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