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/7.3/sys/vm/vm_fault.c 198727 2009-10-31 18:46:55Z 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];
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, tmppindex;
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
497 vm_page_lock_queues();
498 /*
499 * note: partially valid pages cannot be
500 * included in the lookahead - NFS piecemeal
501 * writes will barf on it badly.
502 */
503 for (tmppindex = fs.first_pindex - 1;
504 tmppindex >= firstpindex;
505 --tmppindex) {
506 vm_page_t mt;
507
508 mt = vm_page_lookup(fs.first_object, tmppindex);
509 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL))
510 break;
511 if (mt->busy ||
512 (mt->oflags & VPO_BUSY) ||
513 mt->hold_count ||
514 mt->wire_count)
515 continue;
516 pmap_remove_all(mt);
517 if (mt->dirty) {
518 vm_page_deactivate(mt);
519 } else {
520 vm_page_cache(mt);
521 }
522 }
523 vm_page_unlock_queues();
524 ahead += behind;
525 behind = 0;
526 }
527 if (is_first_object_locked)
528 VM_OBJECT_UNLOCK(fs.first_object);
529 /*
530 * now we find out if any other pages should be paged
531 * in at this time this routine checks to see if the
532 * pages surrounding this fault reside in the same
533 * object as the page for this fault. If they do,
534 * then they are faulted in also into the object. The
535 * array "marray" returned contains an array of
536 * vm_page_t structs where one of them is the
537 * vm_page_t passed to the routine. The reqpage
538 * return value is the index into the marray for the
539 * vm_page_t passed to the routine.
540 *
541 * fs.m plus the additional pages are VPO_BUSY'd.
542 *
543 * XXX vm_fault_additional_pages() can block
544 * without releasing the map lock.
545 */
546 faultcount = vm_fault_additional_pages(
547 fs.m, behind, ahead, marray, &reqpage);
548
549 /*
550 * update lastr imperfectly (we do not know how much
551 * getpages will actually read), but good enough.
552 *
553 * XXX The following assignment modifies the map
554 * without holding a write lock on it.
555 */
556 fs.entry->lastr = fs.pindex + faultcount - behind;
557
558 /*
559 * Call the pager to retrieve the data, if any, after
560 * releasing the lock on the map. We hold a ref on
561 * fs.object and the pages are VPO_BUSY'd.
562 */
563 unlock_map(&fs);
564
565 rv = faultcount ?
566 vm_pager_get_pages(fs.object, marray, faultcount,
567 reqpage) : VM_PAGER_FAIL;
568
569 if (rv == VM_PAGER_OK) {
570 /*
571 * Found the page. Leave it busy while we play
572 * with it.
573 */
574
575 /*
576 * Relookup in case pager changed page. Pager
577 * is responsible for disposition of old page
578 * if moved.
579 */
580 fs.m = vm_page_lookup(fs.object, fs.pindex);
581 if (!fs.m) {
582 unlock_and_deallocate(&fs);
583 goto RetryFault;
584 }
585
586 hardfault++;
587 break; /* break to PAGE HAS BEEN FOUND */
588 }
589 /*
590 * Remove the bogus page (which does not exist at this
591 * object/offset); before doing so, we must get back
592 * our object lock to preserve our invariant.
593 *
594 * Also wake up any other process that may want to bring
595 * in this page.
596 *
597 * If this is the top-level object, we must leave the
598 * busy page to prevent another process from rushing
599 * past us, and inserting the page in that object at
600 * the same time that we are.
601 */
602 if (rv == VM_PAGER_ERROR)
603 printf("vm_fault: pager read error, pid %d (%s)\n",
604 curproc->p_pid, curproc->p_comm);
605 /*
606 * Data outside the range of the pager or an I/O error
607 */
608 /*
609 * XXX - the check for kernel_map is a kludge to work
610 * around having the machine panic on a kernel space
611 * fault w/ I/O error.
612 */
613 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) ||
614 (rv == VM_PAGER_BAD)) {
615 vm_page_lock_queues();
616 vm_page_free(fs.m);
617 vm_page_unlock_queues();
618 fs.m = NULL;
619 unlock_and_deallocate(&fs);
620 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE);
621 }
622 if (fs.object != fs.first_object) {
623 vm_page_lock_queues();
624 vm_page_free(fs.m);
625 vm_page_unlock_queues();
626 fs.m = NULL;
627 /*
628 * XXX - we cannot just fall out at this
629 * point, m has been freed and is invalid!
630 */
631 }
632 }
633
634 /*
635 * We get here if the object has default pager (or unwiring)
636 * or the pager doesn't have the page.
637 */
638 if (fs.object == fs.first_object)
639 fs.first_m = fs.m;
640
641 /*
642 * Move on to the next object. Lock the next object before
643 * unlocking the current one.
644 */
645 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset);
646 next_object = fs.object->backing_object;
647 if (next_object == NULL) {
648 /*
649 * If there's no object left, fill the page in the top
650 * object with zeros.
651 */
652 if (fs.object != fs.first_object) {
653 vm_object_pip_wakeup(fs.object);
654 VM_OBJECT_UNLOCK(fs.object);
655
656 fs.object = fs.first_object;
657 fs.pindex = fs.first_pindex;
658 fs.m = fs.first_m;
659 VM_OBJECT_LOCK(fs.object);
660 }
661 fs.first_m = NULL;
662
663 /*
664 * Zero the page if necessary and mark it valid.
665 */
666 if ((fs.m->flags & PG_ZERO) == 0) {
667 pmap_zero_page(fs.m);
668 } else {
669 PCPU_INC(cnt.v_ozfod);
670 }
671 PCPU_INC(cnt.v_zfod);
672 fs.m->valid = VM_PAGE_BITS_ALL;
673 break; /* break to PAGE HAS BEEN FOUND */
674 } else {
675 KASSERT(fs.object != next_object,
676 ("object loop %p", next_object));
677 VM_OBJECT_LOCK(next_object);
678 vm_object_pip_add(next_object, 1);
679 if (fs.object != fs.first_object)
680 vm_object_pip_wakeup(fs.object);
681 VM_OBJECT_UNLOCK(fs.object);
682 fs.object = next_object;
683 }
684 }
685
686 KASSERT((fs.m->oflags & VPO_BUSY) != 0,
687 ("vm_fault: not busy after main loop"));
688
689 /*
690 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
691 * is held.]
692 */
693
694 /*
695 * If the page is being written, but isn't already owned by the
696 * top-level object, we have to copy it into a new page owned by the
697 * top-level object.
698 */
699 if (fs.object != fs.first_object) {
700 /*
701 * We only really need to copy if we want to write it.
702 */
703 if (fault_type & VM_PROT_WRITE) {
704 /*
705 * This allows pages to be virtually copied from a
706 * backing_object into the first_object, where the
707 * backing object has no other refs to it, and cannot
708 * gain any more refs. Instead of a bcopy, we just
709 * move the page from the backing object to the
710 * first object. Note that we must mark the page
711 * dirty in the first object so that it will go out
712 * to swap when needed.
713 */
714 is_first_object_locked = FALSE;
715 if (
716 /*
717 * Only one shadow object
718 */
719 (fs.object->shadow_count == 1) &&
720 /*
721 * No COW refs, except us
722 */
723 (fs.object->ref_count == 1) &&
724 /*
725 * No one else can look this object up
726 */
727 (fs.object->handle == NULL) &&
728 /*
729 * No other ways to look the object up
730 */
731 ((fs.object->type == OBJT_DEFAULT) ||
732 (fs.object->type == OBJT_SWAP)) &&
733 (is_first_object_locked = VM_OBJECT_TRYLOCK(fs.first_object)) &&
734 /*
735 * We don't chase down the shadow chain
736 */
737 fs.object == fs.first_object->backing_object) {
738 vm_page_lock_queues();
739 /*
740 * get rid of the unnecessary page
741 */
742 vm_page_free(fs.first_m);
743 /*
744 * grab the page and put it into the
745 * process'es object. The page is
746 * automatically made dirty.
747 */
748 vm_page_rename(fs.m, fs.first_object, fs.first_pindex);
749 vm_page_unlock_queues();
750 vm_page_busy(fs.m);
751 fs.first_m = fs.m;
752 fs.m = NULL;
753 PCPU_INC(cnt.v_cow_optim);
754 } else {
755 /*
756 * Oh, well, lets copy it.
757 */
758 pmap_copy_page(fs.m, fs.first_m);
759 fs.first_m->valid = VM_PAGE_BITS_ALL;
760 }
761 if (fs.m) {
762 /*
763 * We no longer need the old page or object.
764 */
765 release_page(&fs);
766 }
767 /*
768 * fs.object != fs.first_object due to above
769 * conditional
770 */
771 vm_object_pip_wakeup(fs.object);
772 VM_OBJECT_UNLOCK(fs.object);
773 /*
774 * Only use the new page below...
775 */
776 fs.object = fs.first_object;
777 fs.pindex = fs.first_pindex;
778 fs.m = fs.first_m;
779 if (!is_first_object_locked)
780 VM_OBJECT_LOCK(fs.object);
781 PCPU_INC(cnt.v_cow_faults);
782 } else {
783 prot &= ~VM_PROT_WRITE;
784 }
785 }
786
787 /*
788 * We must verify that the maps have not changed since our last
789 * lookup.
790 */
791 if (!fs.lookup_still_valid) {
792 vm_object_t retry_object;
793 vm_pindex_t retry_pindex;
794 vm_prot_t retry_prot;
795
796 if (!vm_map_trylock_read(fs.map)) {
797 release_page(&fs);
798 unlock_and_deallocate(&fs);
799 goto RetryFault;
800 }
801 fs.lookup_still_valid = TRUE;
802 if (fs.map->timestamp != map_generation) {
803 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type,
804 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired);
805
806 /*
807 * If we don't need the page any longer, put it on the inactive
808 * list (the easiest thing to do here). If no one needs it,
809 * pageout will grab it eventually.
810 */
811 if (result != KERN_SUCCESS) {
812 release_page(&fs);
813 unlock_and_deallocate(&fs);
814
815 /*
816 * If retry of map lookup would have blocked then
817 * retry fault from start.
818 */
819 if (result == KERN_FAILURE)
820 goto RetryFault;
821 return (result);
822 }
823 if ((retry_object != fs.first_object) ||
824 (retry_pindex != fs.first_pindex)) {
825 release_page(&fs);
826 unlock_and_deallocate(&fs);
827 goto RetryFault;
828 }
829
830 /*
831 * Check whether the protection has changed or the object has
832 * been copied while we left the map unlocked. Changing from
833 * read to write permission is OK - we leave the page
834 * write-protected, and catch the write fault. Changing from
835 * write to read permission means that we can't mark the page
836 * write-enabled after all.
837 */
838 prot &= retry_prot;
839 }
840 }
841 if (prot & VM_PROT_WRITE) {
842 vm_object_set_writeable_dirty(fs.object);
843
844 /*
845 * If the fault is a write, we know that this page is being
846 * written NOW so dirty it explicitly to save on
847 * pmap_is_modified() calls later.
848 *
849 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC
850 * if the page is already dirty to prevent data written with
851 * the expectation of being synced from not being synced.
852 * Likewise if this entry does not request NOSYNC then make
853 * sure the page isn't marked NOSYNC. Applications sharing
854 * data should use the same flags to avoid ping ponging.
855 *
856 * Also tell the backing pager, if any, that it should remove
857 * any swap backing since the page is now dirty.
858 */
859 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) {
860 if (fs.m->dirty == 0)
861 fs.m->oflags |= VPO_NOSYNC;
862 } else {
863 fs.m->oflags &= ~VPO_NOSYNC;
864 }
865 if (fault_flags & VM_FAULT_DIRTY) {
866 vm_page_dirty(fs.m);
867 vm_pager_page_unswapped(fs.m);
868 }
869 }
870
871 /*
872 * Page had better still be busy
873 */
874 KASSERT(fs.m->oflags & VPO_BUSY,
875 ("vm_fault: page %p not busy!", fs.m));
876 /*
877 * Sanity check: page must be completely valid or it is not fit to
878 * map into user space. vm_pager_get_pages() ensures this.
879 */
880 if (fs.m->valid != VM_PAGE_BITS_ALL) {
881 vm_page_zero_invalid(fs.m, TRUE);
882 printf("Warning: page %p partially invalid on fault\n", fs.m);
883 }
884 VM_OBJECT_UNLOCK(fs.object);
885
886 /*
887 * Put this page into the physical map. We had to do the unlock above
888 * because pmap_enter() may sleep. We don't put the page
889 * back on the active queue until later so that the pageout daemon
890 * won't find it (yet).
891 */
892 pmap_enter(fs.map->pmap, vaddr, fault_type, fs.m, prot, wired);
893 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) {
894 vm_fault_prefault(fs.map->pmap, vaddr, fs.entry);
895 }
896 VM_OBJECT_LOCK(fs.object);
897 vm_page_lock_queues();
898 vm_page_flag_set(fs.m, PG_REFERENCED);
899
900 /*
901 * If the page is not wired down, then put it where the pageout daemon
902 * can find it.
903 */
904 if (fault_flags & VM_FAULT_WIRE_MASK) {
905 if (wired)
906 vm_page_wire(fs.m);
907 else
908 vm_page_unwire(fs.m, 1);
909 } else {
910 vm_page_activate(fs.m);
911 }
912 vm_page_unlock_queues();
913 vm_page_wakeup(fs.m);
914
915 /*
916 * Unlock everything, and return
917 */
918 unlock_and_deallocate(&fs);
919 if (hardfault)
920 curthread->td_ru.ru_majflt++;
921 else
922 curthread->td_ru.ru_minflt++;
923
924 return (KERN_SUCCESS);
925 }
926
927 /*
928 * vm_fault_prefault provides a quick way of clustering
929 * pagefaults into a processes address space. It is a "cousin"
930 * of vm_map_pmap_enter, except it runs at page fault time instead
931 * of mmap time.
932 */
933 static void
934 vm_fault_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
935 {
936 int i;
937 vm_offset_t addr, starta;
938 vm_pindex_t pindex;
939 vm_page_t m;
940 vm_object_t object;
941
942 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
943 return;
944
945 object = entry->object.vm_object;
946
947 starta = addra - PFBAK * PAGE_SIZE;
948 if (starta < entry->start) {
949 starta = entry->start;
950 } else if (starta > addra) {
951 starta = 0;
952 }
953
954 for (i = 0; i < PAGEORDER_SIZE; i++) {
955 vm_object_t backing_object, lobject;
956
957 addr = addra + prefault_pageorder[i];
958 if (addr > addra + (PFFOR * PAGE_SIZE))
959 addr = 0;
960
961 if (addr < starta || addr >= entry->end)
962 continue;
963
964 if (!pmap_is_prefaultable(pmap, addr))
965 continue;
966
967 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
968 lobject = object;
969 VM_OBJECT_LOCK(lobject);
970 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
971 lobject->type == OBJT_DEFAULT &&
972 (backing_object = lobject->backing_object) != NULL) {
973 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
974 0, ("vm_fault_prefault: unaligned object offset"));
975 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
976 VM_OBJECT_LOCK(backing_object);
977 VM_OBJECT_UNLOCK(lobject);
978 lobject = backing_object;
979 }
980 /*
981 * give-up when a page is not in memory
982 */
983 if (m == NULL) {
984 VM_OBJECT_UNLOCK(lobject);
985 break;
986 }
987 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
988 (m->busy == 0) &&
989 (m->flags & PG_FICTITIOUS) == 0) {
990
991 vm_page_lock_queues();
992 pmap_enter_quick(pmap, addr, m, entry->protection);
993 vm_page_unlock_queues();
994 }
995 VM_OBJECT_UNLOCK(lobject);
996 }
997 }
998
999 /*
1000 * vm_fault_quick:
1001 *
1002 * Ensure that the requested virtual address, which may be in userland,
1003 * is valid. Fault-in the page if necessary. Return -1 on failure.
1004 */
1005 int
1006 vm_fault_quick(caddr_t v, int prot)
1007 {
1008 int r;
1009
1010 if (prot & VM_PROT_WRITE)
1011 r = subyte(v, fubyte(v));
1012 else
1013 r = fubyte(v);
1014 return(r);
1015 }
1016
1017 /*
1018 * vm_fault_wire:
1019 *
1020 * Wire down a range of virtual addresses in a map.
1021 */
1022 int
1023 vm_fault_wire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1024 boolean_t user_wire, boolean_t fictitious)
1025 {
1026 vm_offset_t va;
1027 int rv;
1028
1029 /*
1030 * We simulate a fault to get the page and enter it in the physical
1031 * map. For user wiring, we only ask for read access on currently
1032 * read-only sections.
1033 */
1034 for (va = start; va < end; va += PAGE_SIZE) {
1035 rv = vm_fault(map, va,
1036 user_wire ? VM_PROT_READ : VM_PROT_READ | VM_PROT_WRITE,
1037 user_wire ? VM_FAULT_USER_WIRE : VM_FAULT_CHANGE_WIRING);
1038 if (rv) {
1039 if (va != start)
1040 vm_fault_unwire(map, start, va, fictitious);
1041 return (rv);
1042 }
1043 }
1044 return (KERN_SUCCESS);
1045 }
1046
1047 /*
1048 * vm_fault_unwire:
1049 *
1050 * Unwire a range of virtual addresses in a map.
1051 */
1052 void
1053 vm_fault_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end,
1054 boolean_t fictitious)
1055 {
1056 vm_paddr_t pa;
1057 vm_offset_t va;
1058 pmap_t pmap;
1059
1060 pmap = vm_map_pmap(map);
1061
1062 /*
1063 * Since the pages are wired down, we must be able to get their
1064 * mappings from the physical map system.
1065 */
1066 for (va = start; va < end; va += PAGE_SIZE) {
1067 pa = pmap_extract(pmap, va);
1068 if (pa != 0) {
1069 pmap_change_wiring(pmap, va, FALSE);
1070 if (!fictitious) {
1071 vm_page_lock_queues();
1072 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1073 vm_page_unlock_queues();
1074 }
1075 }
1076 }
1077 }
1078
1079 /*
1080 * Routine:
1081 * vm_fault_copy_entry
1082 * Function:
1083 * Copy all of the pages from a wired-down map entry to another.
1084 *
1085 * In/out conditions:
1086 * The source and destination maps must be locked for write.
1087 * The source map entry must be wired down (or be a sharing map
1088 * entry corresponding to a main map entry that is wired down).
1089 */
1090 void
1091 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry)
1092 vm_map_t dst_map;
1093 vm_map_t src_map;
1094 vm_map_entry_t dst_entry;
1095 vm_map_entry_t src_entry;
1096 {
1097 vm_object_t backing_object, dst_object, object;
1098 vm_object_t src_object;
1099 vm_ooffset_t dst_offset;
1100 vm_ooffset_t src_offset;
1101 vm_pindex_t pindex;
1102 vm_prot_t prot;
1103 vm_offset_t vaddr;
1104 vm_page_t dst_m;
1105 vm_page_t src_m;
1106
1107 #ifdef lint
1108 src_map++;
1109 #endif /* lint */
1110
1111 src_object = src_entry->object.vm_object;
1112 src_offset = src_entry->offset;
1113
1114 /*
1115 * Create the top-level object for the destination entry. (Doesn't
1116 * actually shadow anything - we copy the pages directly.)
1117 */
1118 dst_object = vm_object_allocate(OBJT_DEFAULT,
1119 OFF_TO_IDX(dst_entry->end - dst_entry->start));
1120 #if VM_NRESERVLEVEL > 0
1121 dst_object->flags |= OBJ_COLORED;
1122 dst_object->pg_color = atop(dst_entry->start);
1123 #endif
1124
1125 VM_OBJECT_LOCK(dst_object);
1126 dst_entry->object.vm_object = dst_object;
1127 dst_entry->offset = 0;
1128
1129 prot = dst_entry->max_protection;
1130
1131 /*
1132 * Loop through all of the pages in the entry's range, copying each
1133 * one from the source object (it should be there) to the destination
1134 * object.
1135 */
1136 for (vaddr = dst_entry->start, dst_offset = 0;
1137 vaddr < dst_entry->end;
1138 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1139
1140 /*
1141 * Allocate a page in the destination object
1142 */
1143 do {
1144 dst_m = vm_page_alloc(dst_object,
1145 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1146 if (dst_m == NULL) {
1147 VM_OBJECT_UNLOCK(dst_object);
1148 VM_WAIT;
1149 VM_OBJECT_LOCK(dst_object);
1150 }
1151 } while (dst_m == NULL);
1152
1153 /*
1154 * Find the page in the source object, and copy it in.
1155 * (Because the source is wired down, the page will be in
1156 * memory.)
1157 */
1158 VM_OBJECT_LOCK(src_object);
1159 object = src_object;
1160 pindex = 0;
1161 while ((src_m = vm_page_lookup(object, pindex +
1162 OFF_TO_IDX(dst_offset + src_offset))) == NULL &&
1163 (src_entry->protection & VM_PROT_WRITE) == 0 &&
1164 (backing_object = object->backing_object) != NULL) {
1165 /*
1166 * Allow fallback to backing objects if we are reading.
1167 */
1168 VM_OBJECT_LOCK(backing_object);
1169 pindex += OFF_TO_IDX(object->backing_object_offset);
1170 VM_OBJECT_UNLOCK(object);
1171 object = backing_object;
1172 }
1173 if (src_m == NULL)
1174 panic("vm_fault_copy_wired: page missing");
1175 pmap_copy_page(src_m, dst_m);
1176 VM_OBJECT_UNLOCK(object);
1177 dst_m->valid = VM_PAGE_BITS_ALL;
1178 VM_OBJECT_UNLOCK(dst_object);
1179
1180 /*
1181 * Enter it in the pmap as a read and/or execute access.
1182 */
1183 pmap_enter(dst_map->pmap, vaddr, prot & ~VM_PROT_WRITE, dst_m,
1184 prot, FALSE);
1185
1186 /*
1187 * Mark it no longer busy, and put it on the active list.
1188 */
1189 VM_OBJECT_LOCK(dst_object);
1190 vm_page_lock_queues();
1191 vm_page_activate(dst_m);
1192 vm_page_unlock_queues();
1193 vm_page_wakeup(dst_m);
1194 }
1195 VM_OBJECT_UNLOCK(dst_object);
1196 }
1197
1198
1199 /*
1200 * This routine checks around the requested page for other pages that
1201 * might be able to be faulted in. This routine brackets the viable
1202 * pages for the pages to be paged in.
1203 *
1204 * Inputs:
1205 * m, rbehind, rahead
1206 *
1207 * Outputs:
1208 * marray (array of vm_page_t), reqpage (index of requested page)
1209 *
1210 * Return value:
1211 * number of pages in marray
1212 *
1213 * This routine can't block.
1214 */
1215 static int
1216 vm_fault_additional_pages(m, rbehind, rahead, marray, reqpage)
1217 vm_page_t m;
1218 int rbehind;
1219 int rahead;
1220 vm_page_t *marray;
1221 int *reqpage;
1222 {
1223 int i,j;
1224 vm_object_t object;
1225 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1226 vm_page_t rtm;
1227 int cbehind, cahead;
1228
1229 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1230
1231 object = m->object;
1232 pindex = m->pindex;
1233 cbehind = cahead = 0;
1234
1235 /*
1236 * if the requested page is not available, then give up now
1237 */
1238 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
1239 return 0;
1240 }
1241
1242 if ((cbehind == 0) && (cahead == 0)) {
1243 *reqpage = 0;
1244 marray[0] = m;
1245 return 1;
1246 }
1247
1248 if (rahead > cahead) {
1249 rahead = cahead;
1250 }
1251
1252 if (rbehind > cbehind) {
1253 rbehind = cbehind;
1254 }
1255
1256 /*
1257 * scan backward for the read behind pages -- in memory
1258 */
1259 if (pindex > 0) {
1260 if (rbehind > pindex) {
1261 rbehind = pindex;
1262 startpindex = 0;
1263 } else {
1264 startpindex = pindex - rbehind;
1265 }
1266
1267 if ((rtm = TAILQ_PREV(m, pglist, listq)) != NULL &&
1268 rtm->pindex >= startpindex)
1269 startpindex = rtm->pindex + 1;
1270
1271 /* tpindex is unsigned; beware of numeric underflow. */
1272 for (i = 0, tpindex = pindex - 1; tpindex >= startpindex &&
1273 tpindex < pindex; i++, tpindex--) {
1274
1275 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1276 VM_ALLOC_IFNOTCACHED);
1277 if (rtm == NULL) {
1278 /*
1279 * Shift the allocated pages to the
1280 * beginning of the array.
1281 */
1282 for (j = 0; j < i; j++) {
1283 marray[j] = marray[j + tpindex + 1 -
1284 startpindex];
1285 }
1286 break;
1287 }
1288
1289 marray[tpindex - startpindex] = rtm;
1290 }
1291 } else {
1292 startpindex = 0;
1293 i = 0;
1294 }
1295
1296 marray[i] = m;
1297 /* page offset of the required page */
1298 *reqpage = i;
1299
1300 tpindex = pindex + 1;
1301 i++;
1302
1303 /*
1304 * scan forward for the read ahead pages
1305 */
1306 endpindex = tpindex + rahead;
1307 if ((rtm = TAILQ_NEXT(m, listq)) != NULL && rtm->pindex < endpindex)
1308 endpindex = rtm->pindex;
1309 if (endpindex > object->size)
1310 endpindex = object->size;
1311
1312 for (; tpindex < endpindex; i++, tpindex++) {
1313
1314 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL |
1315 VM_ALLOC_IFNOTCACHED);
1316 if (rtm == NULL) {
1317 break;
1318 }
1319
1320 marray[i] = rtm;
1321 }
1322
1323 /* return number of pages */
1324 return i;
1325 }
Cache object: 82814ee83e8cc2ca56cf0742081a2282
|