXX The following assignment modifies the map without
268 * holding a write lock on it.
269 */
270 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
271 fs.entry->max_protection &= ~VM_PROT_WRITE;
272 }
273
274 map_generation = fs.map->timestamp;
275
276 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
277 panic("vm_fault: fault on nofault entry, addr: %lx",
278 (u_long)vaddr);
279 }
280
281 /*
282 * Make a reference to this object to prevent its disposal while we
283 * are messing with it. Once we have the reference, the map is free
284 * to be diddled. Since objects reference their shadows (and copies),
285 * they will stay around as well.
286 *
287 * Bump the paging-in-progress count to prevent size changes (e.g.
288 * truncation operations) during I/O. This must be done after
289 * obtaining the vnode lock in order to avoid possible deadlocks.
290 *
291 * XXX vnode_pager_lock() can block without releasing the map lock.
292 */
293 if (fs.first_object->flags & OBJ_NEEDGIANT)
294 mtx_lock(&Giant);
295 VM_OBJECT_LOCK(fs.first_object);
296 vm_object_reference_locked(fs.first_object);
297 fs.vp = vnode_pager_lock(fs.first_object);
298 KASSERT(fs.vp == NULL || !fs.map->system_map,
299 ("vm_fault: vnode-backed object mapped by system map"));
300 KASSERT((fs.first_object->flags & OBJ_NEEDGIANT) == 0 ||
301 !fs.map->system_map,
302 ("vm_fault: Object requiring giant mapped by system map"));
303 if (fs.first_object->flags & OBJ_NEEDGIANT)
304 mtx_unlock(&Giant);
305 vm_object_pip_add(fs.first_object, 1);
306
307 fs.lookup_still_valid = TRUE;
308
309 if (wired)
310 fault_type = prot;
311
312 fs.first_m = NULL;
313
314 /*
315 * Search for the page at object/offset.
316 */
317 fs.object = fs.first_object;
318 fs.pindex = fs.first_pindex;
319 while (TRUE) {
320 /*
321 * If the object is dead, we stop here
322 */
323 if (fs.object->flags & OBJ_DEAD) {
324 unlock_and_deallocate(&fs);
325 return (KERN_PROTECTION_FAILURE);
326 }
327
328 /*
329 * See if page is resident
330 */
331 fs.m = vm_page_lookup(fs.object, fs.pindex);
332 if (fs.m != NULL) {
333 /*
334 * check for page-based copy on write.
335 * We check fs.object == fs.first_object so
336 * as to ensure the legacy COW mechanism is
337 * used when the page in question is part of
338 * a shadow object. Otherwise, vm_page_cowfault()
339 * removes the page from the backing object,
340 * which is not what we want.
341 */
342 vm_page_lock_queues();
343 if ((fs.m->cow) &&
344 (fault_type & VM_PROT_WRITE) &&
345 (fs.object == fs.first_object)) {
346 vm_page_cowfault(fs.m);
347 vm_page_unlock_queues();
348 unlock_and_deallocate(&fs);
349 goto RetryFault;
350 }
351
352 /*
353 * Wait/Retry if the page is busy. We have to do this
354 * if the page is busy via either VPO_BUSY or
355 * vm_page_t->busy because the vm_pager may be using
356 * vm_page_t->busy for pageouts ( and even pageins if
357 * it is the vnode pager ), and we could end up trying
358 * to pagein and pageout the same page simultaneously.
359 *
360 * We can theoretically allow the busy case on a read
361 * fault if the page is marked valid, but since such
362 * pages are typically already pmap'd, putting that
363 * special case in might be more effort then it is
364 * worth. We cannot under any circumstances mess
365 * around with a vm_page_t->busy page except, perhaps,
366 * to pmap it.
367 */
368 if ((fs.m->oflags & VPO_BUSY) || fs.m->busy) {
369 vm_page_unlock_queues();
370 VM_OBJECT_UNLOCK(fs.object);
371 if (fs.object != fs.first_object) {
372 VM_OBJECT_LOCK(fs.first_object);
373 vm_page_lock_queues();
374 vm_page_free(fs.first_m);
375 vm_page_unlock_queues();
376 vm_object_pip_wakeup(fs.first_object);
377 VM_OBJECT_UNLOCK(fs.first_object);
378 fs.first_m = NULL;
379 }
380 unlock_map(&fs);
381 if (fs.vp != NULL) {
382 int vfslck;
383
384 vfslck = VFS_LOCK_GIANT(fs.vp->v_mount);
385 vput(fs.vp);
386 fs.vp = NULL;
387 VFS_UNLOCK_GIANT(vfslck);
388 }
389 VM_OBJECT_LOCK(fs.object);
390 if (fs.m == vm_page_lookup(fs.object,
391 fs.pindex)) {
392 vm_page_sleep_if_busy(fs.m, TRUE,
393 "vmpfw");
394 }
395 vm_object_pip_wakeup(fs.object);
396 VM_OBJECT_UNLOCK(fs.object);
397 PCPU_INC(cnt.v_intrans);
398 vm_object_deallocate(fs.first_object);
399 goto RetryFault;
400 }
401 vm_pageq_remove(fs.m);
402 vm_page_unlock_queues();
403
404 /*
405 * Mark page busy for other processes, and the
406 * pagedaemon. If it still isn't completely valid
407 * (readable), jump to readrest, else break-out ( we
408 * found the page ).
409 */
410 vm_page_busy(fs.m);
411 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) &&
412 fs.m->object != kernel_object && fs.m->object != kmem_object) {
413 goto readrest;
414 }
415
416 break;
417 }
418
419 /*
420 * Page is not resident, If this is the search termination
421 * or the pager might contain the page, allocate a new page.
422 */
423 if (TRYPAGER || fs.object == fs.first_object) {
424 if (fs.pindex >= fs.object->size) {
425 unlock_and_deallocate(&fs);
426 return (KERN_PROTECTION_FAILURE);
427 }
428
429 /*
430 * Allocate a new page for this object/offset pair.
431 */
432 fs.m = NULL;
433 if (!vm_page_count_severe()) {
434 #if VM_NRESERVLEVEL > 0
435 if ((fs.object->flags & OBJ_COLORED) == 0) {
436 fs.object->flags |= OBJ_COLORED;
437 fs.object->pg_color = atop(vaddr) -
438 fs.pindex;
439 }
440 #endif
441 fs.m = vm_page_alloc(fs.object, fs.pindex,
442 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_ZERO);
443 }
444 if (fs.m == NULL) {
445 unlock_and_deallocate(&fs);
446 VM_WAITPFAULT;
447 goto RetryFault;
448 } else if ((fs.m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL)
449 break;
450 }
451
452 readrest:
453 /*
454 * We have found a valid page or we have allocated a new page.
455 * The page thus may not be valid or may not be entirely
456 * valid.
457 *
458 * Attempt to fault-in the page if there is a chance that the
459 * pager has it, and potentially fault in additional pages
460 * at the same time.
461 */
462 if (TRYPAGER) {
463 int rv;
464 int reqpage = 0;
465 int ahead, behind;
466 u_char behavior = vm_map_entry_behavior(fs.entry);
467
468 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
469 ahead = 0;
470 behind = 0;
471 } else {
472 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT;
473 if (behind > VM_FAULT_READ_BEHIND)
474 behind = VM_FAULT_READ_BEHIND;
475
476 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1;
477 if (ahead > VM_FAULT_READ_AHEAD)
478 ahead = VM_FAULT_READ_AHEAD;
479 }
480 is_first_object_locked = FALSE;
481 if ((behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
482 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
483 fs.pindex >= fs.entry->lastr &&
484 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) &&
485 (fs.first_object == fs.object ||
486 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object))) &&
487 fs.first_object->type != OBJT_DEVICE &&
488 fs.first_object->type != OBJT_PHYS &&
489 fs.first_object->type != OBJT_SG) {
490 vm_pindex_t firstpindex;
491
492 if (fs.first_pindex < 2 * VM_FAULT_READ)
493 firstpindex = 0;
494 else
495 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ;
496 mt = fs.first_object != fs.object ?
497 fs.first_m : fs.m;
498 KASSERT(mt != NULL, ("vm_fault: missing mt"));
499 KASSERT((mt->oflags & VPO_BUSY) != 0,
500 ("vm_fault: mt %p not busy", mt));
501 mt_prev = vm_page_prev(mt);
502
503 vm_page_lock_queues();
504 /*
505 * note: partially valid pages cannot be
506 * included in the lookahead - NFS piecemeal
507 * writes will barf on it badly.
508 */
509 while ((mt = mt_prev) != NULL &&
510 mt->pindex >= firstpindex &&
511 mt->valid == VM_PAGE_BITS_ALL) {
512 mt_prev = vm_page_prev(mt);
513 if (mt->busy ||
514 (mt->oflags & VPO_BUSY) ||
515 mt->hold_count ||
516 mt->wire_count)
517 continue;
518 pmap_remove_all(mt);
519 if (mt->dirty) {
520 vm_page_deactivate(mt);
521 } else {
522 vm_page_cache(mt);
523 }
524 }
525 vm_page_unlock_queues();
526 ahead += behind;
527 behind = 0;
528 }
529 if (is_first_object_locked)
530 VM_OBJECT_UNLOCK(fs.first_object);
531 /*
532 * now we find out if any other pages should be paged
533 * in at this time this routine checks to see if the
534 * pages surrounding this fault reside in the same
535 * object as the page for this fault. If they do,
536 * then they are faulted in also into the object. The
537 * array "marray" returned contains an array of
538 * vm_page_t structs where one of them is the
539 * vm_page_t passed to the routine. The reqpage
540 * return value is the index into the marray for the
541 * vm_page_t passed to the routine.
542 *
543 * fs.m plus the additional pages are VPO_BUSY'd.
544 *
545 * XXX vm_fault_additional_pages() can block
546 * without releasing the map lock.
547 */
548 faultcount = vm_fault_additional_pages(
549 fs.m, behind, ahead, marray, &reqpage);
550
551 /*
552 * update lastr imperfectly (we do not know how much
553 * getpages will actually read), but good enough.
554 *
555 * XXX The following assignment modifies the map
556 * without holding a write lock on it.
557 */
558 fs.entry->lastr = fs.pindex + faultcount - behind;
559
560 /*
561 * Call the pager to retrieve the data, if any, after
562 * releasing the lock on the map. We hold a ref on
563 * fs.object and the pages are VPO_BUSY'd.
564 */
565 unlock_map(&fs);
566
567 rv = faultcount ?
568 vm_pager_get_pages(fs.object, marray, faultcount,
569 reqpage) : VM_PAGER_FAIL;
570
571 if (rv == VM_PAGER_OK) {
572 /*
573 * Found the page. Leave it busy while we play
574 * with it.
575 */
576
577 /*
578 * Relookup in case pager changed page. Pager
579 * is responsible for disposition of old page
580 * if moved.
581 */
582 fs.m = vm_page_lookup(fs.object, fs.pindex);
583 if (!fs.m) {
584 unlock_and_deallocate(&fs);
585 goto RetryFault;
586 }
587
588 hardfault++;
589 break; /* break to PAGE HAS BEEN FOUND */
590 }
591 /*
592 * Remove the bogus page (which does not exist at this
593 * object/offset); before doing so, we must get back
594 * our object lock to preserve our invariant.
595 *
596 * Also wake up any other process that may want to bring
597 * in this page.
598 *
599 * If this is the top-level object, we must leave the
600 * busy page to prevent another process from rushing
601 * past us, and inserting the page in that object at
602 * the same time that we are.
603 */
604 if (rv == VM_PAGER_ERROR)
605 printf("vm_fault: pager read error, pid %d (%s)\n",
606 curproc->p_pid, curproc->p_comm);
607 /*
608 * Data outside the range of the pager or an I/O error
609 */
610 /*
611 * XXX - the check for kernel_map is a kludge to work
612 * around having the machine panic on a kernel space
613 * fault w/ I/O error.
614 */
615 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
616 (rv == VM_PAGER_BAD)) {
617 vm_page_lock_queues();
618 vm_page_free(fs.m);
619 vm_page_unlock_queues();
620 fs.m = NULL;
621 unlock_and_deallocate(&fs);
622 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
623 }
624 if (fs.object != fs.first_object) {
625 vm_page_lock_queues();
626 vm_page_free(fs.m);
627 vm_page_unlock_queues();
628 fs.m = NULL;
629 /*
630 * XXX - we cannot just fall out at this
631 * point, m has been freed and is invalid!
632 */
633 }
634 }
635
636 /*
637 * We get here if the object has default pager (or unwiring)
638 * or the pager doesn't have the page.
639 */
640 if (fs.object == fs.first_object)
641 fs.first_m = fs.m;
642
643 /*
644 * Move on to the next object. Lock the next object before
645 * unlocking the current one.
646 */
647 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
648 next_object = fs.object->backing_object;
649 if (next_object == NULL) {
650 /*
651 * If there's no object left, fill the page in the top
652 * object with zeros.
653 */
654 if (fs.object != fs.first_object) {
655 vm_object_pip_wakeup(fs.object);
656 VM_OBJECT_UNLOCK(fs.object);
657
658 fs.object = fs.first_object;
659 fs.pindex = fs.first_pindex;
660 fs.m = fs.first_m;
661 VM_OBJECT_LOCK(fs.object);
662 }
663 fs.first_m = NULL;
664
665 /*
666 * Zero the page if necessary and mark it valid.
667 */
668 if ((fs.m->flags & PG_ZERO) == 0) {
669 pmap_zero_page(fs.m);
670 } else {
671 PCPU_INC(cnt.v_ozfod);
672 }
673 PCPU_INC(cnt.v_zfod);
674 fs.m->valid = VM_PAGE_BITS_ALL;
675 break; /* break to PAGE HAS BEEN FOUND */
676 } else {
677 KASSERT(fs.object != next_object,
678 ("object loop %p", next_object));
679 VM_OBJECT_LOCK(next_object);
680 vm_object_pip_add(next_object, 1);
681 if (fs.object != fs.first_object)
682 vm_object_pip_wakeup(fs.object);
683 VM_OBJECT_UNLOCK(fs.object);
684 fs.object = next_object;
685 }
686 }
687
688 KASSERT((fs.m->oflags & VPO_BUSY) != 0,
689 ("vm_fault: not busy after main loop"));
690
691 /*
692 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
693 * is held.]
694 */
695
696 /*
697 * If the page is being written, but isn't already owned by the
698 * top-level object, we have to copy it into a new page owned by the
699 * top-level object.
700 */
701 if (fs.object != fs.first_object) {
702 /*
703 * We only really need to copy if we want to write it.
704 */
705 if (fault_type & VM_PROT_WRITE) {
706 /*
707 * This allows pages to be virtually copied from a
708 * backing_object into the first_object, where the
709 * backing object has no other refs to it, and cannot
710 * gain any more refs. Instead of a bcopy, we just
711 * move the page from the backing object to the
712 * first object. Note that we must mark the page
713 * dirty in the first object so that it will go out
714 * to swap when needed.
715 */
716 is_first_object_locked = FALSE;
717 if (
718 /*
719 * Only one shadow object
720 */
721 (fs.object->shadow_count == 1) &&
722 /*
723 * No COW refs, except us
724 */
725 (fs.object->ref_count == 1) &&
726 /*
727 * No one else can look this object up
728 */
729 (fs.object->handle == NULL) &&
730 /*
731 * No other ways to look the object up
732 */
733 ((fs.object->type == OBJT_DEFAULT) ||
734 (fs.object->type == OBJT_SWAP)) &&
735 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
736 /*
737 * We don't chase down the shadow chain
738 */
739 fs.object == fs.first_object->backing_object) {
740 vm_page_lock_queues();
741 /*
742 * get rid of the unnecessary page
743 */
744 vm_page_free(fs.first_m);
745 /*
746 * grab the page and put it into the
747 * process'es object. The page is
748 * automatically made dirty.
749 */
750 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
751 vm_page_unlock_queues();
752 vm_page_busy(fs.m);
753 fs.first_m = fs.m;
754 fs.m = NULL;
755 PCPU_INC(cnt.v_cow_optim);
756 } else {
757 /*
758 * Oh, well, lets copy it.
759 */
760 pmap_copy_page(fs.m, fs.first_m);
761 fs.first_m->valid = VM_PAGE_BITS_ALL;
762 }
763 if (fs.m) {
764 /*
765 * We no longer need the old page or object.
766 */
767 release_page(&fs);
768 }
769 /*
770 * fs.object != fs.first_object due to above
771 * conditional
772 */
773 vm_object_pip_wakeup(fs.object);
774 VM_OBJECT_UNLOCK(fs.object);
775 /*
776 * Only use the new page below...
777 */
778 fs.object = fs.first_object;
779 fs.pindex = fs.first_pindex;
780 fs.m = fs.first_m;
781 if (!is_first_object_locked)
782 VM_OBJECT_LOCK(fs.object);
783 PCPU_INC(cnt.v_cow_faults);
784 } else {
785 prot &= ~VM_PROT_WRITE;
786 }
787 }
788
789 /*
790 * We must verify that the maps have not changed since our last
791 * lookup.
792 */
793 if (!fs.lookup_still_valid) {
794 vm_object_t retry_object;
795 vm_pindex_t retry_pindex;
796 vm_prot_t retry_prot;
797
798 if (!vm_map_trylock_read(fs.map)) {
799 release_page(&fs);
800 unlock_and_deallocate(&fs);
801 goto RetryFault;
802 }
803 fs.lookup_still_valid = TRUE;
804 if (fs.map->timestamp != map_generation) {
805 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
806 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
807
808 /*
809 * If we don't need the page any longer, put it on the inactive
810 * list (the easiest thing to do here). If no one needs it,
811 * pageout will grab it eventually.
812 */
813 if (result != KERN_SUCCESS) {
814 release_page(&fs);
815 unlock_and_deallocate(&fs);
816
817 /*
818 * If retry of map lookup would have blocked then
819 * retry fault from start.
820 */
821 if (result == KERN_FAILURE)
822 goto RetryFault;
823 return (result);
824 }
825 if ((retry_object != fs.first_object) ||
826 (retry_pindex != fs.first_pindex)) {
827 release_page(&fs);
828 unlock_and_deallocate(&fs);
829 goto RetryFault;
830 }
831
832 /*
833 * Check whether the protection has changed or the object has
834 * been copied while we left the map unlocked. Changing from
835 * read to write permission is OK - we leave the page
836 * write-protected, and catch the write fault. Changing from
837 * write to read permission means that we can't mark the page
838 * write-enabled after all.
839 */
840 prot &= retry_prot;
841 }
842 }
843 if (prot & VM_PROT_WRITE) {
844 vm_object_set_writeable_dirty(fs.object);
845
846 /*
847 * If the fault is a write, we know that this page is being
848 * written NOW so dirty it explicitly to save on
849 * pmap_is_modified() calls later.
850 *
851 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
852 * if the page is already dirty to prevent data written with
853 * the expectation of being synced from not being synced.
854 * Likewise if this entry does not request NOSYNC then make
855 * sure the page isn't marked NOSYNC. Applications sharing
856 * data should use the same flags to avoid ping ponging.
857 *
858 * Also tell the backing pager, if any, that it should remove
859 * any swap backing since the page is now dirty.
860 */
861 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
862 if (fs.m->dirty == 0)
863 fs.m->oflags |= VPO_NOSYNC;
864 } else {
865 fs.m->oflags &= ~VPO_NOSYNC;
866 }
867 if (fault_flags & VM_FAULT_DIRTY) {
868 vm_page_dirty(fs.m);
869 vm_pager_page_unswapped(fs.m);
870 }
871 }
872
873 /*
874 * Page had better still be busy
875 */
876 KASSERT(fs.m->oflags & VPO_BUSY,
877 ("vm_fault: page %p not busy!", fs.m));
878 /*
879 * Sanity check: page must be completely valid or it is not fit to
880 * map into user space. vm_pager_get_pages() ensures this.
881 */
882 if (fs.m->valid != VM_PAGE_BITS_ALL) {
883 vm_page_zero_invalid(fs.m, TRUE);
884 printf("Warning: page %p partially invalid on fault\n", fs.m);
885 }
886 VM_OBJECT_UNLOCK(fs.object);
887
888 /*
889 * Put this page into the physical map. We had to do the unlock above
890 * because pmap_enter() may sleep. We don't put the page
891 * back on the active queue until later so that the pageout daemon
892 * won't find it (yet).
893 */
894 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
895 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
896 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
897 }
898 VM_OBJECT_LOCK(fs.object);
899 vm_page_lock_queues();
900 vm_page_flag_set(fs.m, PG_REFERENCED);
901
902 /*
903 * If the page is not wired down, then put it where the pageout daemon
904 * can find it.
905 */
906 if (fault_flags & VM_FAULT_WIRE_MASK) {
907 if (wired)
908 vm_page_wire(fs.m);
909 else
910 vm_page_unwire(fs.m, 1);
911 } else {
912 vm_page_activate(fs.m);
913 }
914 vm_page_unlock_queues();
915 vm_page_wakeup(fs.m);
916
917 /*
918 * Unlock everything, and return
919 */
920 unlock_and_deallocate(&fs);
921 if (hardfault)
922 curthread->td_ru.ru_majflt++;
923 else
924 curthread->td_ru.ru_minflt++;
925
926 return (KERN_SUCCESS);
927 }
928
929 /*
930 * vm_fault_prefault provides a quick way of clustering
931 * pagefaults into a processes address space. It is a "cousin"
932 * of vm_map_pmap_enter, except it runs at page fault time instead
933 * of mmap time.
934 */
935 static void
936 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
937 {
938 int i;
939 vm_offset_t addr, starta;
940 vm_pindex_t pindex;
941 vm_page_t m;
942 vm_object_t object;
943
944 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
945 return;
946
947 object = entry->object.vm_object;
948
949 starta = addra - PFBAK * PAGE_SIZE;
950 if (starta < entry->start) {
951 starta = entry->start;
952 } else if (starta > addra) {
953 starta = 0;
954 }
955
956 for (i = 0; i < PAGEORDER_SIZE; i++) {
957 vm_object_t backing_object, lobject;
958
959 addr = addra + prefault_pageorder[i];
960 if (addr > addra + (PFFOR * PAGE_SIZE))
961 addr = 0;
962
963 if (addr < starta || addr >= entry->end)
964 continue;
965
966 if (!pmap_is_prefaultable(pmap, addr))
967 continue;
968
969 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
970 lobject = object;
971 VM_OBJECT_LOCK(lobject);
972 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
973 lobject->type == OBJT_DEFAULT &&
974 (backing_object = lobject->backing_object) != NULL) {
975 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
976 0, ("vm_fault_prefault: unaligned object offset"));
977 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
978 VM_OBJECT_LOCK(backing_object);
979 VM_OBJECT_UNLOCK(lobject);
980 lobject = backing_object;
981 }
982 /*
983 * give-up when a page is not in memory
984 */
985 if (m == NULL) {
986 VM_OBJECT_UNLOCK(lobject);
987 break;
988 }
989 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
990 (m->busy == 0) &&
991 (m->flags & PG_FICTITIOUS) == 0) {
992
993 vm_page_lock_queues();
994 pmap_enter_quick(pmap, addr, m, entry->protection);
995 vm_page_unlock_queues();
996 }
997 VM_OBJECT_UNLOCK(lobject);
998 }
999 }
1000
1001 /*
1002 * vm_fault_quick:
1003 *
1004 * Ensure that the requested virtual address, which may be in userland,
1005 * is valid. Fault-in the page if necessary. Return -1 on failure.
1006 */
1007 int
1008 vm_fault_quick(caddr_t v, int prot)
1009 {
1010 int r;
1011
1012 if (prot & VM_PROT_WRITE)
1013 r = subyte(v, fubyte(v));
1014 else
1015 r = fubyte(v);
1016 return(r);
1017 }
1018
1019 /*
1020 * vm_fault_wire:
1021 *
1022 * Wire down a range of virtual addresses in a map.
1023 */
1024 int
1025 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1026 boolean_t user_wire, boolean_t fictitious)
1027 {
1028 vm_offset_t va;
1029 int rv;
1030
1031 /*
1032 * We simulate a fault to get the page and enter it in the physical
1033 * map. For user wiring, we only ask for read access on currently
1034 * read-only sections.
1035 */
1036 for (va = start; va < end; va += PAGE_SIZE) {
1037 rv = vm_fault(map, va,
1038 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1039 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1040 if (rv) {
1041 if (va != start)
1042 vm_fault_unwire(map, start, va, fictitious);
1043 return (rv);
1044 }
1045 }
1046 return (KERN_SUCCESS);
1047 }
1048
1049 /*
1050 * vm_fault_unwire:
1051 *
1052 * Unwire a range of virtual addresses in a map.
1053 */
1054 void
1055 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1056 boolean_t fictitious)
1057 {
1058 vm_paddr_t pa;
1059 vm_offset_t va;
1060 pmap_t pmap;
1061
1062 pmap = vm_map_pmap(map);
1063
1064 /*
1065 * Since the pages are wired down, we must be able to get their
1066 * mappings from the physical map system.
1067 */
1068 for (va = start; va < end; va += PAGE_SIZE) {
1069 pa = pmap_extract(pmap, va);
1070 if (pa != 0) {
1071 pmap_change_wiring(pmap, va, FALSE);
1072 if (!fictitious) {
1073 vm_page_lock_queues();
1074 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1075 vm_page_unlock_queues();
1076 }
1077 }
1078 }
1079 }
1080
1081 /*
1082 * Routine:
1083 * vm_fault_copy_entry
1084 * Function:
1085 * Copy all of the pages from a wired-down map entry to another.
1086 *
1087 * In/out conditions:
1088 * The source and destination maps must be locked for write.
1089 * The source map entry must be wired down (or be a sharing map
1090 * entry corresponding to a main map entry that is wired down).
1091 */
1092 void
1093 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1094 vm_map_t dst_map;
1095 vm_map_t src_map;
1096 vm_map_entry_t dst_entry;
1097 vm_map_entry_t src_entry;
1098 {
1099 vm_object_t backing_object, dst_object, object;
1100 vm_object_t src_object;
1101 vm_ooffset_t dst_offset;
1102 vm_ooffset_t src_offset;
1103 vm_pindex_t pindex;
1104 vm_prot_t prot;
1105 vm_offset_t vaddr;
1106 vm_page_t dst_m;
1107 vm_page_t src_m;
1108
1109 #ifdef lint
1110 src_map++;
1111 #endif /* lint */
1112
1113 src_object = src_entry->object.vm_object;
1114 src_offset = src_entry->offset;
1115
1116 /*
1117 * Create the top-level object for the destination entry. (Doesn't
1118 * actually shadow anything - we copy the pages directly.)
1119 */
1120 dst_object = vm_object_allocate(OBJT_DEFAULT,
1121 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1122 #if VM_NRESERVLEVEL > 0
1123 dst_object->flags |= OBJ_COLORED;
1124 dst_object->pg_color = atop(dst_entry->start);
1125 #endif
1126
1127 VM_OBJECT_LOCK(dst_object);
1128 dst_entry->object.vm_object = dst_object;
1129 dst_entry->offset = 0;
1130
1131 prot = dst_entry->max_protection;
1132
1133 /*
1134 * Loop through all of the pages in the entry's range, copying each
1135 * one from the source object (it should be there) to the destination
1136 * object.
1137 */
1138 for (vaddr = dst_entry->start, dst_offset = 0;
1139 vaddr < dst_entry->end;
1140 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1141
1142 /*
1143 * Allocate a page in the destination object
1144 */
1145 do {
1146 dst_m = vm_page_alloc(dst_object,
1147 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1148 if (dst_m == NULL) {
1149 VM_OBJECT_UNLOCK(dst_object);
1150 VM_WAIT;
1151 VM_OBJECT_LOCK(dst_object);
1152 }
1153 } while (dst_m == NULL);
1154
1155 /*
1156 * Find the page in the source object, and copy it in.
1157 * (Because the source is wired down, the page will be in
1158 * memory.)
1159 */
1160 VM_OBJECT_LOCK(src_object);
1161 object = src_object;
1162 pindex = 0;
1163 while ((src_m = vm_page_lookup(object, pindex +
1164 OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1165 (src_entry->protection & VM_PROT_WRITE) == 0 &&
1166 (backing_object = object->backing_object) != NULL) {
1167 /*
1168 * Allow fallback to backing objects if we are reading.
1169 */
1170 VM_OBJECT_LOCK(backing_object);
1171 pindex += OFF_TO_IDX(object->backing_object_offset);
1172 VM_OBJECT_UNLOCK(object);
1173 object = backing_object;
1174 }
1175 if (src_m == NULL)
1176 panic("vm_fault_copy_wired: page missing");
1177 pmap_copy_page(src_m, dst_m);
1178 VM_OBJECT_UNLOCK(object);
1179 dst_m->valid = VM_PAGE_BITS_ALL;
1180 VM_OBJECT_UNLOCK(dst_object);
1181
1182 /*
1183 * Enter it in the pmap as a read and/or execute access.
1184 */
1185 pmap_enter(dst_map->pmap, vaddr, prot & ~VM_PROT_WRITE, dst_m,
1186 prot, FALSE);
1187
1188 /*
1189 * Mark it no longer busy, and put it on the active list.
1190 */
1191 VM_OBJECT_LOCK(dst_object);
1192 vm_page_lock_queues();
1193 vm_page_activate(dst_m);
1194 vm_page_unlock_queues();
1195 vm_page_wakeup(dst_m);
1196 }
1197 VM_OBJECT_UNLOCK(dst_object);
1198 }
1199
1200
1201 /*
1202 * This routine checks around the requested page for other pages that
1203 * might be able to be faulted in. This routine brackets the viable
1204 * pages for the pages to be paged in.
1205 *
1206 * Inputs:
1207 * m, rbehind, rahead
1208 *
1209 * Outputs:
1210 * marray (array of vm_page_t), reqpage (index of requested page)
1211 *
1212 * Return value:
1213 * number of pages in marray
1214 *
1215 * This routine can't block.
1216 */
1217 static int
1218 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1219 vm_page_t m;
1220 int rbehind;
1221 int rahead;
1222 vm_page_t *marray;
1223 int *reqpage;
1224 {
1225 int i,j;
1226 vm_object_t object;
1227 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1228 vm_page_t rtm;
1229 int cbehind, cahead;
1230
1231 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1232
1233 object = m->object;
1234 pindex = m->pindex;
1235 cbehind = cahead = 0;
1236
1237 /*
1238 * if the requested page is not available, then give up now
1239 */
1240 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1241 return 0;
1242 }
1243
1244 if ((cbehind == 0) && (cahead == 0)) {
1245 *reqpage = 0;
1246 marray[0] = m;
1247 return 1;
1248 }
1249
1250 if (rahead > cahead) {
1251 rahead = cahead;
1252 }
1253
1254 if (rbehind > cbehind) {
1255 rbehind = cbehind;
1256 }
1257
1258 /*
1259 * scan backward for the read behind pages -- in memory
1260 */
1261 if (pindex > 0) {
1262 if (rbehind > pindex) {
1263 rbehind = pindex;
1264 startpindex = 0;
1265 } else {
1266 startpindex = pindex - rbehind;
1267 }
1268
1269 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1270 rtm->pindex >= startpindex)
1271 startpindex = rtm->pindex + 1;
1272
1273 /* tpindex is unsigned; beware of numeric underflow. */
1274 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1275 tpindex < pindex; i++, tpindex--) {
1276
1277 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1278 VM_ALLOC_IFNOTCACHED);
1279 if (rtm == NULL) {
1280 /*
1281 * Shift the allocated pages to the
1282 * beginning of the array.
1283 */
1284 for (j = 0; j < i; j++) {
1285 marray[j] = marray[j + tpindex + 1 -
1286 startpindex];
1287 }
1288 break;
1289 }
1290
1291 marray[tpindex - startpindex] = rtm;
1292 }
1293 } else {
1294 startpindex = 0;
1295 i = 0;
1296 }
1297
1298 marray[i] = m;
1299 /* page offset of the required page */
1300 *reqpage = i;
1301
1302 tpindex = pindex + 1;
1303 i++;
1304
1305 /*
1306 * scan forward for the read ahead pages
1307 */
1308 endpindex = tpindex + rahead;
1309 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1310 endpindex = rtm->pindex;
1311 if (endpindex > object->size)
1312 endpindex = object->size;
1313
1314 for (; tpindex < endpindex; i++, tpindex++) {
1315
1316 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1317 VM_ALLOC_IFNOTCACHED);
1318 if (rtm == NULL) {
1319 break;
1320 }
1321
1322 marray[i] = rtm;
1323 }
1324
1325 /* return number of pages */
1326 return i;
1327 }
Cache object: 1c01e6f111e70c17ce90d6cea8dcd5f8
|