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 "opt_ktrace.h"
75 #include "opt_vm.h"
76
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/lock.h>
80 #include <sys/mman.h>
81 #include <sys/mutex.h>
82 #include <sys/pctrie.h>
83 #include <sys/proc.h>
84 #include <sys/racct.h>
85 #include <sys/refcount.h>
86 #include <sys/resourcevar.h>
87 #include <sys/rwlock.h>
88 #include <sys/signalvar.h>
89 #include <sys/sysctl.h>
90 #include <sys/sysent.h>
91 #include <sys/vmmeter.h>
92 #include <sys/vnode.h>
93 #ifdef KTRACE
94 #include <sys/ktrace.h>
95 #endif
96
97 #include <vm/vm.h>
98 #include <vm/vm_param.h>
99 #include <vm/pmap.h>
100 #include <vm/vm_map.h>
101 #include <vm/vm_object.h>
102 #include <vm/vm_page.h>
103 #include <vm/vm_pageout.h>
104 #include <vm/vm_kern.h>
105 #include <vm/vm_pager.h>
106 #include <vm/vm_radix.h>
107 #include <vm/vm_extern.h>
108 #include <vm/vm_reserv.h>
109
110 #define PFBAK 4
111 #define PFFOR 4
112
113 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT)
114
115 #define VM_FAULT_DONTNEED_MIN 1048576
116
117 struct faultstate {
118 /* Fault parameters. */
119 vm_offset_t vaddr;
120 vm_page_t *m_hold;
121 vm_prot_t fault_type;
122 vm_prot_t prot;
123 int fault_flags;
124 boolean_t wired;
125
126 /* Control state. */
127 struct timeval oom_start_time;
128 bool oom_started;
129 int nera;
130 bool can_read_lock;
131
132 /* Page reference for cow. */
133 vm_page_t m_cow;
134
135 /* Current object. */
136 vm_object_t object;
137 vm_pindex_t pindex;
138 vm_page_t m;
139
140 /* Top-level map object. */
141 vm_object_t first_object;
142 vm_pindex_t first_pindex;
143 vm_page_t first_m;
144
145 /* Map state. */
146 vm_map_t map;
147 vm_map_entry_t entry;
148 int map_generation;
149 bool lookup_still_valid;
150
151 /* Vnode if locked. */
152 struct vnode *vp;
153 };
154
155 /*
156 * Return codes for internal fault routines.
157 */
158 enum fault_status {
159 FAULT_SUCCESS = 10000, /* Return success to user. */
160 FAULT_FAILURE, /* Return failure to user. */
161 FAULT_CONTINUE, /* Continue faulting. */
162 FAULT_RESTART, /* Restart fault. */
163 FAULT_OUT_OF_BOUNDS, /* Invalid address for pager. */
164 FAULT_HARD, /* Performed I/O. */
165 FAULT_SOFT, /* Found valid page. */
166 FAULT_PROTECTION_FAILURE, /* Invalid access. */
167 };
168
169 enum fault_next_status {
170 FAULT_NEXT_GOTOBJ = 1,
171 FAULT_NEXT_NOOBJ,
172 FAULT_NEXT_RESTART,
173 };
174
175 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr,
176 int ahead);
177 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
178 int backward, int forward, bool obj_locked);
179
180 static int vm_pfault_oom_attempts = 3;
181 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN,
182 &vm_pfault_oom_attempts, 0,
183 "Number of page allocation attempts in page fault handler before it "
184 "triggers OOM handling");
185
186 static int vm_pfault_oom_wait = 10;
187 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN,
188 &vm_pfault_oom_wait, 0,
189 "Number of seconds to wait for free pages before retrying "
190 "the page fault handler");
191
192 static inline void
vm_fault_page_release(vm_page_t * mp)193 vm_fault_page_release(vm_page_t *mp)
194 {
195 vm_page_t m;
196
197 m = *mp;
198 if (m != NULL) {
199 /*
200 * We are likely to loop around again and attempt to busy
201 * this page. Deactivating it leaves it available for
202 * pageout while optimizing fault restarts.
203 */
204 vm_page_deactivate(m);
205 if (vm_page_xbusied(m))
206 vm_page_xunbusy(m);
207 else
208 vm_page_sunbusy(m);
209 *mp = NULL;
210 }
211 }
212
213 static inline void
vm_fault_page_free(vm_page_t * mp)214 vm_fault_page_free(vm_page_t *mp)
215 {
216 vm_page_t m;
217
218 m = *mp;
219 if (m != NULL) {
220 VM_OBJECT_ASSERT_WLOCKED(m->object);
221 if (!vm_page_wired(m))
222 vm_page_free(m);
223 else
224 vm_page_xunbusy(m);
225 *mp = NULL;
226 }
227 }
228
229 /*
230 * Return true if a vm_pager_get_pages() call is needed in order to check
231 * whether the pager might have a particular page, false if it can be determined
232 * immediately that the pager can not have a copy. For swap objects, this can
233 * be checked quickly.
234 */
235 static inline bool
vm_fault_object_needs_getpages(vm_object_t object)236 vm_fault_object_needs_getpages(vm_object_t object)
237 {
238 VM_OBJECT_ASSERT_LOCKED(object);
239
240 return ((object->flags & OBJ_SWAP) == 0 ||
241 !pctrie_is_empty(&object->un_pager.swp.swp_blks));
242 }
243
244 static inline void
vm_fault_unlock_map(struct faultstate * fs)245 vm_fault_unlock_map(struct faultstate *fs)
246 {
247
248 if (fs->lookup_still_valid) {
249 vm_map_lookup_done(fs->map, fs->entry);
250 fs->lookup_still_valid = false;
251 }
252 }
253
254 static void
vm_fault_unlock_vp(struct faultstate * fs)255 vm_fault_unlock_vp(struct faultstate *fs)
256 {
257
258 if (fs->vp != NULL) {
259 vput(fs->vp);
260 fs->vp = NULL;
261 }
262 }
263
264 static bool
vm_fault_might_be_cow(struct faultstate * fs)265 vm_fault_might_be_cow(struct faultstate *fs)
266 {
267 return (fs->object != fs->first_object);
268 }
269
270 static void
vm_fault_deallocate(struct faultstate * fs)271 vm_fault_deallocate(struct faultstate *fs)
272 {
273
274 vm_fault_page_release(&fs->m_cow);
275 vm_fault_page_release(&fs->m);
276 vm_object_pip_wakeup(fs->object);
277 if (vm_fault_might_be_cow(fs)) {
278 VM_OBJECT_WLOCK(fs->first_object);
279 vm_fault_page_free(&fs->first_m);
280 VM_OBJECT_WUNLOCK(fs->first_object);
281 vm_object_pip_wakeup(fs->first_object);
282 }
283 vm_object_deallocate(fs->first_object);
284 vm_fault_unlock_map(fs);
285 vm_fault_unlock_vp(fs);
286 }
287
288 static void
vm_fault_unlock_and_deallocate(struct faultstate * fs)289 vm_fault_unlock_and_deallocate(struct faultstate *fs)
290 {
291
292 VM_OBJECT_UNLOCK(fs->object);
293 vm_fault_deallocate(fs);
294 }
295
296 static void
vm_fault_dirty(struct faultstate * fs,vm_page_t m)297 vm_fault_dirty(struct faultstate *fs, vm_page_t m)
298 {
299 bool need_dirty;
300
301 if (((fs->prot & VM_PROT_WRITE) == 0 &&
302 (fs->fault_flags & VM_FAULT_DIRTY) == 0) ||
303 (m->oflags & VPO_UNMANAGED) != 0)
304 return;
305
306 VM_PAGE_OBJECT_BUSY_ASSERT(m);
307
308 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 &&
309 (fs->fault_flags & VM_FAULT_WIRE) == 0) ||
310 (fs->fault_flags & VM_FAULT_DIRTY) != 0;
311
312 vm_object_set_writeable_dirty(m->object);
313
314 /*
315 * If the fault is a write, we know that this page is being
316 * written NOW so dirty it explicitly to save on
317 * pmap_is_modified() calls later.
318 *
319 * Also, since the page is now dirty, we can possibly tell
320 * the pager to release any swap backing the page.
321 */
322 if (need_dirty && vm_page_set_dirty(m) == 0) {
323 /*
324 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC
325 * if the page is already dirty to prevent data written with
326 * the expectation of being synced from not being synced.
327 * Likewise if this entry does not request NOSYNC then make
328 * sure the page isn't marked NOSYNC. Applications sharing
329 * data should use the same flags to avoid ping ponging.
330 */
331 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0)
332 vm_page_aflag_set(m, PGA_NOSYNC);
333 else
334 vm_page_aflag_clear(m, PGA_NOSYNC);
335 }
336
337 }
338
339 static bool
vm_fault_is_read(const struct faultstate * fs)340 vm_fault_is_read(const struct faultstate *fs)
341 {
342 return ((fs->prot & VM_PROT_WRITE) == 0 &&
343 (fs->fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) == 0);
344 }
345
346 /*
347 * Unlocks fs.first_object and fs.map on success.
348 */
349 static enum fault_status
vm_fault_soft_fast(struct faultstate * fs)350 vm_fault_soft_fast(struct faultstate *fs)
351 {
352 vm_page_t m, m_map;
353 #if VM_NRESERVLEVEL > 0
354 vm_page_t m_super;
355 int flags;
356 #endif
357 int psind;
358 vm_offset_t vaddr;
359
360 MPASS(fs->vp == NULL);
361
362 /*
363 * If we fail, vast majority of the time it is because the page is not
364 * there to begin with. Opportunistically perform the lookup and
365 * subsequent checks without the object lock, revalidate later.
366 *
367 * Note: a busy page can be mapped for read|execute access.
368 */
369 m = vm_page_lookup_unlocked(fs->first_object, fs->first_pindex);
370 if (m == NULL || !vm_page_all_valid(m) ||
371 ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m))) {
372 VM_OBJECT_WLOCK(fs->first_object);
373 return (FAULT_FAILURE);
374 }
375
376 vaddr = fs->vaddr;
377
378 VM_OBJECT_RLOCK(fs->first_object);
379
380 /*
381 * Now that we stabilized the state, revalidate the page is in the shape
382 * we encountered above.
383 */
384
385 if (m->object != fs->first_object || m->pindex != fs->first_pindex)
386 goto fail;
387
388 vm_object_busy(fs->first_object);
389
390 if (!vm_page_all_valid(m) ||
391 ((fs->prot & VM_PROT_WRITE) != 0 && vm_page_busied(m)))
392 goto fail_busy;
393
394 m_map = m;
395 psind = 0;
396 #if VM_NRESERVLEVEL > 0
397 if ((m->flags & PG_FICTITIOUS) == 0 &&
398 (m_super = vm_reserv_to_superpage(m)) != NULL) {
399 psind = m_super->psind;
400 KASSERT(psind > 0,
401 ("psind %d of m_super %p < 1", psind, m_super));
402 flags = PS_ALL_VALID;
403 if ((fs->prot & VM_PROT_WRITE) != 0) {
404 /*
405 * Create a superpage mapping allowing write access
406 * only if none of the constituent pages are busy and
407 * all of them are already dirty (except possibly for
408 * the page that was faulted on).
409 */
410 flags |= PS_NONE_BUSY;
411 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0)
412 flags |= PS_ALL_DIRTY;
413 }
414 while (rounddown2(vaddr, pagesizes[psind]) < fs->entry->start ||
415 roundup2(vaddr + 1, pagesizes[psind]) > fs->entry->end ||
416 (vaddr & (pagesizes[psind] - 1)) !=
417 (VM_PAGE_TO_PHYS(m) & (pagesizes[psind] - 1)) ||
418 !vm_page_ps_test(m_super, psind, flags, m) ||
419 !pmap_ps_enabled(fs->map->pmap)) {
420 psind--;
421 if (psind == 0)
422 break;
423 m_super += rounddown2(m - m_super,
424 atop(pagesizes[psind]));
425 KASSERT(m_super->psind >= psind,
426 ("psind %d of m_super %p < %d", m_super->psind,
427 m_super, psind));
428 }
429 if (psind > 0) {
430 m_map = m_super;
431 vaddr = rounddown2(vaddr, pagesizes[psind]);
432 /* Preset the modified bit for dirty superpages. */
433 if ((flags & PS_ALL_DIRTY) != 0)
434 fs->fault_type |= VM_PROT_WRITE;
435 }
436 }
437 #endif
438 if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type |
439 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) !=
440 KERN_SUCCESS)
441 goto fail_busy;
442 if (fs->m_hold != NULL) {
443 (*fs->m_hold) = m;
444 vm_page_wire(m);
445 }
446 if (psind == 0 && !fs->wired)
447 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true);
448 VM_OBJECT_RUNLOCK(fs->first_object);
449 vm_fault_dirty(fs, m);
450 vm_object_unbusy(fs->first_object);
451 vm_map_lookup_done(fs->map, fs->entry);
452 curthread->td_ru.ru_minflt++;
453 return (FAULT_SUCCESS);
454 fail_busy:
455 vm_object_unbusy(fs->first_object);
456 fail:
457 if (!VM_OBJECT_TRYUPGRADE(fs->first_object)) {
458 VM_OBJECT_RUNLOCK(fs->first_object);
459 VM_OBJECT_WLOCK(fs->first_object);
460 }
461 return (FAULT_FAILURE);
462 }
463
464 static void
vm_fault_restore_map_lock(struct faultstate * fs)465 vm_fault_restore_map_lock(struct faultstate *fs)
466 {
467
468 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
469 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
470
471 if (!vm_map_trylock_read(fs->map)) {
472 VM_OBJECT_WUNLOCK(fs->first_object);
473 vm_map_lock_read(fs->map);
474 VM_OBJECT_WLOCK(fs->first_object);
475 }
476 fs->lookup_still_valid = true;
477 }
478
479 static void
vm_fault_populate_check_page(vm_page_t m)480 vm_fault_populate_check_page(vm_page_t m)
481 {
482
483 /*
484 * Check each page to ensure that the pager is obeying the
485 * interface: the page must be installed in the object, fully
486 * valid, and exclusively busied.
487 */
488 MPASS(m != NULL);
489 MPASS(vm_page_all_valid(m));
490 MPASS(vm_page_xbusied(m));
491 }
492
493 static void
vm_fault_populate_cleanup(vm_object_t object,vm_pindex_t first,vm_pindex_t last)494 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first,
495 vm_pindex_t last)
496 {
497 struct pctrie_iter pages;
498 vm_page_t m;
499
500 VM_OBJECT_ASSERT_WLOCKED(object);
501 MPASS(first <= last);
502 vm_page_iter_limit_init(&pages, object, last + 1);
503 VM_RADIX_FORALL_FROM(m, &pages, first) {
504 vm_fault_populate_check_page(m);
505 vm_page_deactivate(m);
506 vm_page_xunbusy(m);
507 }
508 KASSERT(pages.index == last, ("%s: pindex mismatch", __func__));
509 }
510
511 static enum fault_status
vm_fault_populate(struct faultstate * fs)512 vm_fault_populate(struct faultstate *fs)
513 {
514 vm_offset_t vaddr;
515 vm_page_t m;
516 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx;
517 int bdry_idx, i, npages, psind, rv;
518 enum fault_status res;
519
520 MPASS(fs->object == fs->first_object);
521 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
522 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0);
523 MPASS(fs->first_object->backing_object == NULL);
524 MPASS(fs->lookup_still_valid);
525
526 pager_first = OFF_TO_IDX(fs->entry->offset);
527 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1;
528 vm_fault_unlock_map(fs);
529 vm_fault_unlock_vp(fs);
530
531 res = FAULT_SUCCESS;
532
533 /*
534 * Call the pager (driver) populate() method.
535 *
536 * There is no guarantee that the method will be called again
537 * if the current fault is for read, and a future fault is
538 * for write. Report the entry's maximum allowed protection
539 * to the driver.
540 */
541 rv = vm_pager_populate(fs->first_object, fs->first_pindex,
542 fs->fault_type, fs->entry->max_protection, &pager_first,
543 &pager_last);
544
545 VM_OBJECT_ASSERT_WLOCKED(fs->first_object);
546 if (rv == VM_PAGER_BAD) {
547 /*
548 * VM_PAGER_BAD is the backdoor for a pager to request
549 * normal fault handling.
550 */
551 vm_fault_restore_map_lock(fs);
552 if (fs->map->timestamp != fs->map_generation)
553 return (FAULT_RESTART);
554 return (FAULT_CONTINUE);
555 }
556 if (rv != VM_PAGER_OK)
557 return (FAULT_FAILURE); /* AKA SIGSEGV */
558
559 /* Ensure that the driver is obeying the interface. */
560 MPASS(pager_first <= pager_last);
561 MPASS(fs->first_pindex <= pager_last);
562 MPASS(fs->first_pindex >= pager_first);
563 MPASS(pager_last < fs->first_object->size);
564
565 vm_fault_restore_map_lock(fs);
566 bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(fs->entry);
567 if (fs->map->timestamp != fs->map_generation) {
568 if (bdry_idx == 0) {
569 vm_fault_populate_cleanup(fs->first_object, pager_first,
570 pager_last);
571 } else {
572 m = vm_page_lookup(fs->first_object, pager_first);
573 if (m != fs->m)
574 vm_page_xunbusy(m);
575 }
576 return (FAULT_RESTART);
577 }
578
579 /*
580 * The map is unchanged after our last unlock. Process the fault.
581 *
582 * First, the special case of largepage mappings, where
583 * populate only busies the first page in superpage run.
584 */
585 if (bdry_idx != 0) {
586 KASSERT(PMAP_HAS_LARGEPAGES,
587 ("missing pmap support for large pages"));
588 m = vm_page_lookup(fs->first_object, pager_first);
589 vm_fault_populate_check_page(m);
590 VM_OBJECT_WUNLOCK(fs->first_object);
591 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) -
592 fs->entry->offset;
593 /* assert alignment for entry */
594 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0,
595 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx",
596 (uintmax_t)fs->entry->start, (uintmax_t)pager_first,
597 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr));
598 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0,
599 ("unaligned superpage m %p %#jx", m,
600 (uintmax_t)VM_PAGE_TO_PHYS(m)));
601 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot,
602 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) |
603 PMAP_ENTER_LARGEPAGE, bdry_idx);
604 VM_OBJECT_WLOCK(fs->first_object);
605 vm_page_xunbusy(m);
606 if (rv != KERN_SUCCESS) {
607 res = FAULT_FAILURE;
608 goto out;
609 }
610 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) {
611 for (i = 0; i < atop(pagesizes[bdry_idx]); i++)
612 vm_page_wire(m + i);
613 }
614 if (fs->m_hold != NULL) {
615 *fs->m_hold = m + (fs->first_pindex - pager_first);
616 vm_page_wire(*fs->m_hold);
617 }
618 goto out;
619 }
620
621 /*
622 * The range [pager_first, pager_last] that is given to the
623 * pager is only a hint. The pager may populate any range
624 * within the object that includes the requested page index.
625 * In case the pager expanded the range, clip it to fit into
626 * the map entry.
627 */
628 map_first = OFF_TO_IDX(fs->entry->offset);
629 if (map_first > pager_first) {
630 vm_fault_populate_cleanup(fs->first_object, pager_first,
631 map_first - 1);
632 pager_first = map_first;
633 }
634 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1;
635 if (map_last < pager_last) {
636 vm_fault_populate_cleanup(fs->first_object, map_last + 1,
637 pager_last);
638 pager_last = map_last;
639 }
640 for (pidx = pager_first; pidx <= pager_last; pidx += npages) {
641 m = vm_page_lookup(fs->first_object, pidx);
642 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset;
643 KASSERT(m != NULL && m->pindex == pidx,
644 ("%s: pindex mismatch", __func__));
645 psind = m->psind;
646 while (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 ||
647 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last ||
648 !pmap_ps_enabled(fs->map->pmap)))
649 psind--;
650
651 npages = atop(pagesizes[psind]);
652 for (i = 0; i < npages; i++) {
653 vm_fault_populate_check_page(&m[i]);
654 vm_fault_dirty(fs, &m[i]);
655 }
656 VM_OBJECT_WUNLOCK(fs->first_object);
657 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type |
658 (fs->wired ? PMAP_ENTER_WIRED : 0), psind);
659
660 /*
661 * pmap_enter() may fail for a superpage mapping if additional
662 * protection policies prevent the full mapping.
663 * For example, this will happen on amd64 if the entire
664 * address range does not share the same userspace protection
665 * key. Revert to single-page mappings if this happens.
666 */
667 MPASS(rv == KERN_SUCCESS ||
668 (psind > 0 && rv == KERN_PROTECTION_FAILURE));
669 if (__predict_false(psind > 0 &&
670 rv == KERN_PROTECTION_FAILURE)) {
671 MPASS(!fs->wired);
672 for (i = 0; i < npages; i++) {
673 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i),
674 &m[i], fs->prot, fs->fault_type, 0);
675 MPASS(rv == KERN_SUCCESS);
676 }
677 }
678
679 VM_OBJECT_WLOCK(fs->first_object);
680 for (i = 0; i < npages; i++) {
681 if ((fs->fault_flags & VM_FAULT_WIRE) != 0 &&
682 m[i].pindex == fs->first_pindex)
683 vm_page_wire(&m[i]);
684 else
685 vm_page_activate(&m[i]);
686 if (fs->m_hold != NULL &&
687 m[i].pindex == fs->first_pindex) {
688 (*fs->m_hold) = &m[i];
689 vm_page_wire(&m[i]);
690 }
691 vm_page_xunbusy(&m[i]);
692 }
693 }
694 out:
695 curthread->td_ru.ru_majflt++;
696 return (res);
697 }
698
699 static int prot_fault_translation;
700 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN,
701 &prot_fault_translation, 0,
702 "Control signal to deliver on protection fault");
703
704 /* compat definition to keep common code for signal translation */
705 #define UCODE_PAGEFLT 12
706 #ifdef T_PAGEFLT
707 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT");
708 #endif
709
710 /*
711 * vm_fault_trap:
712 *
713 * Handle a page fault occurring at the given address,
714 * requiring the given permissions, in the map specified.
715 * If successful, the page is inserted into the
716 * associated physical map.
717 *
718 * NOTE: the given address should be truncated to the
719 * proper page address.
720 *
721 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
722 * a standard error specifying why the fault is fatal is returned.
723 *
724 * The map in question must be referenced, and remains so.
725 * Caller may hold no locks.
726 */
727 int
vm_fault_trap(vm_map_t map,vm_offset_t vaddr,vm_prot_t fault_type,int fault_flags,int * signo,int * ucode)728 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
729 int fault_flags, int *signo, int *ucode)
730 {
731 int result;
732
733 MPASS(signo == NULL || ucode != NULL);
734 #ifdef KTRACE
735 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT))
736 ktrfault(vaddr, fault_type);
737 #endif
738 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags,
739 NULL);
740 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE ||
741 result == KERN_INVALID_ADDRESS ||
742 result == KERN_RESOURCE_SHORTAGE ||
743 result == KERN_PROTECTION_FAILURE ||
744 result == KERN_OUT_OF_BOUNDS,
745 ("Unexpected Mach error %d from vm_fault()", result));
746 #ifdef KTRACE
747 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND))
748 ktrfaultend(result);
749 #endif
750 if (result != KERN_SUCCESS && signo != NULL) {
751 switch (result) {
752 case KERN_FAILURE:
753 case KERN_INVALID_ADDRESS:
754 *signo = SIGSEGV;
755 *ucode = SEGV_MAPERR;
756 break;
757 case KERN_RESOURCE_SHORTAGE:
758 *signo = SIGBUS;
759 *ucode = BUS_OOMERR;
760 break;
761 case KERN_OUT_OF_BOUNDS:
762 *signo = SIGBUS;
763 *ucode = BUS_OBJERR;
764 break;
765 case KERN_PROTECTION_FAILURE:
766 if (prot_fault_translation == 0) {
767 /*
768 * Autodetect. This check also covers
769 * the images without the ABI-tag ELF
770 * note.
771 */
772 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD &&
773 curproc->p_osrel >= P_OSREL_SIGSEGV) {
774 *signo = SIGSEGV;
775 *ucode = SEGV_ACCERR;
776 } else {
777 *signo = SIGBUS;
778 *ucode = UCODE_PAGEFLT;
779 }
780 } else if (prot_fault_translation == 1) {
781 /* Always compat mode. */
782 *signo = SIGBUS;
783 *ucode = UCODE_PAGEFLT;
784 } else {
785 /* Always SIGSEGV mode. */
786 *signo = SIGSEGV;
787 *ucode = SEGV_ACCERR;
788 }
789 break;
790 default:
791 KASSERT(0, ("Unexpected Mach error %d from vm_fault()",
792 result));
793 break;
794 }
795 }
796 return (result);
797 }
798
799 static bool
vm_fault_object_ensure_wlocked(struct faultstate * fs)800 vm_fault_object_ensure_wlocked(struct faultstate *fs)
801 {
802 if (fs->object == fs->first_object)
803 VM_OBJECT_ASSERT_WLOCKED(fs->object);
804
805 if (!fs->can_read_lock) {
806 VM_OBJECT_ASSERT_WLOCKED(fs->object);
807 return (true);
808 }
809
810 if (VM_OBJECT_WOWNED(fs->object))
811 return (true);
812
813 if (VM_OBJECT_TRYUPGRADE(fs->object))
814 return (true);
815
816 return (false);
817 }
818
819 static enum fault_status
vm_fault_lock_vnode(struct faultstate * fs,bool objlocked)820 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked)
821 {
822 struct vnode *vp;
823 int error, locked;
824
825 if (fs->object->type != OBJT_VNODE)
826 return (FAULT_CONTINUE);
827 vp = fs->object->handle;
828 if (vp == fs->vp) {
829 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked");
830 return (FAULT_CONTINUE);
831 }
832
833 /*
834 * Perform an unlock in case the desired vnode changed while
835 * the map was unlocked during a retry.
836 */
837 vm_fault_unlock_vp(fs);
838
839 locked = VOP_ISLOCKED(vp);
840 if (locked != LK_EXCLUSIVE)
841 locked = LK_SHARED;
842
843 /*
844 * We must not sleep acquiring the vnode lock while we have
845 * the page exclusive busied or the object's
846 * paging-in-progress count incremented. Otherwise, we could
847 * deadlock.
848 */
849 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT);
850 if (error == 0) {
851 fs->vp = vp;
852 return (FAULT_CONTINUE);
853 }
854
855 vhold(vp);
856 if (objlocked)
857 vm_fault_unlock_and_deallocate(fs);
858 else
859 vm_fault_deallocate(fs);
860 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE);
861 vdrop(vp);
862 fs->vp = vp;
863 KASSERT(error == 0, ("vm_fault: vget failed %d", error));
864 return (FAULT_RESTART);
865 }
866
867 /*
868 * Calculate the desired readahead. Handle drop-behind.
869 *
870 * Returns the number of readahead blocks to pass to the pager.
871 */
872 static int
vm_fault_readahead(struct faultstate * fs)873 vm_fault_readahead(struct faultstate *fs)
874 {
875 int era, nera;
876 u_char behavior;
877
878 KASSERT(fs->lookup_still_valid, ("map unlocked"));
879 era = fs->entry->read_ahead;
880 behavior = vm_map_entry_behavior(fs->entry);
881 if (behavior == MAP_ENTRY_BEHAV_RANDOM) {
882 nera = 0;
883 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) {
884 nera = VM_FAULT_READ_AHEAD_MAX;
885 if (fs->vaddr == fs->entry->next_read)
886 vm_fault_dontneed(fs, fs->vaddr, nera);
887 } else if (fs->vaddr == fs->entry->next_read) {
888 /*
889 * This is a sequential fault. Arithmetically
890 * increase the requested number of pages in
891 * the read-ahead window. The requested
892 * number of pages is "# of sequential faults
893 * x (read ahead min + 1) + read ahead min"
894 */
895 nera = VM_FAULT_READ_AHEAD_MIN;
896 if (era > 0) {
897 nera += era + 1;
898 if (nera > VM_FAULT_READ_AHEAD_MAX)
899 nera = VM_FAULT_READ_AHEAD_MAX;
900 }
901 if (era == VM_FAULT_READ_AHEAD_MAX)
902 vm_fault_dontneed(fs, fs->vaddr, nera);
903 } else {
904 /*
905 * This is a non-sequential fault.
906 */
907 nera = 0;
908 }
909 if (era != nera) {
910 /*
911 * A read lock on the map suffices to update
912 * the read ahead count safely.
913 */
914 fs->entry->read_ahead = nera;
915 }
916
917 return (nera);
918 }
919
920 static int
vm_fault_lookup(struct faultstate * fs)921 vm_fault_lookup(struct faultstate *fs)
922 {
923 int result;
924
925 KASSERT(!fs->lookup_still_valid,
926 ("vm_fault_lookup: Map already locked."));
927 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type |
928 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object,
929 &fs->first_pindex, &fs->prot, &fs->wired);
930 if (result != KERN_SUCCESS) {
931 vm_fault_unlock_vp(fs);
932 return (result);
933 }
934
935 fs->map_generation = fs->map->timestamp;
936
937 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) {
938 panic("%s: fault on nofault entry, addr: %#lx",
939 __func__, (u_long)fs->vaddr);
940 }
941
942 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION &&
943 fs->entry->wiring_thread != curthread) {
944 vm_map_unlock_read(fs->map);
945 vm_map_lock(fs->map);
946 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) &&
947 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) {
948 vm_fault_unlock_vp(fs);
949 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP;
950 vm_map_unlock_and_wait(fs->map, 0);
951 } else
952 vm_map_unlock(fs->map);
953 return (KERN_RESOURCE_SHORTAGE);
954 }
955
956 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0);
957
958 if (fs->wired)
959 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY);
960 else
961 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0,
962 ("!fs->wired && VM_FAULT_WIRE"));
963 fs->lookup_still_valid = true;
964
965 return (KERN_SUCCESS);
966 }
967
968 static int
vm_fault_relookup(struct faultstate * fs)969 vm_fault_relookup(struct faultstate *fs)
970 {
971 vm_object_t retry_object;
972 vm_pindex_t retry_pindex;
973 vm_prot_t retry_prot;
974 int result;
975
976 if (!vm_map_trylock_read(fs->map))
977 return (KERN_RESTART);
978
979 fs->lookup_still_valid = true;
980 if (fs->map->timestamp == fs->map_generation)
981 return (KERN_SUCCESS);
982
983 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type,
984 &fs->entry, &retry_object, &retry_pindex, &retry_prot,
985 &fs->wired);
986 if (result != KERN_SUCCESS) {
987 /*
988 * If retry of map lookup would have blocked then
989 * retry fault from start.
990 */
991 if (result == KERN_FAILURE)
992 return (KERN_RESTART);
993 return (result);
994 }
995 if (retry_object != fs->first_object ||
996 retry_pindex != fs->first_pindex)
997 return (KERN_RESTART);
998
999 /*
1000 * Check whether the protection has changed or the object has
1001 * been copied while we left the map unlocked. Changing from
1002 * read to write permission is OK - we leave the page
1003 * write-protected, and catch the write fault. Changing from
1004 * write to read permission means that we can't mark the page
1005 * write-enabled after all.
1006 */
1007 fs->prot &= retry_prot;
1008 fs->fault_type &= retry_prot;
1009 if (fs->prot == 0)
1010 return (KERN_RESTART);
1011
1012 /* Reassert because wired may have changed. */
1013 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0,
1014 ("!wired && VM_FAULT_WIRE"));
1015
1016 return (KERN_SUCCESS);
1017 }
1018
1019 static bool
vm_fault_can_cow_rename(struct faultstate * fs)1020 vm_fault_can_cow_rename(struct faultstate *fs)
1021 {
1022 return (
1023 /* Only one shadow object and no other refs. */
1024 fs->object->shadow_count == 1 && fs->object->ref_count == 1 &&
1025 /* No other ways to look the object up. */
1026 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0);
1027 }
1028
1029 static void
vm_fault_cow(struct faultstate * fs)1030 vm_fault_cow(struct faultstate *fs)
1031 {
1032 bool is_first_object_locked, rename_cow;
1033
1034 KASSERT(vm_fault_might_be_cow(fs),
1035 ("source and target COW objects are identical"));
1036
1037 /*
1038 * This allows pages to be virtually copied from a backing_object
1039 * into the first_object, where the backing object has no other
1040 * refs to it, and cannot gain any more refs. Instead of a bcopy,
1041 * we just move the page from the backing object to the first
1042 * object. Note that we must mark the page dirty in the first
1043 * object so that it will go out to swap when needed.
1044 */
1045 is_first_object_locked = false;
1046 rename_cow = false;
1047
1048 if (vm_fault_can_cow_rename(fs) && vm_page_xbusied(fs->m)) {
1049 /*
1050 * Check that we don't chase down the shadow chain and
1051 * we can acquire locks. Recheck the conditions for
1052 * rename after the shadow chain is stable after the
1053 * object locking.
1054 */
1055 is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object);
1056 if (is_first_object_locked &&
1057 fs->object == fs->first_object->backing_object) {
1058 if (VM_OBJECT_TRYWLOCK(fs->object)) {
1059 rename_cow = vm_fault_can_cow_rename(fs);
1060 if (!rename_cow)
1061 VM_OBJECT_WUNLOCK(fs->object);
1062 }
1063 }
1064 }
1065
1066 if (rename_cow) {
1067 vm_page_assert_xbusied(fs->m);
1068
1069 /*
1070 * Remove but keep xbusy for replace. fs->m is moved into
1071 * fs->first_object and left busy while fs->first_m is
1072 * conditionally freed.
1073 */
1074 vm_page_remove_xbusy(fs->m);
1075 vm_page_replace(fs->m, fs->first_object, fs->first_pindex,
1076 fs->first_m);
1077 vm_page_dirty(fs->m);
1078 #if VM_NRESERVLEVEL > 0
1079 /*
1080 * Rename the reservation.
1081 */
1082 vm_reserv_rename(fs->m, fs->first_object, fs->object,
1083 OFF_TO_IDX(fs->first_object->backing_object_offset));
1084 #endif
1085 VM_OBJECT_WUNLOCK(fs->object);
1086 VM_OBJECT_WUNLOCK(fs->first_object);
1087 fs->first_m = fs->m;
1088 fs->m = NULL;
1089 VM_CNT_INC(v_cow_optim);
1090 } else {
1091 if (is_first_object_locked)
1092 VM_OBJECT_WUNLOCK(fs->first_object);
1093 /*
1094 * Oh, well, lets copy it.
1095 */
1096 pmap_copy_page(fs->m, fs->first_m);
1097 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) {
1098 vm_page_wire(fs->first_m);
1099 vm_page_unwire(fs->m, PQ_INACTIVE);
1100 }
1101 /*
1102 * Save the COW page to be released after pmap_enter is
1103 * complete. The new copy will be marked valid when we're ready
1104 * to map it.
1105 */
1106 fs->m_cow = fs->m;
1107 fs->m = NULL;
1108
1109 /*
1110 * Typically, the shadow object is either private to this
1111 * address space (OBJ_ONEMAPPING) or its pages are read only.
1112 * In the highly unusual case where the pages of a shadow object
1113 * are read/write shared between this and other address spaces,
1114 * we need to ensure that any pmap-level mappings to the
1115 * original, copy-on-write page from the backing object are
1116 * removed from those other address spaces.
1117 *
1118 * The flag check is racy, but this is tolerable: if
1119 * OBJ_ONEMAPPING is cleared after the check, the busy state
1120 * ensures that new mappings of m_cow can't be created.
1121 * pmap_enter() will replace an existing mapping in the current
1122 * address space. If OBJ_ONEMAPPING is set after the check,
1123 * removing mappings will at worse trigger some unnecessary page
1124 * faults.
1125 *
1126 * In the fs->m shared busy case, the xbusy state of
1127 * fs->first_m prevents new mappings of fs->m from
1128 * being created because a parallel fault on this
1129 * shadow chain should wait for xbusy on fs->first_m.
1130 */
1131 if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0)
1132 pmap_remove_all(fs->m_cow);
1133 }
1134
1135 vm_object_pip_wakeup(fs->object);
1136
1137 /*
1138 * Only use the new page below...
1139 */
1140 fs->object = fs->first_object;
1141 fs->pindex = fs->first_pindex;
1142 fs->m = fs->first_m;
1143 VM_CNT_INC(v_cow_faults);
1144 curthread->td_cow++;
1145 }
1146
1147 static enum fault_next_status
vm_fault_next(struct faultstate * fs)1148 vm_fault_next(struct faultstate *fs)
1149 {
1150 vm_object_t next_object;
1151
1152 if (fs->object == fs->first_object || !fs->can_read_lock)
1153 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1154 else
1155 VM_OBJECT_ASSERT_LOCKED(fs->object);
1156
1157 /*
1158 * The requested page does not exist at this object/
1159 * offset. Remove the invalid page from the object,
1160 * waking up anyone waiting for it, and continue on to
1161 * the next object. However, if this is the top-level
1162 * object, we must leave the busy page in place to
1163 * prevent another process from rushing past us, and
1164 * inserting the page in that object at the same time
1165 * that we are.
1166 */
1167 if (fs->object == fs->first_object) {
1168 fs->first_m = fs->m;
1169 fs->m = NULL;
1170 } else if (fs->m != NULL) {
1171 if (!vm_fault_object_ensure_wlocked(fs)) {
1172 fs->can_read_lock = false;
1173 vm_fault_unlock_and_deallocate(fs);
1174 return (FAULT_NEXT_RESTART);
1175 }
1176 vm_fault_page_free(&fs->m);
1177 }
1178
1179 /*
1180 * Move on to the next object. Lock the next object before
1181 * unlocking the current one.
1182 */
1183 next_object = fs->object->backing_object;
1184 if (next_object == NULL)
1185 return (FAULT_NEXT_NOOBJ);
1186 MPASS(fs->first_m != NULL);
1187 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1188 if (fs->can_read_lock)
1189 VM_OBJECT_RLOCK(next_object);
1190 else
1191 VM_OBJECT_WLOCK(next_object);
1192 vm_object_pip_add(next_object, 1);
1193 if (fs->object != fs->first_object)
1194 vm_object_pip_wakeup(fs->object);
1195 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset);
1196 VM_OBJECT_UNLOCK(fs->object);
1197 fs->object = next_object;
1198
1199 return (FAULT_NEXT_GOTOBJ);
1200 }
1201
1202 static void
vm_fault_zerofill(struct faultstate * fs)1203 vm_fault_zerofill(struct faultstate *fs)
1204 {
1205
1206 /*
1207 * If there's no object left, fill the page in the top
1208 * object with zeros.
1209 */
1210 if (vm_fault_might_be_cow(fs)) {
1211 vm_object_pip_wakeup(fs->object);
1212 fs->object = fs->first_object;
1213 fs->pindex = fs->first_pindex;
1214 }
1215 MPASS(fs->first_m != NULL);
1216 MPASS(fs->m == NULL);
1217 fs->m = fs->first_m;
1218 fs->first_m = NULL;
1219
1220 /*
1221 * Zero the page if necessary and mark it valid.
1222 */
1223 if ((fs->m->flags & PG_ZERO) == 0) {
1224 pmap_zero_page(fs->m);
1225 } else {
1226 VM_CNT_INC(v_ozfod);
1227 }
1228 VM_CNT_INC(v_zfod);
1229 vm_page_valid(fs->m);
1230 }
1231
1232 /*
1233 * Initiate page fault after timeout. Returns true if caller should
1234 * do vm_waitpfault() after the call.
1235 */
1236 static bool
vm_fault_allocate_oom(struct faultstate * fs)1237 vm_fault_allocate_oom(struct faultstate *fs)
1238 {
1239 struct timeval now;
1240
1241 vm_fault_unlock_and_deallocate(fs);
1242 if (vm_pfault_oom_attempts < 0)
1243 return (true);
1244 if (!fs->oom_started) {
1245 fs->oom_started = true;
1246 getmicrotime(&fs->oom_start_time);
1247 return (true);
1248 }
1249
1250 getmicrotime(&now);
1251 timevalsub(&now, &fs->oom_start_time);
1252 if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait)
1253 return (true);
1254
1255 if (bootverbose)
1256 printf(
1257 "proc %d (%s) failed to alloc page on fault, starting OOM\n",
1258 curproc->p_pid, curproc->p_comm);
1259 vm_pageout_oom(VM_OOM_MEM_PF);
1260 fs->oom_started = false;
1261 return (false);
1262 }
1263
1264 /*
1265 * Allocate a page directly or via the object populate method.
1266 */
1267 static enum fault_status
vm_fault_allocate(struct faultstate * fs,struct pctrie_iter * pages)1268 vm_fault_allocate(struct faultstate *fs, struct pctrie_iter *pages)
1269 {
1270 struct domainset *dset;
1271 enum fault_status res;
1272
1273 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) {
1274 res = vm_fault_lock_vnode(fs, true);
1275 MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART);
1276 if (res == FAULT_RESTART)
1277 return (res);
1278 }
1279
1280 if (fs->pindex >= fs->object->size) {
1281 vm_fault_unlock_and_deallocate(fs);
1282 return (FAULT_OUT_OF_BOUNDS);
1283 }
1284
1285 if (fs->object == fs->first_object &&
1286 (fs->first_object->flags & OBJ_POPULATE) != 0 &&
1287 fs->first_object->shadow_count == 0) {
1288 res = vm_fault_populate(fs);
1289 switch (res) {
1290 case FAULT_SUCCESS:
1291 case FAULT_FAILURE:
1292 case FAULT_RESTART:
1293 vm_fault_unlock_and_deallocate(fs);
1294 return (res);
1295 case FAULT_CONTINUE:
1296 pctrie_iter_reset(pages);
1297 /*
1298 * Pager's populate() method
1299 * returned VM_PAGER_BAD.
1300 */
1301 break;
1302 default:
1303 panic("inconsistent return codes");
1304 }
1305 }
1306
1307 /*
1308 * Allocate a new page for this object/offset pair.
1309 *
1310 * If the process has a fatal signal pending, prioritize the allocation
1311 * with the expectation that the process will exit shortly and free some
1312 * pages. In particular, the signal may have been posted by the page
1313 * daemon in an attempt to resolve an out-of-memory condition.
1314 *
1315 * The unlocked read of the p_flag is harmless. At worst, the P_KILLED
1316 * might be not observed here, and allocation fails, causing a restart
1317 * and new reading of the p_flag.
1318 */
1319 dset = fs->object->domain.dr_policy;
1320 if (dset == NULL)
1321 dset = curthread->td_domain.dr_policy;
1322 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) {
1323 #if VM_NRESERVLEVEL > 0
1324 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex);
1325 #endif
1326 if (!vm_pager_can_alloc_page(fs->object, fs->pindex)) {
1327 vm_fault_unlock_and_deallocate(fs);
1328 return (FAULT_FAILURE);
1329 }
1330 fs->m = vm_page_alloc_iter(fs->object, fs->pindex,
1331 P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0, pages);
1332 }
1333 if (fs->m == NULL) {
1334 if (vm_fault_allocate_oom(fs))
1335 vm_waitpfault(dset, vm_pfault_oom_wait * hz);
1336 return (FAULT_RESTART);
1337 }
1338 fs->oom_started = false;
1339
1340 return (FAULT_CONTINUE);
1341 }
1342
1343 /*
1344 * Call the pager to retrieve the page if there is a chance
1345 * that the pager has it, and potentially retrieve additional
1346 * pages at the same time.
1347 */
1348 static enum fault_status
vm_fault_getpages(struct faultstate * fs,int * behindp,int * aheadp)1349 vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp)
1350 {
1351 vm_offset_t e_end, e_start;
1352 int ahead, behind, cluster_offset, rv;
1353 enum fault_status status;
1354 u_char behavior;
1355
1356 /*
1357 * Prepare for unlocking the map. Save the map
1358 * entry's start and end addresses, which are used to
1359 * optimize the size of the pager operation below.
1360 * Even if the map entry's addresses change after
1361 * unlocking the map, using the saved addresses is
1362 * safe.
1363 */
1364 e_start = fs->entry->start;
1365 e_end = fs->entry->end;
1366 behavior = vm_map_entry_behavior(fs->entry);
1367
1368 /*
1369 * If the pager for the current object might have
1370 * the page, then determine the number of additional
1371 * pages to read and potentially reprioritize
1372 * previously read pages for earlier reclamation.
1373 * These operations should only be performed once per
1374 * page fault. Even if the current pager doesn't
1375 * have the page, the number of additional pages to
1376 * read will apply to subsequent objects in the
1377 * shadow chain.
1378 */
1379 if (fs->nera == -1 && !P_KILLED(curproc))
1380 fs->nera = vm_fault_readahead(fs);
1381
1382 /*
1383 * Release the map lock before locking the vnode or
1384 * sleeping in the pager. (If the current object has
1385 * a shadow, then an earlier iteration of this loop
1386 * may have already unlocked the map.)
1387 */
1388 vm_fault_unlock_map(fs);
1389
1390 status = vm_fault_lock_vnode(fs, false);
1391 MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART);
1392 if (status == FAULT_RESTART)
1393 return (status);
1394 KASSERT(fs->vp == NULL || !vm_map_is_system(fs->map),
1395 ("vm_fault: vnode-backed object mapped by system map"));
1396
1397 /*
1398 * Page in the requested page and hint the pager,
1399 * that it may bring up surrounding pages.
1400 */
1401 if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM ||
1402 P_KILLED(curproc)) {
1403 behind = 0;
1404 ahead = 0;
1405 } else {
1406 /* Is this a sequential fault? */
1407 if (fs->nera > 0) {
1408 behind = 0;
1409 ahead = fs->nera;
1410 } else {
1411 /*
1412 * Request a cluster of pages that is
1413 * aligned to a VM_FAULT_READ_DEFAULT
1414 * page offset boundary within the
1415 * object. Alignment to a page offset
1416 * boundary is more likely to coincide
1417 * with the underlying file system
1418 * block than alignment to a virtual
1419 * address boundary.
1420 */
1421 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT;
1422 behind = ulmin(cluster_offset,
1423 atop(fs->vaddr - e_start));
1424 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset;
1425 }
1426 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1);
1427 }
1428 *behindp = behind;
1429 *aheadp = ahead;
1430 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp);
1431 if (rv == VM_PAGER_OK)
1432 return (FAULT_HARD);
1433 if (rv == VM_PAGER_ERROR)
1434 printf("vm_fault: pager read error, pid %d (%s)\n",
1435 curproc->p_pid, curproc->p_comm);
1436 /*
1437 * If an I/O error occurred or the requested page was
1438 * outside the range of the pager, clean up and return
1439 * an error.
1440 */
1441 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) {
1442 VM_OBJECT_WLOCK(fs->object);
1443 vm_fault_page_free(&fs->m);
1444 vm_fault_unlock_and_deallocate(fs);
1445 return (FAULT_OUT_OF_BOUNDS);
1446 }
1447 KASSERT(rv == VM_PAGER_FAIL,
1448 ("%s: unexpected pager error %d", __func__, rv));
1449 return (FAULT_CONTINUE);
1450 }
1451
1452 /*
1453 * Wait/Retry if the page is busy. We have to do this if the page is
1454 * either exclusive or shared busy because the vm_pager may be using
1455 * read busy for pageouts (and even pageins if it is the vnode pager),
1456 * and we could end up trying to pagein and pageout the same page
1457 * simultaneously.
1458 *
1459 * We can theoretically allow the busy case on a read fault if the page
1460 * is marked valid, but since such pages are typically already pmap'd,
1461 * putting that special case in might be more effort then it is worth.
1462 * We cannot under any circumstances mess around with a shared busied
1463 * page except, perhaps, to pmap it.
1464 */
1465 static void
vm_fault_busy_sleep(struct faultstate * fs,int allocflags)1466 vm_fault_busy_sleep(struct faultstate *fs, int allocflags)
1467 {
1468 /*
1469 * Reference the page before unlocking and
1470 * sleeping so that the page daemon is less
1471 * likely to reclaim it.
1472 */
1473 vm_page_aflag_set(fs->m, PGA_REFERENCED);
1474 if (vm_fault_might_be_cow(fs)) {
1475 vm_fault_page_release(&fs->first_m);
1476 vm_object_pip_wakeup(fs->first_object);
1477 }
1478 vm_object_pip_wakeup(fs->object);
1479 vm_fault_unlock_map(fs);
1480 if (!vm_page_busy_sleep(fs->m, "vmpfw", allocflags))
1481 VM_OBJECT_UNLOCK(fs->object);
1482 VM_CNT_INC(v_intrans);
1483 vm_object_deallocate(fs->first_object);
1484 }
1485
1486 /*
1487 * Handle page lookup, populate, allocate, page-in for the current
1488 * object.
1489 *
1490 * The object is locked on entry and will remain locked with a return
1491 * code of FAULT_CONTINUE so that fault may follow the shadow chain.
1492 * Otherwise, the object will be unlocked upon return.
1493 */
1494 static enum fault_status
vm_fault_object(struct faultstate * fs,int * behindp,int * aheadp)1495 vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp)
1496 {
1497 struct pctrie_iter pages;
1498 enum fault_status res;
1499 bool dead;
1500
1501 if (fs->object == fs->first_object || !fs->can_read_lock)
1502 VM_OBJECT_ASSERT_WLOCKED(fs->object);
1503 else
1504 VM_OBJECT_ASSERT_LOCKED(fs->object);
1505
1506 /*
1507 * If the object is marked for imminent termination, we retry
1508 * here, since the collapse pass has raced with us. Otherwise,
1509 * if we see terminally dead object, return fail.
1510 */
1511 if ((fs->object->flags & OBJ_DEAD) != 0) {
1512 dead = fs->object->type == OBJT_DEAD;
1513 vm_fault_unlock_and_deallocate(fs);
1514 if (dead)
1515 return (FAULT_PROTECTION_FAILURE);
1516 pause("vmf_de", 1);
1517 return (FAULT_RESTART);
1518 }
1519
1520 /*
1521 * See if the page is resident.
1522 */
1523 vm_page_iter_init(&pages, fs->object);
1524 fs->m = vm_radix_iter_lookup(&pages, fs->pindex);
1525 if (fs->m != NULL) {
1526 /*
1527 * If the found page is valid, will be either shadowed
1528 * or mapped read-only, and will not be renamed for
1529 * COW, then busy it in shared mode. This allows
1530 * other faults needing this page to proceed in
1531 * parallel.
1532 *
1533 * Unlocked check for validity, rechecked after busy
1534 * is obtained.
1535 */
1536 if (vm_page_all_valid(fs->m) &&
1537 /*
1538 * No write permissions for the new fs->m mapping,
1539 * or the first object has only one mapping, so
1540 * other writeable COW mappings of fs->m cannot
1541 * appear under us.
1542 */
1543 (vm_fault_is_read(fs) || vm_fault_might_be_cow(fs)) &&
1544 /*
1545 * fs->m cannot be renamed from object to
1546 * first_object. These conditions will be
1547 * re-checked with proper synchronization in
1548 * vm_fault_cow().
1549 */
1550 (!vm_fault_can_cow_rename(fs) ||
1551 fs->object != fs->first_object->backing_object)) {
1552 if (!vm_page_trysbusy(fs->m)) {
1553 vm_fault_busy_sleep(fs, VM_ALLOC_SBUSY);
1554 return (FAULT_RESTART);
1555 }
1556
1557 /*
1558 * Now make sure that racily checked
1559 * conditions are still valid.
1560 */
1561 if (__predict_true(vm_page_all_valid(fs->m) &&
1562 (vm_fault_is_read(fs) ||
1563 vm_fault_might_be_cow(fs)))) {
1564 VM_OBJECT_UNLOCK(fs->object);
1565 return (FAULT_SOFT);
1566 }
1567
1568 vm_page_sunbusy(fs->m);
1569 }
1570
1571 if (!vm_page_tryxbusy(fs->m)) {
1572 vm_fault_busy_sleep(fs, 0);
1573 return (FAULT_RESTART);
1574 }
1575
1576 /*
1577 * The page is marked busy for other processes and the
1578 * pagedaemon. If it is still completely valid we are
1579 * done.
1580 */
1581 if (vm_page_all_valid(fs->m)) {
1582 VM_OBJECT_UNLOCK(fs->object);
1583 return (FAULT_SOFT);
1584 }
1585 }
1586
1587 /*
1588 * Page is not resident. If the pager might contain the page
1589 * or this is the beginning of the search, allocate a new
1590 * page.
1591 */
1592 if (fs->m == NULL && (vm_fault_object_needs_getpages(fs->object) ||
1593 fs->object == fs->first_object)) {
1594 if (!vm_fault_object_ensure_wlocked(fs)) {
1595 fs->can_read_lock = false;
1596 vm_fault_unlock_and_deallocate(fs);
1597 return (FAULT_RESTART);
1598 }
1599 res = vm_fault_allocate(fs, &pages);
1600 if (res != FAULT_CONTINUE)
1601 return (res);
1602 }
1603
1604 /*
1605 * Check to see if the pager can possibly satisfy this fault.
1606 * If not, skip to the next object without dropping the lock to
1607 * preserve atomicity of shadow faults.
1608 */
1609 if (vm_fault_object_needs_getpages(fs->object)) {
1610 /*
1611 * At this point, we have either allocated a new page
1612 * or found an existing page that is only partially
1613 * valid.
1614 *
1615 * We hold a reference on the current object and the
1616 * page is exclusive busied. The exclusive busy
1617 * prevents simultaneous faults and collapses while
1618 * the object lock is dropped.
1619 */
1620 VM_OBJECT_UNLOCK(fs->object);
1621 res = vm_fault_getpages(fs, behindp, aheadp);
1622 if (res == FAULT_CONTINUE)
1623 VM_OBJECT_WLOCK(fs->object);
1624 } else {
1625 res = FAULT_CONTINUE;
1626 }
1627 return (res);
1628 }
1629
1630 int
vm_fault(vm_map_t map,vm_offset_t vaddr,vm_prot_t fault_type,int fault_flags,vm_page_t * m_hold)1631 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
1632 int fault_flags, vm_page_t *m_hold)
1633 {
1634 struct pctrie_iter pages;
1635 struct faultstate fs;
1636 int ahead, behind, faultcount, rv;
1637 enum fault_status res;
1638 enum fault_next_status res_next;
1639 bool hardfault;
1640
1641 VM_CNT_INC(v_vm_faults);
1642
1643 if ((curthread->td_pflags & TDP_NOFAULTING) != 0)
1644 return (KERN_PROTECTION_FAILURE);
1645
1646 fs.vp = NULL;
1647 fs.vaddr = vaddr;
1648 fs.m_hold = m_hold;
1649 fs.fault_flags = fault_flags;
1650 fs.map = map;
1651 fs.lookup_still_valid = false;
1652 fs.oom_started = false;
1653 fs.nera = -1;
1654 fs.can_read_lock = true;
1655 faultcount = 0;
1656 hardfault = false;
1657
1658 RetryFault:
1659 fs.fault_type = fault_type;
1660
1661 /*
1662 * Find the backing store object and offset into it to begin the
1663 * search.
1664 */
1665 rv = vm_fault_lookup(&fs);
1666 if (rv != KERN_SUCCESS) {
1667 if (rv == KERN_RESOURCE_SHORTAGE)
1668 goto RetryFault;
1669 return (rv);
1670 }
1671
1672 /*
1673 * Try to avoid lock contention on the top-level object through
1674 * special-case handling of some types of page faults, specifically,
1675 * those that are mapping an existing page from the top-level object.
1676 * Under this condition, a read lock on the object suffices, allowing
1677 * multiple page faults of a similar type to run in parallel.
1678 */
1679 if (fs.vp == NULL /* avoid locked vnode leak */ &&
1680 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 &&
1681 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) {
1682 res = vm_fault_soft_fast(&fs);
1683 if (res == FAULT_SUCCESS) {
1684 VM_OBJECT_ASSERT_UNLOCKED(fs.first_object);
1685 return (KERN_SUCCESS);
1686 }
1687 VM_OBJECT_ASSERT_WLOCKED(fs.first_object);
1688 } else {
1689 vm_page_iter_init(&pages, fs.first_object);
1690 VM_OBJECT_WLOCK(fs.first_object);
1691 }
1692
1693 /*
1694 * Make a reference to this object to prevent its disposal while we
1695 * are messing with it. Once we have the reference, the map is free
1696 * to be diddled. Since objects reference their shadows (and copies),
1697 * they will stay around as well.
1698 *
1699 * Bump the paging-in-progress count to prevent size changes (e.g.
1700 * truncation operations) during I/O.
1701 */
1702 vm_object_reference_locked(fs.first_object);
1703 vm_object_pip_add(fs.first_object, 1);
1704
1705 fs.m_cow = fs.m = fs.first_m = NULL;
1706
1707 /*
1708 * Search for the page at object/offset.
1709 */
1710 fs.object = fs.first_object;
1711 fs.pindex = fs.first_pindex;
1712
1713 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) {
1714 res = vm_fault_allocate(&fs, &pages);
1715 switch (res) {
1716 case FAULT_RESTART:
1717 goto RetryFault;
1718 case FAULT_SUCCESS:
1719 return (KERN_SUCCESS);
1720 case FAULT_FAILURE:
1721 return (KERN_FAILURE);
1722 case FAULT_OUT_OF_BOUNDS:
1723 return (KERN_OUT_OF_BOUNDS);
1724 case FAULT_CONTINUE:
1725 break;
1726 default:
1727 panic("vm_fault: Unhandled status %d", res);
1728 }
1729 }
1730
1731 while (TRUE) {
1732 KASSERT(fs.m == NULL,
1733 ("page still set %p at loop start", fs.m));
1734
1735 res = vm_fault_object(&fs, &behind, &ahead);
1736 switch (res) {
1737 case FAULT_SOFT:
1738 goto found;
1739 case FAULT_HARD:
1740 faultcount = behind + 1 + ahead;
1741 hardfault = true;
1742 goto found;
1743 case FAULT_RESTART:
1744 goto RetryFault;
1745 case FAULT_SUCCESS:
1746 return (KERN_SUCCESS);
1747 case FAULT_FAILURE:
1748 return (KERN_FAILURE);
1749 case FAULT_OUT_OF_BOUNDS:
1750 return (KERN_OUT_OF_BOUNDS);
1751 case FAULT_PROTECTION_FAILURE:
1752 return (KERN_PROTECTION_FAILURE);
1753 case FAULT_CONTINUE:
1754 break;
1755 default:
1756 panic("vm_fault: Unhandled status %d", res);
1757 }
1758
1759 /*
1760 * The page was not found in the current object. Try to
1761 * traverse into a backing object or zero fill if none is
1762 * found.
1763 */
1764 res_next = vm_fault_next(&fs);
1765 if (res_next == FAULT_NEXT_RESTART)
1766 goto RetryFault;
1767 else if (res_next == FAULT_NEXT_GOTOBJ)
1768 continue;
1769 MPASS(res_next == FAULT_NEXT_NOOBJ);
1770 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) {
1771 if (fs.first_object == fs.object)
1772 vm_fault_page_free(&fs.first_m);
1773 vm_fault_unlock_and_deallocate(&fs);
1774 return (KERN_OUT_OF_BOUNDS);
1775 }
1776 VM_OBJECT_UNLOCK(fs.object);
1777 vm_fault_zerofill(&fs);
1778 /* Don't try to prefault neighboring pages. */
1779 faultcount = 1;
1780 break;
1781 }
1782
1783 found:
1784 /*
1785 * A valid page has been found and busied. The object lock
1786 * must no longer be held if the page was busied.
1787 *
1788 * Regardless of the busy state of fs.m, fs.first_m is always
1789 * exclusively busied after the first iteration of the loop
1790 * calling vm_fault_object(). This is an ordering point for
1791 * the parallel faults occuring in on the same page.
1792 */
1793 vm_page_assert_busied(fs.m);
1794 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1795
1796 /*
1797 * If the page is being written, but isn't already owned by the
1798 * top-level object, we have to copy it into a new page owned by the
1799 * top-level object.
1800 */
1801 if (vm_fault_might_be_cow(&fs)) {
1802 /*
1803 * We only really need to copy if we want to write it.
1804 */
1805 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1806 vm_fault_cow(&fs);
1807 /*
1808 * We only try to prefault read-only mappings to the
1809 * neighboring pages when this copy-on-write fault is
1810 * a hard fault. In other cases, trying to prefault
1811 * is typically wasted effort.
1812 */
1813 if (faultcount == 0)
1814 faultcount = 1;
1815
1816 } else {
1817 fs.prot &= ~VM_PROT_WRITE;
1818 }
1819 }
1820
1821 /*
1822 * We must verify that the maps have not changed since our last
1823 * lookup.
1824 */
1825 if (!fs.lookup_still_valid) {
1826 rv = vm_fault_relookup(&fs);
1827 if (rv != KERN_SUCCESS) {
1828 vm_fault_deallocate(&fs);
1829 if (rv == KERN_RESTART)
1830 goto RetryFault;
1831 return (rv);
1832 }
1833 }
1834 VM_OBJECT_ASSERT_UNLOCKED(fs.object);
1835
1836 /*
1837 * If the page was filled by a pager, save the virtual address that
1838 * should be faulted on next under a sequential access pattern to the
1839 * map entry. A read lock on the map suffices to update this address
1840 * safely.
1841 */
1842 if (hardfault)
1843 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE;
1844
1845 /*
1846 * If the page to be mapped was copied from a backing object, we defer
1847 * marking it valid until here, where the fault handler is guaranteed to
1848 * succeed. Otherwise we can end up with a shadowed, mapped page in the
1849 * backing object, which violates an invariant of vm_object_collapse()
1850 * that shadowed pages are not mapped.
1851 */
1852 if (fs.m_cow != NULL) {
1853 KASSERT(vm_page_none_valid(fs.m),
1854 ("vm_fault: page %p is already valid", fs.m_cow));
1855 vm_page_valid(fs.m);
1856 }
1857
1858 /*
1859 * Page must be completely valid or it is not fit to
1860 * map into user space. vm_pager_get_pages() ensures this.
1861 */
1862 vm_page_assert_busied(fs.m);
1863 KASSERT(vm_page_all_valid(fs.m),
1864 ("vm_fault: page %p partially invalid", fs.m));
1865
1866 vm_fault_dirty(&fs, fs.m);
1867
1868 /*
1869 * Put this page into the physical map. We had to do the unlock above
1870 * because pmap_enter() may sleep. We don't put the page
1871 * back on the active queue until later so that the pageout daemon
1872 * won't find it (yet).
1873 */
1874 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot,
1875 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0);
1876 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 &&
1877 fs.wired == 0)
1878 vm_fault_prefault(&fs, vaddr,
1879 faultcount > 0 ? behind : PFBAK,
1880 faultcount > 0 ? ahead : PFFOR, false);
1881
1882 /*
1883 * If the page is not wired down, then put it where the pageout daemon
1884 * can find it.
1885 */
1886 if ((fs.fault_flags & VM_FAULT_WIRE) != 0)
1887 vm_page_wire(fs.m);
1888 else
1889 vm_page_activate(fs.m);
1890 if (fs.m_hold != NULL) {
1891 (*fs.m_hold) = fs.m;
1892 vm_page_wire(fs.m);
1893 }
1894
1895 KASSERT(fs.first_object == fs.object || vm_page_xbusied(fs.first_m),
1896 ("first_m must be xbusy"));
1897 if (vm_page_xbusied(fs.m))
1898 vm_page_xunbusy(fs.m);
1899 else
1900 vm_page_sunbusy(fs.m);
1901 fs.m = NULL;
1902
1903 /*
1904 * Unlock everything, and return
1905 */
1906 vm_fault_deallocate(&fs);
1907 if (hardfault) {
1908 VM_CNT_INC(v_io_faults);
1909 curthread->td_ru.ru_majflt++;
1910 #ifdef RACCT
1911 if (racct_enable && fs.object->type == OBJT_VNODE) {
1912 PROC_LOCK(curproc);
1913 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) {
1914 racct_add_force(curproc, RACCT_WRITEBPS,
1915 PAGE_SIZE + behind * PAGE_SIZE);
1916 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1917 } else {
1918 racct_add_force(curproc, RACCT_READBPS,
1919 PAGE_SIZE + ahead * PAGE_SIZE);
1920 racct_add_force(curproc, RACCT_READIOPS, 1);
1921 }
1922 PROC_UNLOCK(curproc);
1923 }
1924 #endif
1925 } else
1926 curthread->td_ru.ru_minflt++;
1927
1928 return (KERN_SUCCESS);
1929 }
1930
1931 /*
1932 * Speed up the reclamation of pages that precede the faulting pindex within
1933 * the first object of the shadow chain. Essentially, perform the equivalent
1934 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes
1935 * the faulting pindex by the cluster size when the pages read by vm_fault()
1936 * cross a cluster-size boundary. The cluster size is the greater of the
1937 * smallest superpage size and VM_FAULT_DONTNEED_MIN.
1938 *
1939 * When "fs->first_object" is a shadow object, the pages in the backing object
1940 * that precede the faulting pindex are deactivated by vm_fault(). So, this
1941 * function must only be concerned with pages in the first object.
1942 */
1943 static void
vm_fault_dontneed(const struct faultstate * fs,vm_offset_t vaddr,int ahead)1944 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead)
1945 {
1946 struct pctrie_iter pages;
1947 vm_map_entry_t entry;
1948 vm_object_t first_object;
1949 vm_offset_t end, start;
1950 vm_page_t m;
1951 vm_size_t size;
1952
1953 VM_OBJECT_ASSERT_UNLOCKED(fs->object);
1954 first_object = fs->first_object;
1955 /* Neither fictitious nor unmanaged pages can be reclaimed. */
1956 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) {
1957 VM_OBJECT_RLOCK(first_object);
1958 size = VM_FAULT_DONTNEED_MIN;
1959 if (MAXPAGESIZES > 1 && size < pagesizes[1])
1960 size = pagesizes[1];
1961 end = rounddown2(vaddr, size);
1962 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) &&
1963 (entry = fs->entry)->start < end) {
1964 if (end - entry->start < size)
1965 start = entry->start;
1966 else
1967 start = end - size;
1968 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED);
1969 vm_page_iter_limit_init(&pages, first_object,
1970 OFF_TO_IDX(entry->offset) +
1971 atop(end - entry->start));
1972 VM_RADIX_FOREACH_FROM(m, &pages,
1973 OFF_TO_IDX(entry->offset) +
1974 atop(start - entry->start)) {
1975 if (!vm_page_all_valid(m) ||
1976 vm_page_busied(m))
1977 continue;
1978
1979 /*
1980 * Don't clear PGA_REFERENCED, since it would
1981 * likely represent a reference by a different
1982 * process.
1983 *
1984 * Typically, at this point, prefetched pages
1985 * are still in the inactive queue. Only
1986 * pages that triggered page faults are in the
1987 * active queue. The test for whether the page
1988 * is in the inactive queue is racy; in the
1989 * worst case we will requeue the page
1990 * unnecessarily.
1991 */
1992 if (!vm_page_inactive(m))
1993 vm_page_deactivate(m);
1994 }
1995 }
1996 VM_OBJECT_RUNLOCK(first_object);
1997 }
1998 }
1999
2000 /*
2001 * vm_fault_prefault provides a quick way of clustering
2002 * pagefaults into a processes address space. It is a "cousin"
2003 * of vm_map_pmap_enter, except it runs at page fault time instead
2004 * of mmap time.
2005 */
2006 static void
vm_fault_prefault(const struct faultstate * fs,vm_offset_t addra,int backward,int forward,bool obj_locked)2007 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra,
2008 int backward, int forward, bool obj_locked)
2009 {
2010 pmap_t pmap;
2011 vm_map_entry_t entry;
2012 vm_object_t backing_object, lobject;
2013 vm_offset_t addr, starta;
2014 vm_pindex_t pindex;
2015 vm_page_t m;
2016 vm_prot_t prot;
2017 int i;
2018
2019 pmap = fs->map->pmap;
2020 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))
2021 return;
2022
2023 entry = fs->entry;
2024
2025 if (addra < backward * PAGE_SIZE) {
2026 starta = entry->start;
2027 } else {
2028 starta = addra - backward * PAGE_SIZE;
2029 if (starta < entry->start)
2030 starta = entry->start;
2031 }
2032 prot = entry->protection;
2033
2034 /*
2035 * If pmap_enter() has enabled write access on a nearby mapping, then
2036 * don't attempt promotion, because it will fail.
2037 */
2038 if ((fs->prot & VM_PROT_WRITE) != 0)
2039 prot |= VM_PROT_NO_PROMOTE;
2040
2041 /*
2042 * Generate the sequence of virtual addresses that are candidates for
2043 * prefaulting in an outward spiral from the faulting virtual address,
2044 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra
2045 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ...
2046 * If the candidate address doesn't have a backing physical page, then
2047 * the loop immediately terminates.
2048 */
2049 for (i = 0; i < 2 * imax(backward, forward); i++) {
2050 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE :
2051 PAGE_SIZE);
2052 if (addr > addra + forward * PAGE_SIZE)
2053 addr = 0;
2054
2055 if (addr < starta || addr >= entry->end)
2056 continue;
2057
2058 if (!pmap_is_prefaultable(pmap, addr))
2059 continue;
2060
2061 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2062 lobject = entry->object.vm_object;
2063 if (!obj_locked)
2064 VM_OBJECT_RLOCK(lobject);
2065 while ((m = vm_page_lookup(lobject, pindex)) == NULL &&
2066 !vm_fault_object_needs_getpages(lobject) &&
2067 (backing_object = lobject->backing_object) != NULL) {
2068 KASSERT((lobject->backing_object_offset & PAGE_MASK) ==
2069 0, ("vm_fault_prefault: unaligned object offset"));
2070 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
2071 VM_OBJECT_RLOCK(backing_object);
2072 if (!obj_locked || lobject != entry->object.vm_object)
2073 VM_OBJECT_RUNLOCK(lobject);
2074 lobject = backing_object;
2075 }
2076 if (m == NULL) {
2077 if (!obj_locked || lobject != entry->object.vm_object)
2078 VM_OBJECT_RUNLOCK(lobject);
2079 break;
2080 }
2081 if (vm_page_all_valid(m) &&
2082 (m->flags & PG_FICTITIOUS) == 0)
2083 pmap_enter_quick(pmap, addr, m, prot);
2084 if (!obj_locked || lobject != entry->object.vm_object)
2085 VM_OBJECT_RUNLOCK(lobject);
2086 }
2087 }
2088
2089 /*
2090 * Hold each of the physical pages that are mapped by the specified
2091 * range of virtual addresses, ["addr", "addr" + "len"), if those
2092 * mappings are valid and allow the specified types of access, "prot".
2093 * If all of the implied pages are successfully held, then the number
2094 * of held pages is assigned to *ppages_count, together with pointers
2095 * to those pages in the array "ma". The returned value is zero.
2096 *
2097 * However, if any of the pages cannot be held, an error is returned,
2098 * and no pages are held.
2099 * Error values:
2100 * ENOMEM - the range is not valid
2101 * EINVAL - the provided vm_page array is too small to hold all pages
2102 * EAGAIN - a page was not mapped, and the thread is in nofaulting mode
2103 * EFAULT - a page with requested permissions cannot be mapped
2104 * (more detailed result from vm_fault() is lost)
2105 */
2106 int
vm_fault_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,int * ppages_count)2107 vm_fault_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
2108 vm_prot_t prot, vm_page_t *ma, int max_count, int *ppages_count)
2109 {
2110 vm_offset_t end, va;
2111 vm_page_t *mp;
2112 int count, error;
2113 boolean_t pmap_failed;
2114
2115 if (len == 0) {
2116 *ppages_count = 0;
2117 return (0);
2118 }
2119 end = round_page(addr + len);
2120 addr = trunc_page(addr);
2121
2122 if (!vm_map_range_valid(map, addr, end))
2123 return (ENOMEM);
2124
2125 if (atop(end - addr) > max_count)
2126 return (EINVAL);
2127 count = atop(end - addr);
2128
2129 /*
2130 * Most likely, the physical pages are resident in the pmap, so it is
2131 * faster to try pmap_extract_and_hold() first.
2132 */
2133 pmap_failed = FALSE;
2134 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
2135 *mp = pmap_extract_and_hold(map->pmap, va, prot);
2136 if (*mp == NULL)
2137 pmap_failed = TRUE;
2138 else if ((prot & VM_PROT_WRITE) != 0 &&
2139 (*mp)->dirty != VM_PAGE_BITS_ALL) {
2140 /*
2141 * Explicitly dirty the physical page. Otherwise, the
2142 * caller's changes may go unnoticed because they are
2143 * performed through an unmanaged mapping or by a DMA
2144 * operation.
2145 *
2146 * The object lock is not held here.
2147 * See vm_page_clear_dirty_mask().
2148 */
2149 vm_page_dirty(*mp);
2150 }
2151 }
2152 if (pmap_failed) {
2153 /*
2154 * One or more pages could not be held by the pmap. Either no
2155 * page was mapped at the specified virtual address or that
2156 * mapping had insufficient permissions. Attempt to fault in
2157 * and hold these pages.
2158 *
2159 * If vm_fault_disable_pagefaults() was called,
2160 * i.e., TDP_NOFAULTING is set, we must not sleep nor
2161 * acquire MD VM locks, which means we must not call
2162 * vm_fault(). Some (out of tree) callers mark
2163 * too wide a code area with vm_fault_disable_pagefaults()
2164 * already, use the VM_PROT_QUICK_NOFAULT flag to request
2165 * the proper behaviour explicitly.
2166 */
2167 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 &&
2168 (curthread->td_pflags & TDP_NOFAULTING) != 0) {
2169 error = EAGAIN;
2170 goto fail;
2171 }
2172 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) {
2173 if (*mp == NULL && vm_fault(map, va, prot,
2174 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) {
2175 error = EFAULT;
2176 goto fail;
2177 }
2178 }
2179 }
2180 *ppages_count = count;
2181 return (0);
2182 fail:
2183 for (mp = ma; mp < ma + count; mp++)
2184 if (*mp != NULL)
2185 vm_page_unwire(*mp, PQ_INACTIVE);
2186 return (error);
2187 }
2188
2189 /*
2190 * Hold each of the physical pages that are mapped by the specified range of
2191 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid
2192 * and allow the specified types of access, "prot". If all of the implied
2193 * pages are successfully held, then the number of held pages is returned
2194 * together with pointers to those pages in the array "ma". However, if any
2195 * of the pages cannot be held, -1 is returned.
2196 */
2197 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)2198 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len,
2199 vm_prot_t prot, vm_page_t *ma, int max_count)
2200 {
2201 int error, pages_count;
2202
2203 error = vm_fault_hold_pages(map, addr, len, prot, ma,
2204 max_count, &pages_count);
2205 if (error != 0) {
2206 if (error == EINVAL)
2207 panic("vm_fault_quick_hold_pages: count > max_count");
2208 return (-1);
2209 }
2210 return (pages_count);
2211 }
2212
2213 /*
2214 * Routine:
2215 * vm_fault_copy_entry
2216 * Function:
2217 * Create new object backing dst_entry with private copy of all
2218 * underlying pages. When src_entry is equal to dst_entry, function
2219 * implements COW for wired-down map entry. Otherwise, it forks
2220 * wired entry into dst_map.
2221 *
2222 * In/out conditions:
2223 * The source and destination maps must be locked for write.
2224 * The source map entry must be wired down (or be a sharing map
2225 * entry corresponding to a main map entry that is wired down).
2226 */
2227 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)2228 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map __unused,
2229 vm_map_entry_t dst_entry, vm_map_entry_t src_entry,
2230 vm_ooffset_t *fork_charge)
2231 {
2232 struct pctrie_iter pages;
2233 vm_object_t backing_object, dst_object, object, src_object;
2234 vm_pindex_t dst_pindex, pindex, src_pindex;
2235 vm_prot_t access, prot;
2236 vm_offset_t vaddr;
2237 vm_page_t dst_m;
2238 vm_page_t src_m;
2239 bool upgrade;
2240
2241 upgrade = src_entry == dst_entry;
2242 KASSERT(upgrade || dst_entry->object.vm_object == NULL,
2243 ("vm_fault_copy_entry: vm_object not NULL"));
2244
2245 /*
2246 * If not an upgrade, then enter the mappings in the pmap as
2247 * read and/or execute accesses. Otherwise, enter them as
2248 * write accesses.
2249 *
2250 * A writeable large page mapping is only created if all of
2251 * the constituent small page mappings are modified. Marking
2252 * PTEs as modified on inception allows promotion to happen
2253 * without taking potentially large number of soft faults.
2254 */
2255 access = prot = dst_entry->protection;
2256 if (!upgrade)
2257 access &= ~VM_PROT_WRITE;
2258
2259 src_object = src_entry->object.vm_object;
2260 src_pindex = OFF_TO_IDX(src_entry->offset);
2261
2262 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) {
2263 dst_object = src_object;
2264 vm_object_reference(dst_object);
2265 } else {
2266 /*
2267 * Create the top-level object for the destination entry.
2268 * Doesn't actually shadow anything - we copy the pages
2269 * directly.
2270 */
2271 dst_object = vm_object_allocate_anon(atop(dst_entry->end -
2272 dst_entry->start), NULL, NULL, 0);
2273 #if VM_NRESERVLEVEL > 0
2274 dst_object->flags |= OBJ_COLORED;
2275 dst_object->pg_color = atop(dst_entry->start);
2276 #endif
2277 dst_object->domain = src_object->domain;
2278 dst_object->charge = dst_entry->end - dst_entry->start;
2279
2280 dst_entry->object.vm_object = dst_object;
2281 dst_entry->offset = 0;
2282 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC;
2283 }
2284
2285 VM_OBJECT_WLOCK(dst_object);
2286 if (fork_charge != NULL) {
2287 KASSERT(dst_entry->cred == NULL,
2288 ("vm_fault_copy_entry: leaked swp charge"));
2289 dst_object->cred = curthread->td_ucred;
2290 crhold(dst_object->cred);
2291 *fork_charge += dst_object->charge;
2292 } else if ((dst_object->flags & OBJ_SWAP) != 0 &&
2293 dst_object->cred == NULL) {
2294 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p",
2295 dst_entry));
2296 dst_object->cred = dst_entry->cred;
2297 dst_entry->cred = NULL;
2298 }
2299
2300 /*
2301 * Loop through all of the virtual pages within the entry's
2302 * range, copying each page from the source object to the
2303 * destination object. Since the source is wired, those pages
2304 * must exist. In contrast, the destination is pageable.
2305 * Since the destination object doesn't share any backing storage
2306 * with the source object, all of its pages must be dirtied,
2307 * regardless of whether they can be written.
2308 */
2309 vm_page_iter_init(&pages, dst_object);
2310 for (vaddr = dst_entry->start, dst_pindex = 0;
2311 vaddr < dst_entry->end;
2312 vaddr += PAGE_SIZE, dst_pindex++) {
2313 again:
2314 /*
2315 * Find the page in the source object, and copy it in.
2316 * Because the source is wired down, the page will be
2317 * in memory.
2318 */
2319 if (src_object != dst_object)
2320 VM_OBJECT_RLOCK(src_object);
2321 object = src_object;
2322 pindex = src_pindex + dst_pindex;
2323 while ((src_m = vm_page_lookup(object, pindex)) == NULL &&
2324 (backing_object = object->backing_object) != NULL) {
2325 /*
2326 * Unless the source mapping is read-only or
2327 * it is presently being upgraded from
2328 * read-only, the first object in the shadow
2329 * chain should provide all of the pages. In
2330 * other words, this loop body should never be
2331 * executed when the source mapping is already
2332 * read/write.
2333 */
2334 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 ||
2335 upgrade,
2336 ("vm_fault_copy_entry: main object missing page"));
2337
2338 VM_OBJECT_RLOCK(backing_object);
2339 pindex += OFF_TO_IDX(object->backing_object_offset);
2340 if (object != dst_object)
2341 VM_OBJECT_RUNLOCK(object);
2342 object = backing_object;
2343 }
2344 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing"));
2345
2346 if (object != dst_object) {
2347 /*
2348 * Allocate a page in the destination object.
2349 */
2350 pindex = (src_object == dst_object ? src_pindex : 0) +
2351 dst_pindex;
2352 dst_m = vm_page_alloc_iter(dst_object, pindex,
2353 VM_ALLOC_NORMAL, &pages);
2354 if (dst_m == NULL) {
2355 VM_OBJECT_WUNLOCK(dst_object);
2356 VM_OBJECT_RUNLOCK(object);
2357 vm_wait(dst_object);
2358 VM_OBJECT_WLOCK(dst_object);
2359 pctrie_iter_reset(&pages);
2360 goto again;
2361 }
2362
2363 /*
2364 * See the comment in vm_fault_cow().
2365 */
2366 if (src_object == dst_object &&
2367 (object->flags & OBJ_ONEMAPPING) == 0)
2368 pmap_remove_all(src_m);
2369 pmap_copy_page(src_m, dst_m);
2370
2371 /*
2372 * The object lock does not guarantee that "src_m" will
2373 * transition from invalid to valid, but it does ensure
2374 * that "src_m" will not transition from valid to
2375 * invalid.
2376 */
2377 dst_m->dirty = dst_m->valid = src_m->valid;
2378 VM_OBJECT_RUNLOCK(object);
2379 } else {
2380 dst_m = src_m;
2381 if (vm_page_busy_acquire(
2382 dst_m, VM_ALLOC_WAITFAIL) == 0) {
2383 pctrie_iter_reset(&pages);
2384 goto again;
2385 }
2386 if (dst_m->pindex >= dst_object->size) {
2387 /*
2388 * We are upgrading. Index can occur
2389 * out of bounds if the object type is
2390 * vnode and the file was truncated.
2391 */
2392 vm_page_xunbusy(dst_m);
2393 break;
2394 }
2395 }
2396
2397 /*
2398 * Enter it in the pmap. If a wired, copy-on-write
2399 * mapping is being replaced by a write-enabled
2400 * mapping, then wire that new mapping.
2401 *
2402 * The page can be invalid if the user called
2403 * msync(MS_INVALIDATE) or truncated the backing vnode
2404 * or shared memory object. In this case, do not
2405 * insert it into pmap, but still do the copy so that
2406 * all copies of the wired map entry have similar
2407 * backing pages.
2408 */
2409 if (vm_page_all_valid(dst_m)) {
2410 VM_OBJECT_WUNLOCK(dst_object);
2411 pmap_enter(dst_map->pmap, vaddr, dst_m, prot,
2412 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0);
2413 VM_OBJECT_WLOCK(dst_object);
2414 }
2415
2416 /*
2417 * Mark it no longer busy, and put it on the active list.
2418 */
2419 if (upgrade) {
2420 if (src_m != dst_m) {
2421 vm_page_unwire(src_m, PQ_INACTIVE);
2422 vm_page_wire(dst_m);
2423 } else {
2424 KASSERT(vm_page_wired(dst_m),
2425 ("dst_m %p is not wired", dst_m));
2426 }
2427 } else {
2428 vm_page_activate(dst_m);
2429 }
2430 vm_page_xunbusy(dst_m);
2431 }
2432 VM_OBJECT_WUNLOCK(dst_object);
2433 if (upgrade) {
2434 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY);
2435 vm_object_deallocate(src_object);
2436 }
2437 }
2438
2439 /*
2440 * Block entry into the machine-independent layer's page fault handler by
2441 * the calling thread. Subsequent calls to vm_fault() by that thread will
2442 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of
2443 * spurious page faults.
2444 */
2445 int
vm_fault_disable_pagefaults(void)2446 vm_fault_disable_pagefaults(void)
2447 {
2448
2449 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR));
2450 }
2451
2452 void
vm_fault_enable_pagefaults(int save)2453 vm_fault_enable_pagefaults(int save)
2454 {
2455
2456 curthread_pflags_restore(save);
2457 }
2458