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 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 44 * 45 * 46 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 47 * All rights reserved. 48 * 49 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 50 * 51 * Permission to use, copy, modify and distribute this software and 52 * its documentation is hereby granted, provided that both the copyright 53 * notice and this permission notice appear in all copies of the 54 * software, derivative works or modified versions, and any portions 55 * thereof, and that both notices appear in supporting documentation. 56 * 57 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 58 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 59 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 60 * 61 * Carnegie Mellon requests users of this software to return to 62 * 63 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 64 * School of Computer Science 65 * Carnegie Mellon University 66 * Pittsburgh PA 15213-3890 67 * 68 * any improvements or extensions that they make and grant Carnegie the 69 * rights to redistribute these changes. 70 */ 71 72 /* 73 * Page fault handling module. 74 */ 75 76 #include <sys/cdefs.h> 77 __FBSDID("$FreeBSD$"); 78 79 #include "opt_ktrace.h" 80 #include "opt_vm.h" 81 82 #include <sys/param.h> 83 #include <sys/systm.h> 84 #include <sys/kernel.h> 85 #include <sys/lock.h> 86 #include <sys/mman.h> 87 #include <sys/mutex.h> 88 #include <sys/pctrie.h> 89 #include <sys/proc.h> 90 #include <sys/racct.h> 91 #include <sys/refcount.h> 92 #include <sys/resourcevar.h> 93 #include <sys/rwlock.h> 94 #include <sys/signalvar.h> 95 #include <sys/sysctl.h> 96 #include <sys/sysent.h> 97 #include <sys/vmmeter.h> 98 #include <sys/vnode.h> 99 #ifdef KTRACE 100 #include <sys/ktrace.h> 101 #endif 102 103 #include <vm/vm.h> 104 #include <vm/vm_param.h> 105 #include <vm/pmap.h> 106 #include <vm/vm_map.h> 107 #include <vm/vm_object.h> 108 #include <vm/vm_page.h> 109 #include <vm/vm_pageout.h> 110 #include <vm/vm_kern.h> 111 #include <vm/vm_pager.h> 112 #include <vm/vm_extern.h> 113 #include <vm/vm_reserv.h> 114 115 #define PFBAK 4 116 #define PFFOR 4 117 118 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) 119 120 #define VM_FAULT_DONTNEED_MIN 1048576 121 122 struct faultstate { 123 /* Fault parameters. */ 124 vm_offset_t vaddr; 125 vm_page_t *m_hold; 126 vm_prot_t fault_type; 127 vm_prot_t prot; 128 int fault_flags; 129 boolean_t wired; 130 131 /* Control state. */ 132 struct timeval oom_start_time; 133 bool oom_started; 134 int nera; 135 136 /* Page reference for cow. */ 137 vm_page_t m_cow; 138 139 /* Current object. */ 140 vm_object_t object; 141 vm_pindex_t pindex; 142 vm_page_t m; 143 144 /* Top-level map object. */ 145 vm_object_t first_object; 146 vm_pindex_t first_pindex; 147 vm_page_t first_m; 148 149 /* Map state. */ 150 vm_map_t map; 151 vm_map_entry_t entry; 152 int map_generation; 153 bool lookup_still_valid; 154 155 /* Vnode if locked. */ 156 struct vnode *vp; 157 }; 158 159 /* 160 * Return codes for internal fault routines. 161 */ 162 enum fault_status { 163 FAULT_SUCCESS = 1, /* Return success to user. */ 164 FAULT_FAILURE, /* Return failure to user. */ 165 FAULT_CONTINUE, /* Continue faulting. */ 166 FAULT_RESTART, /* Restart fault. */ 167 FAULT_OUT_OF_BOUNDS, /* Invalid address for pager. */ 168 FAULT_HARD, /* Performed I/O. */ 169 FAULT_SOFT, /* Found valid page. */ 170 FAULT_PROTECTION_FAILURE, /* Invalid access. */ 171 }; 172 173 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, 174 int ahead); 175 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 176 int backward, int forward, bool obj_locked); 177 178 static int vm_pfault_oom_attempts = 3; 179 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, 180 &vm_pfault_oom_attempts, 0, 181 "Number of page allocation attempts in page fault handler before it " 182 "triggers OOM handling"); 183 184 static int vm_pfault_oom_wait = 10; 185 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, 186 &vm_pfault_oom_wait, 0, 187 "Number of seconds to wait for free pages before retrying " 188 "the page fault handler"); 189 190 static inline void 191 fault_page_release(vm_page_t *mp) 192 { 193 vm_page_t m; 194 195 m = *mp; 196 if (m != NULL) { 197 /* 198 * We are likely to loop around again and attempt to busy 199 * this page. Deactivating it leaves it available for 200 * pageout while optimizing fault restarts. 201 */ 202 vm_page_deactivate(m); 203 vm_page_xunbusy(m); 204 *mp = NULL; 205 } 206 } 207 208 static inline void 209 fault_page_free(vm_page_t *mp) 210 { 211 vm_page_t m; 212 213 m = *mp; 214 if (m != NULL) { 215 VM_OBJECT_ASSERT_WLOCKED(m->object); 216 if (!vm_page_wired(m)) 217 vm_page_free(m); 218 else 219 vm_page_xunbusy(m); 220 *mp = NULL; 221 } 222 } 223 224 /* 225 * Return true if a vm_pager_get_pages() call is needed in order to check 226 * whether the pager might have a particular page, false if it can be determined 227 * immediately that the pager can not have a copy. For swap objects, this can 228 * be checked quickly. 229 */ 230 static inline bool 231 fault_object_needs_getpages(vm_object_t object) 232 { 233 VM_OBJECT_ASSERT_LOCKED(object); 234 235 return ((object->flags & OBJ_SWAP) == 0 || 236 !pctrie_is_empty(&object->un_pager.swp.swp_blks)); 237 } 238 239 static inline void 240 unlock_map(struct faultstate *fs) 241 { 242 243 if (fs->lookup_still_valid) { 244 vm_map_lookup_done(fs->map, fs->entry); 245 fs->lookup_still_valid = false; 246 } 247 } 248 249 static void 250 unlock_vp(struct faultstate *fs) 251 { 252 253 if (fs->vp != NULL) { 254 vput(fs->vp); 255 fs->vp = NULL; 256 } 257 } 258 259 static void 260 fault_deallocate(struct faultstate *fs) 261 { 262 263 fault_page_release(&fs->m_cow); 264 fault_page_release(&fs->m); 265 vm_object_pip_wakeup(fs->object); 266 if (fs->object != fs->first_object) { 267 VM_OBJECT_WLOCK(fs->first_object); 268 fault_page_free(&fs->first_m); 269 VM_OBJECT_WUNLOCK(fs->first_object); 270 vm_object_pip_wakeup(fs->first_object); 271 } 272 vm_object_deallocate(fs->first_object); 273 unlock_map(fs); 274 unlock_vp(fs); 275 } 276 277 static void 278 unlock_and_deallocate(struct faultstate *fs) 279 { 280 281 VM_OBJECT_WUNLOCK(fs->object); 282 fault_deallocate(fs); 283 } 284 285 static void 286 vm_fault_dirty(struct faultstate *fs, vm_page_t m) 287 { 288 bool need_dirty; 289 290 if (((fs->prot & VM_PROT_WRITE) == 0 && 291 (fs->fault_flags & VM_FAULT_DIRTY) == 0) || 292 (m->oflags & VPO_UNMANAGED) != 0) 293 return; 294 295 VM_PAGE_OBJECT_BUSY_ASSERT(m); 296 297 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 && 298 (fs->fault_flags & VM_FAULT_WIRE) == 0) || 299 (fs->fault_flags & VM_FAULT_DIRTY) != 0; 300 301 vm_object_set_writeable_dirty(m->object); 302 303 /* 304 * If the fault is a write, we know that this page is being 305 * written NOW so dirty it explicitly to save on 306 * pmap_is_modified() calls later. 307 * 308 * Also, since the page is now dirty, we can possibly tell 309 * the pager to release any swap backing the page. 310 */ 311 if (need_dirty && vm_page_set_dirty(m) == 0) { 312 /* 313 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC 314 * if the page is already dirty to prevent data written with 315 * the expectation of being synced from not being synced. 316 * Likewise if this entry does not request NOSYNC then make 317 * sure the page isn't marked NOSYNC. Applications sharing 318 * data should use the same flags to avoid ping ponging. 319 */ 320 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0) 321 vm_page_aflag_set(m, PGA_NOSYNC); 322 else 323 vm_page_aflag_clear(m, PGA_NOSYNC); 324 } 325 326 } 327 328 /* 329 * Unlocks fs.first_object and fs.map on success. 330 */ 331 static enum fault_status 332 vm_fault_soft_fast(struct faultstate *fs) 333 { 334 vm_page_t m, m_map; 335 #if VM_NRESERVLEVEL > 0 336 vm_page_t m_super; 337 int flags; 338 #endif 339 int psind; 340 vm_offset_t vaddr; 341 enum fault_status res; 342 343 MPASS(fs->vp == NULL); 344 345 res = FAULT_SUCCESS; 346 vaddr = fs->vaddr; 347 vm_object_busy(fs->first_object); 348 m = vm_page_lookup(fs->first_object, fs->first_pindex); 349 /* A busy page can be mapped for read|execute access. */ 350 if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 && 351 vm_page_busied(m)) || !vm_page_all_valid(m)) { 352 res = FAULT_FAILURE; 353 goto out; 354 } 355 m_map = m; 356 psind = 0; 357 #if VM_NRESERVLEVEL > 0 358 if ((m->flags & PG_FICTITIOUS) == 0 && 359 (m_super = vm_reserv_to_superpage(m)) != NULL && 360 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && 361 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && 362 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & 363 (pagesizes[m_super->psind] - 1)) && !fs->wired && 364 pmap_ps_enabled(fs->map->pmap)) { 365 flags = PS_ALL_VALID; 366 if ((fs->prot & VM_PROT_WRITE) != 0) { 367 /* 368 * Create a superpage mapping allowing write access 369 * only if none of the constituent pages are busy and 370 * all of them are already dirty (except possibly for 371 * the page that was faulted on). 372 */ 373 flags |= PS_NONE_BUSY; 374 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) 375 flags |= PS_ALL_DIRTY; 376 } 377 if (vm_page_ps_test(m_super, flags, m)) { 378 m_map = m_super; 379 psind = m_super->psind; 380 vaddr = rounddown2(vaddr, pagesizes[psind]); 381 /* Preset the modified bit for dirty superpages. */ 382 if ((flags & PS_ALL_DIRTY) != 0) 383 fs->fault_type |= VM_PROT_WRITE; 384 } 385 } 386 #endif 387 if (pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type | 388 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind) != 389 KERN_SUCCESS) { 390 res = FAULT_FAILURE; 391 goto out; 392 } 393 if (fs->m_hold != NULL) { 394 (*fs->m_hold) = m; 395 vm_page_wire(m); 396 } 397 if (psind == 0 && !fs->wired) 398 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); 399 VM_OBJECT_RUNLOCK(fs->first_object); 400 vm_fault_dirty(fs, m); 401 vm_map_lookup_done(fs->map, fs->entry); 402 curthread->td_ru.ru_minflt++; 403 404 out: 405 vm_object_unbusy(fs->first_object); 406 return (res); 407 } 408 409 static void 410 vm_fault_restore_map_lock(struct faultstate *fs) 411 { 412 413 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 414 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0); 415 416 if (!vm_map_trylock_read(fs->map)) { 417 VM_OBJECT_WUNLOCK(fs->first_object); 418 vm_map_lock_read(fs->map); 419 VM_OBJECT_WLOCK(fs->first_object); 420 } 421 fs->lookup_still_valid = true; 422 } 423 424 static void 425 vm_fault_populate_check_page(vm_page_t m) 426 { 427 428 /* 429 * Check each page to ensure that the pager is obeying the 430 * interface: the page must be installed in the object, fully 431 * valid, and exclusively busied. 432 */ 433 MPASS(m != NULL); 434 MPASS(vm_page_all_valid(m)); 435 MPASS(vm_page_xbusied(m)); 436 } 437 438 static void 439 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, 440 vm_pindex_t last) 441 { 442 vm_page_t m; 443 vm_pindex_t pidx; 444 445 VM_OBJECT_ASSERT_WLOCKED(object); 446 MPASS(first <= last); 447 for (pidx = first, m = vm_page_lookup(object, pidx); 448 pidx <= last; pidx++, m = vm_page_next(m)) { 449 vm_fault_populate_check_page(m); 450 vm_page_deactivate(m); 451 vm_page_xunbusy(m); 452 } 453 } 454 455 static enum fault_status 456 vm_fault_populate(struct faultstate *fs) 457 { 458 vm_offset_t vaddr; 459 vm_page_t m; 460 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; 461 int bdry_idx, i, npages, psind, rv; 462 enum fault_status res; 463 464 MPASS(fs->object == fs->first_object); 465 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 466 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0); 467 MPASS(fs->first_object->backing_object == NULL); 468 MPASS(fs->lookup_still_valid); 469 470 pager_first = OFF_TO_IDX(fs->entry->offset); 471 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; 472 unlock_map(fs); 473 unlock_vp(fs); 474 475 res = FAULT_SUCCESS; 476 477 /* 478 * Call the pager (driver) populate() method. 479 * 480 * There is no guarantee that the method will be called again 481 * if the current fault is for read, and a future fault is 482 * for write. Report the entry's maximum allowed protection 483 * to the driver. 484 */ 485 rv = vm_pager_populate(fs->first_object, fs->first_pindex, 486 fs->fault_type, fs->entry->max_protection, &pager_first, 487 &pager_last); 488 489 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 490 if (rv == VM_PAGER_BAD) { 491 /* 492 * VM_PAGER_BAD is the backdoor for a pager to request 493 * normal fault handling. 494 */ 495 vm_fault_restore_map_lock(fs); 496 if (fs->map->timestamp != fs->map_generation) 497 return (FAULT_RESTART); 498 return (FAULT_CONTINUE); 499 } 500 if (rv != VM_PAGER_OK) 501 return (FAULT_FAILURE); /* AKA SIGSEGV */ 502 503 /* Ensure that the driver is obeying the interface. */ 504 MPASS(pager_first <= pager_last); 505 MPASS(fs->first_pindex <= pager_last); 506 MPASS(fs->first_pindex >= pager_first); 507 MPASS(pager_last < fs->first_object->size); 508 509 vm_fault_restore_map_lock(fs); 510 bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 511 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 512 if (fs->map->timestamp != fs->map_generation) { 513 if (bdry_idx == 0) { 514 vm_fault_populate_cleanup(fs->first_object, pager_first, 515 pager_last); 516 } else { 517 m = vm_page_lookup(fs->first_object, pager_first); 518 if (m != fs->m) 519 vm_page_xunbusy(m); 520 } 521 return (FAULT_RESTART); 522 } 523 524 /* 525 * The map is unchanged after our last unlock. Process the fault. 526 * 527 * First, the special case of largepage mappings, where 528 * populate only busies the first page in superpage run. 529 */ 530 if (bdry_idx != 0) { 531 KASSERT(PMAP_HAS_LARGEPAGES, 532 ("missing pmap support for large pages")); 533 m = vm_page_lookup(fs->first_object, pager_first); 534 vm_fault_populate_check_page(m); 535 VM_OBJECT_WUNLOCK(fs->first_object); 536 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) - 537 fs->entry->offset; 538 /* assert alignment for entry */ 539 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0, 540 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx", 541 (uintmax_t)fs->entry->start, (uintmax_t)pager_first, 542 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr)); 543 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0, 544 ("unaligned superpage m %p %#jx", m, 545 (uintmax_t)VM_PAGE_TO_PHYS(m))); 546 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, 547 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) | 548 PMAP_ENTER_LARGEPAGE, bdry_idx); 549 VM_OBJECT_WLOCK(fs->first_object); 550 vm_page_xunbusy(m); 551 if (rv != KERN_SUCCESS) { 552 res = FAULT_FAILURE; 553 goto out; 554 } 555 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) { 556 for (i = 0; i < atop(pagesizes[bdry_idx]); i++) 557 vm_page_wire(m + i); 558 } 559 if (fs->m_hold != NULL) { 560 *fs->m_hold = m + (fs->first_pindex - pager_first); 561 vm_page_wire(*fs->m_hold); 562 } 563 goto out; 564 } 565 566 /* 567 * The range [pager_first, pager_last] that is given to the 568 * pager is only a hint. The pager may populate any range 569 * within the object that includes the requested page index. 570 * In case the pager expanded the range, clip it to fit into 571 * the map entry. 572 */ 573 map_first = OFF_TO_IDX(fs->entry->offset); 574 if (map_first > pager_first) { 575 vm_fault_populate_cleanup(fs->first_object, pager_first, 576 map_first - 1); 577 pager_first = map_first; 578 } 579 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; 580 if (map_last < pager_last) { 581 vm_fault_populate_cleanup(fs->first_object, map_last + 1, 582 pager_last); 583 pager_last = map_last; 584 } 585 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); 586 pidx <= pager_last; 587 pidx += npages, m = vm_page_next(&m[npages - 1])) { 588 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; 589 590 psind = m->psind; 591 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || 592 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || 593 !pmap_ps_enabled(fs->map->pmap) || fs->wired)) 594 psind = 0; 595 596 npages = atop(pagesizes[psind]); 597 for (i = 0; i < npages; i++) { 598 vm_fault_populate_check_page(&m[i]); 599 vm_fault_dirty(fs, &m[i]); 600 } 601 VM_OBJECT_WUNLOCK(fs->first_object); 602 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type | 603 (fs->wired ? PMAP_ENTER_WIRED : 0), psind); 604 605 /* 606 * pmap_enter() may fail for a superpage mapping if additional 607 * protection policies prevent the full mapping. 608 * For example, this will happen on amd64 if the entire 609 * address range does not share the same userspace protection 610 * key. Revert to single-page mappings if this happens. 611 */ 612 MPASS(rv == KERN_SUCCESS || 613 (psind > 0 && rv == KERN_PROTECTION_FAILURE)); 614 if (__predict_false(psind > 0 && 615 rv == KERN_PROTECTION_FAILURE)) { 616 MPASS(!fs->wired); 617 for (i = 0; i < npages; i++) { 618 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), 619 &m[i], fs->prot, fs->fault_type, 0); 620 MPASS(rv == KERN_SUCCESS); 621 } 622 } 623 624 VM_OBJECT_WLOCK(fs->first_object); 625 for (i = 0; i < npages; i++) { 626 if ((fs->fault_flags & VM_FAULT_WIRE) != 0 && 627 m[i].pindex == fs->first_pindex) 628 vm_page_wire(&m[i]); 629 else 630 vm_page_activate(&m[i]); 631 if (fs->m_hold != NULL && 632 m[i].pindex == fs->first_pindex) { 633 (*fs->m_hold) = &m[i]; 634 vm_page_wire(&m[i]); 635 } 636 vm_page_xunbusy(&m[i]); 637 } 638 } 639 out: 640 curthread->td_ru.ru_majflt++; 641 return (res); 642 } 643 644 static int prot_fault_translation; 645 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN, 646 &prot_fault_translation, 0, 647 "Control signal to deliver on protection fault"); 648 649 /* compat definition to keep common code for signal translation */ 650 #define UCODE_PAGEFLT 12 651 #ifdef T_PAGEFLT 652 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT"); 653 #endif 654 655 /* 656 * vm_fault_trap: 657 * 658 * Handle a page fault occurring at the given address, 659 * requiring the given permissions, in the map specified. 660 * If successful, the page is inserted into the 661 * associated physical map. 662 * 663 * NOTE: the given address should be truncated to the 664 * proper page address. 665 * 666 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 667 * a standard error specifying why the fault is fatal is returned. 668 * 669 * The map in question must be referenced, and remains so. 670 * Caller may hold no locks. 671 */ 672 int 673 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 674 int fault_flags, int *signo, int *ucode) 675 { 676 int result; 677 678 MPASS(signo == NULL || ucode != NULL); 679 #ifdef KTRACE 680 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT)) 681 ktrfault(vaddr, fault_type); 682 #endif 683 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags, 684 NULL); 685 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE || 686 result == KERN_INVALID_ADDRESS || 687 result == KERN_RESOURCE_SHORTAGE || 688 result == KERN_PROTECTION_FAILURE || 689 result == KERN_OUT_OF_BOUNDS, 690 ("Unexpected Mach error %d from vm_fault()", result)); 691 #ifdef KTRACE 692 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND)) 693 ktrfaultend(result); 694 #endif 695 if (result != KERN_SUCCESS && signo != NULL) { 696 switch (result) { 697 case KERN_FAILURE: 698 case KERN_INVALID_ADDRESS: 699 *signo = SIGSEGV; 700 *ucode = SEGV_MAPERR; 701 break; 702 case KERN_RESOURCE_SHORTAGE: 703 *signo = SIGBUS; 704 *ucode = BUS_OOMERR; 705 break; 706 case KERN_OUT_OF_BOUNDS: 707 *signo = SIGBUS; 708 *ucode = BUS_OBJERR; 709 break; 710 case KERN_PROTECTION_FAILURE: 711 if (prot_fault_translation == 0) { 712 /* 713 * Autodetect. This check also covers 714 * the images without the ABI-tag ELF 715 * note. 716 */ 717 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD && 718 curproc->p_osrel >= P_OSREL_SIGSEGV) { 719 *signo = SIGSEGV; 720 *ucode = SEGV_ACCERR; 721 } else { 722 *signo = SIGBUS; 723 *ucode = UCODE_PAGEFLT; 724 } 725 } else if (prot_fault_translation == 1) { 726 /* Always compat mode. */ 727 *signo = SIGBUS; 728 *ucode = UCODE_PAGEFLT; 729 } else { 730 /* Always SIGSEGV mode. */ 731 *signo = SIGSEGV; 732 *ucode = SEGV_ACCERR; 733 } 734 break; 735 default: 736 KASSERT(0, ("Unexpected Mach error %d from vm_fault()", 737 result)); 738 break; 739 } 740 } 741 return (result); 742 } 743 744 static enum fault_status 745 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked) 746 { 747 struct vnode *vp; 748 int error, locked; 749 750 if (fs->object->type != OBJT_VNODE) 751 return (FAULT_CONTINUE); 752 vp = fs->object->handle; 753 if (vp == fs->vp) { 754 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked"); 755 return (FAULT_CONTINUE); 756 } 757 758 /* 759 * Perform an unlock in case the desired vnode changed while 760 * the map was unlocked during a retry. 761 */ 762 unlock_vp(fs); 763 764 locked = VOP_ISLOCKED(vp); 765 if (locked != LK_EXCLUSIVE) 766 locked = LK_SHARED; 767 768 /* 769 * We must not sleep acquiring the vnode lock while we have 770 * the page exclusive busied or the object's 771 * paging-in-progress count incremented. Otherwise, we could 772 * deadlock. 773 */ 774 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT); 775 if (error == 0) { 776 fs->vp = vp; 777 return (FAULT_CONTINUE); 778 } 779 780 vhold(vp); 781 if (objlocked) 782 unlock_and_deallocate(fs); 783 else 784 fault_deallocate(fs); 785 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE); 786 vdrop(vp); 787 fs->vp = vp; 788 KASSERT(error == 0, ("vm_fault: vget failed %d", error)); 789 return (FAULT_RESTART); 790 } 791 792 /* 793 * Calculate the desired readahead. Handle drop-behind. 794 * 795 * Returns the number of readahead blocks to pass to the pager. 796 */ 797 static int 798 vm_fault_readahead(struct faultstate *fs) 799 { 800 int era, nera; 801 u_char behavior; 802 803 KASSERT(fs->lookup_still_valid, ("map unlocked")); 804 era = fs->entry->read_ahead; 805 behavior = vm_map_entry_behavior(fs->entry); 806 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 807 nera = 0; 808 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 809 nera = VM_FAULT_READ_AHEAD_MAX; 810 if (fs->vaddr == fs->entry->next_read) 811 vm_fault_dontneed(fs, fs->vaddr, nera); 812 } else if (fs->vaddr == fs->entry->next_read) { 813 /* 814 * This is a sequential fault. Arithmetically 815 * increase the requested number of pages in 816 * the read-ahead window. The requested 817 * number of pages is "# of sequential faults 818 * x (read ahead min + 1) + read ahead min" 819 */ 820 nera = VM_FAULT_READ_AHEAD_MIN; 821 if (era > 0) { 822 nera += era + 1; 823 if (nera > VM_FAULT_READ_AHEAD_MAX) 824 nera = VM_FAULT_READ_AHEAD_MAX; 825 } 826 if (era == VM_FAULT_READ_AHEAD_MAX) 827 vm_fault_dontneed(fs, fs->vaddr, nera); 828 } else { 829 /* 830 * This is a non-sequential fault. 831 */ 832 nera = 0; 833 } 834 if (era != nera) { 835 /* 836 * A read lock on the map suffices to update 837 * the read ahead count safely. 838 */ 839 fs->entry->read_ahead = nera; 840 } 841 842 return (nera); 843 } 844 845 static int 846 vm_fault_lookup(struct faultstate *fs) 847 { 848 int result; 849 850 KASSERT(!fs->lookup_still_valid, 851 ("vm_fault_lookup: Map already locked.")); 852 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type | 853 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object, 854 &fs->first_pindex, &fs->prot, &fs->wired); 855 if (result != KERN_SUCCESS) { 856 unlock_vp(fs); 857 return (result); 858 } 859 860 fs->map_generation = fs->map->timestamp; 861 862 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) { 863 panic("%s: fault on nofault entry, addr: %#lx", 864 __func__, (u_long)fs->vaddr); 865 } 866 867 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION && 868 fs->entry->wiring_thread != curthread) { 869 vm_map_unlock_read(fs->map); 870 vm_map_lock(fs->map); 871 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) && 872 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 873 unlock_vp(fs); 874 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 875 vm_map_unlock_and_wait(fs->map, 0); 876 } else 877 vm_map_unlock(fs->map); 878 return (KERN_RESOURCE_SHORTAGE); 879 } 880 881 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0); 882 883 if (fs->wired) 884 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY); 885 else 886 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0, 887 ("!fs->wired && VM_FAULT_WIRE")); 888 fs->lookup_still_valid = true; 889 890 return (KERN_SUCCESS); 891 } 892 893 static int 894 vm_fault_relookup(struct faultstate *fs) 895 { 896 vm_object_t retry_object; 897 vm_pindex_t retry_pindex; 898 vm_prot_t retry_prot; 899 int result; 900 901 if (!vm_map_trylock_read(fs->map)) 902 return (KERN_RESTART); 903 904 fs->lookup_still_valid = true; 905 if (fs->map->timestamp == fs->map_generation) 906 return (KERN_SUCCESS); 907 908 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type, 909 &fs->entry, &retry_object, &retry_pindex, &retry_prot, 910 &fs->wired); 911 if (result != KERN_SUCCESS) { 912 /* 913 * If retry of map lookup would have blocked then 914 * retry fault from start. 915 */ 916 if (result == KERN_FAILURE) 917 return (KERN_RESTART); 918 return (result); 919 } 920 if (retry_object != fs->first_object || 921 retry_pindex != fs->first_pindex) 922 return (KERN_RESTART); 923 924 /* 925 * Check whether the protection has changed or the object has 926 * been copied while we left the map unlocked. Changing from 927 * read to write permission is OK - we leave the page 928 * write-protected, and catch the write fault. Changing from 929 * write to read permission means that we can't mark the page 930 * write-enabled after all. 931 */ 932 fs->prot &= retry_prot; 933 fs->fault_type &= retry_prot; 934 if (fs->prot == 0) 935 return (KERN_RESTART); 936 937 /* Reassert because wired may have changed. */ 938 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0, 939 ("!wired && VM_FAULT_WIRE")); 940 941 return (KERN_SUCCESS); 942 } 943 944 static void 945 vm_fault_cow(struct faultstate *fs) 946 { 947 bool is_first_object_locked; 948 949 KASSERT(fs->object != fs->first_object, 950 ("source and target COW objects are identical")); 951 952 /* 953 * This allows pages to be virtually copied from a backing_object 954 * into the first_object, where the backing object has no other 955 * refs to it, and cannot gain any more refs. Instead of a bcopy, 956 * we just move the page from the backing object to the first 957 * object. Note that we must mark the page dirty in the first 958 * object so that it will go out to swap when needed. 959 */ 960 is_first_object_locked = false; 961 if ( 962 /* 963 * Only one shadow object and no other refs. 964 */ 965 fs->object->shadow_count == 1 && fs->object->ref_count == 1 && 966 /* 967 * No other ways to look the object up 968 */ 969 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 && 970 /* 971 * We don't chase down the shadow chain and we can acquire locks. 972 */ 973 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) && 974 fs->object == fs->first_object->backing_object && 975 VM_OBJECT_TRYWLOCK(fs->object)) { 976 /* 977 * Remove but keep xbusy for replace. fs->m is moved into 978 * fs->first_object and left busy while fs->first_m is 979 * conditionally freed. 980 */ 981 vm_page_remove_xbusy(fs->m); 982 vm_page_replace(fs->m, fs->first_object, fs->first_pindex, 983 fs->first_m); 984 vm_page_dirty(fs->m); 985 #if VM_NRESERVLEVEL > 0 986 /* 987 * Rename the reservation. 988 */ 989 vm_reserv_rename(fs->m, fs->first_object, fs->object, 990 OFF_TO_IDX(fs->first_object->backing_object_offset)); 991 #endif 992 VM_OBJECT_WUNLOCK(fs->object); 993 VM_OBJECT_WUNLOCK(fs->first_object); 994 fs->first_m = fs->m; 995 fs->m = NULL; 996 VM_CNT_INC(v_cow_optim); 997 } else { 998 if (is_first_object_locked) 999 VM_OBJECT_WUNLOCK(fs->first_object); 1000 /* 1001 * Oh, well, lets copy it. 1002 */ 1003 pmap_copy_page(fs->m, fs->first_m); 1004 vm_page_valid(fs->first_m); 1005 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) { 1006 vm_page_wire(fs->first_m); 1007 vm_page_unwire(fs->m, PQ_INACTIVE); 1008 } 1009 /* 1010 * Save the cow page to be released after 1011 * pmap_enter is complete. 1012 */ 1013 fs->m_cow = fs->m; 1014 fs->m = NULL; 1015 1016 /* 1017 * Typically, the shadow object is either private to this 1018 * address space (OBJ_ONEMAPPING) or its pages are read only. 1019 * In the highly unusual case where the pages of a shadow object 1020 * are read/write shared between this and other address spaces, 1021 * we need to ensure that any pmap-level mappings to the 1022 * original, copy-on-write page from the backing object are 1023 * removed from those other address spaces. 1024 * 1025 * The flag check is racy, but this is tolerable: if 1026 * OBJ_ONEMAPPING is cleared after the check, the busy state 1027 * ensures that new mappings of m_cow can't be created. 1028 * pmap_enter() will replace an existing mapping in the current 1029 * address space. If OBJ_ONEMAPPING is set after the check, 1030 * removing mappings will at worse trigger some unnecessary page 1031 * faults. 1032 */ 1033 vm_page_assert_xbusied(fs->m_cow); 1034 if ((fs->first_object->flags & OBJ_ONEMAPPING) == 0) 1035 pmap_remove_all(fs->m_cow); 1036 } 1037 1038 vm_object_pip_wakeup(fs->object); 1039 1040 /* 1041 * Only use the new page below... 1042 */ 1043 fs->object = fs->first_object; 1044 fs->pindex = fs->first_pindex; 1045 fs->m = fs->first_m; 1046 VM_CNT_INC(v_cow_faults); 1047 curthread->td_cow++; 1048 } 1049 1050 static bool 1051 vm_fault_next(struct faultstate *fs) 1052 { 1053 vm_object_t next_object; 1054 1055 /* 1056 * The requested page does not exist at this object/ 1057 * offset. Remove the invalid page from the object, 1058 * waking up anyone waiting for it, and continue on to 1059 * the next object. However, if this is the top-level 1060 * object, we must leave the busy page in place to 1061 * prevent another process from rushing past us, and 1062 * inserting the page in that object at the same time 1063 * that we are. 1064 */ 1065 if (fs->object == fs->first_object) { 1066 fs->first_m = fs->m; 1067 fs->m = NULL; 1068 } else 1069 fault_page_free(&fs->m); 1070 1071 /* 1072 * Move on to the next object. Lock the next object before 1073 * unlocking the current one. 1074 */ 1075 VM_OBJECT_ASSERT_WLOCKED(fs->object); 1076 next_object = fs->object->backing_object; 1077 if (next_object == NULL) 1078 return (false); 1079 MPASS(fs->first_m != NULL); 1080 KASSERT(fs->object != next_object, ("object loop %p", next_object)); 1081 VM_OBJECT_WLOCK(next_object); 1082 vm_object_pip_add(next_object, 1); 1083 if (fs->object != fs->first_object) 1084 vm_object_pip_wakeup(fs->object); 1085 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1086 VM_OBJECT_WUNLOCK(fs->object); 1087 fs->object = next_object; 1088 1089 return (true); 1090 } 1091 1092 static void 1093 vm_fault_zerofill(struct faultstate *fs) 1094 { 1095 1096 /* 1097 * If there's no object left, fill the page in the top 1098 * object with zeros. 1099 */ 1100 if (fs->object != fs->first_object) { 1101 vm_object_pip_wakeup(fs->object); 1102 fs->object = fs->first_object; 1103 fs->pindex = fs->first_pindex; 1104 } 1105 MPASS(fs->first_m != NULL); 1106 MPASS(fs->m == NULL); 1107 fs->m = fs->first_m; 1108 fs->first_m = NULL; 1109 1110 /* 1111 * Zero the page if necessary and mark it valid. 1112 */ 1113 if ((fs->m->flags & PG_ZERO) == 0) { 1114 pmap_zero_page(fs->m); 1115 } else { 1116 VM_CNT_INC(v_ozfod); 1117 } 1118 VM_CNT_INC(v_zfod); 1119 vm_page_valid(fs->m); 1120 } 1121 1122 /* 1123 * Initiate page fault after timeout. Returns true if caller should 1124 * do vm_waitpfault() after the call. 1125 */ 1126 static bool 1127 vm_fault_allocate_oom(struct faultstate *fs) 1128 { 1129 struct timeval now; 1130 1131 unlock_and_deallocate(fs); 1132 if (vm_pfault_oom_attempts < 0) 1133 return (true); 1134 if (!fs->oom_started) { 1135 fs->oom_started = true; 1136 getmicrotime(&fs->oom_start_time); 1137 return (true); 1138 } 1139 1140 getmicrotime(&now); 1141 timevalsub(&now, &fs->oom_start_time); 1142 if (now.tv_sec < vm_pfault_oom_attempts * vm_pfault_oom_wait) 1143 return (true); 1144 1145 if (bootverbose) 1146 printf( 1147 "proc %d (%s) failed to alloc page on fault, starting OOM\n", 1148 curproc->p_pid, curproc->p_comm); 1149 vm_pageout_oom(VM_OOM_MEM_PF); 1150 fs->oom_started = false; 1151 return (false); 1152 } 1153 1154 /* 1155 * Allocate a page directly or via the object populate method. 1156 */ 1157 static enum fault_status 1158 vm_fault_allocate(struct faultstate *fs) 1159 { 1160 struct domainset *dset; 1161 enum fault_status res; 1162 1163 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) { 1164 res = vm_fault_lock_vnode(fs, true); 1165 MPASS(res == FAULT_CONTINUE || res == FAULT_RESTART); 1166 if (res == FAULT_RESTART) 1167 return (res); 1168 } 1169 1170 if (fs->pindex >= fs->object->size) { 1171 unlock_and_deallocate(fs); 1172 return (FAULT_OUT_OF_BOUNDS); 1173 } 1174 1175 if (fs->object == fs->first_object && 1176 (fs->first_object->flags & OBJ_POPULATE) != 0 && 1177 fs->first_object->shadow_count == 0) { 1178 res = vm_fault_populate(fs); 1179 switch (res) { 1180 case FAULT_SUCCESS: 1181 case FAULT_FAILURE: 1182 case FAULT_RESTART: 1183 unlock_and_deallocate(fs); 1184 return (res); 1185 case FAULT_CONTINUE: 1186 /* 1187 * Pager's populate() method 1188 * returned VM_PAGER_BAD. 1189 */ 1190 break; 1191 default: 1192 panic("inconsistent return codes"); 1193 } 1194 } 1195 1196 /* 1197 * Allocate a new page for this object/offset pair. 1198 * 1199 * If the process has a fatal signal pending, prioritize the allocation 1200 * with the expectation that the process will exit shortly and free some 1201 * pages. In particular, the signal may have been posted by the page 1202 * daemon in an attempt to resolve an out-of-memory condition. 1203 * 1204 * The unlocked read of the p_flag is harmless. At worst, the P_KILLED 1205 * might be not observed here, and allocation fails, causing a restart 1206 * and new reading of the p_flag. 1207 */ 1208 dset = fs->object->domain.dr_policy; 1209 if (dset == NULL) 1210 dset = curthread->td_domain.dr_policy; 1211 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) { 1212 #if VM_NRESERVLEVEL > 0 1213 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex); 1214 #endif 1215 if (!vm_pager_can_alloc_page(fs->object, fs->pindex)) { 1216 unlock_and_deallocate(fs); 1217 return (FAULT_FAILURE); 1218 } 1219 fs->m = vm_page_alloc(fs->object, fs->pindex, 1220 P_KILLED(curproc) ? VM_ALLOC_SYSTEM : 0); 1221 } 1222 if (fs->m == NULL) { 1223 if (vm_fault_allocate_oom(fs)) 1224 vm_waitpfault(dset, vm_pfault_oom_wait * hz); 1225 return (FAULT_RESTART); 1226 } 1227 fs->oom_started = false; 1228 1229 return (FAULT_CONTINUE); 1230 } 1231 1232 /* 1233 * Call the pager to retrieve the page if there is a chance 1234 * that the pager has it, and potentially retrieve additional 1235 * pages at the same time. 1236 */ 1237 static enum fault_status 1238 vm_fault_getpages(struct faultstate *fs, int *behindp, int *aheadp) 1239 { 1240 vm_offset_t e_end, e_start; 1241 int ahead, behind, cluster_offset, rv; 1242 enum fault_status status; 1243 u_char behavior; 1244 1245 /* 1246 * Prepare for unlocking the map. Save the map 1247 * entry's start and end addresses, which are used to 1248 * optimize the size of the pager operation below. 1249 * Even if the map entry's addresses change after 1250 * unlocking the map, using the saved addresses is 1251 * safe. 1252 */ 1253 e_start = fs->entry->start; 1254 e_end = fs->entry->end; 1255 behavior = vm_map_entry_behavior(fs->entry); 1256 1257 /* 1258 * If the pager for the current object might have 1259 * the page, then determine the number of additional 1260 * pages to read and potentially reprioritize 1261 * previously read pages for earlier reclamation. 1262 * These operations should only be performed once per 1263 * page fault. Even if the current pager doesn't 1264 * have the page, the number of additional pages to 1265 * read will apply to subsequent objects in the 1266 * shadow chain. 1267 */ 1268 if (fs->nera == -1 && !P_KILLED(curproc)) 1269 fs->nera = vm_fault_readahead(fs); 1270 1271 /* 1272 * Release the map lock before locking the vnode or 1273 * sleeping in the pager. (If the current object has 1274 * a shadow, then an earlier iteration of this loop 1275 * may have already unlocked the map.) 1276 */ 1277 unlock_map(fs); 1278 1279 status = vm_fault_lock_vnode(fs, false); 1280 MPASS(status == FAULT_CONTINUE || status == FAULT_RESTART); 1281 if (status == FAULT_RESTART) 1282 return (status); 1283 KASSERT(fs->vp == NULL || !fs->map->system_map, 1284 ("vm_fault: vnode-backed object mapped by system map")); 1285 1286 /* 1287 * Page in the requested page and hint the pager, 1288 * that it may bring up surrounding pages. 1289 */ 1290 if (fs->nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || 1291 P_KILLED(curproc)) { 1292 behind = 0; 1293 ahead = 0; 1294 } else { 1295 /* Is this a sequential fault? */ 1296 if (fs->nera > 0) { 1297 behind = 0; 1298 ahead = fs->nera; 1299 } else { 1300 /* 1301 * Request a cluster of pages that is 1302 * aligned to a VM_FAULT_READ_DEFAULT 1303 * page offset boundary within the 1304 * object. Alignment to a page offset 1305 * boundary is more likely to coincide 1306 * with the underlying file system 1307 * block than alignment to a virtual 1308 * address boundary. 1309 */ 1310 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT; 1311 behind = ulmin(cluster_offset, 1312 atop(fs->vaddr - e_start)); 1313 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset; 1314 } 1315 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1); 1316 } 1317 *behindp = behind; 1318 *aheadp = ahead; 1319 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp); 1320 if (rv == VM_PAGER_OK) 1321 return (FAULT_HARD); 1322 if (rv == VM_PAGER_ERROR) 1323 printf("vm_fault: pager read error, pid %d (%s)\n", 1324 curproc->p_pid, curproc->p_comm); 1325 /* 1326 * If an I/O error occurred or the requested page was 1327 * outside the range of the pager, clean up and return 1328 * an error. 1329 */ 1330 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { 1331 VM_OBJECT_WLOCK(fs->object); 1332 fault_page_free(&fs->m); 1333 unlock_and_deallocate(fs); 1334 return (FAULT_OUT_OF_BOUNDS); 1335 } 1336 KASSERT(rv == VM_PAGER_FAIL, 1337 ("%s: unexpected pager error %d", __func__, rv)); 1338 return (FAULT_CONTINUE); 1339 } 1340 1341 /* 1342 * Wait/Retry if the page is busy. We have to do this if the page is 1343 * either exclusive or shared busy because the vm_pager may be using 1344 * read busy for pageouts (and even pageins if it is the vnode pager), 1345 * and we could end up trying to pagein and pageout the same page 1346 * simultaneously. 1347 * 1348 * We can theoretically allow the busy case on a read fault if the page 1349 * is marked valid, but since such pages are typically already pmap'd, 1350 * putting that special case in might be more effort then it is worth. 1351 * We cannot under any circumstances mess around with a shared busied 1352 * page except, perhaps, to pmap it. 1353 */ 1354 static void 1355 vm_fault_busy_sleep(struct faultstate *fs) 1356 { 1357 /* 1358 * Reference the page before unlocking and 1359 * sleeping so that the page daemon is less 1360 * likely to reclaim it. 1361 */ 1362 vm_page_aflag_set(fs->m, PGA_REFERENCED); 1363 if (fs->object != fs->first_object) { 1364 fault_page_release(&fs->first_m); 1365 vm_object_pip_wakeup(fs->first_object); 1366 } 1367 vm_object_pip_wakeup(fs->object); 1368 unlock_map(fs); 1369 if (fs->m != vm_page_lookup(fs->object, fs->pindex) || 1370 !vm_page_busy_sleep(fs->m, "vmpfw", 0)) 1371 VM_OBJECT_WUNLOCK(fs->object); 1372 VM_CNT_INC(v_intrans); 1373 vm_object_deallocate(fs->first_object); 1374 } 1375 1376 /* 1377 * Handle page lookup, populate, allocate, page-in for the current 1378 * object. 1379 * 1380 * The object is locked on entry and will remain locked with a return 1381 * code of FAULT_CONTINUE so that fault may follow the shadow chain. 1382 * Otherwise, the object will be unlocked upon return. 1383 */ 1384 static enum fault_status 1385 vm_fault_object(struct faultstate *fs, int *behindp, int *aheadp) 1386 { 1387 enum fault_status res; 1388 bool dead; 1389 1390 /* 1391 * If the object is marked for imminent termination, we retry 1392 * here, since the collapse pass has raced with us. Otherwise, 1393 * if we see terminally dead object, return fail. 1394 */ 1395 if ((fs->object->flags & OBJ_DEAD) != 0) { 1396 dead = fs->object->type == OBJT_DEAD; 1397 unlock_and_deallocate(fs); 1398 if (dead) 1399 return (FAULT_PROTECTION_FAILURE); 1400 pause("vmf_de", 1); 1401 return (FAULT_RESTART); 1402 } 1403 1404 /* 1405 * See if the page is resident. 1406 */ 1407 fs->m = vm_page_lookup(fs->object, fs->pindex); 1408 if (fs->m != NULL) { 1409 if (!vm_page_tryxbusy(fs->m)) { 1410 vm_fault_busy_sleep(fs); 1411 return (FAULT_RESTART); 1412 } 1413 1414 /* 1415 * The page is marked busy for other processes and the 1416 * pagedaemon. If it is still completely valid we are 1417 * done. 1418 */ 1419 if (vm_page_all_valid(fs->m)) { 1420 VM_OBJECT_WUNLOCK(fs->object); 1421 return (FAULT_SOFT); 1422 } 1423 } 1424 VM_OBJECT_ASSERT_WLOCKED(fs->object); 1425 1426 /* 1427 * Page is not resident. If the pager might contain the page 1428 * or this is the beginning of the search, allocate a new 1429 * page. 1430 */ 1431 if (fs->m == NULL && (fault_object_needs_getpages(fs->object) || 1432 fs->object == fs->first_object)) { 1433 res = vm_fault_allocate(fs); 1434 if (res != FAULT_CONTINUE) 1435 return (res); 1436 } 1437 1438 /* 1439 * Default objects have no pager so no exclusive busy exists 1440 * to protect this page in the chain. Skip to the next 1441 * object without dropping the lock to preserve atomicity of 1442 * shadow faults. 1443 */ 1444 if (fault_object_needs_getpages(fs->object)) { 1445 /* 1446 * At this point, we have either allocated a new page 1447 * or found an existing page that is only partially 1448 * valid. 1449 * 1450 * We hold a reference on the current object and the 1451 * page is exclusive busied. The exclusive busy 1452 * prevents simultaneous faults and collapses while 1453 * the object lock is dropped. 1454 */ 1455 VM_OBJECT_WUNLOCK(fs->object); 1456 res = vm_fault_getpages(fs, behindp, aheadp); 1457 if (res == FAULT_CONTINUE) 1458 VM_OBJECT_WLOCK(fs->object); 1459 } else { 1460 res = FAULT_CONTINUE; 1461 } 1462 return (res); 1463 } 1464 1465 int 1466 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 1467 int fault_flags, vm_page_t *m_hold) 1468 { 1469 struct faultstate fs; 1470 int ahead, behind, faultcount, rv; 1471 enum fault_status res; 1472 bool hardfault; 1473 1474 VM_CNT_INC(v_vm_faults); 1475 1476 if ((curthread->td_pflags & TDP_NOFAULTING) != 0) 1477 return (KERN_PROTECTION_FAILURE); 1478 1479 fs.vp = NULL; 1480 fs.vaddr = vaddr; 1481 fs.m_hold = m_hold; 1482 fs.fault_flags = fault_flags; 1483 fs.map = map; 1484 fs.lookup_still_valid = false; 1485 fs.oom_started = false; 1486 fs.nera = -1; 1487 faultcount = 0; 1488 hardfault = false; 1489 1490 RetryFault: 1491 fs.fault_type = fault_type; 1492 1493 /* 1494 * Find the backing store object and offset into it to begin the 1495 * search. 1496 */ 1497 rv = vm_fault_lookup(&fs); 1498 if (rv != KERN_SUCCESS) { 1499 if (rv == KERN_RESOURCE_SHORTAGE) 1500 goto RetryFault; 1501 return (rv); 1502 } 1503 1504 /* 1505 * Try to avoid lock contention on the top-level object through 1506 * special-case handling of some types of page faults, specifically, 1507 * those that are mapping an existing page from the top-level object. 1508 * Under this condition, a read lock on the object suffices, allowing 1509 * multiple page faults of a similar type to run in parallel. 1510 */ 1511 if (fs.vp == NULL /* avoid locked vnode leak */ && 1512 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 && 1513 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) { 1514 VM_OBJECT_RLOCK(fs.first_object); 1515 res = vm_fault_soft_fast(&fs); 1516 if (res == FAULT_SUCCESS) 1517 return (KERN_SUCCESS); 1518 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { 1519 VM_OBJECT_RUNLOCK(fs.first_object); 1520 VM_OBJECT_WLOCK(fs.first_object); 1521 } 1522 } else { 1523 VM_OBJECT_WLOCK(fs.first_object); 1524 } 1525 1526 /* 1527 * Make a reference to this object to prevent its disposal while we 1528 * are messing with it. Once we have the reference, the map is free 1529 * to be diddled. Since objects reference their shadows (and copies), 1530 * they will stay around as well. 1531 * 1532 * Bump the paging-in-progress count to prevent size changes (e.g. 1533 * truncation operations) during I/O. 1534 */ 1535 vm_object_reference_locked(fs.first_object); 1536 vm_object_pip_add(fs.first_object, 1); 1537 1538 fs.m_cow = fs.m = fs.first_m = NULL; 1539 1540 /* 1541 * Search for the page at object/offset. 1542 */ 1543 fs.object = fs.first_object; 1544 fs.pindex = fs.first_pindex; 1545 1546 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) { 1547 res = vm_fault_allocate(&fs); 1548 switch (res) { 1549 case FAULT_RESTART: 1550 goto RetryFault; 1551 case FAULT_SUCCESS: 1552 return (KERN_SUCCESS); 1553 case FAULT_FAILURE: 1554 return (KERN_FAILURE); 1555 case FAULT_OUT_OF_BOUNDS: 1556 return (KERN_OUT_OF_BOUNDS); 1557 case FAULT_CONTINUE: 1558 break; 1559 default: 1560 panic("vm_fault: Unhandled status %d", res); 1561 } 1562 } 1563 1564 while (TRUE) { 1565 KASSERT(fs.m == NULL, 1566 ("page still set %p at loop start", fs.m)); 1567 1568 res = vm_fault_object(&fs, &behind, &ahead); 1569 switch (res) { 1570 case FAULT_SOFT: 1571 goto found; 1572 case FAULT_HARD: 1573 faultcount = behind + 1 + ahead; 1574 hardfault = true; 1575 goto found; 1576 case FAULT_RESTART: 1577 goto RetryFault; 1578 case FAULT_SUCCESS: 1579 return (KERN_SUCCESS); 1580 case FAULT_FAILURE: 1581 return (KERN_FAILURE); 1582 case FAULT_OUT_OF_BOUNDS: 1583 return (KERN_OUT_OF_BOUNDS); 1584 case FAULT_PROTECTION_FAILURE: 1585 return (KERN_PROTECTION_FAILURE); 1586 case FAULT_CONTINUE: 1587 break; 1588 default: 1589 panic("vm_fault: Unhandled status %d", res); 1590 } 1591 1592 /* 1593 * The page was not found in the current object. Try to 1594 * traverse into a backing object or zero fill if none is 1595 * found. 1596 */ 1597 if (vm_fault_next(&fs)) 1598 continue; 1599 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) { 1600 if (fs.first_object == fs.object) 1601 fault_page_free(&fs.first_m); 1602 unlock_and_deallocate(&fs); 1603 return (KERN_OUT_OF_BOUNDS); 1604 } 1605 VM_OBJECT_WUNLOCK(fs.object); 1606 vm_fault_zerofill(&fs); 1607 /* Don't try to prefault neighboring pages. */ 1608 faultcount = 1; 1609 break; 1610 } 1611 1612 found: 1613 /* 1614 * A valid page has been found and exclusively busied. The 1615 * object lock must no longer be held. 1616 */ 1617 vm_page_assert_xbusied(fs.m); 1618 VM_OBJECT_ASSERT_UNLOCKED(fs.object); 1619 1620 /* 1621 * If the page is being written, but isn't already owned by the 1622 * top-level object, we have to copy it into a new page owned by the 1623 * top-level object. 1624 */ 1625 if (fs.object != fs.first_object) { 1626 /* 1627 * We only really need to copy if we want to write it. 1628 */ 1629 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1630 vm_fault_cow(&fs); 1631 /* 1632 * We only try to prefault read-only mappings to the 1633 * neighboring pages when this copy-on-write fault is 1634 * a hard fault. In other cases, trying to prefault 1635 * is typically wasted effort. 1636 */ 1637 if (faultcount == 0) 1638 faultcount = 1; 1639 1640 } else { 1641 fs.prot &= ~VM_PROT_WRITE; 1642 } 1643 } 1644 1645 /* 1646 * We must verify that the maps have not changed since our last 1647 * lookup. 1648 */ 1649 if (!fs.lookup_still_valid) { 1650 rv = vm_fault_relookup(&fs); 1651 if (rv != KERN_SUCCESS) { 1652 fault_deallocate(&fs); 1653 if (rv == KERN_RESTART) 1654 goto RetryFault; 1655 return (rv); 1656 } 1657 } 1658 VM_OBJECT_ASSERT_UNLOCKED(fs.object); 1659 1660 /* 1661 * If the page was filled by a pager, save the virtual address that 1662 * should be faulted on next under a sequential access pattern to the 1663 * map entry. A read lock on the map suffices to update this address 1664 * safely. 1665 */ 1666 if (hardfault) 1667 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; 1668 1669 /* 1670 * Page must be completely valid or it is not fit to 1671 * map into user space. vm_pager_get_pages() ensures this. 1672 */ 1673 vm_page_assert_xbusied(fs.m); 1674 KASSERT(vm_page_all_valid(fs.m), 1675 ("vm_fault: page %p partially invalid", fs.m)); 1676 1677 vm_fault_dirty(&fs, fs.m); 1678 1679 /* 1680 * Put this page into the physical map. We had to do the unlock above 1681 * because pmap_enter() may sleep. We don't put the page 1682 * back on the active queue until later so that the pageout daemon 1683 * won't find it (yet). 1684 */ 1685 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, 1686 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0); 1687 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 && 1688 fs.wired == 0) 1689 vm_fault_prefault(&fs, vaddr, 1690 faultcount > 0 ? behind : PFBAK, 1691 faultcount > 0 ? ahead : PFFOR, false); 1692 1693 /* 1694 * If the page is not wired down, then put it where the pageout daemon 1695 * can find it. 1696 */ 1697 if ((fs.fault_flags & VM_FAULT_WIRE) != 0) 1698 vm_page_wire(fs.m); 1699 else 1700 vm_page_activate(fs.m); 1701 if (fs.m_hold != NULL) { 1702 (*fs.m_hold) = fs.m; 1703 vm_page_wire(fs.m); 1704 } 1705 vm_page_xunbusy(fs.m); 1706 fs.m = NULL; 1707 1708 /* 1709 * Unlock everything, and return 1710 */ 1711 fault_deallocate(&fs); 1712 if (hardfault) { 1713 VM_CNT_INC(v_io_faults); 1714 curthread->td_ru.ru_majflt++; 1715 #ifdef RACCT 1716 if (racct_enable && fs.object->type == OBJT_VNODE) { 1717 PROC_LOCK(curproc); 1718 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1719 racct_add_force(curproc, RACCT_WRITEBPS, 1720 PAGE_SIZE + behind * PAGE_SIZE); 1721 racct_add_force(curproc, RACCT_WRITEIOPS, 1); 1722 } else { 1723 racct_add_force(curproc, RACCT_READBPS, 1724 PAGE_SIZE + ahead * PAGE_SIZE); 1725 racct_add_force(curproc, RACCT_READIOPS, 1); 1726 } 1727 PROC_UNLOCK(curproc); 1728 } 1729 #endif 1730 } else 1731 curthread->td_ru.ru_minflt++; 1732 1733 return (KERN_SUCCESS); 1734 } 1735 1736 /* 1737 * Speed up the reclamation of pages that precede the faulting pindex within 1738 * the first object of the shadow chain. Essentially, perform the equivalent 1739 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes 1740 * the faulting pindex by the cluster size when the pages read by vm_fault() 1741 * cross a cluster-size boundary. The cluster size is the greater of the 1742 * smallest superpage size and VM_FAULT_DONTNEED_MIN. 1743 * 1744 * When "fs->first_object" is a shadow object, the pages in the backing object 1745 * that precede the faulting pindex are deactivated by vm_fault(). So, this 1746 * function must only be concerned with pages in the first object. 1747 */ 1748 static void 1749 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) 1750 { 1751 vm_map_entry_t entry; 1752 vm_object_t first_object; 1753 vm_offset_t end, start; 1754 vm_page_t m, m_next; 1755 vm_pindex_t pend, pstart; 1756 vm_size_t size; 1757 1758 VM_OBJECT_ASSERT_UNLOCKED(fs->object); 1759 first_object = fs->first_object; 1760 /* Neither fictitious nor unmanaged pages can be reclaimed. */ 1761 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 1762 VM_OBJECT_RLOCK(first_object); 1763 size = VM_FAULT_DONTNEED_MIN; 1764 if (MAXPAGESIZES > 1 && size < pagesizes[1]) 1765 size = pagesizes[1]; 1766 end = rounddown2(vaddr, size); 1767 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && 1768 (entry = fs->entry)->start < end) { 1769 if (end - entry->start < size) 1770 start = entry->start; 1771 else 1772 start = end - size; 1773 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); 1774 pstart = OFF_TO_IDX(entry->offset) + atop(start - 1775 entry->start); 1776 m_next = vm_page_find_least(first_object, pstart); 1777 pend = OFF_TO_IDX(entry->offset) + atop(end - 1778 entry->start); 1779 while ((m = m_next) != NULL && m->pindex < pend) { 1780 m_next = TAILQ_NEXT(m, listq); 1781 if (!vm_page_all_valid(m) || 1782 vm_page_busied(m)) 1783 continue; 1784 1785 /* 1786 * Don't clear PGA_REFERENCED, since it would 1787 * likely represent a reference by a different 1788 * process. 1789 * 1790 * Typically, at this point, prefetched pages 1791 * are still in the inactive queue. Only 1792 * pages that triggered page faults are in the 1793 * active queue. The test for whether the page 1794 * is in the inactive queue is racy; in the 1795 * worst case we will requeue the page 1796 * unnecessarily. 1797 */ 1798 if (!vm_page_inactive(m)) 1799 vm_page_deactivate(m); 1800 } 1801 } 1802 VM_OBJECT_RUNLOCK(first_object); 1803 } 1804 } 1805 1806 /* 1807 * vm_fault_prefault provides a quick way of clustering 1808 * pagefaults into a processes address space. It is a "cousin" 1809 * of vm_map_pmap_enter, except it runs at page fault time instead 1810 * of mmap time. 1811 */ 1812 static void 1813 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 1814 int backward, int forward, bool obj_locked) 1815 { 1816 pmap_t pmap; 1817 vm_map_entry_t entry; 1818 vm_object_t backing_object, lobject; 1819 vm_offset_t addr, starta; 1820 vm_pindex_t pindex; 1821 vm_page_t m; 1822 int i; 1823 1824 pmap = fs->map->pmap; 1825 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1826 return; 1827 1828 entry = fs->entry; 1829 1830 if (addra < backward * PAGE_SIZE) { 1831 starta = entry->start; 1832 } else { 1833 starta = addra - backward * PAGE_SIZE; 1834 if (starta < entry->start) 1835 starta = entry->start; 1836 } 1837 1838 /* 1839 * Generate the sequence of virtual addresses that are candidates for 1840 * prefaulting in an outward spiral from the faulting virtual address, 1841 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra 1842 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... 1843 * If the candidate address doesn't have a backing physical page, then 1844 * the loop immediately terminates. 1845 */ 1846 for (i = 0; i < 2 * imax(backward, forward); i++) { 1847 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : 1848 PAGE_SIZE); 1849 if (addr > addra + forward * PAGE_SIZE) 1850 addr = 0; 1851 1852 if (addr < starta || addr >= entry->end) 1853 continue; 1854 1855 if (!pmap_is_prefaultable(pmap, addr)) 1856 continue; 1857 1858 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1859 lobject = entry->object.vm_object; 1860 if (!obj_locked) 1861 VM_OBJECT_RLOCK(lobject); 1862 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1863 !fault_object_needs_getpages(lobject) && 1864 (backing_object = lobject->backing_object) != NULL) { 1865 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1866 0, ("vm_fault_prefault: unaligned object offset")); 1867 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1868 VM_OBJECT_RLOCK(backing_object); 1869 if (!obj_locked || lobject != entry->object.vm_object) 1870 VM_OBJECT_RUNLOCK(lobject); 1871 lobject = backing_object; 1872 } 1873 if (m == NULL) { 1874 if (!obj_locked || lobject != entry->object.vm_object) 1875 VM_OBJECT_RUNLOCK(lobject); 1876 break; 1877 } 1878 if (vm_page_all_valid(m) && 1879 (m->flags & PG_FICTITIOUS) == 0) 1880 pmap_enter_quick(pmap, addr, m, entry->protection); 1881 if (!obj_locked || lobject != entry->object.vm_object) 1882 VM_OBJECT_RUNLOCK(lobject); 1883 } 1884 } 1885 1886 /* 1887 * Hold each of the physical pages that are mapped by the specified range of 1888 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1889 * and allow the specified types of access, "prot". If all of the implied 1890 * pages are successfully held, then the number of held pages is returned 1891 * together with pointers to those pages in the array "ma". However, if any 1892 * of the pages cannot be held, -1 is returned. 1893 */ 1894 int 1895 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1896 vm_prot_t prot, vm_page_t *ma, int max_count) 1897 { 1898 vm_offset_t end, va; 1899 vm_page_t *mp; 1900 int count; 1901 boolean_t pmap_failed; 1902 1903 if (len == 0) 1904 return (0); 1905 end = round_page(addr + len); 1906 addr = trunc_page(addr); 1907 1908 if (!vm_map_range_valid(map, addr, end)) 1909 return (-1); 1910 1911 if (atop(end - addr) > max_count) 1912 panic("vm_fault_quick_hold_pages: count > max_count"); 1913 count = atop(end - addr); 1914 1915 /* 1916 * Most likely, the physical pages are resident in the pmap, so it is 1917 * faster to try pmap_extract_and_hold() first. 1918 */ 1919 pmap_failed = FALSE; 1920 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1921 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1922 if (*mp == NULL) 1923 pmap_failed = TRUE; 1924 else if ((prot & VM_PROT_WRITE) != 0 && 1925 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1926 /* 1927 * Explicitly dirty the physical page. Otherwise, the 1928 * caller's changes may go unnoticed because they are 1929 * performed through an unmanaged mapping or by a DMA 1930 * operation. 1931 * 1932 * The object lock is not held here. 1933 * See vm_page_clear_dirty_mask(). 1934 */ 1935 vm_page_dirty(*mp); 1936 } 1937 } 1938 if (pmap_failed) { 1939 /* 1940 * One or more pages could not be held by the pmap. Either no 1941 * page was mapped at the specified virtual address or that 1942 * mapping had insufficient permissions. Attempt to fault in 1943 * and hold these pages. 1944 * 1945 * If vm_fault_disable_pagefaults() was called, 1946 * i.e., TDP_NOFAULTING is set, we must not sleep nor 1947 * acquire MD VM locks, which means we must not call 1948 * vm_fault(). Some (out of tree) callers mark 1949 * too wide a code area with vm_fault_disable_pagefaults() 1950 * already, use the VM_PROT_QUICK_NOFAULT flag to request 1951 * the proper behaviour explicitly. 1952 */ 1953 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && 1954 (curthread->td_pflags & TDP_NOFAULTING) != 0) 1955 goto error; 1956 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1957 if (*mp == NULL && vm_fault(map, va, prot, 1958 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1959 goto error; 1960 } 1961 return (count); 1962 error: 1963 for (mp = ma; mp < ma + count; mp++) 1964 if (*mp != NULL) 1965 vm_page_unwire(*mp, PQ_INACTIVE); 1966 return (-1); 1967 } 1968 1969 /* 1970 * Routine: 1971 * vm_fault_copy_entry 1972 * Function: 1973 * Create new object backing dst_entry with private copy of all 1974 * underlying pages. When src_entry is equal to dst_entry, function 1975 * implements COW for wired-down map entry. Otherwise, it forks 1976 * wired entry into dst_map. 1977 * 1978 * In/out conditions: 1979 * The source and destination maps must be locked for write. 1980 * The source map entry must be wired down (or be a sharing map 1981 * entry corresponding to a main map entry that is wired down). 1982 */ 1983 void 1984 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map __unused, 1985 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1986 vm_ooffset_t *fork_charge) 1987 { 1988 vm_object_t backing_object, dst_object, object, src_object; 1989 vm_pindex_t dst_pindex, pindex, src_pindex; 1990 vm_prot_t access, prot; 1991 vm_offset_t vaddr; 1992 vm_page_t dst_m; 1993 vm_page_t src_m; 1994 bool upgrade; 1995 1996 upgrade = src_entry == dst_entry; 1997 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1998 ("vm_fault_copy_entry: vm_object not NULL")); 1999 2000 /* 2001 * If not an upgrade, then enter the mappings in the pmap as 2002 * read and/or execute accesses. Otherwise, enter them as 2003 * write accesses. 2004 * 2005 * A writeable large page mapping is only created if all of 2006 * the constituent small page mappings are modified. Marking 2007 * PTEs as modified on inception allows promotion to happen 2008 * without taking potentially large number of soft faults. 2009 */ 2010 access = prot = dst_entry->protection; 2011 if (!upgrade) 2012 access &= ~VM_PROT_WRITE; 2013 2014 src_object = src_entry->object.vm_object; 2015 src_pindex = OFF_TO_IDX(src_entry->offset); 2016 2017 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 2018 dst_object = src_object; 2019 vm_object_reference(dst_object); 2020 } else { 2021 /* 2022 * Create the top-level object for the destination entry. 2023 * Doesn't actually shadow anything - we copy the pages 2024 * directly. 2025 */ 2026 dst_object = vm_object_allocate_anon(atop(dst_entry->end - 2027 dst_entry->start), NULL, NULL, 0); 2028 #if VM_NRESERVLEVEL > 0 2029 dst_object->flags |= OBJ_COLORED; 2030 dst_object->pg_color = atop(dst_entry->start); 2031 #endif 2032 dst_object->domain = src_object->domain; 2033 dst_object->charge = dst_entry->end - dst_entry->start; 2034 2035 dst_entry->object.vm_object = dst_object; 2036 dst_entry->offset = 0; 2037 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; 2038 } 2039 2040 VM_OBJECT_WLOCK(dst_object); 2041 if (fork_charge != NULL) { 2042 KASSERT(dst_entry->cred == NULL, 2043 ("vm_fault_copy_entry: leaked swp charge")); 2044 dst_object->cred = curthread->td_ucred; 2045 crhold(dst_object->cred); 2046 *fork_charge += dst_object->charge; 2047 } else if ((dst_object->flags & OBJ_SWAP) != 0 && 2048 dst_object->cred == NULL) { 2049 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", 2050 dst_entry)); 2051 dst_object->cred = dst_entry->cred; 2052 dst_entry->cred = NULL; 2053 } 2054 2055 /* 2056 * Loop through all of the virtual pages within the entry's 2057 * range, copying each page from the source object to the 2058 * destination object. Since the source is wired, those pages 2059 * must exist. In contrast, the destination is pageable. 2060 * Since the destination object doesn't share any backing storage 2061 * with the source object, all of its pages must be dirtied, 2062 * regardless of whether they can be written. 2063 */ 2064 for (vaddr = dst_entry->start, dst_pindex = 0; 2065 vaddr < dst_entry->end; 2066 vaddr += PAGE_SIZE, dst_pindex++) { 2067 again: 2068 /* 2069 * Find the page in the source object, and copy it in. 2070 * Because the source is wired down, the page will be 2071 * in memory. 2072 */ 2073 if (src_object != dst_object) 2074 VM_OBJECT_RLOCK(src_object); 2075 object = src_object; 2076 pindex = src_pindex + dst_pindex; 2077 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 2078 (backing_object = object->backing_object) != NULL) { 2079 /* 2080 * Unless the source mapping is read-only or 2081 * it is presently being upgraded from 2082 * read-only, the first object in the shadow 2083 * chain should provide all of the pages. In 2084 * other words, this loop body should never be 2085 * executed when the source mapping is already 2086 * read/write. 2087 */ 2088 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || 2089 upgrade, 2090 ("vm_fault_copy_entry: main object missing page")); 2091 2092 VM_OBJECT_RLOCK(backing_object); 2093 pindex += OFF_TO_IDX(object->backing_object_offset); 2094 if (object != dst_object) 2095 VM_OBJECT_RUNLOCK(object); 2096 object = backing_object; 2097 } 2098 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); 2099 2100 if (object != dst_object) { 2101 /* 2102 * Allocate a page in the destination object. 2103 */ 2104 dst_m = vm_page_alloc(dst_object, (src_object == 2105 dst_object ? src_pindex : 0) + dst_pindex, 2106 VM_ALLOC_NORMAL); 2107 if (dst_m == NULL) { 2108 VM_OBJECT_WUNLOCK(dst_object); 2109 VM_OBJECT_RUNLOCK(object); 2110 vm_wait(dst_object); 2111 VM_OBJECT_WLOCK(dst_object); 2112 goto again; 2113 } 2114 2115 /* 2116 * See the comment in vm_fault_cow(). 2117 */ 2118 if (src_object == dst_object && 2119 (object->flags & OBJ_ONEMAPPING) == 0) 2120 pmap_remove_all(src_m); 2121 pmap_copy_page(src_m, dst_m); 2122 2123 /* 2124 * The object lock does not guarantee that "src_m" will 2125 * transition from invalid to valid, but it does ensure 2126 * that "src_m" will not transition from valid to 2127 * invalid. 2128 */ 2129 dst_m->dirty = dst_m->valid = src_m->valid; 2130 VM_OBJECT_RUNLOCK(object); 2131 } else { 2132 dst_m = src_m; 2133 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0) 2134 goto again; 2135 if (dst_m->pindex >= dst_object->size) { 2136 /* 2137 * We are upgrading. Index can occur 2138 * out of bounds if the object type is 2139 * vnode and the file was truncated. 2140 */ 2141 vm_page_xunbusy(dst_m); 2142 break; 2143 } 2144 } 2145 2146 /* 2147 * Enter it in the pmap. If a wired, copy-on-write 2148 * mapping is being replaced by a write-enabled 2149 * mapping, then wire that new mapping. 2150 * 2151 * The page can be invalid if the user called 2152 * msync(MS_INVALIDATE) or truncated the backing vnode 2153 * or shared memory object. In this case, do not 2154 * insert it into pmap, but still do the copy so that 2155 * all copies of the wired map entry have similar 2156 * backing pages. 2157 */ 2158 if (vm_page_all_valid(dst_m)) { 2159 VM_OBJECT_WUNLOCK(dst_object); 2160 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, 2161 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); 2162 VM_OBJECT_WLOCK(dst_object); 2163 } 2164 2165 /* 2166 * Mark it no longer busy, and put it on the active list. 2167 */ 2168 if (upgrade) { 2169 if (src_m != dst_m) { 2170 vm_page_unwire(src_m, PQ_INACTIVE); 2171 vm_page_wire(dst_m); 2172 } else { 2173 KASSERT(vm_page_wired(dst_m), 2174 ("dst_m %p is not wired", dst_m)); 2175 } 2176 } else { 2177 vm_page_activate(dst_m); 2178 } 2179 vm_page_xunbusy(dst_m); 2180 } 2181 VM_OBJECT_WUNLOCK(dst_object); 2182 if (upgrade) { 2183 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 2184 vm_object_deallocate(src_object); 2185 } 2186 } 2187 2188 /* 2189 * Block entry into the machine-independent layer's page fault handler by 2190 * the calling thread. Subsequent calls to vm_fault() by that thread will 2191 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 2192 * spurious page faults. 2193 */ 2194 int 2195 vm_fault_disable_pagefaults(void) 2196 { 2197 2198 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 2199 } 2200 2201 void 2202 vm_fault_enable_pagefaults(int save) 2203 { 2204 2205 curthread_pflags_restore(save); 2206 } 2207