1 /*-
2 * SPDX-License-Identifier: (BSD-4-Clause AND MIT-CMU)
3 *
4 * Copyright (c) 1991, 1993
5 * The Regents of the University of California. All rights reserved.
6 * Copyright (c) 1994 John S. Dyson
7 * All rights reserved.
8 * Copyright (c) 1994 David Greenman
9 * All rights reserved.
10 *
11 *
12 * This code is derived from software contributed to Berkeley by
13 * The Mach Operating System project at Carnegie-Mellon University.
14 *
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
17 * are met:
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
30 *
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * SUCH DAMAGE.
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 #include "opt_ktrace.h"
76 #include "opt_vm.h"
77
78 #include <sys/param.h>
79 #include <sys/systm.h>
80 #include <sys/kernel.h>
81 #include <sys/lock.h>
82 #include <sys/mman.h>
83 #include <sys/mutex.h>
84 #include <sys/pctrie.h>
85 #include <sys/proc.h>
86 #include <sys/racct.h>
87 #include <sys/refcount.h>
88 #include <sys/resourcevar.h>
89 #include <sys/rwlock.h>
90 #include <sys/signalvar.h>
91 #include <sys/sysctl.h>
92 #include <sys/sysent.h>
93 #include <sys/vmmeter.h>
94 #include <sys/vnode.h>
95 #ifdef KTRACE
96 #include <sys/ktrace.h>
97 #endif
98
99 #include <vm/vm.h>
100 #include <vm/vm_param.h>
101 #include <vm/pmap.h>
102 #include <vm/vm_map.h>
103 #include <vm/vm_object.h>
104 #include <vm/vm_page.h>
105 #include <vm/vm_pageout.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_pager.h>
108 #include <vm/vm_extern.h>
109 #include <vm/vm_reserv.h>
110
111 #define PFBAK 4
112 #define PFFOR 4
113
114 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
115
116 #define VM_FAULT_DONTNEED_MIN 1048576
117
118 struct faultstate {
119 /* Fault parameters. */
120 vm_offset_t vaddr;
121 vm_page_t *m_hold;
122 vm_prot_t fault_type;
123 vm_prot_t prot;
124 int fault_flags;
125 boolean_t wired;
126
127 /* Control state. */
128 struct timeval oom_start_time;
129 bool oom_started;
130 int nera;
131 bool can_read_lock;
132
133 /* Page reference for cow. */
134 vm_page_t m_cow;
135
136 /* Current object. */
137 vm_object_t object;
138 vm_pindex_t pindex;
139 vm_page_t m;
140
141 /* Top-level map object. */
142 vm_object_t first_object;
143 vm_pindex_t first_pindex;
144 vm_page_t first_m;
145
146 /* Map state. */
147 vm_map_t map;
148 vm_map_entry_t entry;
149 int map_generation;
150 bool lookup_still_valid;
151
152 /* Vnode if locked. */
153 struct vnode *vp;
154 };
155
156 /*
157 * Return codes for internal fault routines.
158 */
159 enum fault_status {
160 FAULT_SUCCESS = 10000, /* Return success to user. */
161 FAULT_FAILURE, /* Return failure to user. */
162 FAULT_CONTINUE, /* Continue faulting. */
163 FAULT_RESTART, /* Restart fault. */
164 FAULT_OUT_OF_BOUNDS, /* Invalid address for pager. */
165 FAULT_HARD, /* Performed I/O. */
166 FAULT_SOFT, /* Found valid page. */
167 FAULT_PROTECTION_FAILURE, /* Invalid access. */
168 };
169
170 enum fault_next_status {
171 FAULT_NEXT_GOTOBJ = 1,
172 FAULT_NEXT_NOOBJ,
173 FAULT_NEXT_RESTART,
174 };
175
176 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
177 int ahead);
178 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
179 int backward, int forward, bool obj_locked);
180
181 static int vm_pfault_oom_attempts = 3;
182 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
183 &vm_pfault_oom_attempts, 0,
184 "Number of page allocation attempts in page fault handler before it "
185 "triggers OOM handling");
186
187 static int vm_pfault_oom_wait = 10;
188 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
189 &vm_pfault_oom_wait, 0,
190 "Number of seconds to wait for free pages before retrying "
191 "the page fault handler");
192
193 static inline void
vm_fault_page_release(vm_page_t * mp)194 vm_fault_page_release(vm_page_t *mp)
195 {
196 vm_page_t m;
197
198 m = *mp;
199 if (m != NULL) {
200 /*
201 * We are likely to loop around again and attempt to busy
202 * this page. Deactivating it leaves it available for
203 * pageout while optimizing fault restarts.
204 */
205 vm_page_deactivate(m);
206 vm_page_xunbusy(m);
207 *mp = NULL;
208 }
209 }
210
211 static inline void
vm_fault_page_free(vm_page_t * mp)212 vm_fault_page_free(vm_page_t *mp)
213 {
214 vm_page_t m;
215
216 m = *mp;
217 if (m != NULL) {
218 VM_OBJECT_ASSERT_WLOCKED(m->object);
219 if (!vm_page_wired(m))
220 vm_page_free(m);
221 else
222 vm_page_xunbusy(m);
223 *mp = NULL;
224 }
225 }
226
227 /*
228 * Return true if a vm_pager_get_pages() call is needed in order to check
229 * whether the pager might have a particular page, false if it can be determined
230 * immediately that the pager can not have a copy. For swap objects, this can
231 * be checked quickly.
232 */
233 static inline bool
vm_fault_object_needs_getpages(vm_object_t object)234 vm_fault_object_needs_getpages(vm_object_t object)
235 {
236 VM_OBJECT_ASSERT_LOCKED(object);
237
238 return ((object->flags & OBJ_SWAP) == 0 ||
239 !pctrie_is_empty(&object->un_pager.swp.swp_blks));
240 }
241
242 static inline void
vm_fault_unlock_map(struct faultstate * fs)243 vm_fault_unlock_map(struct faultstate *fs)
244 {
245
246 if (fs->lookup_still_valid) {
247 vm_map_lookup_done(fs->map, fs->entry);
248 fs->lookup_still_valid = false;
249 }
250 }
251
252 static void
vm_fault_unlock_vp(struct faultstate * fs)253 vm_fault_unlock_vp(struct faultstate *fs)
254 {
255
256 if (fs->vp != NULL) {
257 vput(fs->vp);
258 fs->vp = NULL;
259 }
260 }
261
262 static void
vm_fault_deallocate(struct faultstate * fs)263 vm_fault_deallocate(struct faultstate *fs)
264 {
265
266 vm_fault_page_release(&fs->m_cow);
267 vm_fault_page_release(&fs->m);
268 vm_object_pip_wakeup(fs->object);
269 if (fs->object != fs->first_object) {
270 VM_OBJECT_WLOCK(fs->first_object);
271 vm_fault_page_free(&fs->first_m);
272 VM_OBJECT_WUNLOCK(fs->first_object);
273 vm_object_pip_wakeup(fs->first_object);
274 }
275 vm_object_deallocate(fs->first_object);
276 vm_fault_unlock_map(fs);
277 vm_fault_unlock_vp(fs);
278 }
279
280 static void
vm_fault_unlock_and_deallocate(struct faultstate * fs)281 vm_fault_unlock_and_deallocate(struct faultstate *fs)
282 {
283
284 VM_OBJECT_UNLOCK(fs->object);
285 vm_fault_deallocate(fs);
286 }
287
288 static void
vm_fault_dirty(struct faultstate * fs,vm_page_t m)289 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
290 {
291 bool need_dirty;
292
293 if (((fs->prot & VM_PROT_WRITE) == 0 &&
294 (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
295 (m->oflags & VPO_UNMANAGED) != 0)
296 return;
297
298 VM_PAGE_OBJECT_BUSY_ASSERT(m);
299
300 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
301 (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
302 (fs->fault_flags & VM_FAULT_DIRTY) != 0;
303
304 vm_object_set_writeable_dirty(m->object);
305
306 /*
307 * If the fault is a write, we know that this page is being
308 * written NOW so dirty it explicitly to save on
309 * pmap_is_modified() calls later.
310 *
311 * Also, since the page is now dirty, we can possibly tell
312 * the pager to release any swap backing the page.
313 */
314 if (need_dirty && vm_page_set_dirty(m) == 0) {
315 /*
316 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
317 * if the page is already dirty to prevent data written with
318 * the expectation of being synced from not being synced.
319 * Likewise if this entry does not request NOSYNC then make
320 * sure the page isn't marked NOSYNC. Applications sharing
321 * data should use the same flags to avoid ping ponging.
322 */
323 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
324 vm_page_aflag_set(m, PGA_NOSYNC);
325 else
326 vm_page_aflag_clear(m, PGA_NOSYNC);
327 }
328
329 }
330
331 /*
332 * Unlocks fs.first_object and fs.map on success.
333 */
334 static enum fault_status
vm_fault_soft_fast(struct faultstate * fs)335 vm_fault_soft_fast(struct faultstate *fs)
336 {
337 vm_page_t m, m_map;
338 #if VM_NRESERVLEVEL > 0
339 vm_page_t m_super;
340 int flags;
341 #endif
342 int psind;
343 vm_offset_t vaddr;
344
345 MPASS(fs->vp == NULL);
346
347 /*
348 * If we fail, vast majority of the time it is because the page is not
349 * there to begin with. Opportunistically perform the lookup and
350 * subsequent checks without the object lock, revalidate later.
351 *
352 * Note: a busy page can be mapped for read|execute access.
353 */
354 m = vm_page_lookup_unlocked(fs->first_object, fs->first_pindex);
355 if (m == NULL || !vm_page_all_valid(m) ||
356 ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m))) {
357 VM_OBJECT_WLOCK(fs->first_object);
358 return (FAULT_FAILURE);
359 }
360
361 vaddr = fs->vaddr;
362
363 VM_OBJECT_RLOCK(fs->first_object);
364
365 /*
366 * Now that we stabilized the state, revalidate the page is in the shape
367 * we encountered above.
368 */
369
370 if (m->object != fs->first_object || m->pindex != fs->first_pindex)
371 goto fail;
372
373 vm_object_busy(fs->first_object);
374
375 if (!vm_page_all_valid(m) ||
376 ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m)))
377 goto fail_busy;
378
379 m_map = m;
380 psind = 0;
381 #if VM_NRESERVLEVEL > 0
382 if ((m->flags & PG_FICTITIOUS) == 0 &&
383 (m_super = vm_reserv_to_superpage(m)) != NULL &&
384 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start &&
385 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end &&
386 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) &
387 (pagesizes[m_super->psind] - 1)) &&
388 pmap_ps_enabled(fs->map->pmap)) {
389 flags = PS_ALL_VALID;
390 if ((fs->prot & VM_PROT_WRITE) != 0) {
391 /*
392 * Create a superpage mapping allowing write access
393 * only if none of the constituent pages are busy and
394 * all of them are already dirty (except possibly for
395 * the page that was faulted on).
396 */
397 flags |= PS_NONE_BUSY;
398 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
399 flags |= PS_ALL_DIRTY;
400 }
401 if (vm_page_ps_test(m_super, flags, m)) {
402 m_map = m_super;
403 psind = m_super->psind;
404 vaddr = rounddown2(vaddr, pagesizes[psind]);
405 /* Preset the modified bit for dirty superpages. */
406 if ((flags & PS_ALL_DIRTY) != 0)
407 fs->fault_type |= VM_PROT_WRITE;
408 }
409 }
410 #endif
411 if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
412 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) !=
413 KERN_SUCCESS)
414 goto fail_busy;
415 if (fs->m_hold != NULL) {
416 (*fs->m_hold) = m;
417 vm_page_wire(m);
418 }
419 if (psind == 0 && !fs->wired)
420 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
421 VM_OBJECT_RUNLOCK(fs->first_object);
422 vm_fault_dirty(fs, m);
423 vm_object_unbusy(fs->first_object);
424 vm_map_lookup_done(fs->map, fs->entry);
425 curthread->td_ru.ru_minflt++;
426 return (FAULT_SUCCESS);
427 fail_busy:
428 vm_object_unbusy(fs->first_object);
429 fail:
430 if (!VM_OBJECT_TRYUPGRADE(fs->first_object)) {
431 VM_OBJECT_RUNLOCK(fs->first_object);
432 VM_OBJECT_WLOCK(fs->first_object);
433 }
434 return (FAULT_FAILURE);
435 }
436
437 static void
vm_fault_restore_map_lock(struct faultstate * fs)438 vm_fault_restore_map_lock(struct faultstate *fs)
439 {
440
441 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
442 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
443
444 if (!vm_map_trylock_read(fs->map)) {
445 VM_OBJECT_WUNLOCK(fs->first_object);
446 vm_map_lock_read(fs->map);
447 VM_OBJECT_WLOCK(fs->first_object);
448 }
449 fs->lookup_still_valid = true;
450 }
451
452 static void
vm_fault_populate_check_page(vm_page_t m)453 vm_fault_populate_check_page(vm_page_t m)
454 {
455
456 /*
457 * Check each page to ensure that the pager is obeying the
458 * interface: the page must be installed in the object, fully
459 * valid, and exclusively busied.
460 */
461 MPASS(m != NULL);
462 MPASS(vm_page_all_valid(m));
463 MPASS(vm_page_xbusied(m));
464 }
465
466 static void
vm_fault_populate_cleanup(vm_object_t object,vm_pindex_t first,vm_pindex_t last)467 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
468 vm_pindex_t last)
469 {
470 vm_page_t m;
471 vm_pindex_t pidx;
472
473 VM_OBJECT_ASSERT_WLOCKED(object);
474 MPASS(first <= last);
475 for (pidx = first, m = vm_page_lookup(object, pidx);
476 pidx <= last; pidx++, m = vm_page_next(m)) {
477 vm_fault_populate_check_page(m);
478 vm_page_deactivate(m);
479 vm_page_xunbusy(m);
480 }
481 }
482
483 static enum fault_status
vm_fault_populate(struct faultstate * fs)484 vm_fault_populate(struct faultstate *fs)
485 {
486 vm_offset_t vaddr;
487 vm_page_t m;
488 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
489 int bdry_idx, i, npages, psind, rv;
490 enum fault_status res;
491
492 MPASS(fs->object == fs->first_object);
493 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
494 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
495 MPASS(fs->first_object->backing_object == NULL);
496 MPASS(fs->lookup_still_valid);
497
498 pager_first = OFF_TO_IDX(fs->entry->offset);
499 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
500 vm_fault_unlock_map(fs);
501 vm_fault_unlock_vp(fs);
502
503 res = FAULT_SUCCESS;
504
505 /*
506 * Call the pager (driver) populate() method.
507 *
508 * There is no guarantee that the method will be called again
509 * if the current fault is for read, and a future fault is
510 * for write. Report the entry's maximum allowed protection
511 * to the driver.
512 */
513 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
514 fs->fault_type, fs->entry->max_protection, &pager_first,
515 &pager_last);
516
517 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
518 if (rv == VM_PAGER_BAD) {
519 /*
520 * VM_PAGER_BAD is the backdoor for a pager to request
521 * normal fault handling.
522 */
523 vm_fault_restore_map_lock(fs);
524 if (fs->map->timestamp != fs->map_generation)
525 return (FAULT_RESTART);
526 return (FAULT_CONTINUE);
527 }
528 if (rv != VM_PAGER_OK)
529 return (FAULT_FAILURE); /* AKA SIGSEGV */
530
531 /* Ensure that the driver is obeying the interface. */
532 MPASS(pager_first <= pager_last);
533 MPASS(fs->first_pindex <= pager_last);
534 MPASS(fs->first_pindex >= pager_first);
535 MPASS(pager_last < fs->first_object->size);
536
537 vm_fault_restore_map_lock(fs);
538 bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(fs->entry);
539 if (fs->map->timestamp != fs->map_generation) {
540 if (bdry_idx == 0) {
541 vm_fault_populate_cleanup(fs->first_object, pager_first,
542 pager_last);
543 } else {
544 m = vm_page_lookup(fs->first_object, pager_first);
545 if (m != fs->m)
546 vm_page_xunbusy(m);
547 }
548 return (FAULT_RESTART);
549 }
550
551 /*
552 * The map is unchanged after our last unlock. Process the fault.
553 *
554 * First, the special case of largepage mappings, where
555 * populate only busies the first page in superpage run.
556 */
557 if (bdry_idx != 0) {
558 KASSERT(PMAP_HAS_LARGEPAGES,
559 ("missing pmap support for large pages"));
560 m = vm_page_lookup(fs->first_object, pager_first);
561 vm_fault_populate_check_page(m);
562 VM_OBJECT_WUNLOCK(fs->first_object);
563 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
564 fs->entry->offset;
565 /* assert alignment for entry */
566 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
567 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
568 (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
569 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
570 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
571 ("unaligned superpage m %p %#jx", m,
572 (uintmax_t)VM_PAGE_TO_PHYS(m)));
573 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
574 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
575 PMAP_ENTER_LARGEPAGE, bdry_idx);
576 VM_OBJECT_WLOCK(fs->first_object);
577 vm_page_xunbusy(m);
578 if (rv != KERN_SUCCESS) {
579 res = FAULT_FAILURE;
580 goto out;
581 }
582 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
583 for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
584 vm_page_wire(m + i);
585 }
586 if (fs->m_hold != NULL) {
587 *fs->m_hold = m + (fs->first_pindex - pager_first);
588 vm_page_wire(*fs->m_hold);
589 }
590 goto out;
591 }
592
593 /*
594 * The range [pager_first, pager_last] that is given to the
595 * pager is only a hint. The pager may populate any range
596 * within the object that includes the requested page index.
597 * In case the pager expanded the range, clip it to fit into
598 * the map entry.
599 */
600 map_first = OFF_TO_IDX(fs->entry->offset);
601 if (map_first > pager_first) {
602 vm_fault_populate_cleanup(fs->first_object, pager_first,
603 map_first - 1);
604 pager_first = map_first;
605 }
606 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
607 if (map_last < pager_last) {
608 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
609 pager_last);
610 pager_last = map_last;
611 }
612 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx);
613 pidx <= pager_last;
614 pidx += npages, m = vm_page_next(&m[npages - 1])) {
615 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
616
617 psind = m->psind;
618 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
619 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
620 !pmap_ps_enabled(fs->map->pmap)))
621 psind = 0;
622
623 npages = atop(pagesizes[psind]);
624 for (i = 0; i < npages; i++) {
625 vm_fault_populate_check_page(&m[i]);
626 vm_fault_dirty(fs, &m[i]);
627 }
628 VM_OBJECT_WUNLOCK(fs->first_object);
629 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
630 (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
631
632 /*
633 * pmap_enter() may fail for a superpage mapping if additional
634 * protection policies prevent the full mapping.
635 * For example, this will happen on amd64 if the entire
636 * address range does not share the same userspace protection
637 * key. Revert to single-page mappings if this happens.
638 */
639 MPASS(rv == KERN_SUCCESS ||
640 (psind > 0 && rv == KERN_PROTECTION_FAILURE));
641 if (__predict_false(psind > 0 &&
642 rv == KERN_PROTECTION_FAILURE)) {
643 MPASS(!fs->wired);
644 for (i = 0; i < npages; i++) {
645 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
646 &m[i], fs->prot, fs->fault_type, 0);
647 MPASS(rv == KERN_SUCCESS);
648 }
649 }
650
651 VM_OBJECT_WLOCK(fs->first_object);
652 for (i = 0; i < npages; i++) {
653 if ((fs->fault_flags & VM_FAULT_WIRE) != 0 &&
654 m[i].pindex == fs->first_pindex)
655 vm_page_wire(&m[i]);
656 else
657 vm_page_activate(&m[i]);
658 if (fs->m_hold != NULL &&
659 m[i].pindex == fs->first_pindex) {
660 (*fs->m_hold) = &m[i];
661 vm_page_wire(&m[i]);
662 }
663 vm_page_xunbusy(&m[i]);
664 }
665 }
666 out:
667 curthread->td_ru.ru_majflt++;
668 return (res);
669 }
670
671 static int prot_fault_translation;
672 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
673 &prot_fault_translation, 0,
674 "Control signal to deliver on protection fault");
675
676 /* compat definition to keep common code for signal translation */
677 #define UCODE_PAGEFLT 12
678 #ifdef T_PAGEFLT
679 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
680 #endif
681
682 /*
683 * vm_fault_trap:
684 *
685 * Handle a page fault occurring at the given address,
686 * requiring the given permissions, in the map specified.
687 * If successful, the page is inserted into the
688 * associated physical map.
689 *
690 * NOTE: the given address should be truncated to the
691 * proper page address.
692 *
693 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
694 * a standard error specifying why the fault is fatal is returned.
695 *
696 * The map in question must be referenced, and remains so.
697 * Caller may hold no locks.
698 */
699 int
vm_fault_trap(vm_map_t map,vm_offset_t vaddr,vm_prot_t fault_type,int fault_flags,int * signo,int * ucode)700 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
701 int fault_flags, int *signo, int *ucode)
702 {
703 int result;
704
705 MPASS(signo == NULL || ucode != NULL);
706 #ifdef KTRACE
707 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
708 ktrfault(vaddr, fault_type);
709 #endif
710 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
711 NULL);
712 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
713 result == KERN_INVALID_ADDRESS ||
714 result == KERN_RESOURCE_SHORTAGE ||
715 result == KERN_PROTECTION_FAILURE ||
716 result == KERN_OUT_OF_BOUNDS,
717 ("Unexpected Mach error %d from vm_fault()", result));
718 #ifdef KTRACE
719 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
720 ktrfaultend(result);
721 #endif
722 if (result != KERN_SUCCESS && signo != NULL) {
723 switch (result) {
724 case KERN_FAILURE:
725 case KERN_INVALID_ADDRESS:
726 *signo = SIGSEGV;
727 *ucode = SEGV_MAPERR;
728 break;
729 case KERN_RESOURCE_SHORTAGE:
730 *signo = SIGBUS;
731 *ucode = BUS_OOMERR;
732 break;
733 case KERN_OUT_OF_BOUNDS:
734 *signo = SIGBUS;
735 *ucode = BUS_OBJERR;
736 break;
737 case KERN_PROTECTION_FAILURE:
738 if (prot_fault_translation == 0) {
739 /*
740 * Autodetect. This check also covers
741 * the images without the ABI-tag ELF
742 * note.
743 */
744 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
745 curproc->p_osrel >= P_OSREL_SIGSEGV) {
746 *signo = SIGSEGV;
747 *ucode = SEGV_ACCERR;
748 } else {
749 *signo = SIGBUS;
750 *ucode = UCODE_PAGEFLT;
751 }
752 } else if (prot_fault_translation == 1) {
753 /* Always compat mode. */
754 *signo = SIGBUS;
755 *ucode = UCODE_PAGEFLT;
756 } else {
757 /* Always SIGSEGV mode. */
758 *signo = SIGSEGV;
759 *ucode = SEGV_ACCERR;
760 }
761 break;
762 default:
763 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
764 result));
765 break;
766 }
767 }
768 return (result);
769 }
770
771 static bool
vm_fault_object_ensure_wlocked(struct faultstate * fs)772 vm_fault_object_ensure_wlocked(struct faultstate *fs)
773 {
774 if (fs->object == fs->first_object)
775 VM_OBJECT_ASSERT_WLOCKED(fs->object);
776
777 if (!fs->can_read_lock) {
778 VM_OBJECT_ASSERT_WLOCKED(fs->object);
779 return (true);
780 }
781
782 if (VM_OBJECT_WOWNED(fs->object))
783 return (true);
784
785 if (VM_OBJECT_TRYUPGRADE(fs->object))
786 return (true);
787
788 return (false);
789 }
790
791 static enum fault_status
vm_fault_lock_vnode(struct faultstate * fs,bool objlocked)792 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
793 {
794 struct vnode *vp;
795 int error, locked;
796
797 if (fs->object->type != OBJT_VNODE)
798 return (FAULT_CONTINUE);
799 vp = fs->object->handle;
800 if (vp == fs->vp) {
801 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
802 return (FAULT_CONTINUE);
803 }
804
805 /*
806 * Perform an unlock in case the desired vnode changed while
807 * the map was unlocked during a retry.
808 */
809 vm_fault_unlock_vp(fs);
810
811 locked = VOP_ISLOCKED(vp);
812 if (locked != LK_EXCLUSIVE)
813 locked = LK_SHARED;
814
815 /*
816 * We must not sleep acquiring the vnode lock while we have
817 * the page exclusive busied or the object's
818 * paging-in-progress count incremented. Otherwise, we could
819 * deadlock.
820 */
821 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
822 if (error == 0) {
823 fs->vp = vp;
824 return (FAULT_CONTINUE);
825 }
826
827 vhold(vp);
828 if (objlocked)
829 vm_fault_unlock_and_deallocate(fs);
830 else
831 vm_fault_deallocate(fs);
832 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
833 vdrop(vp);
834 fs->vp = vp;
835 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
836 return (FAULT_RESTART);
837 }
838
839 /*
840 * Calculate the desired readahead. Handle drop-behind.
841 *
842 * Returns the number of readahead blocks to pass to the pager.
843 */
844 static int
vm_fault_readahead(struct faultstate * fs)845 vm_fault_readahead(struct faultstate *fs)
846 {
847 int era, nera;
848 u_char behavior;
849
850 KASSERT(fs->lookup_still_valid, ("map unlocked"));
851 era = fs->entry->read_ahead;
852 behavior = vm_map_entry_behavior(fs->entry);
853 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
854 nera = 0;
855 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
856 nera = VM_FAULT_READ_AHEAD_MAX;
857 if (fs->vaddr == fs->entry->next_read)
858 vm_fault_dontneed(fs, fs->vaddr, nera);
859 } else if (fs->vaddr == fs->entry->next_read) {
860 /*
861 * This is a sequential fault. Arithmetically
862 * increase the requested number of pages in
863 * the read-ahead window. The requested
864 * number of pages is "# of sequential faults
865 * x (read ahead min + 1) + read ahead min"
866 */
867 nera = VM_FAULT_READ_AHEAD_MIN;
868 if (era > 0) {
869 nera += era + 1;
870 if (nera > VM_FAULT_READ_AHEAD_MAX)
871 nera = VM_FAULT_READ_AHEAD_MAX;
872 }
873 if (era == VM_FAULT_READ_AHEAD_MAX)
874 vm_fault_dontneed(fs, fs->vaddr, nera);
875 } else {
876 /*
877 * This is a non-sequential fault.
878 */
879 nera = 0;
880 }
881 if (era != nera) {
882 /*
883 * A read lock on the map suffices to update
884 * the read ahead count safely.
885 */
886 fs->entry->read_ahead = nera;
887 }
888
889 return (nera);
890 }
891
892 static int
vm_fault_lookup(struct faultstate * fs)893 vm_fault_lookup(struct faultstate *fs)
894 {
895 int result;
896
897 KASSERT(!fs->lookup_still_valid,
898 ("vm_fault_lookup: Map already locked."));
899 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
900 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
901 &fs->first_pindex, &fs->prot, &fs->wired);
902 if (result != KERN_SUCCESS) {
903 vm_fault_unlock_vp(fs);
904 return (result);
905 }
906
907 fs->map_generation = fs->map->timestamp;
908
909 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
910 panic("%s: fault on nofault entry, addr: %#lx",
911 __func__, (u_long)fs->vaddr);
912 }
913
914 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
915 fs->entry->wiring_thread != curthread) {
916 vm_map_unlock_read(fs->map);
917 vm_map_lock(fs->map);
918 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
919 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
920 vm_fault_unlock_vp(fs);
921 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
922 vm_map_unlock_and_wait(fs->map, 0);
923 } else
924 vm_map_unlock(fs->map);
925 return (KERN_RESOURCE_SHORTAGE);
926 }
927
928 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
929
930 if (fs->wired)
931 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
932 else
933 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
934 ("!fs->wired && VM_FAULT_WIRE"));
935 fs->lookup_still_valid = true;
936
937 return (KERN_SUCCESS);
938 }
939
940 static int
vm_fault_relookup(struct faultstate * fs)941 vm_fault_relookup(struct faultstate *fs)
942 {
943 vm_object_t retry_object;
944 vm_pindex_t retry_pindex;
945 vm_prot_t retry_prot;
946 int result;
947
948 if (!vm_map_trylock_read(fs->map))
949 return (KERN_RESTART);
950
951 fs->lookup_still_valid = true;
952 if (fs->map->timestamp == fs->map_generation)
953 return (KERN_SUCCESS);
954
955 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
956 &fs->entry, &retry_object, &retry_pindex, &retry_prot,
957 &fs->wired);
958 if (result != KERN_SUCCESS) {
959 /*
960 * If retry of map lookup would have blocked then
961 * retry fault from start.
962 */
963 if (result == KERN_FAILURE)
964 return (KERN_RESTART);
965 return (result);
966 }
967 if (retry_object != fs->first_object ||
968 retry_pindex != fs->first_pindex)
969 return (KERN_RESTART);
970
971 /*
972 * Check whether the protection has changed or the object has
973 * been copied while we left the map unlocked. Changing from
974 * read to write permission is OK - we leave the page
975 * write-protected, and catch the write fault. Changing from
976 * write to read permission means that we can't mark the page
977 * write-enabled after all.
978 */
979 fs->prot &= retry_prot;
980 fs->fault_type &= retry_prot;
981 if (fs->prot == 0)
982 return (KERN_RESTART);
983
984 /* Reassert because wired may have changed. */
985 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
986 ("!wired && VM_FAULT_WIRE"));
987
988 return (KERN_SUCCESS);
989 }
990
991 static void
vm_fault_cow(struct faultstate * fs)992 vm_fault_cow(struct faultstate *fs)
993 {
994 bool is_first_object_locked;
995
996 KASSERT(fs->object != fs->first_object,
997 ("source and target COW objects are identical"));
998
999 /*
1000 * This allows pages to be virtually copied from a backing_object
1001 * into the first_object, where the backing object has no other
1002 * refs to it, and cannot gain any more refs. Instead of a bcopy,
1003 * we just move the page from the backing object to the first
1004 * object. Note that we must mark the page dirty in the first
1005 * object so that it will go out to swap when needed.
1006 */
1007 is_first_object_locked = false;
1008 if (
1009 /*
1010 * Only one shadow object and no other refs.
1011 */
1012 fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
1013 /*
1014 * No other ways to look the object up
1015 */
1016 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 &&
1017 /*
1018 * We don't chase down the shadow chain and we can acquire locks.
1019 */
1020 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) &&
1021 fs->object == fs->first_object->backing_object &&
1022 VM_OBJECT_TRYWLOCK(fs->object)) {
1023 /*
1024 * Remove but keep xbusy for replace. fs->m is moved into
1025 * fs->first_object and left busy while fs->first_m is
1026 * conditionally freed.
1027 */
1028 vm_page_remove_xbusy(fs->m);
1029 vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
1030 fs->first_m);
1031 vm_page_dirty(fs->m);
1032 #if VM_NRESERVLEVEL > 0
1033 /*
1034 * Rename the reservation.
1035 */
1036 vm_reserv_rename(fs->m, fs->first_object, fs->object,
1037 OFF_TO_IDX(fs->first_object->backing_object_offset));
1038 #endif
1039 VM_OBJECT_WUNLOCK(fs->object);
1040 VM_OBJECT_WUNLOCK(fs->first_object);
1041 fs->first_m = fs->m;
1042 fs->m = NULL;
1043 VM_CNT_INC(v_cow_optim);
1044 } else {
1045 if (is_first_object_locked)
1046 VM_OBJECT_WUNLOCK(fs->first_object);
1047 /*
1048 * Oh, well, lets copy it.
1049 */
1050 pmap_copy_page(fs->m, fs->first_m);
1051 vm_page_valid(fs->first_m);
1052 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
1053 vm_page_wire(fs->first_m);
1054 vm_page_unwire(fs->m, PQ_INACTIVE);
1055 }
1056 /*
1057 * Save the cow page to be released after
1058 * pmap_enter is complete.
1059 */
1060 fs->m_cow = fs->m;
1061 fs->m = NULL;
1062
1063 /*
1064 * Typically, the shadow object is either private to this
1065 * address space (OBJ_ONEMAPPING) or its pages are read only.
1066 * In the highly unusual case where the pages of a shadow object
1067 * are read/write shared between this and other address spaces,
1068 * we need to ensure that any pmap-level mappings to the
1069 * original, copy-on-write page from the backing object are
1070 * removed from those other address spaces.
1071 *
1072 * The flag check is racy, but this is tolerable: if
1073 * OBJ_ONEMAPPING is cleared after the check, the busy state
1074 * ensures that new mappings of m_cow can't be created.
1075 * pmap_enter() will replace an existing mapping in the current
1076 * address space. If OBJ_ONEMAPPING is set after the check,
1077 * removing mappings will at worse trigger some unnecessary page
1078 * faults.
1079 */
1080 vm_page_assert_xbusied(fs->m_cow);
1081 if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
1082 pmap_remove_all(fs->m_cow);
1083 }
1084
1085 vm_object_pip_wakeup(fs->object);
1086
1087 /*
1088 * Only use the new page below...
1089 */
1090 fs->object = fs->first_object;
1091 fs->pindex = fs->first_pindex;
1092 fs->m = fs->first_m;
1093 VM_CNT_INC(v_cow_faults);
1094 curthread->td_cow++;
1095 }
1096
1097 static enum fault_next_status
vm_fault_next(struct faultstate * fs)1098 vm_fault_next(struct faultstate *fs)
1099 {
1100 vm_object_t next_object;
1101
1102 if (fs->object == fs->first_object || !fs->can_read_lock)
1103 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1104 else
1105 VM_OBJECT_ASSERT_LOCKED(fs->object);
1106
1107 /*
1108 * The requested page does not exist at this object/
1109 * offset. Remove the invalid page from the object,
1110 * waking up anyone waiting for it, and continue on to
1111 * the next object. However, if this is the top-level
1112 * object, we must leave the busy page in place to
1113 * prevent another process from rushing past us, and
1114 * inserting the page in that object at the same time
1115 * that we are.
1116 */
1117 if (fs->object == fs->first_object) {
1118 fs->first_m = fs->m;
1119 fs->m = NULL;
1120 } else if (fs->m != NULL) {
1121 if (!vm_fault_object_ensure_wlocked(fs)) {
1122 fs->can_read_lock = false;
1123 vm_fault_unlock_and_deallocate(fs);
1124 return (FAULT_NEXT_RESTART);
1125 }
1126 vm_fault_page_free(&fs->m);
1127 }
1128
1129 /*
1130 * Move on to the next object. Lock the next object before
1131 * unlocking the current one.
1132 */
1133 next_object = fs->object->backing_object;
1134 if (next_object == NULL)
1135 return (FAULT_NEXT_NOOBJ);
1136 MPASS(fs->first_m != NULL);
1137 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1138 if (fs->can_read_lock)
1139 VM_OBJECT_RLOCK(next_object);
1140 else
1141 VM_OBJECT_WLOCK(next_object);
1142 vm_object_pip_add(next_object, 1);
1143 if (fs->object != fs->first_object)
1144 vm_object_pip_wakeup(fs->object);
1145 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1146 VM_OBJECT_UNLOCK(fs->object);
1147 fs->object = next_object;
1148
1149 return (FAULT_NEXT_GOTOBJ);
1150 }
1151
1152 static void
vm_fault_zerofill(struct faultstate * fs)1153 vm_fault_zerofill(struct faultstate *fs)
1154 {
1155
1156 /*
1157 * If there's no object left, fill the page in the top
1158 * object with zeros.
1159 */
1160 if (fs->object != fs->first_object) {
1161 vm_object_pip_wakeup(fs->object);
1162 fs->object = fs->first_object;
1163 fs->pindex = fs->first_pindex;
1164 }
1165 MPASS(fs->first_m != NULL);
1166 MPASS(fs->m == NULL);
1167 fs->m = fs->first_m;
1168 fs->first_m = NULL;
1169
1170 /*
1171 * Zero the page if necessary and mark it valid.
1172 */
1173 if ((fs->m->flags & PG_ZERO) == 0) {
1174 pmap_zero_page(fs->m);
1175 } else {
1176 VM_CNT_INC(v_ozfod);
1177 }
1178 VM_CNT_INC(v_zfod);
1179 vm_page_valid(fs->m);
1180 }
1181
1182 /*
1183 * Initiate page fault after timeout. Returns true if caller should
1184 * do vm_waitpfault() after the call.
1185 */
1186 static bool
vm_fault_allocate_oom(struct faultstate * fs)1187 vm_fault_allocate_oom(struct faultstate *fs)
1188 {
1189 struct timeval now;
1190
1191 vm_fault_unlock_and_deallocate(fs);
1192 if (vm_pfault_oom_attempts < 0)
1193 return (true);
1194 if (!fs->oom_started) {
1195 fs->oom_started = true;
1196 getmicrotime(&fs->oom_start_time);
1197 return (true);
1198 }
1199
1200 getmicrotime(&now);
1201 timevalsub(&now, &fs->oom_start_time);
1202 if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait)
1203 return (true);
1204
1205 if (bootverbose)
1206 printf(
1207 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1208 curproc->p_pid, curproc->p_comm);
1209 vm_pageout_oom(VM_OOM_MEM_PF);
1210 fs->oom_started = false;
1211 return (false);
1212 }
1213
1214 /*
1215 * Allocate a page directly or via the object populate method.
1216 */
1217 static enum fault_status
vm_fault_allocate(struct faultstate * fs)1218 vm_fault_allocate(struct faultstate *fs)
1219 {
1220 struct domainset *dset;
1221 enum fault_status res;
1222
1223 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1224 res = vm_fault_lock_vnode(fs, true);
1225 MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART);
1226 if (res == FAULT_RESTART)
1227 return (res);
1228 }
1229
1230 if (fs->pindex >= fs->object->size) {
1231 vm_fault_unlock_and_deallocate(fs);
1232 return (FAULT_OUT_OF_BOUNDS);
1233 }
1234
1235 if (fs->object == fs->first_object &&
1236 (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1237 fs->first_object->shadow_count == 0) {
1238 res = vm_fault_populate(fs);
1239 switch (res) {
1240 case FAULT_SUCCESS:
1241 case FAULT_FAILURE:
1242 case FAULT_RESTART:
1243 vm_fault_unlock_and_deallocate(fs);
1244 return (res);
1245 case FAULT_CONTINUE:
1246 /*
1247 * Pager's populate() method
1248 * returned VM_PAGER_BAD.
1249 */
1250 break;
1251 default:
1252 panic("inconsistent return codes");
1253 }
1254 }
1255
1256 /*
1257 * Allocate a new page for this object/offset pair.
1258 *
1259 * If the process has a fatal signal pending, prioritize the allocation
1260 * with the expectation that the process will exit shortly and free some
1261 * pages. In particular, the signal may have been posted by the page
1262 * daemon in an attempt to resolve an out-of-memory condition.
1263 *
1264 * The unlocked read of the p_flag is harmless. At worst, the P_KILLED
1265 * might be not observed here, and allocation fails, causing a restart
1266 * and new reading of the p_flag.
1267 */
1268 dset = fs->object->domain.dr_policy;
1269 if (dset == NULL)
1270 dset = curthread->td_domain.dr_policy;
1271 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1272 #if VM_NRESERVLEVEL > 0
1273 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1274 #endif
1275 if (!vm_pager_can_alloc_page(fs->object, fs->pindex)) {
1276 vm_fault_unlock_and_deallocate(fs);
1277 return (FAULT_FAILURE);
1278 }
1279 fs->m = vm_page_alloc(fs->object, fs->pindex,
1280 P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0);
1281 }
1282 if (fs->m == NULL) {
1283 if (vm_fault_allocate_oom(fs))
1284 vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1285 return (FAULT_RESTART);
1286 }
1287 fs->oom_started = false;
1288
1289 return (FAULT_CONTINUE);
1290 }
1291
1292 /*
1293 * Call the pager to retrieve the page if there is a chance
1294 * that the pager has it, and potentially retrieve additional
1295 * pages at the same time.
1296 */
1297 static enum fault_status
vm_fault_getpages(struct faultstate * fs,int * behindp,int * aheadp)1298 vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp)
1299 {
1300 vm_offset_t e_end, e_start;
1301 int ahead, behind, cluster_offset, rv;
1302 enum fault_status status;
1303 u_char behavior;
1304
1305 /*
1306 * Prepare for unlocking the map. Save the map
1307 * entry's start and end addresses, which are used to
1308 * optimize the size of the pager operation below.
1309 * Even if the map entry's addresses change after
1310 * unlocking the map, using the saved addresses is
1311 * safe.
1312 */
1313 e_start = fs->entry->start;
1314 e_end = fs->entry->end;
1315 behavior = vm_map_entry_behavior(fs->entry);
1316
1317 /*
1318 * If the pager for the current object might have
1319 * the page, then determine the number of additional
1320 * pages to read and potentially reprioritize
1321 * previously read pages for earlier reclamation.
1322 * These operations should only be performed once per
1323 * page fault. Even if the current pager doesn't
1324 * have the page, the number of additional pages to
1325 * read will apply to subsequent objects in the
1326 * shadow chain.
1327 */
1328 if (fs->nera == -1 && !P_KILLED(curproc))
1329 fs->nera = vm_fault_readahead(fs);
1330
1331 /*
1332 * Release the map lock before locking the vnode or
1333 * sleeping in the pager. (If the current object has
1334 * a shadow, then an earlier iteration of this loop
1335 * may have already unlocked the map.)
1336 */
1337 vm_fault_unlock_map(fs);
1338
1339 status = vm_fault_lock_vnode(fs, false);
1340 MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART);
1341 if (status == FAULT_RESTART)
1342 return (status);
1343 KASSERT(fs->vp == NULL || !fs->map->system_map,
1344 ("vm_fault: vnode-backed object mapped by system map"));
1345
1346 /*
1347 * Page in the requested page and hint the pager,
1348 * that it may bring up surrounding pages.
1349 */
1350 if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1351 P_KILLED(curproc)) {
1352 behind = 0;
1353 ahead = 0;
1354 } else {
1355 /* Is this a sequential fault? */
1356 if (fs->nera > 0) {
1357 behind = 0;
1358 ahead = fs->nera;
1359 } else {
1360 /*
1361 * Request a cluster of pages that is
1362 * aligned to a VM_FAULT_READ_DEFAULT
1363 * page offset boundary within the
1364 * object. Alignment to a page offset
1365 * boundary is more likely to coincide
1366 * with the underlying file system
1367 * block than alignment to a virtual
1368 * address boundary.
1369 */
1370 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1371 behind = ulmin(cluster_offset,
1372 atop(fs->vaddr - e_start));
1373 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1374 }
1375 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1376 }
1377 *behindp = behind;
1378 *aheadp = ahead;
1379 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1380 if (rv == VM_PAGER_OK)
1381 return (FAULT_HARD);
1382 if (rv == VM_PAGER_ERROR)
1383 printf("vm_fault: pager read error, pid %d (%s)\n",
1384 curproc->p_pid, curproc->p_comm);
1385 /*
1386 * If an I/O error occurred or the requested page was
1387 * outside the range of the pager, clean up and return
1388 * an error.
1389 */
1390 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1391 VM_OBJECT_WLOCK(fs->object);
1392 vm_fault_page_free(&fs->m);
1393 vm_fault_unlock_and_deallocate(fs);
1394 return (FAULT_OUT_OF_BOUNDS);
1395 }
1396 KASSERT(rv == VM_PAGER_FAIL,
1397 ("%s: unexpected pager error %d", __func__, rv));
1398 return (FAULT_CONTINUE);
1399 }
1400
1401 /*
1402 * Wait/Retry if the page is busy. We have to do this if the page is
1403 * either exclusive or shared busy because the vm_pager may be using
1404 * read busy for pageouts (and even pageins if it is the vnode pager),
1405 * and we could end up trying to pagein and pageout the same page
1406 * simultaneously.
1407 *
1408 * We can theoretically allow the busy case on a read fault if the page
1409 * is marked valid, but since such pages are typically already pmap'd,
1410 * putting that special case in might be more effort then it is worth.
1411 * We cannot under any circumstances mess around with a shared busied
1412 * page except, perhaps, to pmap it.
1413 */
1414 static void
vm_fault_busy_sleep(struct faultstate * fs)1415 vm_fault_busy_sleep(struct faultstate *fs)
1416 {
1417 /*
1418 * Reference the page before unlocking and
1419 * sleeping so that the page daemon is less
1420 * likely to reclaim it.
1421 */
1422 vm_page_aflag_set(fs->m, PGA_REFERENCED);
1423 if (fs->object != fs->first_object) {
1424 vm_fault_page_release(&fs->first_m);
1425 vm_object_pip_wakeup(fs->first_object);
1426 }
1427 vm_object_pip_wakeup(fs->object);
1428 vm_fault_unlock_map(fs);
1429 if (fs->m != vm_page_lookup(fs->object, fs->pindex) ||
1430 !vm_page_busy_sleep(fs->m, "vmpfw", 0))
1431 VM_OBJECT_UNLOCK(fs->object);
1432 VM_CNT_INC(v_intrans);
1433 vm_object_deallocate(fs->first_object);
1434 }
1435
1436 /*
1437 * Handle page lookup, populate, allocate, page-in for the current
1438 * object.
1439 *
1440 * The object is locked on entry and will remain locked with a return
1441 * code of FAULT_CONTINUE so that fault may follow the shadow chain.
1442 * Otherwise, the object will be unlocked upon return.
1443 */
1444 static enum fault_status
vm_fault_object(struct faultstate * fs,int * behindp,int * aheadp)1445 vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp)
1446 {
1447 enum fault_status res;
1448 bool dead;
1449
1450 if (fs->object == fs->first_object || !fs->can_read_lock)
1451 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1452 else
1453 VM_OBJECT_ASSERT_LOCKED(fs->object);
1454
1455 /*
1456 * If the object is marked for imminent termination, we retry
1457 * here, since the collapse pass has raced with us. Otherwise,
1458 * if we see terminally dead object, return fail.
1459 */
1460 if ((fs->object->flags & OBJ_DEAD) != 0) {
1461 dead = fs->object->type == OBJT_DEAD;
1462 vm_fault_unlock_and_deallocate(fs);
1463 if (dead)
1464 return (FAULT_PROTECTION_FAILURE);
1465 pause("vmf_de", 1);
1466 return (FAULT_RESTART);
1467 }
1468
1469 /*
1470 * See if the page is resident.
1471 */
1472 fs->m = vm_page_lookup(fs->object, fs->pindex);
1473 if (fs->m != NULL) {
1474 if (!vm_page_tryxbusy(fs->m)) {
1475 vm_fault_busy_sleep(fs);
1476 return (FAULT_RESTART);
1477 }
1478
1479 /*
1480 * The page is marked busy for other processes and the
1481 * pagedaemon. If it is still completely valid we are
1482 * done.
1483 */
1484 if (vm_page_all_valid(fs->m)) {
1485 VM_OBJECT_UNLOCK(fs->object);
1486 return (FAULT_SOFT);
1487 }
1488 }
1489
1490 /*
1491 * Page is not resident. If the pager might contain the page
1492 * or this is the beginning of the search, allocate a new
1493 * page.
1494 */
1495 if (fs->m == NULL && (vm_fault_object_needs_getpages(fs->object) ||
1496 fs->object == fs->first_object)) {
1497 if (!vm_fault_object_ensure_wlocked(fs)) {
1498 fs->can_read_lock = false;
1499 vm_fault_unlock_and_deallocate(fs);
1500 return (FAULT_RESTART);
1501 }
1502 res = vm_fault_allocate(fs);
1503 if (res != FAULT_CONTINUE)
1504 return (res);
1505 }
1506
1507 /*
1508 * Check to see if the pager can possibly satisfy this fault.
1509 * If not, skip to the next object without dropping the lock to
1510 * preserve atomicity of shadow faults.
1511 */
1512 if (vm_fault_object_needs_getpages(fs->object)) {
1513 /*
1514 * At this point, we have either allocated a new page
1515 * or found an existing page that is only partially
1516 * valid.
1517 *
1518 * We hold a reference on the current object and the
1519 * page is exclusive busied. The exclusive busy
1520 * prevents simultaneous faults and collapses while
1521 * the object lock is dropped.
1522 */
1523 VM_OBJECT_UNLOCK(fs->object);
1524 res = vm_fault_getpages(fs, behindp, aheadp);
1525 if (res == FAULT_CONTINUE)
1526 VM_OBJECT_WLOCK(fs->object);
1527 } else {
1528 res = FAULT_CONTINUE;
1529 }
1530 return (res);
1531 }
1532
1533 int
vm_fault(vm_map_t map,vm_offset_t vaddr,vm_prot_t fault_type,int fault_flags,vm_page_t * m_hold)1534 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1535 int fault_flags, vm_page_t *m_hold)
1536 {
1537 struct faultstate fs;
1538 int ahead, behind, faultcount, rv;
1539 enum fault_status res;
1540 enum fault_next_status res_next;
1541 bool hardfault;
1542
1543 VM_CNT_INC(v_vm_faults);
1544
1545 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1546 return (KERN_PROTECTION_FAILURE);
1547
1548 fs.vp = NULL;
1549 fs.vaddr = vaddr;
1550 fs.m_hold = m_hold;
1551 fs.fault_flags = fault_flags;
1552 fs.map = map;
1553 fs.lookup_still_valid = false;
1554 fs.oom_started = false;
1555 fs.nera = -1;
1556 fs.can_read_lock = true;
1557 faultcount = 0;
1558 hardfault = false;
1559
1560 RetryFault:
1561 fs.fault_type = fault_type;
1562
1563 /*
1564 * Find the backing store object and offset into it to begin the
1565 * search.
1566 */
1567 rv = vm_fault_lookup(&fs);
1568 if (rv != KERN_SUCCESS) {
1569 if (rv == KERN_RESOURCE_SHORTAGE)
1570 goto RetryFault;
1571 return (rv);
1572 }
1573
1574 /*
1575 * Try to avoid lock contention on the top-level object through
1576 * special-case handling of some types of page faults, specifically,
1577 * those that are mapping an existing page from the top-level object.
1578 * Under this condition, a read lock on the object suffices, allowing
1579 * multiple page faults of a similar type to run in parallel.
1580 */
1581 if (fs.vp == NULL /* avoid locked vnode leak */ &&
1582 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1583 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1584 res = vm_fault_soft_fast(&fs);
1585 if (res == FAULT_SUCCESS) {
1586 VM_OBJECT_ASSERT_UNLOCKED(fs.first_object);
1587 return (KERN_SUCCESS);
1588 }
1589 VM_OBJECT_ASSERT_WLOCKED(fs.first_object);
1590 } else {
1591 VM_OBJECT_WLOCK(fs.first_object);
1592 }
1593
1594 /*
1595 * Make a reference to this object to prevent its disposal while we
1596 * are messing with it. Once we have the reference, the map is free
1597 * to be diddled. Since objects reference their shadows (and copies),
1598 * they will stay around as well.
1599 *
1600 * Bump the paging-in-progress count to prevent size changes (e.g.
1601 * truncation operations) during I/O.
1602 */
1603 vm_object_reference_locked(fs.first_object);
1604 vm_object_pip_add(fs.first_object, 1);
1605
1606 fs.m_cow = fs.m = fs.first_m = NULL;
1607
1608 /*
1609 * Search for the page at object/offset.
1610 */
1611 fs.object = fs.first_object;
1612 fs.pindex = fs.first_pindex;
1613
1614 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1615 res = vm_fault_allocate(&fs);
1616 switch (res) {
1617 case FAULT_RESTART:
1618 goto RetryFault;
1619 case FAULT_SUCCESS:
1620 return (KERN_SUCCESS);
1621 case FAULT_FAILURE:
1622 return (KERN_FAILURE);
1623 case FAULT_OUT_OF_BOUNDS:
1624 return (KERN_OUT_OF_BOUNDS);
1625 case FAULT_CONTINUE:
1626 break;
1627 default:
1628 panic("vm_fault: Unhandled status %d", res);
1629 }
1630 }
1631
1632 while (TRUE) {
1633 KASSERT(fs.m == NULL,
1634 ("page still set %p at loop start", fs.m));
1635
1636 res = vm_fault_object(&fs, &behind, &ahead);
1637 switch (res) {
1638 case FAULT_SOFT:
1639 goto found;
1640 case FAULT_HARD:
1641 faultcount = behind + 1 + ahead;
1642 hardfault = true;
1643 goto found;
1644 case FAULT_RESTART:
1645 goto RetryFault;
1646 case FAULT_SUCCESS:
1647 return (KERN_SUCCESS);
1648 case FAULT_FAILURE:
1649 return (KERN_FAILURE);
1650 case FAULT_OUT_OF_BOUNDS:
1651 return (KERN_OUT_OF_BOUNDS);
1652 case FAULT_PROTECTION_FAILURE:
1653 return (KERN_PROTECTION_FAILURE);
1654 case FAULT_CONTINUE:
1655 break;
1656 default:
1657 panic("vm_fault: Unhandled status %d", res);
1658 }
1659
1660 /*
1661 * The page was not found in the current object. Try to
1662 * traverse into a backing object or zero fill if none is
1663 * found.
1664 */
1665 res_next = vm_fault_next(&fs);
1666 if (res_next == FAULT_NEXT_RESTART)
1667 goto RetryFault;
1668 else if (res_next == FAULT_NEXT_GOTOBJ)
1669 continue;
1670 MPASS(res_next == FAULT_NEXT_NOOBJ);
1671 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1672 if (fs.first_object == fs.object)
1673 vm_fault_page_free(&fs.first_m);
1674 vm_fault_unlock_and_deallocate(&fs);
1675 return (KERN_OUT_OF_BOUNDS);
1676 }
1677 VM_OBJECT_UNLOCK(fs.object);
1678 vm_fault_zerofill(&fs);
1679 /* Don't try to prefault neighboring pages. */
1680 faultcount = 1;
1681 break;
1682 }
1683
1684 found:
1685 /*
1686 * A valid page has been found and exclusively busied. The
1687 * object lock must no longer be held.
1688 */
1689 vm_page_assert_xbusied(fs.m);
1690 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1691
1692 /*
1693 * If the page is being written, but isn't already owned by the
1694 * top-level object, we have to copy it into a new page owned by the
1695 * top-level object.
1696 */
1697 if (fs.object != fs.first_object) {
1698 /*
1699 * We only really need to copy if we want to write it.
1700 */
1701 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1702 vm_fault_cow(&fs);
1703 /*
1704 * We only try to prefault read-only mappings to the
1705 * neighboring pages when this copy-on-write fault is
1706 * a hard fault. In other cases, trying to prefault
1707 * is typically wasted effort.
1708 */
1709 if (faultcount == 0)
1710 faultcount = 1;
1711
1712 } else {
1713 fs.prot &= ~VM_PROT_WRITE;
1714 }
1715 }
1716
1717 /*
1718 * We must verify that the maps have not changed since our last
1719 * lookup.
1720 */
1721 if (!fs.lookup_still_valid) {
1722 rv = vm_fault_relookup(&fs);
1723 if (rv != KERN_SUCCESS) {
1724 vm_fault_deallocate(&fs);
1725 if (rv == KERN_RESTART)
1726 goto RetryFault;
1727 return (rv);
1728 }
1729 }
1730 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1731
1732 /*
1733 * If the page was filled by a pager, save the virtual address that
1734 * should be faulted on next under a sequential access pattern to the
1735 * map entry. A read lock on the map suffices to update this address
1736 * safely.
1737 */
1738 if (hardfault)
1739 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1740
1741 /*
1742 * Page must be completely valid or it is not fit to
1743 * map into user space. vm_pager_get_pages() ensures this.
1744 */
1745 vm_page_assert_xbusied(fs.m);
1746 KASSERT(vm_page_all_valid(fs.m),
1747 ("vm_fault: page %p partially invalid", fs.m));
1748
1749 vm_fault_dirty(&fs, fs.m);
1750
1751 /*
1752 * Put this page into the physical map. We had to do the unlock above
1753 * because pmap_enter() may sleep. We don't put the page
1754 * back on the active queue until later so that the pageout daemon
1755 * won't find it (yet).
1756 */
1757 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1758 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1759 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1760 fs.wired == 0)
1761 vm_fault_prefault(&fs, vaddr,
1762 faultcount > 0 ? behind : PFBAK,
1763 faultcount > 0 ? ahead : PFFOR, false);
1764
1765 /*
1766 * If the page is not wired down, then put it where the pageout daemon
1767 * can find it.
1768 */
1769 if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1770 vm_page_wire(fs.m);
1771 else
1772 vm_page_activate(fs.m);
1773 if (fs.m_hold != NULL) {
1774 (*fs.m_hold) = fs.m;
1775 vm_page_wire(fs.m);
1776 }
1777 vm_page_xunbusy(fs.m);
1778 fs.m = NULL;
1779
1780 /*
1781 * Unlock everything, and return
1782 */
1783 vm_fault_deallocate(&fs);
1784 if (hardfault) {
1785 VM_CNT_INC(v_io_faults);
1786 curthread->td_ru.ru_majflt++;
1787 #ifdef RACCT
1788 if (racct_enable && fs.object->type == OBJT_VNODE) {
1789 PROC_LOCK(curproc);
1790 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1791 racct_add_force(curproc, RACCT_WRITEBPS,
1792 PAGE_SIZE + behind * PAGE_SIZE);
1793 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1794 } else {
1795 racct_add_force(curproc, RACCT_READBPS,
1796 PAGE_SIZE + ahead * PAGE_SIZE);
1797 racct_add_force(curproc, RACCT_READIOPS, 1);
1798 }
1799 PROC_UNLOCK(curproc);
1800 }
1801 #endif
1802 } else
1803 curthread->td_ru.ru_minflt++;
1804
1805 return (KERN_SUCCESS);
1806 }
1807
1808 /*
1809 * Speed up the reclamation of pages that precede the faulting pindex within
1810 * the first object of the shadow chain. Essentially, perform the equivalent
1811 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1812 * the faulting pindex by the cluster size when the pages read by vm_fault()
1813 * cross a cluster-size boundary. The cluster size is the greater of the
1814 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1815 *
1816 * When "fs->first_object" is a shadow object, the pages in the backing object
1817 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1818 * function must only be concerned with pages in the first object.
1819 */
1820 static void
vm_fault_dontneed(const struct faultstate * fs,vm_offset_t vaddr,int ahead)1821 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1822 {
1823 vm_map_entry_t entry;
1824 vm_object_t first_object;
1825 vm_offset_t end, start;
1826 vm_page_t m, m_next;
1827 vm_pindex_t pend, pstart;
1828 vm_size_t size;
1829
1830 VM_OBJECT_ASSERT_UNLOCKED(fs->object);
1831 first_object = fs->first_object;
1832 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1833 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1834 VM_OBJECT_RLOCK(first_object);
1835 size = VM_FAULT_DONTNEED_MIN;
1836 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1837 size = pagesizes[1];
1838 end = rounddown2(vaddr, size);
1839 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1840 (entry = fs->entry)->start < end) {
1841 if (end - entry->start < size)
1842 start = entry->start;
1843 else
1844 start = end - size;
1845 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1846 pstart = OFF_TO_IDX(entry->offset) + atop(start -
1847 entry->start);
1848 m_next = vm_page_find_least(first_object, pstart);
1849 pend = OFF_TO_IDX(entry->offset) + atop(end -
1850 entry->start);
1851 while ((m = m_next) != NULL && m->pindex < pend) {
1852 m_next = TAILQ_NEXT(m, listq);
1853 if (!vm_page_all_valid(m) ||
1854 vm_page_busied(m))
1855 continue;
1856
1857 /*
1858 * Don't clear PGA_REFERENCED, since it would
1859 * likely represent a reference by a different
1860 * process.
1861 *
1862 * Typically, at this point, prefetched pages
1863 * are still in the inactive queue. Only
1864 * pages that triggered page faults are in the
1865 * active queue. The test for whether the page
1866 * is in the inactive queue is racy; in the
1867 * worst case we will requeue the page
1868 * unnecessarily.
1869 */
1870 if (!vm_page_inactive(m))
1871 vm_page_deactivate(m);
1872 }
1873 }
1874 VM_OBJECT_RUNLOCK(first_object);
1875 }
1876 }
1877
1878 /*
1879 * vm_fault_prefault provides a quick way of clustering
1880 * pagefaults into a processes address space. It is a "cousin"
1881 * of vm_map_pmap_enter, except it runs at page fault time instead
1882 * of mmap time.
1883 */
1884 static void
vm_fault_prefault(const struct faultstate * fs,vm_offset_t addra,int backward,int forward,bool obj_locked)1885 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
1886 int backward, int forward, bool obj_locked)
1887 {
1888 pmap_t pmap;
1889 vm_map_entry_t entry;
1890 vm_object_t backing_object, lobject;
1891 vm_offset_t addr, starta;
1892 vm_pindex_t pindex;
1893 vm_page_t m;
1894 int i;
1895
1896 pmap = fs->map->pmap;
1897 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
1898 return;
1899
1900 entry = fs->entry;
1901
1902 if (addra < backward * PAGE_SIZE) {
1903 starta = entry->start;
1904 } else {
1905 starta = addra - backward * PAGE_SIZE;
1906 if (starta < entry->start)
1907 starta = entry->start;
1908 }
1909
1910 /*
1911 * Generate the sequence of virtual addresses that are candidates for
1912 * prefaulting in an outward spiral from the faulting virtual address,
1913 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
1914 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
1915 * If the candidate address doesn't have a backing physical page, then
1916 * the loop immediately terminates.
1917 */
1918 for (i = 0; i < 2 * imax(backward, forward); i++) {
1919 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
1920 PAGE_SIZE);
1921 if (addr > addra + forward * PAGE_SIZE)
1922 addr = 0;
1923
1924 if (addr < starta || addr >= entry->end)
1925 continue;
1926
1927 if (!pmap_is_prefaultable(pmap, addr))
1928 continue;
1929
1930 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1931 lobject = entry->object.vm_object;
1932 if (!obj_locked)
1933 VM_OBJECT_RLOCK(lobject);
1934 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
1935 !vm_fault_object_needs_getpages(lobject) &&
1936 (backing_object = lobject->backing_object) != NULL) {
1937 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
1938 0, ("vm_fault_prefault: unaligned object offset"));
1939 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1940 VM_OBJECT_RLOCK(backing_object);
1941 if (!obj_locked || lobject != entry->object.vm_object)
1942 VM_OBJECT_RUNLOCK(lobject);
1943 lobject = backing_object;
1944 }
1945 if (m == NULL) {
1946 if (!obj_locked || lobject != entry->object.vm_object)
1947 VM_OBJECT_RUNLOCK(lobject);
1948 break;
1949 }
1950 if (vm_page_all_valid(m) &&
1951 (m->flags & PG_FICTITIOUS) == 0)
1952 pmap_enter_quick(pmap, addr, m, entry->protection);
1953 if (!obj_locked || lobject != entry->object.vm_object)
1954 VM_OBJECT_RUNLOCK(lobject);
1955 }
1956 }
1957
1958 /*
1959 * Hold each of the physical pages that are mapped by the specified range of
1960 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
1961 * and allow the specified types of access, "prot". If all of the implied
1962 * pages are successfully held, then the number of held pages is returned
1963 * together with pointers to those pages in the array "ma". However, if any
1964 * of the pages cannot be held, -1 is returned.
1965 */
1966 int
vm_fault_quick_hold_pages(vm_map_t map,vm_offset_t addr,vm_size_t len,vm_prot_t prot,vm_page_t * ma,int max_count)1967 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
1968 vm_prot_t prot, vm_page_t *ma, int max_count)
1969 {
1970 vm_offset_t end, va;
1971 vm_page_t *mp;
1972 int count;
1973 boolean_t pmap_failed;
1974
1975 if (len == 0)
1976 return (0);
1977 end = round_page(addr + len);
1978 addr = trunc_page(addr);
1979
1980 if (!vm_map_range_valid(map, addr, end))
1981 return (-1);
1982
1983 if (atop(end - addr) > max_count)
1984 panic("vm_fault_quick_hold_pages: count > max_count");
1985 count = atop(end - addr);
1986
1987 /*
1988 * Most likely, the physical pages are resident in the pmap, so it is
1989 * faster to try pmap_extract_and_hold() first.
1990 */
1991 pmap_failed = FALSE;
1992 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
1993 *mp = pmap_extract_and_hold(map->pmap, va, prot);
1994 if (*mp == NULL)
1995 pmap_failed = TRUE;
1996 else if ((prot & VM_PROT_WRITE) != 0 &&
1997 (*mp)->dirty != VM_PAGE_BITS_ALL) {
1998 /*
1999 * Explicitly dirty the physical page. Otherwise, the
2000 * caller's changes may go unnoticed because they are
2001 * performed through an unmanaged mapping or by a DMA
2002 * operation.
2003 *
2004 * The object lock is not held here.
2005 * See vm_page_clear_dirty_mask().
2006 */
2007 vm_page_dirty(*mp);
2008 }
2009 }
2010 if (pmap_failed) {
2011 /*
2012 * One or more pages could not be held by the pmap. Either no
2013 * page was mapped at the specified virtual address or that
2014 * mapping had insufficient permissions. Attempt to fault in
2015 * and hold these pages.
2016 *
2017 * If vm_fault_disable_pagefaults() was called,
2018 * i.e., TDP_NOFAULTING is set, we must not sleep nor
2019 * acquire MD VM locks, which means we must not call
2020 * vm_fault(). Some (out of tree) callers mark
2021 * too wide a code area with vm_fault_disable_pagefaults()
2022 * already, use the VM_PROT_QUICK_NOFAULT flag to request
2023 * the proper behaviour explicitly.
2024 */
2025 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
2026 (curthread->td_pflags & TDP_NOFAULTING) != 0)
2027 goto error;
2028 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE)
2029 if (*mp == NULL && vm_fault(map, va, prot,
2030 VM_FAULT_NORMAL, mp) != KERN_SUCCESS)
2031 goto error;
2032 }
2033 return (count);
2034 error:
2035 for (mp = ma; mp < ma + count; mp++)
2036 if (*mp != NULL)
2037 vm_page_unwire(*mp, PQ_INACTIVE);
2038 return (-1);
2039 }
2040
2041 /*
2042 * Routine:
2043 * vm_fault_copy_entry
2044 * Function:
2045 * Create new object backing dst_entry with private copy of all
2046 * underlying pages. When src_entry is equal to dst_entry, function
2047 * implements COW for wired-down map entry. Otherwise, it forks
2048 * wired entry into dst_map.
2049 *
2050 * In/out conditions:
2051 * The source and destination maps must be locked for write.
2052 * The source map entry must be wired down (or be a sharing map
2053 * entry corresponding to a main map entry that is wired down).
2054 */
2055 void
vm_fault_copy_entry(vm_map_t dst_map,vm_map_t src_map __unused,vm_map_entry_t dst_entry,vm_map_entry_t src_entry,vm_ooffset_t * fork_charge)2056 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map __unused,
2057 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
2058 vm_ooffset_t *fork_charge)
2059 {
2060 vm_object_t backing_object, dst_object, object, src_object;
2061 vm_pindex_t dst_pindex, pindex, src_pindex;
2062 vm_prot_t access, prot;
2063 vm_offset_t vaddr;
2064 vm_page_t dst_m;
2065 vm_page_t src_m;
2066 bool upgrade;
2067
2068 upgrade = src_entry == dst_entry;
2069 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
2070 ("vm_fault_copy_entry: vm_object not NULL"));
2071
2072 /*
2073 * If not an upgrade, then enter the mappings in the pmap as
2074 * read and/or execute accesses. Otherwise, enter them as
2075 * write accesses.
2076 *
2077 * A writeable large page mapping is only created if all of
2078 * the constituent small page mappings are modified. Marking
2079 * PTEs as modified on inception allows promotion to happen
2080 * without taking potentially large number of soft faults.
2081 */
2082 access = prot = dst_entry->protection;
2083 if (!upgrade)
2084 access &= ~VM_PROT_WRITE;
2085
2086 src_object = src_entry->object.vm_object;
2087 src_pindex = OFF_TO_IDX(src_entry->offset);
2088
2089 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2090 dst_object = src_object;
2091 vm_object_reference(dst_object);
2092 } else {
2093 /*
2094 * Create the top-level object for the destination entry.
2095 * Doesn't actually shadow anything - we copy the pages
2096 * directly.
2097 */
2098 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
2099 dst_entry->start), NULL, NULL, 0);
2100 #if VM_NRESERVLEVEL > 0
2101 dst_object->flags |= OBJ_COLORED;
2102 dst_object->pg_color = atop(dst_entry->start);
2103 #endif
2104 dst_object->domain = src_object->domain;
2105 dst_object->charge = dst_entry->end - dst_entry->start;
2106
2107 dst_entry->object.vm_object = dst_object;
2108 dst_entry->offset = 0;
2109 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
2110 }
2111
2112 VM_OBJECT_WLOCK(dst_object);
2113 if (fork_charge != NULL) {
2114 KASSERT(dst_entry->cred == NULL,
2115 ("vm_fault_copy_entry: leaked swp charge"));
2116 dst_object->cred = curthread->td_ucred;
2117 crhold(dst_object->cred);
2118 *fork_charge += dst_object->charge;
2119 } else if ((dst_object->flags & OBJ_SWAP) != 0 &&
2120 dst_object->cred == NULL) {
2121 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
2122 dst_entry));
2123 dst_object->cred = dst_entry->cred;
2124 dst_entry->cred = NULL;
2125 }
2126
2127 /*
2128 * Loop through all of the virtual pages within the entry's
2129 * range, copying each page from the source object to the
2130 * destination object. Since the source is wired, those pages
2131 * must exist. In contrast, the destination is pageable.
2132 * Since the destination object doesn't share any backing storage
2133 * with the source object, all of its pages must be dirtied,
2134 * regardless of whether they can be written.
2135 */
2136 for (vaddr = dst_entry->start, dst_pindex = 0;
2137 vaddr < dst_entry->end;
2138 vaddr += PAGE_SIZE, dst_pindex++) {
2139 again:
2140 /*
2141 * Find the page in the source object, and copy it in.
2142 * Because the source is wired down, the page will be
2143 * in memory.
2144 */
2145 if (src_object != dst_object)
2146 VM_OBJECT_RLOCK(src_object);
2147 object = src_object;
2148 pindex = src_pindex + dst_pindex;
2149 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
2150 (backing_object = object->backing_object) != NULL) {
2151 /*
2152 * Unless the source mapping is read-only or
2153 * it is presently being upgraded from
2154 * read-only, the first object in the shadow
2155 * chain should provide all of the pages. In
2156 * other words, this loop body should never be
2157 * executed when the source mapping is already
2158 * read/write.
2159 */
2160 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
2161 upgrade,
2162 ("vm_fault_copy_entry: main object missing page"));
2163
2164 VM_OBJECT_RLOCK(backing_object);
2165 pindex += OFF_TO_IDX(object->backing_object_offset);
2166 if (object != dst_object)
2167 VM_OBJECT_RUNLOCK(object);
2168 object = backing_object;
2169 }
2170 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2171
2172 if (object != dst_object) {
2173 /*
2174 * Allocate a page in the destination object.
2175 */
2176 dst_m = vm_page_alloc(dst_object, (src_object ==
2177 dst_object ? src_pindex : 0) + dst_pindex,
2178 VM_ALLOC_NORMAL);
2179 if (dst_m == NULL) {
2180 VM_OBJECT_WUNLOCK(dst_object);
2181 VM_OBJECT_RUNLOCK(object);
2182 vm_wait(dst_object);
2183 VM_OBJECT_WLOCK(dst_object);
2184 goto again;
2185 }
2186
2187 /*
2188 * See the comment in vm_fault_cow().
2189 */
2190 if (src_object == dst_object &&
2191 (object->flags & OBJ_ONEMAPPING) == 0)
2192 pmap_remove_all(src_m);
2193 pmap_copy_page(src_m, dst_m);
2194
2195 /*
2196 * The object lock does not guarantee that "src_m" will
2197 * transition from invalid to valid, but it does ensure
2198 * that "src_m" will not transition from valid to
2199 * invalid.
2200 */
2201 dst_m->dirty = dst_m->valid = src_m->valid;
2202 VM_OBJECT_RUNLOCK(object);
2203 } else {
2204 dst_m = src_m;
2205 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0)
2206 goto again;
2207 if (dst_m->pindex >= dst_object->size) {
2208 /*
2209 * We are upgrading. Index can occur
2210 * out of bounds if the object type is
2211 * vnode and the file was truncated.
2212 */
2213 vm_page_xunbusy(dst_m);
2214 break;
2215 }
2216 }
2217
2218 /*
2219 * Enter it in the pmap. If a wired, copy-on-write
2220 * mapping is being replaced by a write-enabled
2221 * mapping, then wire that new mapping.
2222 *
2223 * The page can be invalid if the user called
2224 * msync(MS_INVALIDATE) or truncated the backing vnode
2225 * or shared memory object. In this case, do not
2226 * insert it into pmap, but still do the copy so that
2227 * all copies of the wired map entry have similar
2228 * backing pages.
2229 */
2230 if (vm_page_all_valid(dst_m)) {
2231 VM_OBJECT_WUNLOCK(dst_object);
2232 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2233 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2234 VM_OBJECT_WLOCK(dst_object);
2235 }
2236
2237 /*
2238 * Mark it no longer busy, and put it on the active list.
2239 */
2240 if (upgrade) {
2241 if (src_m != dst_m) {
2242 vm_page_unwire(src_m, PQ_INACTIVE);
2243 vm_page_wire(dst_m);
2244 } else {
2245 KASSERT(vm_page_wired(dst_m),
2246 ("dst_m %p is not wired", dst_m));
2247 }
2248 } else {
2249 vm_page_activate(dst_m);
2250 }
2251 vm_page_xunbusy(dst_m);
2252 }
2253 VM_OBJECT_WUNLOCK(dst_object);
2254 if (upgrade) {
2255 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2256 vm_object_deallocate(src_object);
2257 }
2258 }
2259
2260 /*
2261 * Block entry into the machine-independent layer's page fault handler by
2262 * the calling thread. Subsequent calls to vm_fault() by that thread will
2263 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
2264 * spurious page faults.
2265 */
2266 int
vm_fault_disable_pagefaults(void)2267 vm_fault_disable_pagefaults(void)
2268 {
2269
2270 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2271 }
2272
2273 void
vm_fault_enable_pagefaults(int save)2274 vm_fault_enable_pagefaults(int save)
2275 {
2276
2277 curthread_pflags_restore(save);
2278 }
2279