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 [email protected] 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/proc.h> 88 #include <sys/racct.h> 89 #include <sys/resourcevar.h> 90 #include <sys/rwlock.h> 91 #include <sys/sysctl.h> 92 #include <sys/vmmeter.h> 93 #include <sys/vnode.h> 94 #ifdef KTRACE 95 #include <sys/ktrace.h> 96 #endif 97 98 #include <vm/vm.h> 99 #include <vm/vm_param.h> 100 #include <vm/pmap.h> 101 #include <vm/vm_map.h> 102 #include <vm/vm_object.h> 103 #include <vm/vm_page.h> 104 #include <vm/vm_pageout.h> 105 #include <vm/vm_kern.h> 106 #include <vm/vm_pager.h> 107 #include <vm/vm_extern.h> 108 #include <vm/vm_reserv.h> 109 110 #define PFBAK 4 111 #define PFFOR 4 112 113 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) 114 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) 115 116 #define VM_FAULT_DONTNEED_MIN 1048576 117 118 struct faultstate { 119 vm_page_t m; 120 vm_object_t object; 121 vm_pindex_t pindex; 122 vm_page_t first_m; 123 vm_object_t first_object; 124 vm_pindex_t first_pindex; 125 vm_map_t map; 126 vm_map_entry_t entry; 127 int map_generation; 128 bool lookup_still_valid; 129 struct vnode *vp; 130 }; 131 132 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, 133 int ahead); 134 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 135 int backward, int forward, bool obj_locked); 136 137 static int vm_pfault_oom_attempts = 3; 138 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, 139 &vm_pfault_oom_attempts, 0, 140 "Number of page allocation attempts in page fault handler before it " 141 "triggers OOM handling"); 142 143 static int vm_pfault_oom_wait = 10; 144 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, 145 &vm_pfault_oom_wait, 0, 146 "Number of seconds to wait for free pages before retrying " 147 "the page fault handler"); 148 149 static inline void 150 release_page(struct faultstate *fs) 151 { 152 153 vm_page_xunbusy(fs->m); 154 vm_page_lock(fs->m); 155 vm_page_deactivate(fs->m); 156 vm_page_unlock(fs->m); 157 fs->m = NULL; 158 } 159 160 static inline void 161 unlock_map(struct faultstate *fs) 162 { 163 164 if (fs->lookup_still_valid) { 165 vm_map_lookup_done(fs->map, fs->entry); 166 fs->lookup_still_valid = false; 167 } 168 } 169 170 static void 171 unlock_vp(struct faultstate *fs) 172 { 173 174 if (fs->vp != NULL) { 175 vput(fs->vp); 176 fs->vp = NULL; 177 } 178 } 179 180 static void 181 unlock_and_deallocate(struct faultstate *fs) 182 { 183 184 vm_object_pip_wakeup(fs->object); 185 VM_OBJECT_WUNLOCK(fs->object); 186 if (fs->object != fs->first_object) { 187 VM_OBJECT_WLOCK(fs->first_object); 188 vm_page_lock(fs->first_m); 189 vm_page_free(fs->first_m); 190 vm_page_unlock(fs->first_m); 191 vm_object_pip_wakeup(fs->first_object); 192 VM_OBJECT_WUNLOCK(fs->first_object); 193 fs->first_m = NULL; 194 } 195 vm_object_deallocate(fs->first_object); 196 unlock_map(fs); 197 unlock_vp(fs); 198 } 199 200 static void 201 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, 202 vm_prot_t fault_type, int fault_flags, bool set_wd) 203 { 204 bool need_dirty; 205 206 if (((prot & VM_PROT_WRITE) == 0 && 207 (fault_flags & VM_FAULT_DIRTY) == 0) || 208 (m->oflags & VPO_UNMANAGED) != 0) 209 return; 210 211 VM_OBJECT_ASSERT_LOCKED(m->object); 212 213 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && 214 (fault_flags & VM_FAULT_WIRE) == 0) || 215 (fault_flags & VM_FAULT_DIRTY) != 0; 216 217 if (set_wd) 218 vm_object_set_writeable_dirty(m->object); 219 else 220 /* 221 * If two callers of vm_fault_dirty() with set_wd == 222 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC 223 * flag set, other with flag clear, race, it is 224 * possible for the no-NOSYNC thread to see m->dirty 225 * != 0 and not clear VPO_NOSYNC. Take vm_page lock 226 * around manipulation of VPO_NOSYNC and 227 * vm_page_dirty() call, to avoid the race and keep 228 * m->oflags consistent. 229 */ 230 vm_page_lock(m); 231 232 /* 233 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 234 * if the page is already dirty to prevent data written with 235 * the expectation of being synced from not being synced. 236 * Likewise if this entry does not request NOSYNC then make 237 * sure the page isn't marked NOSYNC. Applications sharing 238 * data should use the same flags to avoid ping ponging. 239 */ 240 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) { 241 if (m->dirty == 0) { 242 m->oflags |= VPO_NOSYNC; 243 } 244 } else { 245 m->oflags &= ~VPO_NOSYNC; 246 } 247 248 /* 249 * If the fault is a write, we know that this page is being 250 * written NOW so dirty it explicitly to save on 251 * pmap_is_modified() calls later. 252 * 253 * Also, since the page is now dirty, we can possibly tell 254 * the pager to release any swap backing the page. Calling 255 * the pager requires a write lock on the object. 256 */ 257 if (need_dirty) 258 vm_page_dirty(m); 259 if (!set_wd) 260 vm_page_unlock(m); 261 else if (need_dirty) 262 vm_pager_page_unswapped(m); 263 } 264 265 static void 266 vm_fault_fill_hold(vm_page_t *m_hold, vm_page_t m) 267 { 268 269 if (m_hold != NULL) { 270 *m_hold = m; 271 vm_page_lock(m); 272 vm_page_hold(m); 273 vm_page_unlock(m); 274 } 275 } 276 277 /* 278 * Unlocks fs.first_object and fs.map on success. 279 */ 280 static int 281 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot, 282 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) 283 { 284 vm_page_t m, m_map; 285 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 286 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ 287 VM_NRESERVLEVEL > 0 288 vm_page_t m_super; 289 int flags; 290 #endif 291 int psind, rv; 292 293 MPASS(fs->vp == NULL); 294 m = vm_page_lookup(fs->first_object, fs->first_pindex); 295 /* A busy page can be mapped for read|execute access. */ 296 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && 297 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL) 298 return (KERN_FAILURE); 299 m_map = m; 300 psind = 0; 301 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 302 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ 303 VM_NRESERVLEVEL > 0 304 if ((m->flags & PG_FICTITIOUS) == 0 && 305 (m_super = vm_reserv_to_superpage(m)) != NULL && 306 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && 307 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && 308 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & 309 (pagesizes[m_super->psind] - 1)) && 310 pmap_ps_enabled(fs->map->pmap)) { 311 flags = PS_ALL_VALID; 312 if ((prot & VM_PROT_WRITE) != 0) { 313 /* 314 * Create a superpage mapping allowing write access 315 * only if none of the constituent pages are busy and 316 * all of them are already dirty (except possibly for 317 * the page that was faulted on). 318 */ 319 flags |= PS_NONE_BUSY; 320 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) 321 flags |= PS_ALL_DIRTY; 322 } 323 if (vm_page_ps_test(m_super, flags, m)) { 324 m_map = m_super; 325 psind = m_super->psind; 326 vaddr = rounddown2(vaddr, pagesizes[psind]); 327 /* Preset the modified bit for dirty superpages. */ 328 if ((flags & PS_ALL_DIRTY) != 0) 329 fault_type |= VM_PROT_WRITE; 330 } 331 } 332 #endif 333 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type | 334 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind); 335 if (rv != KERN_SUCCESS) 336 return (rv); 337 vm_fault_fill_hold(m_hold, m); 338 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false); 339 if (psind == 0 && !wired) 340 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); 341 VM_OBJECT_RUNLOCK(fs->first_object); 342 vm_map_lookup_done(fs->map, fs->entry); 343 curthread->td_ru.ru_minflt++; 344 return (KERN_SUCCESS); 345 } 346 347 static void 348 vm_fault_restore_map_lock(struct faultstate *fs) 349 { 350 351 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 352 MPASS(fs->first_object->paging_in_progress > 0); 353 354 if (!vm_map_trylock_read(fs->map)) { 355 VM_OBJECT_WUNLOCK(fs->first_object); 356 vm_map_lock_read(fs->map); 357 VM_OBJECT_WLOCK(fs->first_object); 358 } 359 fs->lookup_still_valid = true; 360 } 361 362 static void 363 vm_fault_populate_check_page(vm_page_t m) 364 { 365 366 /* 367 * Check each page to ensure that the pager is obeying the 368 * interface: the page must be installed in the object, fully 369 * valid, and exclusively busied. 370 */ 371 MPASS(m != NULL); 372 MPASS(m->valid == VM_PAGE_BITS_ALL); 373 MPASS(vm_page_xbusied(m)); 374 } 375 376 static void 377 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, 378 vm_pindex_t last) 379 { 380 vm_page_t m; 381 vm_pindex_t pidx; 382 383 VM_OBJECT_ASSERT_WLOCKED(object); 384 MPASS(first <= last); 385 for (pidx = first, m = vm_page_lookup(object, pidx); 386 pidx <= last; pidx++, m = vm_page_next(m)) { 387 vm_fault_populate_check_page(m); 388 vm_page_lock(m); 389 vm_page_deactivate(m); 390 vm_page_unlock(m); 391 vm_page_xunbusy(m); 392 } 393 } 394 395 static int 396 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type, 397 int fault_flags, boolean_t wired, vm_page_t *m_hold) 398 { 399 struct mtx *m_mtx; 400 vm_offset_t vaddr; 401 vm_page_t m; 402 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; 403 int i, npages, psind, rv; 404 405 MPASS(fs->object == fs->first_object); 406 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 407 MPASS(fs->first_object->paging_in_progress > 0); 408 MPASS(fs->first_object->backing_object == NULL); 409 MPASS(fs->lookup_still_valid); 410 411 pager_first = OFF_TO_IDX(fs->entry->offset); 412 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; 413 unlock_map(fs); 414 unlock_vp(fs); 415 416 /* 417 * Call the pager (driver) populate() method. 418 * 419 * There is no guarantee that the method will be called again 420 * if the current fault is for read, and a future fault is 421 * for write. Report the entry's maximum allowed protection 422 * to the driver. 423 */ 424 rv = vm_pager_populate(fs->first_object, fs->first_pindex, 425 fault_type, fs->entry->max_protection, &pager_first, &pager_last); 426 427 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 428 if (rv == VM_PAGER_BAD) { 429 /* 430 * VM_PAGER_BAD is the backdoor for a pager to request 431 * normal fault handling. 432 */ 433 vm_fault_restore_map_lock(fs); 434 if (fs->map->timestamp != fs->map_generation) 435 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ 436 return (KERN_NOT_RECEIVER); 437 } 438 if (rv != VM_PAGER_OK) 439 return (KERN_FAILURE); /* AKA SIGSEGV */ 440 441 /* Ensure that the driver is obeying the interface. */ 442 MPASS(pager_first <= pager_last); 443 MPASS(fs->first_pindex <= pager_last); 444 MPASS(fs->first_pindex >= pager_first); 445 MPASS(pager_last < fs->first_object->size); 446 447 vm_fault_restore_map_lock(fs); 448 if (fs->map->timestamp != fs->map_generation) { 449 vm_fault_populate_cleanup(fs->first_object, pager_first, 450 pager_last); 451 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ 452 } 453 454 /* 455 * The map is unchanged after our last unlock. Process the fault. 456 * 457 * The range [pager_first, pager_last] that is given to the 458 * pager is only a hint. The pager may populate any range 459 * within the object that includes the requested page index. 460 * In case the pager expanded the range, clip it to fit into 461 * the map entry. 462 */ 463 map_first = OFF_TO_IDX(fs->entry->offset); 464 if (map_first > pager_first) { 465 vm_fault_populate_cleanup(fs->first_object, pager_first, 466 map_first - 1); 467 pager_first = map_first; 468 } 469 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; 470 if (map_last < pager_last) { 471 vm_fault_populate_cleanup(fs->first_object, map_last + 1, 472 pager_last); 473 pager_last = map_last; 474 } 475 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); 476 pidx <= pager_last; 477 pidx += npages, m = vm_page_next(&m[npages - 1])) { 478 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; 479 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 480 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) 481 psind = m->psind; 482 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || 483 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || 484 !pmap_ps_enabled(fs->map->pmap))) 485 psind = 0; 486 #else 487 psind = 0; 488 #endif 489 npages = atop(pagesizes[psind]); 490 for (i = 0; i < npages; i++) { 491 vm_fault_populate_check_page(&m[i]); 492 vm_fault_dirty(fs->entry, &m[i], prot, fault_type, 493 fault_flags, true); 494 } 495 VM_OBJECT_WUNLOCK(fs->first_object); 496 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | 497 (wired ? PMAP_ENTER_WIRED : 0), psind); 498 #if defined(__amd64__) 499 if (psind > 0 && rv == KERN_FAILURE) { 500 for (i = 0; i < npages; i++) { 501 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), 502 &m[i], prot, fault_type | 503 (wired ? PMAP_ENTER_WIRED : 0), 0); 504 MPASS(rv == KERN_SUCCESS); 505 } 506 } 507 #else 508 MPASS(rv == KERN_SUCCESS); 509 #endif 510 VM_OBJECT_WLOCK(fs->first_object); 511 m_mtx = NULL; 512 for (i = 0; i < npages; i++) { 513 vm_page_change_lock(&m[i], &m_mtx); 514 if ((fault_flags & VM_FAULT_WIRE) != 0) 515 vm_page_wire(&m[i]); 516 else 517 vm_page_activate(&m[i]); 518 if (m_hold != NULL && m[i].pindex == fs->first_pindex) { 519 *m_hold = &m[i]; 520 vm_page_hold(&m[i]); 521 } 522 vm_page_xunbusy_maybelocked(&m[i]); 523 } 524 if (m_mtx != NULL) 525 mtx_unlock(m_mtx); 526 } 527 curthread->td_ru.ru_majflt++; 528 return (KERN_SUCCESS); 529 } 530 531 /* 532 * vm_fault: 533 * 534 * Handle a page fault occurring at the given address, 535 * requiring the given permissions, in the map specified. 536 * If successful, the page is inserted into the 537 * associated physical map. 538 * 539 * NOTE: the given address should be truncated to the 540 * proper page address. 541 * 542 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 543 * a standard error specifying why the fault is fatal is returned. 544 * 545 * The map in question must be referenced, and remains so. 546 * Caller may hold no locks. 547 */ 548 int 549 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 550 int fault_flags) 551 { 552 struct thread *td; 553 int result; 554 555 td = curthread; 556 if ((td->td_pflags & TDP_NOFAULTING) != 0) 557 return (KERN_PROTECTION_FAILURE); 558 #ifdef KTRACE 559 if (map != kernel_map && KTRPOINT(td, KTR_FAULT)) 560 ktrfault(vaddr, fault_type); 561 #endif 562 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags, 563 NULL); 564 #ifdef KTRACE 565 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND)) 566 ktrfaultend(result); 567 #endif 568 return (result); 569 } 570 571 int 572 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 573 int fault_flags, vm_page_t *m_hold) 574 { 575 struct faultstate fs; 576 struct vnode *vp; 577 struct domainset *dset; 578 vm_object_t next_object, retry_object; 579 vm_offset_t e_end, e_start; 580 vm_pindex_t retry_pindex; 581 vm_prot_t prot, retry_prot; 582 int ahead, alloc_req, behind, cluster_offset, error, era, faultcount; 583 int locked, nera, oom, result, rv; 584 u_char behavior; 585 boolean_t wired; /* Passed by reference. */ 586 bool dead, hardfault, is_first_object_locked; 587 588 VM_CNT_INC(v_vm_faults); 589 fs.vp = NULL; 590 faultcount = 0; 591 nera = -1; 592 hardfault = false; 593 594 RetryFault: 595 oom = 0; 596 RetryFault_oom: 597 598 /* 599 * Find the backing store object and offset into it to begin the 600 * search. 601 */ 602 fs.map = map; 603 result = vm_map_lookup(&fs.map, vaddr, fault_type | 604 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object, 605 &fs.first_pindex, &prot, &wired); 606 if (result != KERN_SUCCESS) { 607 unlock_vp(&fs); 608 return (result); 609 } 610 611 fs.map_generation = fs.map->timestamp; 612 613 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 614 panic("%s: fault on nofault entry, addr: %#lx", 615 __func__, (u_long)vaddr); 616 } 617 618 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && 619 fs.entry->wiring_thread != curthread) { 620 vm_map_unlock_read(fs.map); 621 vm_map_lock(fs.map); 622 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && 623 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 624 unlock_vp(&fs); 625 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 626 vm_map_unlock_and_wait(fs.map, 0); 627 } else 628 vm_map_unlock(fs.map); 629 goto RetryFault; 630 } 631 632 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0); 633 634 if (wired) 635 fault_type = prot | (fault_type & VM_PROT_COPY); 636 else 637 KASSERT((fault_flags & VM_FAULT_WIRE) == 0, 638 ("!wired && VM_FAULT_WIRE")); 639 640 /* 641 * Try to avoid lock contention on the top-level object through 642 * special-case handling of some types of page faults, specifically, 643 * those that are both (1) mapping an existing page from the top- 644 * level object and (2) not having to mark that object as containing 645 * dirty pages. Under these conditions, a read lock on the top-level 646 * object suffices, allowing multiple page faults of a similar type to 647 * run in parallel on the same top-level object. 648 */ 649 if (fs.vp == NULL /* avoid locked vnode leak */ && 650 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 && 651 /* avoid calling vm_object_set_writeable_dirty() */ 652 ((prot & VM_PROT_WRITE) == 0 || 653 (fs.first_object->type != OBJT_VNODE && 654 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 655 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) { 656 VM_OBJECT_RLOCK(fs.first_object); 657 if ((prot & VM_PROT_WRITE) == 0 || 658 (fs.first_object->type != OBJT_VNODE && 659 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 660 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) { 661 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type, 662 fault_flags, wired, m_hold); 663 if (rv == KERN_SUCCESS) 664 return (rv); 665 } 666 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { 667 VM_OBJECT_RUNLOCK(fs.first_object); 668 VM_OBJECT_WLOCK(fs.first_object); 669 } 670 } else { 671 VM_OBJECT_WLOCK(fs.first_object); 672 } 673 674 /* 675 * Make a reference to this object to prevent its disposal while we 676 * are messing with it. Once we have the reference, the map is free 677 * to be diddled. Since objects reference their shadows (and copies), 678 * they will stay around as well. 679 * 680 * Bump the paging-in-progress count to prevent size changes (e.g. 681 * truncation operations) during I/O. 682 */ 683 vm_object_reference_locked(fs.first_object); 684 vm_object_pip_add(fs.first_object, 1); 685 686 fs.lookup_still_valid = true; 687 688 fs.first_m = NULL; 689 690 /* 691 * Search for the page at object/offset. 692 */ 693 fs.object = fs.first_object; 694 fs.pindex = fs.first_pindex; 695 while (TRUE) { 696 /* 697 * If the object is marked for imminent termination, 698 * we retry here, since the collapse pass has raced 699 * with us. Otherwise, if we see terminally dead 700 * object, return fail. 701 */ 702 if ((fs.object->flags & OBJ_DEAD) != 0) { 703 dead = fs.object->type == OBJT_DEAD; 704 unlock_and_deallocate(&fs); 705 if (dead) 706 return (KERN_PROTECTION_FAILURE); 707 pause("vmf_de", 1); 708 goto RetryFault; 709 } 710 711 /* 712 * See if page is resident 713 */ 714 fs.m = vm_page_lookup(fs.object, fs.pindex); 715 if (fs.m != NULL) { 716 /* 717 * Wait/Retry if the page is busy. We have to do this 718 * if the page is either exclusive or shared busy 719 * because the vm_pager may be using read busy for 720 * pageouts (and even pageins if it is the vnode 721 * pager), and we could end up trying to pagein and 722 * pageout the same page simultaneously. 723 * 724 * We can theoretically allow the busy case on a read 725 * fault if the page is marked valid, but since such 726 * pages are typically already pmap'd, putting that 727 * special case in might be more effort then it is 728 * worth. We cannot under any circumstances mess 729 * around with a shared busied page except, perhaps, 730 * to pmap it. 731 */ 732 if (vm_page_busied(fs.m)) { 733 /* 734 * Reference the page before unlocking and 735 * sleeping so that the page daemon is less 736 * likely to reclaim it. 737 */ 738 vm_page_aflag_set(fs.m, PGA_REFERENCED); 739 if (fs.object != fs.first_object) { 740 if (!VM_OBJECT_TRYWLOCK( 741 fs.first_object)) { 742 VM_OBJECT_WUNLOCK(fs.object); 743 VM_OBJECT_WLOCK(fs.first_object); 744 VM_OBJECT_WLOCK(fs.object); 745 } 746 vm_page_lock(fs.first_m); 747 vm_page_free(fs.first_m); 748 vm_page_unlock(fs.first_m); 749 vm_object_pip_wakeup(fs.first_object); 750 VM_OBJECT_WUNLOCK(fs.first_object); 751 fs.first_m = NULL; 752 } 753 unlock_map(&fs); 754 if (fs.m == vm_page_lookup(fs.object, 755 fs.pindex)) { 756 vm_page_sleep_if_busy(fs.m, "vmpfw"); 757 } 758 vm_object_pip_wakeup(fs.object); 759 VM_OBJECT_WUNLOCK(fs.object); 760 VM_CNT_INC(v_intrans); 761 vm_object_deallocate(fs.first_object); 762 goto RetryFault; 763 } 764 765 /* 766 * Mark page busy for other processes, and the 767 * pagedaemon. If it still isn't completely valid 768 * (readable), jump to readrest, else break-out ( we 769 * found the page ). 770 */ 771 vm_page_xbusy(fs.m); 772 if (fs.m->valid != VM_PAGE_BITS_ALL) 773 goto readrest; 774 break; /* break to PAGE HAS BEEN FOUND */ 775 } 776 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m)); 777 778 /* 779 * Page is not resident. If the pager might contain the page 780 * or this is the beginning of the search, allocate a new 781 * page. (Default objects are zero-fill, so there is no real 782 * pager for them.) 783 */ 784 if (fs.object->type != OBJT_DEFAULT || 785 fs.object == fs.first_object) { 786 if (fs.pindex >= fs.object->size) { 787 unlock_and_deallocate(&fs); 788 return (KERN_PROTECTION_FAILURE); 789 } 790 791 if (fs.object == fs.first_object && 792 (fs.first_object->flags & OBJ_POPULATE) != 0 && 793 fs.first_object->shadow_count == 0) { 794 rv = vm_fault_populate(&fs, prot, fault_type, 795 fault_flags, wired, m_hold); 796 switch (rv) { 797 case KERN_SUCCESS: 798 case KERN_FAILURE: 799 unlock_and_deallocate(&fs); 800 return (rv); 801 case KERN_RESOURCE_SHORTAGE: 802 unlock_and_deallocate(&fs); 803 goto RetryFault; 804 case KERN_NOT_RECEIVER: 805 /* 806 * Pager's populate() method 807 * returned VM_PAGER_BAD. 808 */ 809 break; 810 default: 811 panic("inconsistent return codes"); 812 } 813 } 814 815 /* 816 * Allocate a new page for this object/offset pair. 817 * 818 * Unlocked read of the p_flag is harmless. At 819 * worst, the P_KILLED might be not observed 820 * there, and allocation can fail, causing 821 * restart and new reading of the p_flag. 822 */ 823 dset = fs.object->domain.dr_policy; 824 if (dset == NULL) 825 dset = curthread->td_domain.dr_policy; 826 if (!vm_page_count_severe_set(&dset->ds_mask) || 827 P_KILLED(curproc)) { 828 #if VM_NRESERVLEVEL > 0 829 vm_object_color(fs.object, atop(vaddr) - 830 fs.pindex); 831 #endif 832 alloc_req = P_KILLED(curproc) ? 833 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 834 if (fs.object->type != OBJT_VNODE && 835 fs.object->backing_object == NULL) 836 alloc_req |= VM_ALLOC_ZERO; 837 fs.m = vm_page_alloc(fs.object, fs.pindex, 838 alloc_req); 839 } 840 if (fs.m == NULL) { 841 unlock_and_deallocate(&fs); 842 if (vm_pfault_oom_attempts < 0 || 843 oom < vm_pfault_oom_attempts) { 844 oom++; 845 vm_waitpfault(dset, 846 vm_pfault_oom_wait * hz); 847 goto RetryFault_oom; 848 } 849 if (bootverbose) 850 printf( 851 "proc %d (%s) failed to alloc page on fault, starting OOM\n", 852 curproc->p_pid, curproc->p_comm); 853 vm_pageout_oom(VM_OOM_MEM_PF); 854 goto RetryFault; 855 } 856 } 857 858 readrest: 859 /* 860 * At this point, we have either allocated a new page or found 861 * an existing page that is only partially valid. 862 * 863 * We hold a reference on the current object and the page is 864 * exclusive busied. 865 */ 866 867 /* 868 * If the pager for the current object might have the page, 869 * then determine the number of additional pages to read and 870 * potentially reprioritize previously read pages for earlier 871 * reclamation. These operations should only be performed 872 * once per page fault. Even if the current pager doesn't 873 * have the page, the number of additional pages to read will 874 * apply to subsequent objects in the shadow chain. 875 */ 876 if (fs.object->type != OBJT_DEFAULT && nera == -1 && 877 !P_KILLED(curproc)) { 878 KASSERT(fs.lookup_still_valid, ("map unlocked")); 879 era = fs.entry->read_ahead; 880 behavior = vm_map_entry_behavior(fs.entry); 881 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 882 nera = 0; 883 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 884 nera = VM_FAULT_READ_AHEAD_MAX; 885 if (vaddr == fs.entry->next_read) 886 vm_fault_dontneed(&fs, vaddr, nera); 887 } else if (vaddr == fs.entry->next_read) { 888 /* 889 * This is a sequential fault. Arithmetically 890 * increase the requested number of pages in 891 * the read-ahead window. The requested 892 * number of pages is "# of sequential faults 893 * x (read ahead min + 1) + read ahead min" 894 */ 895 nera = VM_FAULT_READ_AHEAD_MIN; 896 if (era > 0) { 897 nera += era + 1; 898 if (nera > VM_FAULT_READ_AHEAD_MAX) 899 nera = VM_FAULT_READ_AHEAD_MAX; 900 } 901 if (era == VM_FAULT_READ_AHEAD_MAX) 902 vm_fault_dontneed(&fs, vaddr, nera); 903 } else { 904 /* 905 * This is a non-sequential fault. 906 */ 907 nera = 0; 908 } 909 if (era != nera) { 910 /* 911 * A read lock on the map suffices to update 912 * the read ahead count safely. 913 */ 914 fs.entry->read_ahead = nera; 915 } 916 917 /* 918 * Prepare for unlocking the map. Save the map 919 * entry's start and end addresses, which are used to 920 * optimize the size of the pager operation below. 921 * Even if the map entry's addresses change after 922 * unlocking the map, using the saved addresses is 923 * safe. 924 */ 925 e_start = fs.entry->start; 926 e_end = fs.entry->end; 927 } 928 929 /* 930 * Call the pager to retrieve the page if there is a chance 931 * that the pager has it, and potentially retrieve additional 932 * pages at the same time. 933 */ 934 if (fs.object->type != OBJT_DEFAULT) { 935 /* 936 * Release the map lock before locking the vnode or 937 * sleeping in the pager. (If the current object has 938 * a shadow, then an earlier iteration of this loop 939 * may have already unlocked the map.) 940 */ 941 unlock_map(&fs); 942 943 if (fs.object->type == OBJT_VNODE && 944 (vp = fs.object->handle) != fs.vp) { 945 /* 946 * Perform an unlock in case the desired vnode 947 * changed while the map was unlocked during a 948 * retry. 949 */ 950 unlock_vp(&fs); 951 952 locked = VOP_ISLOCKED(vp); 953 if (locked != LK_EXCLUSIVE) 954 locked = LK_SHARED; 955 956 /* 957 * We must not sleep acquiring the vnode lock 958 * while we have the page exclusive busied or 959 * the object's paging-in-progress count 960 * incremented. Otherwise, we could deadlock. 961 */ 962 error = vget(vp, locked | LK_CANRECURSE | 963 LK_NOWAIT, curthread); 964 if (error != 0) { 965 vhold(vp); 966 release_page(&fs); 967 unlock_and_deallocate(&fs); 968 error = vget(vp, locked | LK_RETRY | 969 LK_CANRECURSE, curthread); 970 vdrop(vp); 971 fs.vp = vp; 972 KASSERT(error == 0, 973 ("vm_fault: vget failed")); 974 goto RetryFault; 975 } 976 fs.vp = vp; 977 } 978 KASSERT(fs.vp == NULL || !fs.map->system_map, 979 ("vm_fault: vnode-backed object mapped by system map")); 980 981 /* 982 * Page in the requested page and hint the pager, 983 * that it may bring up surrounding pages. 984 */ 985 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || 986 P_KILLED(curproc)) { 987 behind = 0; 988 ahead = 0; 989 } else { 990 /* Is this a sequential fault? */ 991 if (nera > 0) { 992 behind = 0; 993 ahead = nera; 994 } else { 995 /* 996 * Request a cluster of pages that is 997 * aligned to a VM_FAULT_READ_DEFAULT 998 * page offset boundary within the 999 * object. Alignment to a page offset 1000 * boundary is more likely to coincide 1001 * with the underlying file system 1002 * block than alignment to a virtual 1003 * address boundary. 1004 */ 1005 cluster_offset = fs.pindex % 1006 VM_FAULT_READ_DEFAULT; 1007 behind = ulmin(cluster_offset, 1008 atop(vaddr - e_start)); 1009 ahead = VM_FAULT_READ_DEFAULT - 1 - 1010 cluster_offset; 1011 } 1012 ahead = ulmin(ahead, atop(e_end - vaddr) - 1); 1013 } 1014 rv = vm_pager_get_pages(fs.object, &fs.m, 1, 1015 &behind, &ahead); 1016 if (rv == VM_PAGER_OK) { 1017 faultcount = behind + 1 + ahead; 1018 hardfault = true; 1019 break; /* break to PAGE HAS BEEN FOUND */ 1020 } 1021 if (rv == VM_PAGER_ERROR) 1022 printf("vm_fault: pager read error, pid %d (%s)\n", 1023 curproc->p_pid, curproc->p_comm); 1024 1025 /* 1026 * If an I/O error occurred or the requested page was 1027 * outside the range of the pager, clean up and return 1028 * an error. 1029 */ 1030 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { 1031 vm_page_lock(fs.m); 1032 if (!vm_page_wired(fs.m)) 1033 vm_page_free(fs.m); 1034 else 1035 vm_page_xunbusy_maybelocked(fs.m); 1036 vm_page_unlock(fs.m); 1037 fs.m = NULL; 1038 unlock_and_deallocate(&fs); 1039 return (rv == VM_PAGER_ERROR ? KERN_FAILURE : 1040 KERN_PROTECTION_FAILURE); 1041 } 1042 1043 /* 1044 * The requested page does not exist at this object/ 1045 * offset. Remove the invalid page from the object, 1046 * waking up anyone waiting for it, and continue on to 1047 * the next object. However, if this is the top-level 1048 * object, we must leave the busy page in place to 1049 * prevent another process from rushing past us, and 1050 * inserting the page in that object at the same time 1051 * that we are. 1052 */ 1053 if (fs.object != fs.first_object) { 1054 vm_page_lock(fs.m); 1055 if (!vm_page_wired(fs.m)) 1056 vm_page_free(fs.m); 1057 else 1058 vm_page_xunbusy_maybelocked(fs.m); 1059 vm_page_unlock(fs.m); 1060 fs.m = NULL; 1061 } 1062 } 1063 1064 /* 1065 * We get here if the object has default pager (or unwiring) 1066 * or the pager doesn't have the page. 1067 */ 1068 if (fs.object == fs.first_object) 1069 fs.first_m = fs.m; 1070 1071 /* 1072 * Move on to the next object. Lock the next object before 1073 * unlocking the current one. 1074 */ 1075 next_object = fs.object->backing_object; 1076 if (next_object == NULL) { 1077 /* 1078 * If there's no object left, fill the page in the top 1079 * object with zeros. 1080 */ 1081 if (fs.object != fs.first_object) { 1082 vm_object_pip_wakeup(fs.object); 1083 VM_OBJECT_WUNLOCK(fs.object); 1084 1085 fs.object = fs.first_object; 1086 fs.pindex = fs.first_pindex; 1087 fs.m = fs.first_m; 1088 VM_OBJECT_WLOCK(fs.object); 1089 } 1090 fs.first_m = NULL; 1091 1092 /* 1093 * Zero the page if necessary and mark it valid. 1094 */ 1095 if ((fs.m->flags & PG_ZERO) == 0) { 1096 pmap_zero_page(fs.m); 1097 } else { 1098 VM_CNT_INC(v_ozfod); 1099 } 1100 VM_CNT_INC(v_zfod); 1101 fs.m->valid = VM_PAGE_BITS_ALL; 1102 /* Don't try to prefault neighboring pages. */ 1103 faultcount = 1; 1104 break; /* break to PAGE HAS BEEN FOUND */ 1105 } else { 1106 KASSERT(fs.object != next_object, 1107 ("object loop %p", next_object)); 1108 VM_OBJECT_WLOCK(next_object); 1109 vm_object_pip_add(next_object, 1); 1110 if (fs.object != fs.first_object) 1111 vm_object_pip_wakeup(fs.object); 1112 fs.pindex += 1113 OFF_TO_IDX(fs.object->backing_object_offset); 1114 VM_OBJECT_WUNLOCK(fs.object); 1115 fs.object = next_object; 1116 } 1117 } 1118 1119 vm_page_assert_xbusied(fs.m); 1120 1121 /* 1122 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1123 * is held.] 1124 */ 1125 1126 /* 1127 * If the page is being written, but isn't already owned by the 1128 * top-level object, we have to copy it into a new page owned by the 1129 * top-level object. 1130 */ 1131 if (fs.object != fs.first_object) { 1132 /* 1133 * We only really need to copy if we want to write it. 1134 */ 1135 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1136 /* 1137 * This allows pages to be virtually copied from a 1138 * backing_object into the first_object, where the 1139 * backing object has no other refs to it, and cannot 1140 * gain any more refs. Instead of a bcopy, we just 1141 * move the page from the backing object to the 1142 * first object. Note that we must mark the page 1143 * dirty in the first object so that it will go out 1144 * to swap when needed. 1145 */ 1146 is_first_object_locked = false; 1147 if ( 1148 /* 1149 * Only one shadow object 1150 */ 1151 (fs.object->shadow_count == 1) && 1152 /* 1153 * No COW refs, except us 1154 */ 1155 (fs.object->ref_count == 1) && 1156 /* 1157 * No one else can look this object up 1158 */ 1159 (fs.object->handle == NULL) && 1160 /* 1161 * No other ways to look the object up 1162 */ 1163 ((fs.object->type == OBJT_DEFAULT) || 1164 (fs.object->type == OBJT_SWAP)) && 1165 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && 1166 /* 1167 * We don't chase down the shadow chain 1168 */ 1169 fs.object == fs.first_object->backing_object) { 1170 vm_page_lock(fs.m); 1171 vm_page_dequeue(fs.m); 1172 (void)vm_page_remove(fs.m); 1173 vm_page_unlock(fs.m); 1174 vm_page_lock(fs.first_m); 1175 vm_page_replace_checked(fs.m, fs.first_object, 1176 fs.first_pindex, fs.first_m); 1177 vm_page_free(fs.first_m); 1178 vm_page_unlock(fs.first_m); 1179 vm_page_dirty(fs.m); 1180 #if VM_NRESERVLEVEL > 0 1181 /* 1182 * Rename the reservation. 1183 */ 1184 vm_reserv_rename(fs.m, fs.first_object, 1185 fs.object, OFF_TO_IDX( 1186 fs.first_object->backing_object_offset)); 1187 #endif 1188 /* 1189 * Removing the page from the backing object 1190 * unbusied it. 1191 */ 1192 vm_page_xbusy(fs.m); 1193 fs.first_m = fs.m; 1194 fs.m = NULL; 1195 VM_CNT_INC(v_cow_optim); 1196 } else { 1197 /* 1198 * Oh, well, lets copy it. 1199 */ 1200 pmap_copy_page(fs.m, fs.first_m); 1201 fs.first_m->valid = VM_PAGE_BITS_ALL; 1202 if (wired && (fault_flags & 1203 VM_FAULT_WIRE) == 0) { 1204 vm_page_lock(fs.first_m); 1205 vm_page_wire(fs.first_m); 1206 vm_page_unlock(fs.first_m); 1207 1208 vm_page_lock(fs.m); 1209 vm_page_unwire(fs.m, PQ_INACTIVE); 1210 vm_page_unlock(fs.m); 1211 } 1212 /* 1213 * We no longer need the old page or object. 1214 */ 1215 release_page(&fs); 1216 } 1217 /* 1218 * fs.object != fs.first_object due to above 1219 * conditional 1220 */ 1221 vm_object_pip_wakeup(fs.object); 1222 VM_OBJECT_WUNLOCK(fs.object); 1223 /* 1224 * Only use the new page below... 1225 */ 1226 fs.object = fs.first_object; 1227 fs.pindex = fs.first_pindex; 1228 fs.m = fs.first_m; 1229 if (!is_first_object_locked) 1230 VM_OBJECT_WLOCK(fs.object); 1231 VM_CNT_INC(v_cow_faults); 1232 curthread->td_cow++; 1233 } else { 1234 prot &= ~VM_PROT_WRITE; 1235 } 1236 } 1237 1238 /* 1239 * We must verify that the maps have not changed since our last 1240 * lookup. 1241 */ 1242 if (!fs.lookup_still_valid) { 1243 if (!vm_map_trylock_read(fs.map)) { 1244 release_page(&fs); 1245 unlock_and_deallocate(&fs); 1246 goto RetryFault; 1247 } 1248 fs.lookup_still_valid = true; 1249 if (fs.map->timestamp != fs.map_generation) { 1250 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 1251 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 1252 1253 /* 1254 * If we don't need the page any longer, put it on the inactive 1255 * list (the easiest thing to do here). If no one needs it, 1256 * pageout will grab it eventually. 1257 */ 1258 if (result != KERN_SUCCESS) { 1259 release_page(&fs); 1260 unlock_and_deallocate(&fs); 1261 1262 /* 1263 * If retry of map lookup would have blocked then 1264 * retry fault from start. 1265 */ 1266 if (result == KERN_FAILURE) 1267 goto RetryFault; 1268 return (result); 1269 } 1270 if ((retry_object != fs.first_object) || 1271 (retry_pindex != fs.first_pindex)) { 1272 release_page(&fs); 1273 unlock_and_deallocate(&fs); 1274 goto RetryFault; 1275 } 1276 1277 /* 1278 * Check whether the protection has changed or the object has 1279 * been copied while we left the map unlocked. Changing from 1280 * read to write permission is OK - we leave the page 1281 * write-protected, and catch the write fault. Changing from 1282 * write to read permission means that we can't mark the page 1283 * write-enabled after all. 1284 */ 1285 prot &= retry_prot; 1286 fault_type &= retry_prot; 1287 if (prot == 0) { 1288 release_page(&fs); 1289 unlock_and_deallocate(&fs); 1290 goto RetryFault; 1291 } 1292 1293 /* Reassert because wired may have changed. */ 1294 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0, 1295 ("!wired && VM_FAULT_WIRE")); 1296 } 1297 } 1298 1299 /* 1300 * If the page was filled by a pager, save the virtual address that 1301 * should be faulted on next under a sequential access pattern to the 1302 * map entry. A read lock on the map suffices to update this address 1303 * safely. 1304 */ 1305 if (hardfault) 1306 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; 1307 1308 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true); 1309 vm_page_assert_xbusied(fs.m); 1310 1311 /* 1312 * Page must be completely valid or it is not fit to 1313 * map into user space. vm_pager_get_pages() ensures this. 1314 */ 1315 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, 1316 ("vm_fault: page %p partially invalid", fs.m)); 1317 VM_OBJECT_WUNLOCK(fs.object); 1318 1319 /* 1320 * Put this page into the physical map. We had to do the unlock above 1321 * because pmap_enter() may sleep. We don't put the page 1322 * back on the active queue until later so that the pageout daemon 1323 * won't find it (yet). 1324 */ 1325 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, 1326 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); 1327 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 && 1328 wired == 0) 1329 vm_fault_prefault(&fs, vaddr, 1330 faultcount > 0 ? behind : PFBAK, 1331 faultcount > 0 ? ahead : PFFOR, false); 1332 VM_OBJECT_WLOCK(fs.object); 1333 vm_page_lock(fs.m); 1334 1335 /* 1336 * If the page is not wired down, then put it where the pageout daemon 1337 * can find it. 1338 */ 1339 if ((fault_flags & VM_FAULT_WIRE) != 0) 1340 vm_page_wire(fs.m); 1341 else 1342 vm_page_activate(fs.m); 1343 if (m_hold != NULL) { 1344 *m_hold = fs.m; 1345 vm_page_hold(fs.m); 1346 } 1347 vm_page_unlock(fs.m); 1348 vm_page_xunbusy(fs.m); 1349 1350 /* 1351 * Unlock everything, and return 1352 */ 1353 unlock_and_deallocate(&fs); 1354 if (hardfault) { 1355 VM_CNT_INC(v_io_faults); 1356 curthread->td_ru.ru_majflt++; 1357 #ifdef RACCT 1358 if (racct_enable && fs.object->type == OBJT_VNODE) { 1359 PROC_LOCK(curproc); 1360 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1361 racct_add_force(curproc, RACCT_WRITEBPS, 1362 PAGE_SIZE + behind * PAGE_SIZE); 1363 racct_add_force(curproc, RACCT_WRITEIOPS, 1); 1364 } else { 1365 racct_add_force(curproc, RACCT_READBPS, 1366 PAGE_SIZE + ahead * PAGE_SIZE); 1367 racct_add_force(curproc, RACCT_READIOPS, 1); 1368 } 1369 PROC_UNLOCK(curproc); 1370 } 1371 #endif 1372 } else 1373 curthread->td_ru.ru_minflt++; 1374 1375 return (KERN_SUCCESS); 1376 } 1377 1378 /* 1379 * Speed up the reclamation of pages that precede the faulting pindex within 1380 * the first object of the shadow chain. Essentially, perform the equivalent 1381 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes 1382 * the faulting pindex by the cluster size when the pages read by vm_fault() 1383 * cross a cluster-size boundary. The cluster size is the greater of the 1384 * smallest superpage size and VM_FAULT_DONTNEED_MIN. 1385 * 1386 * When "fs->first_object" is a shadow object, the pages in the backing object 1387 * that precede the faulting pindex are deactivated by vm_fault(). So, this 1388 * function must only be concerned with pages in the first object. 1389 */ 1390 static void 1391 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) 1392 { 1393 vm_map_entry_t entry; 1394 vm_object_t first_object, object; 1395 vm_offset_t end, start; 1396 vm_page_t m, m_next; 1397 vm_pindex_t pend, pstart; 1398 vm_size_t size; 1399 1400 object = fs->object; 1401 VM_OBJECT_ASSERT_WLOCKED(object); 1402 first_object = fs->first_object; 1403 if (first_object != object) { 1404 if (!VM_OBJECT_TRYWLOCK(first_object)) { 1405 VM_OBJECT_WUNLOCK(object); 1406 VM_OBJECT_WLOCK(first_object); 1407 VM_OBJECT_WLOCK(object); 1408 } 1409 } 1410 /* Neither fictitious nor unmanaged pages can be reclaimed. */ 1411 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 1412 size = VM_FAULT_DONTNEED_MIN; 1413 if (MAXPAGESIZES > 1 && size < pagesizes[1]) 1414 size = pagesizes[1]; 1415 end = rounddown2(vaddr, size); 1416 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && 1417 (entry = fs->entry)->start < end) { 1418 if (end - entry->start < size) 1419 start = entry->start; 1420 else 1421 start = end - size; 1422 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); 1423 pstart = OFF_TO_IDX(entry->offset) + atop(start - 1424 entry->start); 1425 m_next = vm_page_find_least(first_object, pstart); 1426 pend = OFF_TO_IDX(entry->offset) + atop(end - 1427 entry->start); 1428 while ((m = m_next) != NULL && m->pindex < pend) { 1429 m_next = TAILQ_NEXT(m, listq); 1430 if (m->valid != VM_PAGE_BITS_ALL || 1431 vm_page_busied(m)) 1432 continue; 1433 1434 /* 1435 * Don't clear PGA_REFERENCED, since it would 1436 * likely represent a reference by a different 1437 * process. 1438 * 1439 * Typically, at this point, prefetched pages 1440 * are still in the inactive queue. Only 1441 * pages that triggered page faults are in the 1442 * active queue. 1443 */ 1444 vm_page_lock(m); 1445 if (!vm_page_inactive(m)) 1446 vm_page_deactivate(m); 1447 vm_page_unlock(m); 1448 } 1449 } 1450 } 1451 if (first_object != object) 1452 VM_OBJECT_WUNLOCK(first_object); 1453 } 1454 1455 /* 1456 * vm_fault_prefault provides a quick way of clustering 1457 * pagefaults into a processes address space. It is a "cousin" 1458 * of vm_map_pmap_enter, except it runs at page fault time instead 1459 * of mmap time. 1460 */ 1461 static void 1462 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 1463 int backward, int forward, bool obj_locked) 1464 { 1465 pmap_t pmap; 1466 vm_map_entry_t entry; 1467 vm_object_t backing_object, lobject; 1468 vm_offset_t addr, starta; 1469 vm_pindex_t pindex; 1470 vm_page_t m; 1471 int i; 1472 1473 pmap = fs->map->pmap; 1474 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1475 return; 1476 1477 entry = fs->entry; 1478 1479 if (addra < backward * PAGE_SIZE) { 1480 starta = entry->start; 1481 } else { 1482 starta = addra - backward * PAGE_SIZE; 1483 if (starta < entry->start) 1484 starta = entry->start; 1485 } 1486 1487 /* 1488 * Generate the sequence of virtual addresses that are candidates for 1489 * prefaulting in an outward spiral from the faulting virtual address, 1490 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra 1491 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... 1492 * If the candidate address doesn't have a backing physical page, then 1493 * the loop immediately terminates. 1494 */ 1495 for (i = 0; i < 2 * imax(backward, forward); i++) { 1496 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : 1497 PAGE_SIZE); 1498 if (addr > addra + forward * PAGE_SIZE) 1499 addr = 0; 1500 1501 if (addr < starta || addr >= entry->end) 1502 continue; 1503 1504 if (!pmap_is_prefaultable(pmap, addr)) 1505 continue; 1506 1507 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1508 lobject = entry->object.vm_object; 1509 if (!obj_locked) 1510 VM_OBJECT_RLOCK(lobject); 1511 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1512 lobject->type == OBJT_DEFAULT && 1513 (backing_object = lobject->backing_object) != NULL) { 1514 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1515 0, ("vm_fault_prefault: unaligned object offset")); 1516 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1517 VM_OBJECT_RLOCK(backing_object); 1518 if (!obj_locked || lobject != entry->object.vm_object) 1519 VM_OBJECT_RUNLOCK(lobject); 1520 lobject = backing_object; 1521 } 1522 if (m == NULL) { 1523 if (!obj_locked || lobject != entry->object.vm_object) 1524 VM_OBJECT_RUNLOCK(lobject); 1525 break; 1526 } 1527 if (m->valid == VM_PAGE_BITS_ALL && 1528 (m->flags & PG_FICTITIOUS) == 0) 1529 pmap_enter_quick(pmap, addr, m, entry->protection); 1530 if (!obj_locked || lobject != entry->object.vm_object) 1531 VM_OBJECT_RUNLOCK(lobject); 1532 } 1533 } 1534 1535 /* 1536 * Hold each of the physical pages that are mapped by the specified range of 1537 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1538 * and allow the specified types of access, "prot". If all of the implied 1539 * pages are successfully held, then the number of held pages is returned 1540 * together with pointers to those pages in the array "ma". However, if any 1541 * of the pages cannot be held, -1 is returned. 1542 */ 1543 int 1544 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1545 vm_prot_t prot, vm_page_t *ma, int max_count) 1546 { 1547 vm_offset_t end, va; 1548 vm_page_t *mp; 1549 int count; 1550 boolean_t pmap_failed; 1551 1552 if (len == 0) 1553 return (0); 1554 end = round_page(addr + len); 1555 addr = trunc_page(addr); 1556 1557 /* 1558 * Check for illegal addresses. 1559 */ 1560 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) 1561 return (-1); 1562 1563 if (atop(end - addr) > max_count) 1564 panic("vm_fault_quick_hold_pages: count > max_count"); 1565 count = atop(end - addr); 1566 1567 /* 1568 * Most likely, the physical pages are resident in the pmap, so it is 1569 * faster to try pmap_extract_and_hold() first. 1570 */ 1571 pmap_failed = FALSE; 1572 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1573 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1574 if (*mp == NULL) 1575 pmap_failed = TRUE; 1576 else if ((prot & VM_PROT_WRITE) != 0 && 1577 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1578 /* 1579 * Explicitly dirty the physical page. Otherwise, the 1580 * caller's changes may go unnoticed because they are 1581 * performed through an unmanaged mapping or by a DMA 1582 * operation. 1583 * 1584 * The object lock is not held here. 1585 * See vm_page_clear_dirty_mask(). 1586 */ 1587 vm_page_dirty(*mp); 1588 } 1589 } 1590 if (pmap_failed) { 1591 /* 1592 * One or more pages could not be held by the pmap. Either no 1593 * page was mapped at the specified virtual address or that 1594 * mapping had insufficient permissions. Attempt to fault in 1595 * and hold these pages. 1596 * 1597 * If vm_fault_disable_pagefaults() was called, 1598 * i.e., TDP_NOFAULTING is set, we must not sleep nor 1599 * acquire MD VM locks, which means we must not call 1600 * vm_fault_hold(). Some (out of tree) callers mark 1601 * too wide a code area with vm_fault_disable_pagefaults() 1602 * already, use the VM_PROT_QUICK_NOFAULT flag to request 1603 * the proper behaviour explicitly. 1604 */ 1605 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && 1606 (curthread->td_pflags & TDP_NOFAULTING) != 0) 1607 goto error; 1608 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1609 if (*mp == NULL && vm_fault_hold(map, va, prot, 1610 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1611 goto error; 1612 } 1613 return (count); 1614 error: 1615 for (mp = ma; mp < ma + count; mp++) 1616 if (*mp != NULL) { 1617 vm_page_lock(*mp); 1618 vm_page_unhold(*mp); 1619 vm_page_unlock(*mp); 1620 } 1621 return (-1); 1622 } 1623 1624 /* 1625 * Routine: 1626 * vm_fault_copy_entry 1627 * Function: 1628 * Create new shadow object backing dst_entry with private copy of 1629 * all underlying pages. When src_entry is equal to dst_entry, 1630 * function implements COW for wired-down map entry. Otherwise, 1631 * it forks wired entry into dst_map. 1632 * 1633 * In/out conditions: 1634 * The source and destination maps must be locked for write. 1635 * The source map entry must be wired down (or be a sharing map 1636 * entry corresponding to a main map entry that is wired down). 1637 */ 1638 void 1639 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1640 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1641 vm_ooffset_t *fork_charge) 1642 { 1643 vm_object_t backing_object, dst_object, object, src_object; 1644 vm_pindex_t dst_pindex, pindex, src_pindex; 1645 vm_prot_t access, prot; 1646 vm_offset_t vaddr; 1647 vm_page_t dst_m; 1648 vm_page_t src_m; 1649 boolean_t upgrade; 1650 1651 #ifdef lint 1652 src_map++; 1653 #endif /* lint */ 1654 1655 upgrade = src_entry == dst_entry; 1656 access = prot = dst_entry->protection; 1657 1658 src_object = src_entry->object.vm_object; 1659 src_pindex = OFF_TO_IDX(src_entry->offset); 1660 1661 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 1662 dst_object = src_object; 1663 vm_object_reference(dst_object); 1664 } else { 1665 /* 1666 * Create the top-level object for the destination entry. (Doesn't 1667 * actually shadow anything - we copy the pages directly.) 1668 */ 1669 dst_object = vm_object_allocate(OBJT_DEFAULT, 1670 atop(dst_entry->end - dst_entry->start)); 1671 #if VM_NRESERVLEVEL > 0 1672 dst_object->flags |= OBJ_COLORED; 1673 dst_object->pg_color = atop(dst_entry->start); 1674 #endif 1675 dst_object->domain = src_object->domain; 1676 dst_object->charge = dst_entry->end - dst_entry->start; 1677 } 1678 1679 VM_OBJECT_WLOCK(dst_object); 1680 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1681 ("vm_fault_copy_entry: vm_object not NULL")); 1682 if (src_object != dst_object) { 1683 dst_entry->object.vm_object = dst_object; 1684 dst_entry->offset = 0; 1685 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; 1686 } 1687 if (fork_charge != NULL) { 1688 KASSERT(dst_entry->cred == NULL, 1689 ("vm_fault_copy_entry: leaked swp charge")); 1690 dst_object->cred = curthread->td_ucred; 1691 crhold(dst_object->cred); 1692 *fork_charge += dst_object->charge; 1693 } else if ((dst_object->type == OBJT_DEFAULT || 1694 dst_object->type == OBJT_SWAP) && 1695 dst_object->cred == NULL) { 1696 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", 1697 dst_entry)); 1698 dst_object->cred = dst_entry->cred; 1699 dst_entry->cred = NULL; 1700 } 1701 1702 /* 1703 * If not an upgrade, then enter the mappings in the pmap as 1704 * read and/or execute accesses. Otherwise, enter them as 1705 * write accesses. 1706 * 1707 * A writeable large page mapping is only created if all of 1708 * the constituent small page mappings are modified. Marking 1709 * PTEs as modified on inception allows promotion to happen 1710 * without taking potentially large number of soft faults. 1711 */ 1712 if (!upgrade) 1713 access &= ~VM_PROT_WRITE; 1714 1715 /* 1716 * Loop through all of the virtual pages within the entry's 1717 * range, copying each page from the source object to the 1718 * destination object. Since the source is wired, those pages 1719 * must exist. In contrast, the destination is pageable. 1720 * Since the destination object doesn't share any backing storage 1721 * with the source object, all of its pages must be dirtied, 1722 * regardless of whether they can be written. 1723 */ 1724 for (vaddr = dst_entry->start, dst_pindex = 0; 1725 vaddr < dst_entry->end; 1726 vaddr += PAGE_SIZE, dst_pindex++) { 1727 again: 1728 /* 1729 * Find the page in the source object, and copy it in. 1730 * Because the source is wired down, the page will be 1731 * in memory. 1732 */ 1733 if (src_object != dst_object) 1734 VM_OBJECT_RLOCK(src_object); 1735 object = src_object; 1736 pindex = src_pindex + dst_pindex; 1737 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1738 (backing_object = object->backing_object) != NULL) { 1739 /* 1740 * Unless the source mapping is read-only or 1741 * it is presently being upgraded from 1742 * read-only, the first object in the shadow 1743 * chain should provide all of the pages. In 1744 * other words, this loop body should never be 1745 * executed when the source mapping is already 1746 * read/write. 1747 */ 1748 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || 1749 upgrade, 1750 ("vm_fault_copy_entry: main object missing page")); 1751 1752 VM_OBJECT_RLOCK(backing_object); 1753 pindex += OFF_TO_IDX(object->backing_object_offset); 1754 if (object != dst_object) 1755 VM_OBJECT_RUNLOCK(object); 1756 object = backing_object; 1757 } 1758 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); 1759 1760 if (object != dst_object) { 1761 /* 1762 * Allocate a page in the destination object. 1763 */ 1764 dst_m = vm_page_alloc(dst_object, (src_object == 1765 dst_object ? src_pindex : 0) + dst_pindex, 1766 VM_ALLOC_NORMAL); 1767 if (dst_m == NULL) { 1768 VM_OBJECT_WUNLOCK(dst_object); 1769 VM_OBJECT_RUNLOCK(object); 1770 vm_wait(dst_object); 1771 VM_OBJECT_WLOCK(dst_object); 1772 goto again; 1773 } 1774 pmap_copy_page(src_m, dst_m); 1775 VM_OBJECT_RUNLOCK(object); 1776 dst_m->dirty = dst_m->valid = src_m->valid; 1777 } else { 1778 dst_m = src_m; 1779 if (vm_page_sleep_if_busy(dst_m, "fltupg")) 1780 goto again; 1781 if (dst_m->pindex >= dst_object->size) 1782 /* 1783 * We are upgrading. Index can occur 1784 * out of bounds if the object type is 1785 * vnode and the file was truncated. 1786 */ 1787 break; 1788 vm_page_xbusy(dst_m); 1789 } 1790 VM_OBJECT_WUNLOCK(dst_object); 1791 1792 /* 1793 * Enter it in the pmap. If a wired, copy-on-write 1794 * mapping is being replaced by a write-enabled 1795 * mapping, then wire that new mapping. 1796 * 1797 * The page can be invalid if the user called 1798 * msync(MS_INVALIDATE) or truncated the backing vnode 1799 * or shared memory object. In this case, do not 1800 * insert it into pmap, but still do the copy so that 1801 * all copies of the wired map entry have similar 1802 * backing pages. 1803 */ 1804 if (dst_m->valid == VM_PAGE_BITS_ALL) { 1805 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, 1806 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); 1807 } 1808 1809 /* 1810 * Mark it no longer busy, and put it on the active list. 1811 */ 1812 VM_OBJECT_WLOCK(dst_object); 1813 1814 if (upgrade) { 1815 if (src_m != dst_m) { 1816 vm_page_lock(src_m); 1817 vm_page_unwire(src_m, PQ_INACTIVE); 1818 vm_page_unlock(src_m); 1819 vm_page_lock(dst_m); 1820 vm_page_wire(dst_m); 1821 vm_page_unlock(dst_m); 1822 } else { 1823 KASSERT(vm_page_wired(dst_m), 1824 ("dst_m %p is not wired", dst_m)); 1825 } 1826 } else { 1827 vm_page_lock(dst_m); 1828 vm_page_activate(dst_m); 1829 vm_page_unlock(dst_m); 1830 } 1831 vm_page_xunbusy(dst_m); 1832 } 1833 VM_OBJECT_WUNLOCK(dst_object); 1834 if (upgrade) { 1835 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1836 vm_object_deallocate(src_object); 1837 } 1838 } 1839 1840 /* 1841 * Block entry into the machine-independent layer's page fault handler by 1842 * the calling thread. Subsequent calls to vm_fault() by that thread will 1843 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 1844 * spurious page faults. 1845 */ 1846 int 1847 vm_fault_disable_pagefaults(void) 1848 { 1849 1850 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 1851 } 1852 1853 void 1854 vm_fault_enable_pagefaults(int save) 1855 { 1856 1857 curthread_pflags_restore(save); 1858 } 1859