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/proc.h> 89 #include <sys/racct.h> 90 #include <sys/resourcevar.h> 91 #include <sys/rwlock.h> 92 #include <sys/sysctl.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 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) 116 117 #define VM_FAULT_DONTNEED_MIN 1048576 118 119 struct faultstate { 120 vm_page_t m; 121 vm_object_t object; 122 vm_pindex_t pindex; 123 vm_page_t first_m; 124 vm_object_t first_object; 125 vm_pindex_t first_pindex; 126 vm_map_t map; 127 vm_map_entry_t entry; 128 int map_generation; 129 bool lookup_still_valid; 130 struct vnode *vp; 131 }; 132 133 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, 134 int ahead); 135 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 136 int backward, int forward, bool obj_locked); 137 138 static int vm_pfault_oom_attempts = 3; 139 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, 140 &vm_pfault_oom_attempts, 0, 141 "Number of page allocation attempts in page fault handler before it " 142 "triggers OOM handling"); 143 144 static int vm_pfault_oom_wait = 10; 145 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, 146 &vm_pfault_oom_wait, 0, 147 "Number of seconds to wait for free pages before retrying " 148 "the page fault handler"); 149 150 static inline void 151 release_page(struct faultstate *fs) 152 { 153 154 vm_page_xunbusy(fs->m); 155 vm_page_lock(fs->m); 156 vm_page_deactivate(fs->m); 157 vm_page_unlock(fs->m); 158 fs->m = NULL; 159 } 160 161 static inline void 162 unlock_map(struct faultstate *fs) 163 { 164 165 if (fs->lookup_still_valid) { 166 vm_map_lookup_done(fs->map, fs->entry); 167 fs->lookup_still_valid = false; 168 } 169 } 170 171 static void 172 unlock_vp(struct faultstate *fs) 173 { 174 175 if (fs->vp != NULL) { 176 vput(fs->vp); 177 fs->vp = NULL; 178 } 179 } 180 181 static void 182 unlock_and_deallocate(struct faultstate *fs) 183 { 184 185 vm_object_pip_wakeup(fs->object); 186 VM_OBJECT_WUNLOCK(fs->object); 187 if (fs->object != fs->first_object) { 188 VM_OBJECT_WLOCK(fs->first_object); 189 vm_page_lock(fs->first_m); 190 vm_page_free(fs->first_m); 191 vm_page_unlock(fs->first_m); 192 vm_object_pip_wakeup(fs->first_object); 193 VM_OBJECT_WUNLOCK(fs->first_object); 194 fs->first_m = NULL; 195 } 196 vm_object_deallocate(fs->first_object); 197 unlock_map(fs); 198 unlock_vp(fs); 199 } 200 201 static void 202 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, 203 vm_prot_t fault_type, int fault_flags, bool set_wd) 204 { 205 bool need_dirty; 206 207 if (((prot & VM_PROT_WRITE) == 0 && 208 (fault_flags & VM_FAULT_DIRTY) == 0) || 209 (m->oflags & VPO_UNMANAGED) != 0) 210 return; 211 212 VM_OBJECT_ASSERT_LOCKED(m->object); 213 214 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && 215 (fault_flags & VM_FAULT_WIRE) == 0) || 216 (fault_flags & VM_FAULT_DIRTY) != 0; 217 218 if (set_wd) 219 vm_object_set_writeable_dirty(m->object); 220 else 221 /* 222 * If two callers of vm_fault_dirty() with set_wd == 223 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC 224 * flag set, other with flag clear, race, it is 225 * possible for the no-NOSYNC thread to see m->dirty 226 * != 0 and not clear VPO_NOSYNC. Take vm_page lock 227 * around manipulation of VPO_NOSYNC and 228 * vm_page_dirty() call, to avoid the race and keep 229 * m->oflags consistent. 230 */ 231 vm_page_lock(m); 232 233 /* 234 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 235 * if the page is already dirty to prevent data written with 236 * the expectation of being synced from not being synced. 237 * Likewise if this entry does not request NOSYNC then make 238 * sure the page isn't marked NOSYNC. Applications sharing 239 * data should use the same flags to avoid ping ponging. 240 */ 241 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) { 242 if (m->dirty == 0) { 243 m->oflags |= VPO_NOSYNC; 244 } 245 } else { 246 m->oflags &= ~VPO_NOSYNC; 247 } 248 249 /* 250 * If the fault is a write, we know that this page is being 251 * written NOW so dirty it explicitly to save on 252 * pmap_is_modified() calls later. 253 * 254 * Also, since the page is now dirty, we can possibly tell 255 * the pager to release any swap backing the page. Calling 256 * the pager requires a write lock on the object. 257 */ 258 if (need_dirty) 259 vm_page_dirty(m); 260 if (!set_wd) 261 vm_page_unlock(m); 262 else if (need_dirty) 263 vm_pager_page_unswapped(m); 264 } 265 266 static void 267 vm_fault_fill_hold(vm_page_t *m_hold, vm_page_t m) 268 { 269 270 if (m_hold != NULL) { 271 *m_hold = m; 272 vm_page_lock(m); 273 vm_page_wire(m); 274 vm_page_unlock(m); 275 } 276 } 277 278 /* 279 * Unlocks fs.first_object and fs.map on success. 280 */ 281 static int 282 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot, 283 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) 284 { 285 vm_page_t m, m_map; 286 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 287 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ 288 VM_NRESERVLEVEL > 0 289 vm_page_t m_super; 290 int flags; 291 #endif 292 int psind, rv; 293 294 MPASS(fs->vp == NULL); 295 m = vm_page_lookup(fs->first_object, fs->first_pindex); 296 /* A busy page can be mapped for read|execute access. */ 297 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && 298 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL) 299 return (KERN_FAILURE); 300 m_map = m; 301 psind = 0; 302 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 303 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ 304 VM_NRESERVLEVEL > 0 305 if ((m->flags & PG_FICTITIOUS) == 0 && 306 (m_super = vm_reserv_to_superpage(m)) != NULL && 307 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && 308 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && 309 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & 310 (pagesizes[m_super->psind] - 1)) && !wired && 311 pmap_ps_enabled(fs->map->pmap)) { 312 flags = PS_ALL_VALID; 313 if ((prot & VM_PROT_WRITE) != 0) { 314 /* 315 * Create a superpage mapping allowing write access 316 * only if none of the constituent pages are busy and 317 * all of them are already dirty (except possibly for 318 * the page that was faulted on). 319 */ 320 flags |= PS_NONE_BUSY; 321 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) 322 flags |= PS_ALL_DIRTY; 323 } 324 if (vm_page_ps_test(m_super, flags, m)) { 325 m_map = m_super; 326 psind = m_super->psind; 327 vaddr = rounddown2(vaddr, pagesizes[psind]); 328 /* Preset the modified bit for dirty superpages. */ 329 if ((flags & PS_ALL_DIRTY) != 0) 330 fault_type |= VM_PROT_WRITE; 331 } 332 } 333 #endif 334 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type | 335 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind); 336 if (rv != KERN_SUCCESS) 337 return (rv); 338 vm_fault_fill_hold(m_hold, m); 339 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false); 340 if (psind == 0 && !wired) 341 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); 342 VM_OBJECT_RUNLOCK(fs->first_object); 343 vm_map_lookup_done(fs->map, fs->entry); 344 curthread->td_ru.ru_minflt++; 345 return (KERN_SUCCESS); 346 } 347 348 static void 349 vm_fault_restore_map_lock(struct faultstate *fs) 350 { 351 352 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 353 MPASS(fs->first_object->paging_in_progress > 0); 354 355 if (!vm_map_trylock_read(fs->map)) { 356 VM_OBJECT_WUNLOCK(fs->first_object); 357 vm_map_lock_read(fs->map); 358 VM_OBJECT_WLOCK(fs->first_object); 359 } 360 fs->lookup_still_valid = true; 361 } 362 363 static void 364 vm_fault_populate_check_page(vm_page_t m) 365 { 366 367 /* 368 * Check each page to ensure that the pager is obeying the 369 * interface: the page must be installed in the object, fully 370 * valid, and exclusively busied. 371 */ 372 MPASS(m != NULL); 373 MPASS(m->valid == VM_PAGE_BITS_ALL); 374 MPASS(vm_page_xbusied(m)); 375 } 376 377 static void 378 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, 379 vm_pindex_t last) 380 { 381 vm_page_t m; 382 vm_pindex_t pidx; 383 384 VM_OBJECT_ASSERT_WLOCKED(object); 385 MPASS(first <= last); 386 for (pidx = first, m = vm_page_lookup(object, pidx); 387 pidx <= last; pidx++, m = vm_page_next(m)) { 388 vm_fault_populate_check_page(m); 389 vm_page_lock(m); 390 vm_page_deactivate(m); 391 vm_page_unlock(m); 392 vm_page_xunbusy(m); 393 } 394 } 395 396 static int 397 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type, 398 int fault_flags, boolean_t wired, vm_page_t *m_hold) 399 { 400 struct mtx *m_mtx; 401 vm_offset_t vaddr; 402 vm_page_t m; 403 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; 404 int i, npages, psind, rv; 405 406 MPASS(fs->object == fs->first_object); 407 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 408 MPASS(fs->first_object->paging_in_progress > 0); 409 MPASS(fs->first_object->backing_object == NULL); 410 MPASS(fs->lookup_still_valid); 411 412 pager_first = OFF_TO_IDX(fs->entry->offset); 413 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; 414 unlock_map(fs); 415 unlock_vp(fs); 416 417 /* 418 * Call the pager (driver) populate() method. 419 * 420 * There is no guarantee that the method will be called again 421 * if the current fault is for read, and a future fault is 422 * for write. Report the entry's maximum allowed protection 423 * to the driver. 424 */ 425 rv = vm_pager_populate(fs->first_object, fs->first_pindex, 426 fault_type, fs->entry->max_protection, &pager_first, &pager_last); 427 428 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 429 if (rv == VM_PAGER_BAD) { 430 /* 431 * VM_PAGER_BAD is the backdoor for a pager to request 432 * normal fault handling. 433 */ 434 vm_fault_restore_map_lock(fs); 435 if (fs->map->timestamp != fs->map_generation) 436 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ 437 return (KERN_NOT_RECEIVER); 438 } 439 if (rv != VM_PAGER_OK) 440 return (KERN_FAILURE); /* AKA SIGSEGV */ 441 442 /* Ensure that the driver is obeying the interface. */ 443 MPASS(pager_first <= pager_last); 444 MPASS(fs->first_pindex <= pager_last); 445 MPASS(fs->first_pindex >= pager_first); 446 MPASS(pager_last < fs->first_object->size); 447 448 vm_fault_restore_map_lock(fs); 449 if (fs->map->timestamp != fs->map_generation) { 450 vm_fault_populate_cleanup(fs->first_object, pager_first, 451 pager_last); 452 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ 453 } 454 455 /* 456 * The map is unchanged after our last unlock. Process the fault. 457 * 458 * The range [pager_first, pager_last] that is given to the 459 * pager is only a hint. The pager may populate any range 460 * within the object that includes the requested page index. 461 * In case the pager expanded the range, clip it to fit into 462 * the map entry. 463 */ 464 map_first = OFF_TO_IDX(fs->entry->offset); 465 if (map_first > pager_first) { 466 vm_fault_populate_cleanup(fs->first_object, pager_first, 467 map_first - 1); 468 pager_first = map_first; 469 } 470 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; 471 if (map_last < pager_last) { 472 vm_fault_populate_cleanup(fs->first_object, map_last + 1, 473 pager_last); 474 pager_last = map_last; 475 } 476 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); 477 pidx <= pager_last; 478 pidx += npages, m = vm_page_next(&m[npages - 1])) { 479 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; 480 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 481 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) 482 psind = m->psind; 483 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || 484 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || 485 !pmap_ps_enabled(fs->map->pmap) || wired)) 486 psind = 0; 487 #else 488 psind = 0; 489 #endif 490 npages = atop(pagesizes[psind]); 491 for (i = 0; i < npages; i++) { 492 vm_fault_populate_check_page(&m[i]); 493 vm_fault_dirty(fs->entry, &m[i], prot, fault_type, 494 fault_flags, true); 495 } 496 VM_OBJECT_WUNLOCK(fs->first_object); 497 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | 498 (wired ? PMAP_ENTER_WIRED : 0), psind); 499 #if defined(__amd64__) 500 if (psind > 0 && rv == KERN_FAILURE) { 501 for (i = 0; i < npages; i++) { 502 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), 503 &m[i], prot, fault_type | 504 (wired ? PMAP_ENTER_WIRED : 0), 0); 505 MPASS(rv == KERN_SUCCESS); 506 } 507 } 508 #else 509 MPASS(rv == KERN_SUCCESS); 510 #endif 511 VM_OBJECT_WLOCK(fs->first_object); 512 m_mtx = NULL; 513 for (i = 0; i < npages; i++) { 514 vm_page_change_lock(&m[i], &m_mtx); 515 if ((fault_flags & VM_FAULT_WIRE) != 0) 516 vm_page_wire(&m[i]); 517 else 518 vm_page_activate(&m[i]); 519 if (m_hold != NULL && m[i].pindex == fs->first_pindex) { 520 *m_hold = &m[i]; 521 vm_page_wire(&m[i]); 522 } 523 vm_page_xunbusy_maybelocked(&m[i]); 524 } 525 if (m_mtx != NULL) 526 mtx_unlock(m_mtx); 527 } 528 curthread->td_ru.ru_majflt++; 529 return (KERN_SUCCESS); 530 } 531 532 /* 533 * vm_fault: 534 * 535 * Handle a page fault occurring at the given address, 536 * requiring the given permissions, in the map specified. 537 * If successful, the page is inserted into the 538 * associated physical map. 539 * 540 * NOTE: the given address should be truncated to the 541 * proper page address. 542 * 543 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 544 * a standard error specifying why the fault is fatal is returned. 545 * 546 * The map in question must be referenced, and remains so. 547 * Caller may hold no locks. 548 */ 549 int 550 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 551 int fault_flags) 552 { 553 struct thread *td; 554 int result; 555 556 td = curthread; 557 if ((td->td_pflags & TDP_NOFAULTING) != 0) 558 return (KERN_PROTECTION_FAILURE); 559 #ifdef KTRACE 560 if (map != kernel_map && KTRPOINT(td, KTR_FAULT)) 561 ktrfault(vaddr, fault_type); 562 #endif 563 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags, 564 NULL); 565 #ifdef KTRACE 566 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND)) 567 ktrfaultend(result); 568 #endif 569 return (result); 570 } 571 572 int 573 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 574 int fault_flags, vm_page_t *m_hold) 575 { 576 struct faultstate fs; 577 struct vnode *vp; 578 struct domainset *dset; 579 struct mtx *mtx; 580 vm_object_t next_object, retry_object; 581 vm_offset_t e_end, e_start; 582 vm_pindex_t retry_pindex; 583 vm_prot_t prot, retry_prot; 584 int ahead, alloc_req, behind, cluster_offset, error, era, faultcount; 585 int locked, nera, oom, result, rv; 586 u_char behavior; 587 boolean_t wired; /* Passed by reference. */ 588 bool dead, hardfault, is_first_object_locked; 589 590 VM_CNT_INC(v_vm_faults); 591 fs.vp = NULL; 592 faultcount = 0; 593 nera = -1; 594 hardfault = false; 595 596 RetryFault: 597 oom = 0; 598 RetryFault_oom: 599 600 /* 601 * Find the backing store object and offset into it to begin the 602 * search. 603 */ 604 fs.map = map; 605 result = vm_map_lookup(&fs.map, vaddr, fault_type | 606 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object, 607 &fs.first_pindex, &prot, &wired); 608 if (result != KERN_SUCCESS) { 609 unlock_vp(&fs); 610 return (result); 611 } 612 613 fs.map_generation = fs.map->timestamp; 614 615 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 616 panic("%s: fault on nofault entry, addr: %#lx", 617 __func__, (u_long)vaddr); 618 } 619 620 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && 621 fs.entry->wiring_thread != curthread) { 622 vm_map_unlock_read(fs.map); 623 vm_map_lock(fs.map); 624 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && 625 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 626 unlock_vp(&fs); 627 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 628 vm_map_unlock_and_wait(fs.map, 0); 629 } else 630 vm_map_unlock(fs.map); 631 goto RetryFault; 632 } 633 634 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0); 635 636 if (wired) 637 fault_type = prot | (fault_type & VM_PROT_COPY); 638 else 639 KASSERT((fault_flags & VM_FAULT_WIRE) == 0, 640 ("!wired && VM_FAULT_WIRE")); 641 642 /* 643 * Try to avoid lock contention on the top-level object through 644 * special-case handling of some types of page faults, specifically, 645 * those that are both (1) mapping an existing page from the top- 646 * level object and (2) not having to mark that object as containing 647 * dirty pages. Under these conditions, a read lock on the top-level 648 * object suffices, allowing multiple page faults of a similar type to 649 * run in parallel on the same top-level object. 650 */ 651 if (fs.vp == NULL /* avoid locked vnode leak */ && 652 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 && 653 /* avoid calling vm_object_set_writeable_dirty() */ 654 ((prot & VM_PROT_WRITE) == 0 || 655 (fs.first_object->type != OBJT_VNODE && 656 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 657 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) { 658 VM_OBJECT_RLOCK(fs.first_object); 659 if ((prot & VM_PROT_WRITE) == 0 || 660 (fs.first_object->type != OBJT_VNODE && 661 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 662 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) { 663 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type, 664 fault_flags, wired, m_hold); 665 if (rv == KERN_SUCCESS) 666 return (rv); 667 } 668 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { 669 VM_OBJECT_RUNLOCK(fs.first_object); 670 VM_OBJECT_WLOCK(fs.first_object); 671 } 672 } else { 673 VM_OBJECT_WLOCK(fs.first_object); 674 } 675 676 /* 677 * Make a reference to this object to prevent its disposal while we 678 * are messing with it. Once we have the reference, the map is free 679 * to be diddled. Since objects reference their shadows (and copies), 680 * they will stay around as well. 681 * 682 * Bump the paging-in-progress count to prevent size changes (e.g. 683 * truncation operations) during I/O. 684 */ 685 vm_object_reference_locked(fs.first_object); 686 vm_object_pip_add(fs.first_object, 1); 687 688 fs.lookup_still_valid = true; 689 690 fs.first_m = NULL; 691 692 /* 693 * Search for the page at object/offset. 694 */ 695 fs.object = fs.first_object; 696 fs.pindex = fs.first_pindex; 697 while (TRUE) { 698 /* 699 * If the object is marked for imminent termination, 700 * we retry here, since the collapse pass has raced 701 * with us. Otherwise, if we see terminally dead 702 * object, return fail. 703 */ 704 if ((fs.object->flags & OBJ_DEAD) != 0) { 705 dead = fs.object->type == OBJT_DEAD; 706 unlock_and_deallocate(&fs); 707 if (dead) 708 return (KERN_PROTECTION_FAILURE); 709 pause("vmf_de", 1); 710 goto RetryFault; 711 } 712 713 /* 714 * See if page is resident 715 */ 716 fs.m = vm_page_lookup(fs.object, fs.pindex); 717 if (fs.m != NULL) { 718 /* 719 * Wait/Retry if the page is busy. We have to do this 720 * if the page is either exclusive or shared busy 721 * because the vm_pager may be using read busy for 722 * pageouts (and even pageins if it is the vnode 723 * pager), and we could end up trying to pagein and 724 * pageout the same page simultaneously. 725 * 726 * We can theoretically allow the busy case on a read 727 * fault if the page is marked valid, but since such 728 * pages are typically already pmap'd, putting that 729 * special case in might be more effort then it is 730 * worth. We cannot under any circumstances mess 731 * around with a shared busied page except, perhaps, 732 * to pmap it. 733 */ 734 if (vm_page_busied(fs.m)) { 735 /* 736 * Reference the page before unlocking and 737 * sleeping so that the page daemon is less 738 * likely to reclaim it. 739 */ 740 vm_page_aflag_set(fs.m, PGA_REFERENCED); 741 if (fs.object != fs.first_object) { 742 if (!VM_OBJECT_TRYWLOCK( 743 fs.first_object)) { 744 VM_OBJECT_WUNLOCK(fs.object); 745 VM_OBJECT_WLOCK(fs.first_object); 746 VM_OBJECT_WLOCK(fs.object); 747 } 748 vm_page_lock(fs.first_m); 749 vm_page_free(fs.first_m); 750 vm_page_unlock(fs.first_m); 751 vm_object_pip_wakeup(fs.first_object); 752 VM_OBJECT_WUNLOCK(fs.first_object); 753 fs.first_m = NULL; 754 } 755 unlock_map(&fs); 756 if (fs.m == vm_page_lookup(fs.object, 757 fs.pindex)) { 758 vm_page_sleep_if_busy(fs.m, "vmpfw"); 759 } 760 vm_object_pip_wakeup(fs.object); 761 VM_OBJECT_WUNLOCK(fs.object); 762 VM_CNT_INC(v_intrans); 763 vm_object_deallocate(fs.first_object); 764 goto RetryFault; 765 } 766 767 /* 768 * Mark page busy for other processes, and the 769 * pagedaemon. If it still isn't completely valid 770 * (readable), jump to readrest, else break-out ( we 771 * found the page ). 772 */ 773 vm_page_xbusy(fs.m); 774 if (fs.m->valid != VM_PAGE_BITS_ALL) 775 goto readrest; 776 break; /* break to PAGE HAS BEEN FOUND */ 777 } 778 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m)); 779 780 /* 781 * Page is not resident. If the pager might contain the page 782 * or this is the beginning of the search, allocate a new 783 * page. (Default objects are zero-fill, so there is no real 784 * pager for them.) 785 */ 786 if (fs.object->type != OBJT_DEFAULT || 787 fs.object == fs.first_object) { 788 if (fs.pindex >= fs.object->size) { 789 unlock_and_deallocate(&fs); 790 return (KERN_PROTECTION_FAILURE); 791 } 792 793 if (fs.object == fs.first_object && 794 (fs.first_object->flags & OBJ_POPULATE) != 0 && 795 fs.first_object->shadow_count == 0) { 796 rv = vm_fault_populate(&fs, prot, fault_type, 797 fault_flags, wired, m_hold); 798 switch (rv) { 799 case KERN_SUCCESS: 800 case KERN_FAILURE: 801 unlock_and_deallocate(&fs); 802 return (rv); 803 case KERN_RESOURCE_SHORTAGE: 804 unlock_and_deallocate(&fs); 805 goto RetryFault; 806 case KERN_NOT_RECEIVER: 807 /* 808 * Pager's populate() method 809 * returned VM_PAGER_BAD. 810 */ 811 break; 812 default: 813 panic("inconsistent return codes"); 814 } 815 } 816 817 /* 818 * Allocate a new page for this object/offset pair. 819 * 820 * Unlocked read of the p_flag is harmless. At 821 * worst, the P_KILLED might be not observed 822 * there, and allocation can fail, causing 823 * restart and new reading of the p_flag. 824 */ 825 dset = fs.object->domain.dr_policy; 826 if (dset == NULL) 827 dset = curthread->td_domain.dr_policy; 828 if (!vm_page_count_severe_set(&dset->ds_mask) || 829 P_KILLED(curproc)) { 830 #if VM_NRESERVLEVEL > 0 831 vm_object_color(fs.object, atop(vaddr) - 832 fs.pindex); 833 #endif 834 alloc_req = P_KILLED(curproc) ? 835 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 836 if (fs.object->type != OBJT_VNODE && 837 fs.object->backing_object == NULL) 838 alloc_req |= VM_ALLOC_ZERO; 839 fs.m = vm_page_alloc(fs.object, fs.pindex, 840 alloc_req); 841 } 842 if (fs.m == NULL) { 843 unlock_and_deallocate(&fs); 844 if (vm_pfault_oom_attempts < 0 || 845 oom < vm_pfault_oom_attempts) { 846 oom++; 847 vm_waitpfault(dset, 848 vm_pfault_oom_wait * hz); 849 goto RetryFault_oom; 850 } 851 if (bootverbose) 852 printf( 853 "proc %d (%s) failed to alloc page on fault, starting OOM\n", 854 curproc->p_pid, curproc->p_comm); 855 vm_pageout_oom(VM_OOM_MEM_PF); 856 goto RetryFault; 857 } 858 } 859 860 readrest: 861 /* 862 * At this point, we have either allocated a new page or found 863 * an existing page that is only partially valid. 864 * 865 * We hold a reference on the current object and the page is 866 * exclusive busied. 867 */ 868 869 /* 870 * If the pager for the current object might have the page, 871 * then determine the number of additional pages to read and 872 * potentially reprioritize previously read pages for earlier 873 * reclamation. These operations should only be performed 874 * once per page fault. Even if the current pager doesn't 875 * have the page, the number of additional pages to read will 876 * apply to subsequent objects in the shadow chain. 877 */ 878 if (fs.object->type != OBJT_DEFAULT && nera == -1 && 879 !P_KILLED(curproc)) { 880 KASSERT(fs.lookup_still_valid, ("map unlocked")); 881 era = fs.entry->read_ahead; 882 behavior = vm_map_entry_behavior(fs.entry); 883 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 884 nera = 0; 885 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 886 nera = VM_FAULT_READ_AHEAD_MAX; 887 if (vaddr == fs.entry->next_read) 888 vm_fault_dontneed(&fs, vaddr, nera); 889 } else if (vaddr == fs.entry->next_read) { 890 /* 891 * This is a sequential fault. Arithmetically 892 * increase the requested number of pages in 893 * the read-ahead window. The requested 894 * number of pages is "# of sequential faults 895 * x (read ahead min + 1) + read ahead min" 896 */ 897 nera = VM_FAULT_READ_AHEAD_MIN; 898 if (era > 0) { 899 nera += era + 1; 900 if (nera > VM_FAULT_READ_AHEAD_MAX) 901 nera = VM_FAULT_READ_AHEAD_MAX; 902 } 903 if (era == VM_FAULT_READ_AHEAD_MAX) 904 vm_fault_dontneed(&fs, vaddr, nera); 905 } else { 906 /* 907 * This is a non-sequential fault. 908 */ 909 nera = 0; 910 } 911 if (era != nera) { 912 /* 913 * A read lock on the map suffices to update 914 * the read ahead count safely. 915 */ 916 fs.entry->read_ahead = nera; 917 } 918 919 /* 920 * Prepare for unlocking the map. Save the map 921 * entry's start and end addresses, which are used to 922 * optimize the size of the pager operation below. 923 * Even if the map entry's addresses change after 924 * unlocking the map, using the saved addresses is 925 * safe. 926 */ 927 e_start = fs.entry->start; 928 e_end = fs.entry->end; 929 } 930 931 /* 932 * Call the pager to retrieve the page if there is a chance 933 * that the pager has it, and potentially retrieve additional 934 * pages at the same time. 935 */ 936 if (fs.object->type != OBJT_DEFAULT) { 937 /* 938 * Release the map lock before locking the vnode or 939 * sleeping in the pager. (If the current object has 940 * a shadow, then an earlier iteration of this loop 941 * may have already unlocked the map.) 942 */ 943 unlock_map(&fs); 944 945 if (fs.object->type == OBJT_VNODE && 946 (vp = fs.object->handle) != fs.vp) { 947 /* 948 * Perform an unlock in case the desired vnode 949 * changed while the map was unlocked during a 950 * retry. 951 */ 952 unlock_vp(&fs); 953 954 locked = VOP_ISLOCKED(vp); 955 if (locked != LK_EXCLUSIVE) 956 locked = LK_SHARED; 957 958 /* 959 * We must not sleep acquiring the vnode lock 960 * while we have the page exclusive busied or 961 * the object's paging-in-progress count 962 * incremented. Otherwise, we could deadlock. 963 */ 964 error = vget(vp, locked | LK_CANRECURSE | 965 LK_NOWAIT, curthread); 966 if (error != 0) { 967 vhold(vp); 968 release_page(&fs); 969 unlock_and_deallocate(&fs); 970 error = vget(vp, locked | LK_RETRY | 971 LK_CANRECURSE, curthread); 972 vdrop(vp); 973 fs.vp = vp; 974 KASSERT(error == 0, 975 ("vm_fault: vget failed")); 976 goto RetryFault; 977 } 978 fs.vp = vp; 979 } 980 KASSERT(fs.vp == NULL || !fs.map->system_map, 981 ("vm_fault: vnode-backed object mapped by system map")); 982 983 /* 984 * Page in the requested page and hint the pager, 985 * that it may bring up surrounding pages. 986 */ 987 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || 988 P_KILLED(curproc)) { 989 behind = 0; 990 ahead = 0; 991 } else { 992 /* Is this a sequential fault? */ 993 if (nera > 0) { 994 behind = 0; 995 ahead = nera; 996 } else { 997 /* 998 * Request a cluster of pages that is 999 * aligned to a VM_FAULT_READ_DEFAULT 1000 * page offset boundary within the 1001 * object. Alignment to a page offset 1002 * boundary is more likely to coincide 1003 * with the underlying file system 1004 * block than alignment to a virtual 1005 * address boundary. 1006 */ 1007 cluster_offset = fs.pindex % 1008 VM_FAULT_READ_DEFAULT; 1009 behind = ulmin(cluster_offset, 1010 atop(vaddr - e_start)); 1011 ahead = VM_FAULT_READ_DEFAULT - 1 - 1012 cluster_offset; 1013 } 1014 ahead = ulmin(ahead, atop(e_end - vaddr) - 1); 1015 } 1016 rv = vm_pager_get_pages(fs.object, &fs.m, 1, 1017 &behind, &ahead); 1018 if (rv == VM_PAGER_OK) { 1019 faultcount = behind + 1 + ahead; 1020 hardfault = true; 1021 break; /* break to PAGE HAS BEEN FOUND */ 1022 } 1023 if (rv == VM_PAGER_ERROR) 1024 printf("vm_fault: pager read error, pid %d (%s)\n", 1025 curproc->p_pid, curproc->p_comm); 1026 1027 /* 1028 * If an I/O error occurred or the requested page was 1029 * outside the range of the pager, clean up and return 1030 * an error. 1031 */ 1032 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { 1033 vm_page_lock(fs.m); 1034 if (!vm_page_wired(fs.m)) 1035 vm_page_free(fs.m); 1036 else 1037 vm_page_xunbusy_maybelocked(fs.m); 1038 vm_page_unlock(fs.m); 1039 fs.m = NULL; 1040 unlock_and_deallocate(&fs); 1041 return (rv == VM_PAGER_ERROR ? KERN_FAILURE : 1042 KERN_PROTECTION_FAILURE); 1043 } 1044 1045 /* 1046 * The requested page does not exist at this object/ 1047 * offset. Remove the invalid page from the object, 1048 * waking up anyone waiting for it, and continue on to 1049 * the next object. However, if this is the top-level 1050 * object, we must leave the busy page in place to 1051 * prevent another process from rushing past us, and 1052 * inserting the page in that object at the same time 1053 * that we are. 1054 */ 1055 if (fs.object != fs.first_object) { 1056 vm_page_lock(fs.m); 1057 if (!vm_page_wired(fs.m)) 1058 vm_page_free(fs.m); 1059 else 1060 vm_page_xunbusy_maybelocked(fs.m); 1061 vm_page_unlock(fs.m); 1062 fs.m = NULL; 1063 } 1064 } 1065 1066 /* 1067 * We get here if the object has default pager (or unwiring) 1068 * or the pager doesn't have the page. 1069 */ 1070 if (fs.object == fs.first_object) 1071 fs.first_m = fs.m; 1072 1073 /* 1074 * Move on to the next object. Lock the next object before 1075 * unlocking the current one. 1076 */ 1077 next_object = fs.object->backing_object; 1078 if (next_object == NULL) { 1079 /* 1080 * If there's no object left, fill the page in the top 1081 * object with zeros. 1082 */ 1083 if (fs.object != fs.first_object) { 1084 vm_object_pip_wakeup(fs.object); 1085 VM_OBJECT_WUNLOCK(fs.object); 1086 1087 fs.object = fs.first_object; 1088 fs.pindex = fs.first_pindex; 1089 fs.m = fs.first_m; 1090 VM_OBJECT_WLOCK(fs.object); 1091 } 1092 fs.first_m = NULL; 1093 1094 /* 1095 * Zero the page if necessary and mark it valid. 1096 */ 1097 if ((fs.m->flags & PG_ZERO) == 0) { 1098 pmap_zero_page(fs.m); 1099 } else { 1100 VM_CNT_INC(v_ozfod); 1101 } 1102 VM_CNT_INC(v_zfod); 1103 fs.m->valid = VM_PAGE_BITS_ALL; 1104 /* Don't try to prefault neighboring pages. */ 1105 faultcount = 1; 1106 break; /* break to PAGE HAS BEEN FOUND */ 1107 } else { 1108 KASSERT(fs.object != next_object, 1109 ("object loop %p", next_object)); 1110 VM_OBJECT_WLOCK(next_object); 1111 vm_object_pip_add(next_object, 1); 1112 if (fs.object != fs.first_object) 1113 vm_object_pip_wakeup(fs.object); 1114 fs.pindex += 1115 OFF_TO_IDX(fs.object->backing_object_offset); 1116 VM_OBJECT_WUNLOCK(fs.object); 1117 fs.object = next_object; 1118 } 1119 } 1120 1121 vm_page_assert_xbusied(fs.m); 1122 1123 /* 1124 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1125 * is held.] 1126 */ 1127 1128 /* 1129 * If the page is being written, but isn't already owned by the 1130 * top-level object, we have to copy it into a new page owned by the 1131 * top-level object. 1132 */ 1133 if (fs.object != fs.first_object) { 1134 /* 1135 * We only really need to copy if we want to write it. 1136 */ 1137 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1138 /* 1139 * This allows pages to be virtually copied from a 1140 * backing_object into the first_object, where the 1141 * backing object has no other refs to it, and cannot 1142 * gain any more refs. Instead of a bcopy, we just 1143 * move the page from the backing object to the 1144 * first object. Note that we must mark the page 1145 * dirty in the first object so that it will go out 1146 * to swap when needed. 1147 */ 1148 is_first_object_locked = false; 1149 if ( 1150 /* 1151 * Only one shadow object 1152 */ 1153 (fs.object->shadow_count == 1) && 1154 /* 1155 * No COW refs, except us 1156 */ 1157 (fs.object->ref_count == 1) && 1158 /* 1159 * No one else can look this object up 1160 */ 1161 (fs.object->handle == NULL) && 1162 /* 1163 * No other ways to look the object up 1164 */ 1165 ((fs.object->type == OBJT_DEFAULT) || 1166 (fs.object->type == OBJT_SWAP)) && 1167 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && 1168 /* 1169 * We don't chase down the shadow chain 1170 */ 1171 fs.object == fs.first_object->backing_object) { 1172 /* 1173 * Keep the page wired to ensure that it is not 1174 * freed by another thread, such as the page 1175 * daemon, while it is disassociated from an 1176 * object. 1177 */ 1178 mtx = NULL; 1179 vm_page_change_lock(fs.m, &mtx); 1180 vm_page_wire(fs.m); 1181 (void)vm_page_remove(fs.m); 1182 vm_page_change_lock(fs.first_m, &mtx); 1183 vm_page_replace_checked(fs.m, fs.first_object, 1184 fs.first_pindex, fs.first_m); 1185 vm_page_free(fs.first_m); 1186 vm_page_change_lock(fs.m, &mtx); 1187 vm_page_unwire(fs.m, PQ_ACTIVE); 1188 mtx_unlock(mtx); 1189 vm_page_dirty(fs.m); 1190 #if VM_NRESERVLEVEL > 0 1191 /* 1192 * Rename the reservation. 1193 */ 1194 vm_reserv_rename(fs.m, fs.first_object, 1195 fs.object, OFF_TO_IDX( 1196 fs.first_object->backing_object_offset)); 1197 #endif 1198 /* 1199 * Removing the page from the backing object 1200 * unbusied it. 1201 */ 1202 vm_page_xbusy(fs.m); 1203 fs.first_m = fs.m; 1204 fs.m = NULL; 1205 VM_CNT_INC(v_cow_optim); 1206 } else { 1207 /* 1208 * Oh, well, lets copy it. 1209 */ 1210 pmap_copy_page(fs.m, fs.first_m); 1211 fs.first_m->valid = VM_PAGE_BITS_ALL; 1212 if (wired && (fault_flags & 1213 VM_FAULT_WIRE) == 0) { 1214 vm_page_lock(fs.first_m); 1215 vm_page_wire(fs.first_m); 1216 vm_page_unlock(fs.first_m); 1217 1218 vm_page_lock(fs.m); 1219 vm_page_unwire(fs.m, PQ_INACTIVE); 1220 vm_page_unlock(fs.m); 1221 } 1222 /* 1223 * We no longer need the old page or object. 1224 */ 1225 release_page(&fs); 1226 } 1227 /* 1228 * fs.object != fs.first_object due to above 1229 * conditional 1230 */ 1231 vm_object_pip_wakeup(fs.object); 1232 VM_OBJECT_WUNLOCK(fs.object); 1233 1234 /* 1235 * We only try to prefault read-only mappings to the 1236 * neighboring pages when this copy-on-write fault is 1237 * a hard fault. In other cases, trying to prefault 1238 * is typically wasted effort. 1239 */ 1240 if (faultcount == 0) 1241 faultcount = 1; 1242 1243 /* 1244 * Only use the new page below... 1245 */ 1246 fs.object = fs.first_object; 1247 fs.pindex = fs.first_pindex; 1248 fs.m = fs.first_m; 1249 if (!is_first_object_locked) 1250 VM_OBJECT_WLOCK(fs.object); 1251 VM_CNT_INC(v_cow_faults); 1252 curthread->td_cow++; 1253 } else { 1254 prot &= ~VM_PROT_WRITE; 1255 } 1256 } 1257 1258 /* 1259 * We must verify that the maps have not changed since our last 1260 * lookup. 1261 */ 1262 if (!fs.lookup_still_valid) { 1263 if (!vm_map_trylock_read(fs.map)) { 1264 release_page(&fs); 1265 unlock_and_deallocate(&fs); 1266 goto RetryFault; 1267 } 1268 fs.lookup_still_valid = true; 1269 if (fs.map->timestamp != fs.map_generation) { 1270 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 1271 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 1272 1273 /* 1274 * If we don't need the page any longer, put it on the inactive 1275 * list (the easiest thing to do here). If no one needs it, 1276 * pageout will grab it eventually. 1277 */ 1278 if (result != KERN_SUCCESS) { 1279 release_page(&fs); 1280 unlock_and_deallocate(&fs); 1281 1282 /* 1283 * If retry of map lookup would have blocked then 1284 * retry fault from start. 1285 */ 1286 if (result == KERN_FAILURE) 1287 goto RetryFault; 1288 return (result); 1289 } 1290 if ((retry_object != fs.first_object) || 1291 (retry_pindex != fs.first_pindex)) { 1292 release_page(&fs); 1293 unlock_and_deallocate(&fs); 1294 goto RetryFault; 1295 } 1296 1297 /* 1298 * Check whether the protection has changed or the object has 1299 * been copied while we left the map unlocked. Changing from 1300 * read to write permission is OK - we leave the page 1301 * write-protected, and catch the write fault. Changing from 1302 * write to read permission means that we can't mark the page 1303 * write-enabled after all. 1304 */ 1305 prot &= retry_prot; 1306 fault_type &= retry_prot; 1307 if (prot == 0) { 1308 release_page(&fs); 1309 unlock_and_deallocate(&fs); 1310 goto RetryFault; 1311 } 1312 1313 /* Reassert because wired may have changed. */ 1314 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0, 1315 ("!wired && VM_FAULT_WIRE")); 1316 } 1317 } 1318 1319 /* 1320 * If the page was filled by a pager, save the virtual address that 1321 * should be faulted on next under a sequential access pattern to the 1322 * map entry. A read lock on the map suffices to update this address 1323 * safely. 1324 */ 1325 if (hardfault) 1326 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; 1327 1328 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true); 1329 vm_page_assert_xbusied(fs.m); 1330 1331 /* 1332 * Page must be completely valid or it is not fit to 1333 * map into user space. vm_pager_get_pages() ensures this. 1334 */ 1335 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, 1336 ("vm_fault: page %p partially invalid", fs.m)); 1337 VM_OBJECT_WUNLOCK(fs.object); 1338 1339 /* 1340 * Put this page into the physical map. We had to do the unlock above 1341 * because pmap_enter() may sleep. We don't put the page 1342 * back on the active queue until later so that the pageout daemon 1343 * won't find it (yet). 1344 */ 1345 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, 1346 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); 1347 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 && 1348 wired == 0) 1349 vm_fault_prefault(&fs, vaddr, 1350 faultcount > 0 ? behind : PFBAK, 1351 faultcount > 0 ? ahead : PFFOR, false); 1352 VM_OBJECT_WLOCK(fs.object); 1353 vm_page_lock(fs.m); 1354 1355 /* 1356 * If the page is not wired down, then put it where the pageout daemon 1357 * can find it. 1358 */ 1359 if ((fault_flags & VM_FAULT_WIRE) != 0) 1360 vm_page_wire(fs.m); 1361 else 1362 vm_page_activate(fs.m); 1363 if (m_hold != NULL) { 1364 *m_hold = fs.m; 1365 vm_page_wire(fs.m); 1366 } 1367 vm_page_unlock(fs.m); 1368 vm_page_xunbusy(fs.m); 1369 1370 /* 1371 * Unlock everything, and return 1372 */ 1373 unlock_and_deallocate(&fs); 1374 if (hardfault) { 1375 VM_CNT_INC(v_io_faults); 1376 curthread->td_ru.ru_majflt++; 1377 #ifdef RACCT 1378 if (racct_enable && fs.object->type == OBJT_VNODE) { 1379 PROC_LOCK(curproc); 1380 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1381 racct_add_force(curproc, RACCT_WRITEBPS, 1382 PAGE_SIZE + behind * PAGE_SIZE); 1383 racct_add_force(curproc, RACCT_WRITEIOPS, 1); 1384 } else { 1385 racct_add_force(curproc, RACCT_READBPS, 1386 PAGE_SIZE + ahead * PAGE_SIZE); 1387 racct_add_force(curproc, RACCT_READIOPS, 1); 1388 } 1389 PROC_UNLOCK(curproc); 1390 } 1391 #endif 1392 } else 1393 curthread->td_ru.ru_minflt++; 1394 1395 return (KERN_SUCCESS); 1396 } 1397 1398 /* 1399 * Speed up the reclamation of pages that precede the faulting pindex within 1400 * the first object of the shadow chain. Essentially, perform the equivalent 1401 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes 1402 * the faulting pindex by the cluster size when the pages read by vm_fault() 1403 * cross a cluster-size boundary. The cluster size is the greater of the 1404 * smallest superpage size and VM_FAULT_DONTNEED_MIN. 1405 * 1406 * When "fs->first_object" is a shadow object, the pages in the backing object 1407 * that precede the faulting pindex are deactivated by vm_fault(). So, this 1408 * function must only be concerned with pages in the first object. 1409 */ 1410 static void 1411 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) 1412 { 1413 vm_map_entry_t entry; 1414 vm_object_t first_object, object; 1415 vm_offset_t end, start; 1416 vm_page_t m, m_next; 1417 vm_pindex_t pend, pstart; 1418 vm_size_t size; 1419 1420 object = fs->object; 1421 VM_OBJECT_ASSERT_WLOCKED(object); 1422 first_object = fs->first_object; 1423 if (first_object != object) { 1424 if (!VM_OBJECT_TRYWLOCK(first_object)) { 1425 VM_OBJECT_WUNLOCK(object); 1426 VM_OBJECT_WLOCK(first_object); 1427 VM_OBJECT_WLOCK(object); 1428 } 1429 } 1430 /* Neither fictitious nor unmanaged pages can be reclaimed. */ 1431 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 1432 size = VM_FAULT_DONTNEED_MIN; 1433 if (MAXPAGESIZES > 1 && size < pagesizes[1]) 1434 size = pagesizes[1]; 1435 end = rounddown2(vaddr, size); 1436 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && 1437 (entry = fs->entry)->start < end) { 1438 if (end - entry->start < size) 1439 start = entry->start; 1440 else 1441 start = end - size; 1442 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); 1443 pstart = OFF_TO_IDX(entry->offset) + atop(start - 1444 entry->start); 1445 m_next = vm_page_find_least(first_object, pstart); 1446 pend = OFF_TO_IDX(entry->offset) + atop(end - 1447 entry->start); 1448 while ((m = m_next) != NULL && m->pindex < pend) { 1449 m_next = TAILQ_NEXT(m, listq); 1450 if (m->valid != VM_PAGE_BITS_ALL || 1451 vm_page_busied(m)) 1452 continue; 1453 1454 /* 1455 * Don't clear PGA_REFERENCED, since it would 1456 * likely represent a reference by a different 1457 * process. 1458 * 1459 * Typically, at this point, prefetched pages 1460 * are still in the inactive queue. Only 1461 * pages that triggered page faults are in the 1462 * active queue. 1463 */ 1464 vm_page_lock(m); 1465 if (!vm_page_inactive(m)) 1466 vm_page_deactivate(m); 1467 vm_page_unlock(m); 1468 } 1469 } 1470 } 1471 if (first_object != object) 1472 VM_OBJECT_WUNLOCK(first_object); 1473 } 1474 1475 /* 1476 * vm_fault_prefault provides a quick way of clustering 1477 * pagefaults into a processes address space. It is a "cousin" 1478 * of vm_map_pmap_enter, except it runs at page fault time instead 1479 * of mmap time. 1480 */ 1481 static void 1482 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 1483 int backward, int forward, bool obj_locked) 1484 { 1485 pmap_t pmap; 1486 vm_map_entry_t entry; 1487 vm_object_t backing_object, lobject; 1488 vm_offset_t addr, starta; 1489 vm_pindex_t pindex; 1490 vm_page_t m; 1491 int i; 1492 1493 pmap = fs->map->pmap; 1494 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1495 return; 1496 1497 entry = fs->entry; 1498 1499 if (addra < backward * PAGE_SIZE) { 1500 starta = entry->start; 1501 } else { 1502 starta = addra - backward * PAGE_SIZE; 1503 if (starta < entry->start) 1504 starta = entry->start; 1505 } 1506 1507 /* 1508 * Generate the sequence of virtual addresses that are candidates for 1509 * prefaulting in an outward spiral from the faulting virtual address, 1510 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra 1511 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... 1512 * If the candidate address doesn't have a backing physical page, then 1513 * the loop immediately terminates. 1514 */ 1515 for (i = 0; i < 2 * imax(backward, forward); i++) { 1516 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : 1517 PAGE_SIZE); 1518 if (addr > addra + forward * PAGE_SIZE) 1519 addr = 0; 1520 1521 if (addr < starta || addr >= entry->end) 1522 continue; 1523 1524 if (!pmap_is_prefaultable(pmap, addr)) 1525 continue; 1526 1527 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1528 lobject = entry->object.vm_object; 1529 if (!obj_locked) 1530 VM_OBJECT_RLOCK(lobject); 1531 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1532 lobject->type == OBJT_DEFAULT && 1533 (backing_object = lobject->backing_object) != NULL) { 1534 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1535 0, ("vm_fault_prefault: unaligned object offset")); 1536 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1537 VM_OBJECT_RLOCK(backing_object); 1538 if (!obj_locked || lobject != entry->object.vm_object) 1539 VM_OBJECT_RUNLOCK(lobject); 1540 lobject = backing_object; 1541 } 1542 if (m == NULL) { 1543 if (!obj_locked || lobject != entry->object.vm_object) 1544 VM_OBJECT_RUNLOCK(lobject); 1545 break; 1546 } 1547 if (m->valid == VM_PAGE_BITS_ALL && 1548 (m->flags & PG_FICTITIOUS) == 0) 1549 pmap_enter_quick(pmap, addr, m, entry->protection); 1550 if (!obj_locked || lobject != entry->object.vm_object) 1551 VM_OBJECT_RUNLOCK(lobject); 1552 } 1553 } 1554 1555 /* 1556 * Hold each of the physical pages that are mapped by the specified range of 1557 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1558 * and allow the specified types of access, "prot". If all of the implied 1559 * pages are successfully held, then the number of held pages is returned 1560 * together with pointers to those pages in the array "ma". However, if any 1561 * of the pages cannot be held, -1 is returned. 1562 */ 1563 int 1564 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1565 vm_prot_t prot, vm_page_t *ma, int max_count) 1566 { 1567 vm_offset_t end, va; 1568 vm_page_t *mp; 1569 int count; 1570 boolean_t pmap_failed; 1571 1572 if (len == 0) 1573 return (0); 1574 end = round_page(addr + len); 1575 addr = trunc_page(addr); 1576 1577 /* 1578 * Check for illegal addresses. 1579 */ 1580 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) 1581 return (-1); 1582 1583 if (atop(end - addr) > max_count) 1584 panic("vm_fault_quick_hold_pages: count > max_count"); 1585 count = atop(end - addr); 1586 1587 /* 1588 * Most likely, the physical pages are resident in the pmap, so it is 1589 * faster to try pmap_extract_and_hold() first. 1590 */ 1591 pmap_failed = FALSE; 1592 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1593 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1594 if (*mp == NULL) 1595 pmap_failed = TRUE; 1596 else if ((prot & VM_PROT_WRITE) != 0 && 1597 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1598 /* 1599 * Explicitly dirty the physical page. Otherwise, the 1600 * caller's changes may go unnoticed because they are 1601 * performed through an unmanaged mapping or by a DMA 1602 * operation. 1603 * 1604 * The object lock is not held here. 1605 * See vm_page_clear_dirty_mask(). 1606 */ 1607 vm_page_dirty(*mp); 1608 } 1609 } 1610 if (pmap_failed) { 1611 /* 1612 * One or more pages could not be held by the pmap. Either no 1613 * page was mapped at the specified virtual address or that 1614 * mapping had insufficient permissions. Attempt to fault in 1615 * and hold these pages. 1616 * 1617 * If vm_fault_disable_pagefaults() was called, 1618 * i.e., TDP_NOFAULTING is set, we must not sleep nor 1619 * acquire MD VM locks, which means we must not call 1620 * vm_fault_hold(). Some (out of tree) callers mark 1621 * too wide a code area with vm_fault_disable_pagefaults() 1622 * already, use the VM_PROT_QUICK_NOFAULT flag to request 1623 * the proper behaviour explicitly. 1624 */ 1625 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && 1626 (curthread->td_pflags & TDP_NOFAULTING) != 0) 1627 goto error; 1628 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1629 if (*mp == NULL && vm_fault_hold(map, va, prot, 1630 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1631 goto error; 1632 } 1633 return (count); 1634 error: 1635 for (mp = ma; mp < ma + count; mp++) 1636 if (*mp != NULL) { 1637 vm_page_lock(*mp); 1638 if (vm_page_unwire(*mp, PQ_INACTIVE) && 1639 (*mp)->object == NULL) 1640 vm_page_free(*mp); 1641 vm_page_unlock(*mp); 1642 } 1643 return (-1); 1644 } 1645 1646 /* 1647 * Routine: 1648 * vm_fault_copy_entry 1649 * Function: 1650 * Create new shadow object backing dst_entry with private copy of 1651 * all underlying pages. When src_entry is equal to dst_entry, 1652 * function implements COW for wired-down map entry. Otherwise, 1653 * it forks wired entry into dst_map. 1654 * 1655 * In/out conditions: 1656 * The source and destination maps must be locked for write. 1657 * The source map entry must be wired down (or be a sharing map 1658 * entry corresponding to a main map entry that is wired down). 1659 */ 1660 void 1661 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1662 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1663 vm_ooffset_t *fork_charge) 1664 { 1665 vm_object_t backing_object, dst_object, object, src_object; 1666 vm_pindex_t dst_pindex, pindex, src_pindex; 1667 vm_prot_t access, prot; 1668 vm_offset_t vaddr; 1669 vm_page_t dst_m; 1670 vm_page_t src_m; 1671 boolean_t upgrade; 1672 1673 #ifdef lint 1674 src_map++; 1675 #endif /* lint */ 1676 1677 upgrade = src_entry == dst_entry; 1678 access = prot = dst_entry->protection; 1679 1680 src_object = src_entry->object.vm_object; 1681 src_pindex = OFF_TO_IDX(src_entry->offset); 1682 1683 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 1684 dst_object = src_object; 1685 vm_object_reference(dst_object); 1686 } else { 1687 /* 1688 * Create the top-level object for the destination entry. (Doesn't 1689 * actually shadow anything - we copy the pages directly.) 1690 */ 1691 dst_object = vm_object_allocate(OBJT_DEFAULT, 1692 atop(dst_entry->end - dst_entry->start)); 1693 #if VM_NRESERVLEVEL > 0 1694 dst_object->flags |= OBJ_COLORED; 1695 dst_object->pg_color = atop(dst_entry->start); 1696 #endif 1697 dst_object->domain = src_object->domain; 1698 dst_object->charge = dst_entry->end - dst_entry->start; 1699 } 1700 1701 VM_OBJECT_WLOCK(dst_object); 1702 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1703 ("vm_fault_copy_entry: vm_object not NULL")); 1704 if (src_object != dst_object) { 1705 dst_entry->object.vm_object = dst_object; 1706 dst_entry->offset = 0; 1707 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; 1708 } 1709 if (fork_charge != NULL) { 1710 KASSERT(dst_entry->cred == NULL, 1711 ("vm_fault_copy_entry: leaked swp charge")); 1712 dst_object->cred = curthread->td_ucred; 1713 crhold(dst_object->cred); 1714 *fork_charge += dst_object->charge; 1715 } else if ((dst_object->type == OBJT_DEFAULT || 1716 dst_object->type == OBJT_SWAP) && 1717 dst_object->cred == NULL) { 1718 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", 1719 dst_entry)); 1720 dst_object->cred = dst_entry->cred; 1721 dst_entry->cred = NULL; 1722 } 1723 1724 /* 1725 * If not an upgrade, then enter the mappings in the pmap as 1726 * read and/or execute accesses. Otherwise, enter them as 1727 * write accesses. 1728 * 1729 * A writeable large page mapping is only created if all of 1730 * the constituent small page mappings are modified. Marking 1731 * PTEs as modified on inception allows promotion to happen 1732 * without taking potentially large number of soft faults. 1733 */ 1734 if (!upgrade) 1735 access &= ~VM_PROT_WRITE; 1736 1737 /* 1738 * Loop through all of the virtual pages within the entry's 1739 * range, copying each page from the source object to the 1740 * destination object. Since the source is wired, those pages 1741 * must exist. In contrast, the destination is pageable. 1742 * Since the destination object doesn't share any backing storage 1743 * with the source object, all of its pages must be dirtied, 1744 * regardless of whether they can be written. 1745 */ 1746 for (vaddr = dst_entry->start, dst_pindex = 0; 1747 vaddr < dst_entry->end; 1748 vaddr += PAGE_SIZE, dst_pindex++) { 1749 again: 1750 /* 1751 * Find the page in the source object, and copy it in. 1752 * Because the source is wired down, the page will be 1753 * in memory. 1754 */ 1755 if (src_object != dst_object) 1756 VM_OBJECT_RLOCK(src_object); 1757 object = src_object; 1758 pindex = src_pindex + dst_pindex; 1759 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1760 (backing_object = object->backing_object) != NULL) { 1761 /* 1762 * Unless the source mapping is read-only or 1763 * it is presently being upgraded from 1764 * read-only, the first object in the shadow 1765 * chain should provide all of the pages. In 1766 * other words, this loop body should never be 1767 * executed when the source mapping is already 1768 * read/write. 1769 */ 1770 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || 1771 upgrade, 1772 ("vm_fault_copy_entry: main object missing page")); 1773 1774 VM_OBJECT_RLOCK(backing_object); 1775 pindex += OFF_TO_IDX(object->backing_object_offset); 1776 if (object != dst_object) 1777 VM_OBJECT_RUNLOCK(object); 1778 object = backing_object; 1779 } 1780 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); 1781 1782 if (object != dst_object) { 1783 /* 1784 * Allocate a page in the destination object. 1785 */ 1786 dst_m = vm_page_alloc(dst_object, (src_object == 1787 dst_object ? src_pindex : 0) + dst_pindex, 1788 VM_ALLOC_NORMAL); 1789 if (dst_m == NULL) { 1790 VM_OBJECT_WUNLOCK(dst_object); 1791 VM_OBJECT_RUNLOCK(object); 1792 vm_wait(dst_object); 1793 VM_OBJECT_WLOCK(dst_object); 1794 goto again; 1795 } 1796 pmap_copy_page(src_m, dst_m); 1797 VM_OBJECT_RUNLOCK(object); 1798 dst_m->dirty = dst_m->valid = src_m->valid; 1799 } else { 1800 dst_m = src_m; 1801 if (vm_page_sleep_if_busy(dst_m, "fltupg")) 1802 goto again; 1803 if (dst_m->pindex >= dst_object->size) 1804 /* 1805 * We are upgrading. Index can occur 1806 * out of bounds if the object type is 1807 * vnode and the file was truncated. 1808 */ 1809 break; 1810 vm_page_xbusy(dst_m); 1811 } 1812 VM_OBJECT_WUNLOCK(dst_object); 1813 1814 /* 1815 * Enter it in the pmap. If a wired, copy-on-write 1816 * mapping is being replaced by a write-enabled 1817 * mapping, then wire that new mapping. 1818 * 1819 * The page can be invalid if the user called 1820 * msync(MS_INVALIDATE) or truncated the backing vnode 1821 * or shared memory object. In this case, do not 1822 * insert it into pmap, but still do the copy so that 1823 * all copies of the wired map entry have similar 1824 * backing pages. 1825 */ 1826 if (dst_m->valid == VM_PAGE_BITS_ALL) { 1827 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, 1828 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); 1829 } 1830 1831 /* 1832 * Mark it no longer busy, and put it on the active list. 1833 */ 1834 VM_OBJECT_WLOCK(dst_object); 1835 1836 if (upgrade) { 1837 if (src_m != dst_m) { 1838 vm_page_lock(src_m); 1839 vm_page_unwire(src_m, PQ_INACTIVE); 1840 vm_page_unlock(src_m); 1841 vm_page_lock(dst_m); 1842 vm_page_wire(dst_m); 1843 vm_page_unlock(dst_m); 1844 } else { 1845 KASSERT(vm_page_wired(dst_m), 1846 ("dst_m %p is not wired", dst_m)); 1847 } 1848 } else { 1849 vm_page_lock(dst_m); 1850 vm_page_activate(dst_m); 1851 vm_page_unlock(dst_m); 1852 } 1853 vm_page_xunbusy(dst_m); 1854 } 1855 VM_OBJECT_WUNLOCK(dst_object); 1856 if (upgrade) { 1857 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1858 vm_object_deallocate(src_object); 1859 } 1860 } 1861 1862 /* 1863 * Block entry into the machine-independent layer's page fault handler by 1864 * the calling thread. Subsequent calls to vm_fault() by that thread will 1865 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 1866 * spurious page faults. 1867 */ 1868 int 1869 vm_fault_disable_pagefaults(void) 1870 { 1871 1872 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 1873 } 1874 1875 void 1876 vm_fault_enable_pagefaults(int save) 1877 { 1878 1879 curthread_pflags_restore(save); 1880 } 1881