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/refcount.h> 91 #include <sys/resourcevar.h> 92 #include <sys/rwlock.h> 93 #include <sys/signalvar.h> 94 #include <sys/sysctl.h> 95 #include <sys/sysent.h> 96 #include <sys/vmmeter.h> 97 #include <sys/vnode.h> 98 #ifdef KTRACE 99 #include <sys/ktrace.h> 100 #endif 101 102 #include <vm/vm.h> 103 #include <vm/vm_param.h> 104 #include <vm/pmap.h> 105 #include <vm/vm_map.h> 106 #include <vm/vm_object.h> 107 #include <vm/vm_page.h> 108 #include <vm/vm_pageout.h> 109 #include <vm/vm_kern.h> 110 #include <vm/vm_pager.h> 111 #include <vm/vm_extern.h> 112 #include <vm/vm_reserv.h> 113 114 #define PFBAK 4 115 #define PFFOR 4 116 117 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) 118 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) 119 120 #define VM_FAULT_DONTNEED_MIN 1048576 121 122 struct faultstate { 123 vm_page_t m; 124 vm_object_t object; 125 vm_pindex_t pindex; 126 vm_page_t first_m; 127 vm_object_t first_object; 128 vm_pindex_t first_pindex; 129 vm_map_t map; 130 vm_map_entry_t entry; 131 int map_generation; 132 bool lookup_still_valid; 133 struct vnode *vp; 134 }; 135 136 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, 137 int ahead); 138 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 139 int backward, int forward, bool obj_locked); 140 141 static int vm_pfault_oom_attempts = 3; 142 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, 143 &vm_pfault_oom_attempts, 0, 144 "Number of page allocation attempts in page fault handler before it " 145 "triggers OOM handling"); 146 147 static int vm_pfault_oom_wait = 10; 148 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, 149 &vm_pfault_oom_wait, 0, 150 "Number of seconds to wait for free pages before retrying " 151 "the page fault handler"); 152 153 static inline void 154 release_page(struct faultstate *fs) 155 { 156 157 vm_page_xunbusy(fs->m); 158 vm_page_lock(fs->m); 159 vm_page_deactivate(fs->m); 160 vm_page_unlock(fs->m); 161 fs->m = NULL; 162 } 163 164 static inline void 165 unlock_map(struct faultstate *fs) 166 { 167 168 if (fs->lookup_still_valid) { 169 vm_map_lookup_done(fs->map, fs->entry); 170 fs->lookup_still_valid = false; 171 } 172 } 173 174 static void 175 unlock_vp(struct faultstate *fs) 176 { 177 178 if (fs->vp != NULL) { 179 vput(fs->vp); 180 fs->vp = NULL; 181 } 182 } 183 184 static void 185 unlock_and_deallocate(struct faultstate *fs) 186 { 187 188 vm_object_pip_wakeup(fs->object); 189 VM_OBJECT_WUNLOCK(fs->object); 190 if (fs->object != fs->first_object) { 191 VM_OBJECT_WLOCK(fs->first_object); 192 vm_page_free(fs->first_m); 193 vm_object_pip_wakeup(fs->first_object); 194 VM_OBJECT_WUNLOCK(fs->first_object); 195 fs->first_m = NULL; 196 } 197 vm_object_deallocate(fs->first_object); 198 unlock_map(fs); 199 unlock_vp(fs); 200 } 201 202 static void 203 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, 204 vm_prot_t fault_type, int fault_flags, bool set_wd) 205 { 206 bool need_dirty; 207 208 if (((prot & VM_PROT_WRITE) == 0 && 209 (fault_flags & VM_FAULT_DIRTY) == 0) || 210 (m->oflags & VPO_UNMANAGED) != 0) 211 return; 212 213 VM_OBJECT_ASSERT_LOCKED(m->object); 214 VM_PAGE_OBJECT_BUSY_ASSERT(m); 215 216 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && 217 (fault_flags & VM_FAULT_WIRE) == 0) || 218 (fault_flags & VM_FAULT_DIRTY) != 0; 219 220 if (set_wd) 221 vm_object_set_writeable_dirty(m->object); 222 else 223 /* 224 * If two callers of vm_fault_dirty() with set_wd == 225 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC 226 * flag set, other with flag clear, race, it is 227 * possible for the no-NOSYNC thread to see m->dirty 228 * != 0 and not clear PGA_NOSYNC. Take vm_page lock 229 * around manipulation of PGA_NOSYNC and 230 * vm_page_dirty() call to avoid the race. 231 */ 232 vm_page_lock(m); 233 234 /* 235 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC 236 * if the page is already dirty to prevent data written with 237 * the expectation of being synced from not being synced. 238 * Likewise if this entry does not request NOSYNC then make 239 * sure the page isn't marked NOSYNC. Applications sharing 240 * data should use the same flags to avoid ping ponging. 241 */ 242 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) { 243 if (m->dirty == 0) { 244 vm_page_aflag_set(m, PGA_NOSYNC); 245 } 246 } else { 247 vm_page_aflag_clear(m, PGA_NOSYNC); 248 } 249 250 /* 251 * If the fault is a write, we know that this page is being 252 * written NOW so dirty it explicitly to save on 253 * pmap_is_modified() calls later. 254 * 255 * Also, since the page is now dirty, we can possibly tell 256 * the pager to release any swap backing the page. Calling 257 * the pager requires a write lock on the object. 258 */ 259 if (need_dirty) 260 vm_page_dirty(m); 261 if (!set_wd) 262 vm_page_unlock(m); 263 else if (need_dirty) 264 vm_pager_page_unswapped(m); 265 } 266 267 /* 268 * Unlocks fs.first_object and fs.map on success. 269 */ 270 static int 271 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot, 272 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) 273 { 274 vm_page_t m, m_map; 275 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 276 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ 277 VM_NRESERVLEVEL > 0 278 vm_page_t m_super; 279 int flags; 280 #endif 281 int psind, rv; 282 283 MPASS(fs->vp == NULL); 284 vm_object_busy(fs->first_object); 285 m = vm_page_lookup(fs->first_object, fs->first_pindex); 286 /* A busy page can be mapped for read|execute access. */ 287 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && 288 vm_page_busied(m)) || !vm_page_all_valid(m)) { 289 rv = KERN_FAILURE; 290 goto out; 291 } 292 m_map = m; 293 psind = 0; 294 #if (defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 295 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv)) && \ 296 VM_NRESERVLEVEL > 0 297 if ((m->flags & PG_FICTITIOUS) == 0 && 298 (m_super = vm_reserv_to_superpage(m)) != NULL && 299 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && 300 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && 301 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & 302 (pagesizes[m_super->psind] - 1)) && !wired && 303 pmap_ps_enabled(fs->map->pmap)) { 304 flags = PS_ALL_VALID; 305 if ((prot & VM_PROT_WRITE) != 0) { 306 /* 307 * Create a superpage mapping allowing write access 308 * only if none of the constituent pages are busy and 309 * all of them are already dirty (except possibly for 310 * the page that was faulted on). 311 */ 312 flags |= PS_NONE_BUSY; 313 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) 314 flags |= PS_ALL_DIRTY; 315 } 316 if (vm_page_ps_test(m_super, flags, m)) { 317 m_map = m_super; 318 psind = m_super->psind; 319 vaddr = rounddown2(vaddr, pagesizes[psind]); 320 /* Preset the modified bit for dirty superpages. */ 321 if ((flags & PS_ALL_DIRTY) != 0) 322 fault_type |= VM_PROT_WRITE; 323 } 324 } 325 #endif 326 rv = pmap_enter(fs->map->pmap, vaddr, m_map, prot, fault_type | 327 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), psind); 328 if (rv != KERN_SUCCESS) 329 goto out; 330 if (m_hold != NULL) { 331 *m_hold = m; 332 vm_page_wire(m); 333 } 334 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false); 335 if (psind == 0 && !wired) 336 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); 337 VM_OBJECT_RUNLOCK(fs->first_object); 338 vm_map_lookup_done(fs->map, fs->entry); 339 curthread->td_ru.ru_minflt++; 340 341 out: 342 vm_object_unbusy(fs->first_object); 343 return (rv); 344 } 345 346 static void 347 vm_fault_restore_map_lock(struct faultstate *fs) 348 { 349 350 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 351 MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0); 352 353 if (!vm_map_trylock_read(fs->map)) { 354 VM_OBJECT_WUNLOCK(fs->first_object); 355 vm_map_lock_read(fs->map); 356 VM_OBJECT_WLOCK(fs->first_object); 357 } 358 fs->lookup_still_valid = true; 359 } 360 361 static void 362 vm_fault_populate_check_page(vm_page_t m) 363 { 364 365 /* 366 * Check each page to ensure that the pager is obeying the 367 * interface: the page must be installed in the object, fully 368 * valid, and exclusively busied. 369 */ 370 MPASS(m != NULL); 371 MPASS(vm_page_all_valid(m)); 372 MPASS(vm_page_xbusied(m)); 373 } 374 375 static void 376 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, 377 vm_pindex_t last) 378 { 379 vm_page_t m; 380 vm_pindex_t pidx; 381 382 VM_OBJECT_ASSERT_WLOCKED(object); 383 MPASS(first <= last); 384 for (pidx = first, m = vm_page_lookup(object, pidx); 385 pidx <= last; pidx++, m = vm_page_next(m)) { 386 vm_fault_populate_check_page(m); 387 vm_page_lock(m); 388 vm_page_deactivate(m); 389 vm_page_unlock(m); 390 vm_page_xunbusy(m); 391 } 392 } 393 394 static int 395 vm_fault_populate(struct faultstate *fs, vm_prot_t prot, int fault_type, 396 int fault_flags, boolean_t wired, vm_page_t *m_hold) 397 { 398 struct mtx *m_mtx; 399 vm_offset_t vaddr; 400 vm_page_t m; 401 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; 402 int i, npages, psind, rv; 403 404 MPASS(fs->object == fs->first_object); 405 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 406 MPASS(REFCOUNT_COUNT(fs->first_object->paging_in_progress) > 0); 407 MPASS(fs->first_object->backing_object == NULL); 408 MPASS(fs->lookup_still_valid); 409 410 pager_first = OFF_TO_IDX(fs->entry->offset); 411 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; 412 unlock_map(fs); 413 unlock_vp(fs); 414 415 /* 416 * Call the pager (driver) populate() method. 417 * 418 * There is no guarantee that the method will be called again 419 * if the current fault is for read, and a future fault is 420 * for write. Report the entry's maximum allowed protection 421 * to the driver. 422 */ 423 rv = vm_pager_populate(fs->first_object, fs->first_pindex, 424 fault_type, fs->entry->max_protection, &pager_first, &pager_last); 425 426 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 427 if (rv == VM_PAGER_BAD) { 428 /* 429 * VM_PAGER_BAD is the backdoor for a pager to request 430 * normal fault handling. 431 */ 432 vm_fault_restore_map_lock(fs); 433 if (fs->map->timestamp != fs->map_generation) 434 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ 435 return (KERN_NOT_RECEIVER); 436 } 437 if (rv != VM_PAGER_OK) 438 return (KERN_FAILURE); /* AKA SIGSEGV */ 439 440 /* Ensure that the driver is obeying the interface. */ 441 MPASS(pager_first <= pager_last); 442 MPASS(fs->first_pindex <= pager_last); 443 MPASS(fs->first_pindex >= pager_first); 444 MPASS(pager_last < fs->first_object->size); 445 446 vm_fault_restore_map_lock(fs); 447 if (fs->map->timestamp != fs->map_generation) { 448 vm_fault_populate_cleanup(fs->first_object, pager_first, 449 pager_last); 450 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ 451 } 452 453 /* 454 * The map is unchanged after our last unlock. Process the fault. 455 * 456 * The range [pager_first, pager_last] that is given to the 457 * pager is only a hint. The pager may populate any range 458 * within the object that includes the requested page index. 459 * In case the pager expanded the range, clip it to fit into 460 * the map entry. 461 */ 462 map_first = OFF_TO_IDX(fs->entry->offset); 463 if (map_first > pager_first) { 464 vm_fault_populate_cleanup(fs->first_object, pager_first, 465 map_first - 1); 466 pager_first = map_first; 467 } 468 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; 469 if (map_last < pager_last) { 470 vm_fault_populate_cleanup(fs->first_object, map_last + 1, 471 pager_last); 472 pager_last = map_last; 473 } 474 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); 475 pidx <= pager_last; 476 pidx += npages, m = vm_page_next(&m[npages - 1])) { 477 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; 478 #if defined(__aarch64__) || defined(__amd64__) || (defined(__arm__) && \ 479 __ARM_ARCH >= 6) || defined(__i386__) || defined(__riscv) 480 psind = m->psind; 481 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || 482 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || 483 !pmap_ps_enabled(fs->map->pmap) || wired)) 484 psind = 0; 485 #else 486 psind = 0; 487 #endif 488 npages = atop(pagesizes[psind]); 489 for (i = 0; i < npages; i++) { 490 vm_fault_populate_check_page(&m[i]); 491 vm_fault_dirty(fs->entry, &m[i], prot, fault_type, 492 fault_flags, true); 493 } 494 VM_OBJECT_WUNLOCK(fs->first_object); 495 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | 496 (wired ? PMAP_ENTER_WIRED : 0), psind); 497 #if defined(__amd64__) 498 if (psind > 0 && rv == KERN_FAILURE) { 499 for (i = 0; i < npages; i++) { 500 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), 501 &m[i], prot, fault_type | 502 (wired ? PMAP_ENTER_WIRED : 0), 0); 503 MPASS(rv == KERN_SUCCESS); 504 } 505 } 506 #else 507 MPASS(rv == KERN_SUCCESS); 508 #endif 509 VM_OBJECT_WLOCK(fs->first_object); 510 m_mtx = NULL; 511 for (i = 0; i < npages; i++) { 512 if ((fault_flags & VM_FAULT_WIRE) != 0) { 513 vm_page_wire(&m[i]); 514 } else { 515 vm_page_change_lock(&m[i], &m_mtx); 516 vm_page_activate(&m[i]); 517 } 518 if (m_hold != NULL && m[i].pindex == fs->first_pindex) { 519 *m_hold = &m[i]; 520 vm_page_wire(&m[i]); 521 } 522 vm_page_xunbusy(&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 static int prot_fault_translation; 532 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN, 533 &prot_fault_translation, 0, 534 "Control signal to deliver on protection fault"); 535 536 /* compat definition to keep common code for signal translation */ 537 #define UCODE_PAGEFLT 12 538 #ifdef T_PAGEFLT 539 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT"); 540 #endif 541 542 /* 543 * vm_fault_trap: 544 * 545 * Handle a page fault occurring at the given address, 546 * requiring the given permissions, in the map specified. 547 * If successful, the page is inserted into the 548 * associated physical map. 549 * 550 * NOTE: the given address should be truncated to the 551 * proper page address. 552 * 553 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 554 * a standard error specifying why the fault is fatal is returned. 555 * 556 * The map in question must be referenced, and remains so. 557 * Caller may hold no locks. 558 */ 559 int 560 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 561 int fault_flags, int *signo, int *ucode) 562 { 563 int result; 564 565 MPASS(signo == NULL || ucode != NULL); 566 #ifdef KTRACE 567 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT)) 568 ktrfault(vaddr, fault_type); 569 #endif 570 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags, 571 NULL); 572 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE || 573 result == KERN_INVALID_ADDRESS || 574 result == KERN_RESOURCE_SHORTAGE || 575 result == KERN_PROTECTION_FAILURE || 576 result == KERN_OUT_OF_BOUNDS, 577 ("Unexpected Mach error %d from vm_fault()", result)); 578 #ifdef KTRACE 579 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND)) 580 ktrfaultend(result); 581 #endif 582 if (result != KERN_SUCCESS && signo != NULL) { 583 switch (result) { 584 case KERN_FAILURE: 585 case KERN_INVALID_ADDRESS: 586 *signo = SIGSEGV; 587 *ucode = SEGV_MAPERR; 588 break; 589 case KERN_RESOURCE_SHORTAGE: 590 *signo = SIGBUS; 591 *ucode = BUS_OOMERR; 592 break; 593 case KERN_OUT_OF_BOUNDS: 594 *signo = SIGBUS; 595 *ucode = BUS_OBJERR; 596 break; 597 case KERN_PROTECTION_FAILURE: 598 if (prot_fault_translation == 0) { 599 /* 600 * Autodetect. This check also covers 601 * the images without the ABI-tag ELF 602 * note. 603 */ 604 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD && 605 curproc->p_osrel >= P_OSREL_SIGSEGV) { 606 *signo = SIGSEGV; 607 *ucode = SEGV_ACCERR; 608 } else { 609 *signo = SIGBUS; 610 *ucode = UCODE_PAGEFLT; 611 } 612 } else if (prot_fault_translation == 1) { 613 /* Always compat mode. */ 614 *signo = SIGBUS; 615 *ucode = UCODE_PAGEFLT; 616 } else { 617 /* Always SIGSEGV mode. */ 618 *signo = SIGSEGV; 619 *ucode = SEGV_ACCERR; 620 } 621 break; 622 default: 623 KASSERT(0, ("Unexpected Mach error %d from vm_fault()", 624 result)); 625 break; 626 } 627 } 628 return (result); 629 } 630 631 int 632 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 633 int fault_flags, vm_page_t *m_hold) 634 { 635 struct faultstate fs; 636 struct vnode *vp; 637 struct domainset *dset; 638 vm_object_t next_object, retry_object; 639 vm_offset_t e_end, e_start; 640 vm_pindex_t retry_pindex; 641 vm_prot_t prot, retry_prot; 642 int ahead, alloc_req, behind, cluster_offset, error, era, faultcount; 643 int locked, nera, oom, result, rv; 644 u_char behavior; 645 boolean_t wired; /* Passed by reference. */ 646 bool dead, hardfault, is_first_object_locked; 647 648 VM_CNT_INC(v_vm_faults); 649 650 if ((curthread->td_pflags & TDP_NOFAULTING) != 0) 651 return (KERN_PROTECTION_FAILURE); 652 653 fs.vp = NULL; 654 faultcount = 0; 655 nera = -1; 656 hardfault = false; 657 658 RetryFault: 659 oom = 0; 660 RetryFault_oom: 661 662 /* 663 * Find the backing store object and offset into it to begin the 664 * search. 665 */ 666 fs.map = map; 667 result = vm_map_lookup(&fs.map, vaddr, fault_type | 668 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object, 669 &fs.first_pindex, &prot, &wired); 670 if (result != KERN_SUCCESS) { 671 unlock_vp(&fs); 672 return (result); 673 } 674 675 fs.map_generation = fs.map->timestamp; 676 677 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 678 panic("%s: fault on nofault entry, addr: %#lx", 679 __func__, (u_long)vaddr); 680 } 681 682 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && 683 fs.entry->wiring_thread != curthread) { 684 vm_map_unlock_read(fs.map); 685 vm_map_lock(fs.map); 686 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && 687 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 688 unlock_vp(&fs); 689 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 690 vm_map_unlock_and_wait(fs.map, 0); 691 } else 692 vm_map_unlock(fs.map); 693 goto RetryFault; 694 } 695 696 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0); 697 698 if (wired) 699 fault_type = prot | (fault_type & VM_PROT_COPY); 700 else 701 KASSERT((fault_flags & VM_FAULT_WIRE) == 0, 702 ("!wired && VM_FAULT_WIRE")); 703 704 /* 705 * Try to avoid lock contention on the top-level object through 706 * special-case handling of some types of page faults, specifically, 707 * those that are both (1) mapping an existing page from the top- 708 * level object and (2) not having to mark that object as containing 709 * dirty pages. Under these conditions, a read lock on the top-level 710 * object suffices, allowing multiple page faults of a similar type to 711 * run in parallel on the same top-level object. 712 */ 713 if (fs.vp == NULL /* avoid locked vnode leak */ && 714 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 && 715 /* avoid calling vm_object_set_writeable_dirty() */ 716 ((prot & VM_PROT_WRITE) == 0 || 717 (fs.first_object->type != OBJT_VNODE && 718 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 719 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) { 720 VM_OBJECT_RLOCK(fs.first_object); 721 if ((prot & VM_PROT_WRITE) == 0 || 722 (fs.first_object->type != OBJT_VNODE && 723 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 724 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) { 725 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type, 726 fault_flags, wired, m_hold); 727 if (rv == KERN_SUCCESS) 728 return (rv); 729 } 730 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { 731 VM_OBJECT_RUNLOCK(fs.first_object); 732 VM_OBJECT_WLOCK(fs.first_object); 733 } 734 } else { 735 VM_OBJECT_WLOCK(fs.first_object); 736 } 737 738 /* 739 * Make a reference to this object to prevent its disposal while we 740 * are messing with it. Once we have the reference, the map is free 741 * to be diddled. Since objects reference their shadows (and copies), 742 * they will stay around as well. 743 * 744 * Bump the paging-in-progress count to prevent size changes (e.g. 745 * truncation operations) during I/O. 746 */ 747 vm_object_reference_locked(fs.first_object); 748 vm_object_pip_add(fs.first_object, 1); 749 750 fs.lookup_still_valid = true; 751 752 fs.first_m = NULL; 753 754 /* 755 * Search for the page at object/offset. 756 */ 757 fs.object = fs.first_object; 758 fs.pindex = fs.first_pindex; 759 while (TRUE) { 760 /* 761 * If the object is marked for imminent termination, 762 * we retry here, since the collapse pass has raced 763 * with us. Otherwise, if we see terminally dead 764 * object, return fail. 765 */ 766 if ((fs.object->flags & OBJ_DEAD) != 0) { 767 dead = fs.object->type == OBJT_DEAD; 768 unlock_and_deallocate(&fs); 769 if (dead) 770 return (KERN_PROTECTION_FAILURE); 771 pause("vmf_de", 1); 772 goto RetryFault; 773 } 774 775 /* 776 * See if page is resident 777 */ 778 fs.m = vm_page_lookup(fs.object, fs.pindex); 779 if (fs.m != NULL) { 780 /* 781 * Wait/Retry if the page is busy. We have to do this 782 * if the page is either exclusive or shared busy 783 * because the vm_pager may be using read busy for 784 * pageouts (and even pageins if it is the vnode 785 * pager), and we could end up trying to pagein and 786 * pageout the same page simultaneously. 787 * 788 * We can theoretically allow the busy case on a read 789 * fault if the page is marked valid, but since such 790 * pages are typically already pmap'd, putting that 791 * special case in might be more effort then it is 792 * worth. We cannot under any circumstances mess 793 * around with a shared busied page except, perhaps, 794 * to pmap it. 795 */ 796 if (vm_page_tryxbusy(fs.m) == 0) { 797 /* 798 * Reference the page before unlocking and 799 * sleeping so that the page daemon is less 800 * likely to reclaim it. 801 */ 802 vm_page_aflag_set(fs.m, PGA_REFERENCED); 803 if (fs.object != fs.first_object) { 804 if (!VM_OBJECT_TRYWLOCK( 805 fs.first_object)) { 806 VM_OBJECT_WUNLOCK(fs.object); 807 VM_OBJECT_WLOCK(fs.first_object); 808 VM_OBJECT_WLOCK(fs.object); 809 } 810 vm_page_free(fs.first_m); 811 vm_object_pip_wakeup(fs.first_object); 812 VM_OBJECT_WUNLOCK(fs.first_object); 813 fs.first_m = NULL; 814 } 815 unlock_map(&fs); 816 if (fs.m == vm_page_lookup(fs.object, 817 fs.pindex)) { 818 vm_page_sleep_if_busy(fs.m, "vmpfw"); 819 } 820 vm_object_pip_wakeup(fs.object); 821 VM_OBJECT_WUNLOCK(fs.object); 822 VM_CNT_INC(v_intrans); 823 vm_object_deallocate(fs.first_object); 824 goto RetryFault; 825 } 826 827 /* 828 * The page is marked busy for other processes and the 829 * pagedaemon. If it still isn't completely valid 830 * (readable), jump to readrest, else break-out ( we 831 * found the page ). 832 */ 833 if (!vm_page_all_valid(fs.m)) 834 goto readrest; 835 break; /* break to PAGE HAS BEEN FOUND */ 836 } 837 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m)); 838 839 /* 840 * Page is not resident. If the pager might contain the page 841 * or this is the beginning of the search, allocate a new 842 * page. (Default objects are zero-fill, so there is no real 843 * pager for them.) 844 */ 845 if (fs.object->type != OBJT_DEFAULT || 846 fs.object == fs.first_object) { 847 if (fs.pindex >= fs.object->size) { 848 unlock_and_deallocate(&fs); 849 return (KERN_OUT_OF_BOUNDS); 850 } 851 852 if (fs.object == fs.first_object && 853 (fs.first_object->flags & OBJ_POPULATE) != 0 && 854 fs.first_object->shadow_count == 0) { 855 rv = vm_fault_populate(&fs, prot, fault_type, 856 fault_flags, wired, m_hold); 857 switch (rv) { 858 case KERN_SUCCESS: 859 case KERN_FAILURE: 860 unlock_and_deallocate(&fs); 861 return (rv); 862 case KERN_RESOURCE_SHORTAGE: 863 unlock_and_deallocate(&fs); 864 goto RetryFault; 865 case KERN_NOT_RECEIVER: 866 /* 867 * Pager's populate() method 868 * returned VM_PAGER_BAD. 869 */ 870 break; 871 default: 872 panic("inconsistent return codes"); 873 } 874 } 875 876 /* 877 * Allocate a new page for this object/offset pair. 878 * 879 * Unlocked read of the p_flag is harmless. At 880 * worst, the P_KILLED might be not observed 881 * there, and allocation can fail, causing 882 * restart and new reading of the p_flag. 883 */ 884 dset = fs.object->domain.dr_policy; 885 if (dset == NULL) 886 dset = curthread->td_domain.dr_policy; 887 if (!vm_page_count_severe_set(&dset->ds_mask) || 888 P_KILLED(curproc)) { 889 #if VM_NRESERVLEVEL > 0 890 vm_object_color(fs.object, atop(vaddr) - 891 fs.pindex); 892 #endif 893 alloc_req = P_KILLED(curproc) ? 894 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 895 if (fs.object->type != OBJT_VNODE && 896 fs.object->backing_object == NULL) 897 alloc_req |= VM_ALLOC_ZERO; 898 fs.m = vm_page_alloc(fs.object, fs.pindex, 899 alloc_req); 900 } 901 if (fs.m == NULL) { 902 unlock_and_deallocate(&fs); 903 if (vm_pfault_oom_attempts < 0 || 904 oom < vm_pfault_oom_attempts) { 905 oom++; 906 vm_waitpfault(dset, 907 vm_pfault_oom_wait * hz); 908 goto RetryFault_oom; 909 } 910 if (bootverbose) 911 printf( 912 "proc %d (%s) failed to alloc page on fault, starting OOM\n", 913 curproc->p_pid, curproc->p_comm); 914 vm_pageout_oom(VM_OOM_MEM_PF); 915 goto RetryFault; 916 } 917 } 918 919 readrest: 920 /* 921 * At this point, we have either allocated a new page or found 922 * an existing page that is only partially valid. 923 * 924 * We hold a reference on the current object and the page is 925 * exclusive busied. 926 */ 927 928 /* 929 * If the pager for the current object might have the page, 930 * then determine the number of additional pages to read and 931 * potentially reprioritize previously read pages for earlier 932 * reclamation. These operations should only be performed 933 * once per page fault. Even if the current pager doesn't 934 * have the page, the number of additional pages to read will 935 * apply to subsequent objects in the shadow chain. 936 */ 937 if (fs.object->type != OBJT_DEFAULT && nera == -1 && 938 !P_KILLED(curproc)) { 939 KASSERT(fs.lookup_still_valid, ("map unlocked")); 940 era = fs.entry->read_ahead; 941 behavior = vm_map_entry_behavior(fs.entry); 942 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 943 nera = 0; 944 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 945 nera = VM_FAULT_READ_AHEAD_MAX; 946 if (vaddr == fs.entry->next_read) 947 vm_fault_dontneed(&fs, vaddr, nera); 948 } else if (vaddr == fs.entry->next_read) { 949 /* 950 * This is a sequential fault. Arithmetically 951 * increase the requested number of pages in 952 * the read-ahead window. The requested 953 * number of pages is "# of sequential faults 954 * x (read ahead min + 1) + read ahead min" 955 */ 956 nera = VM_FAULT_READ_AHEAD_MIN; 957 if (era > 0) { 958 nera += era + 1; 959 if (nera > VM_FAULT_READ_AHEAD_MAX) 960 nera = VM_FAULT_READ_AHEAD_MAX; 961 } 962 if (era == VM_FAULT_READ_AHEAD_MAX) 963 vm_fault_dontneed(&fs, vaddr, nera); 964 } else { 965 /* 966 * This is a non-sequential fault. 967 */ 968 nera = 0; 969 } 970 if (era != nera) { 971 /* 972 * A read lock on the map suffices to update 973 * the read ahead count safely. 974 */ 975 fs.entry->read_ahead = nera; 976 } 977 978 /* 979 * Prepare for unlocking the map. Save the map 980 * entry's start and end addresses, which are used to 981 * optimize the size of the pager operation below. 982 * Even if the map entry's addresses change after 983 * unlocking the map, using the saved addresses is 984 * safe. 985 */ 986 e_start = fs.entry->start; 987 e_end = fs.entry->end; 988 } 989 990 /* 991 * Call the pager to retrieve the page if there is a chance 992 * that the pager has it, and potentially retrieve additional 993 * pages at the same time. 994 */ 995 if (fs.object->type != OBJT_DEFAULT) { 996 /* 997 * Release the map lock before locking the vnode or 998 * sleeping in the pager. (If the current object has 999 * a shadow, then an earlier iteration of this loop 1000 * may have already unlocked the map.) 1001 */ 1002 unlock_map(&fs); 1003 1004 if (fs.object->type == OBJT_VNODE && 1005 (vp = fs.object->handle) != fs.vp) { 1006 /* 1007 * Perform an unlock in case the desired vnode 1008 * changed while the map was unlocked during a 1009 * retry. 1010 */ 1011 unlock_vp(&fs); 1012 1013 locked = VOP_ISLOCKED(vp); 1014 if (locked != LK_EXCLUSIVE) 1015 locked = LK_SHARED; 1016 1017 /* 1018 * We must not sleep acquiring the vnode lock 1019 * while we have the page exclusive busied or 1020 * the object's paging-in-progress count 1021 * incremented. Otherwise, we could deadlock. 1022 */ 1023 error = vget(vp, locked | LK_CANRECURSE | 1024 LK_NOWAIT, curthread); 1025 if (error != 0) { 1026 vhold(vp); 1027 release_page(&fs); 1028 unlock_and_deallocate(&fs); 1029 error = vget(vp, locked | LK_RETRY | 1030 LK_CANRECURSE, curthread); 1031 vdrop(vp); 1032 fs.vp = vp; 1033 KASSERT(error == 0, 1034 ("vm_fault: vget failed")); 1035 goto RetryFault; 1036 } 1037 fs.vp = vp; 1038 } 1039 KASSERT(fs.vp == NULL || !fs.map->system_map, 1040 ("vm_fault: vnode-backed object mapped by system map")); 1041 1042 /* 1043 * Page in the requested page and hint the pager, 1044 * that it may bring up surrounding pages. 1045 */ 1046 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || 1047 P_KILLED(curproc)) { 1048 behind = 0; 1049 ahead = 0; 1050 } else { 1051 /* Is this a sequential fault? */ 1052 if (nera > 0) { 1053 behind = 0; 1054 ahead = nera; 1055 } else { 1056 /* 1057 * Request a cluster of pages that is 1058 * aligned to a VM_FAULT_READ_DEFAULT 1059 * page offset boundary within the 1060 * object. Alignment to a page offset 1061 * boundary is more likely to coincide 1062 * with the underlying file system 1063 * block than alignment to a virtual 1064 * address boundary. 1065 */ 1066 cluster_offset = fs.pindex % 1067 VM_FAULT_READ_DEFAULT; 1068 behind = ulmin(cluster_offset, 1069 atop(vaddr - e_start)); 1070 ahead = VM_FAULT_READ_DEFAULT - 1 - 1071 cluster_offset; 1072 } 1073 ahead = ulmin(ahead, atop(e_end - vaddr) - 1); 1074 } 1075 rv = vm_pager_get_pages(fs.object, &fs.m, 1, 1076 &behind, &ahead); 1077 if (rv == VM_PAGER_OK) { 1078 faultcount = behind + 1 + ahead; 1079 hardfault = true; 1080 break; /* break to PAGE HAS BEEN FOUND */ 1081 } 1082 if (rv == VM_PAGER_ERROR) 1083 printf("vm_fault: pager read error, pid %d (%s)\n", 1084 curproc->p_pid, curproc->p_comm); 1085 1086 /* 1087 * If an I/O error occurred or the requested page was 1088 * outside the range of the pager, clean up and return 1089 * an error. 1090 */ 1091 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { 1092 if (!vm_page_wired(fs.m)) 1093 vm_page_free(fs.m); 1094 else 1095 vm_page_xunbusy(fs.m); 1096 fs.m = NULL; 1097 unlock_and_deallocate(&fs); 1098 return (KERN_OUT_OF_BOUNDS); 1099 } 1100 1101 /* 1102 * The requested page does not exist at this object/ 1103 * offset. Remove the invalid page from the object, 1104 * waking up anyone waiting for it, and continue on to 1105 * the next object. However, if this is the top-level 1106 * object, we must leave the busy page in place to 1107 * prevent another process from rushing past us, and 1108 * inserting the page in that object at the same time 1109 * that we are. 1110 */ 1111 if (fs.object != fs.first_object) { 1112 if (!vm_page_wired(fs.m)) 1113 vm_page_free(fs.m); 1114 else 1115 vm_page_xunbusy(fs.m); 1116 fs.m = NULL; 1117 } 1118 } 1119 1120 /* 1121 * We get here if the object has default pager (or unwiring) 1122 * or the pager doesn't have the page. 1123 */ 1124 if (fs.object == fs.first_object) 1125 fs.first_m = fs.m; 1126 1127 /* 1128 * Move on to the next object. Lock the next object before 1129 * unlocking the current one. 1130 */ 1131 next_object = fs.object->backing_object; 1132 if (next_object == NULL) { 1133 /* 1134 * If there's no object left, fill the page in the top 1135 * object with zeros. 1136 */ 1137 if (fs.object != fs.first_object) { 1138 vm_object_pip_wakeup(fs.object); 1139 VM_OBJECT_WUNLOCK(fs.object); 1140 1141 fs.object = fs.first_object; 1142 fs.pindex = fs.first_pindex; 1143 fs.m = fs.first_m; 1144 VM_OBJECT_WLOCK(fs.object); 1145 } 1146 fs.first_m = NULL; 1147 1148 /* 1149 * Zero the page if necessary and mark it valid. 1150 */ 1151 if ((fs.m->flags & PG_ZERO) == 0) { 1152 pmap_zero_page(fs.m); 1153 } else { 1154 VM_CNT_INC(v_ozfod); 1155 } 1156 VM_CNT_INC(v_zfod); 1157 vm_page_valid(fs.m); 1158 /* Don't try to prefault neighboring pages. */ 1159 faultcount = 1; 1160 break; /* break to PAGE HAS BEEN FOUND */ 1161 } else { 1162 KASSERT(fs.object != next_object, 1163 ("object loop %p", next_object)); 1164 VM_OBJECT_WLOCK(next_object); 1165 vm_object_pip_add(next_object, 1); 1166 if (fs.object != fs.first_object) 1167 vm_object_pip_wakeup(fs.object); 1168 fs.pindex += 1169 OFF_TO_IDX(fs.object->backing_object_offset); 1170 VM_OBJECT_WUNLOCK(fs.object); 1171 fs.object = next_object; 1172 } 1173 } 1174 1175 vm_page_assert_xbusied(fs.m); 1176 1177 /* 1178 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1179 * is held.] 1180 */ 1181 1182 /* 1183 * If the page is being written, but isn't already owned by the 1184 * top-level object, we have to copy it into a new page owned by the 1185 * top-level object. 1186 */ 1187 if (fs.object != fs.first_object) { 1188 /* 1189 * We only really need to copy if we want to write it. 1190 */ 1191 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1192 /* 1193 * This allows pages to be virtually copied from a 1194 * backing_object into the first_object, where the 1195 * backing object has no other refs to it, and cannot 1196 * gain any more refs. Instead of a bcopy, we just 1197 * move the page from the backing object to the 1198 * first object. Note that we must mark the page 1199 * dirty in the first object so that it will go out 1200 * to swap when needed. 1201 */ 1202 is_first_object_locked = false; 1203 if ( 1204 /* 1205 * Only one shadow object 1206 */ 1207 (fs.object->shadow_count == 1) && 1208 /* 1209 * No COW refs, except us 1210 */ 1211 (fs.object->ref_count == 1) && 1212 /* 1213 * No one else can look this object up 1214 */ 1215 (fs.object->handle == NULL) && 1216 /* 1217 * No other ways to look the object up 1218 */ 1219 ((fs.object->type == OBJT_DEFAULT) || 1220 (fs.object->type == OBJT_SWAP)) && 1221 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && 1222 /* 1223 * We don't chase down the shadow chain 1224 */ 1225 fs.object == fs.first_object->backing_object) { 1226 1227 (void)vm_page_remove(fs.m); 1228 vm_page_replace_checked(fs.m, fs.first_object, 1229 fs.first_pindex, fs.first_m); 1230 vm_page_free(fs.first_m); 1231 vm_page_dirty(fs.m); 1232 #if VM_NRESERVLEVEL > 0 1233 /* 1234 * Rename the reservation. 1235 */ 1236 vm_reserv_rename(fs.m, fs.first_object, 1237 fs.object, OFF_TO_IDX( 1238 fs.first_object->backing_object_offset)); 1239 #endif 1240 fs.first_m = fs.m; 1241 fs.m = NULL; 1242 VM_CNT_INC(v_cow_optim); 1243 } else { 1244 /* 1245 * Oh, well, lets copy it. 1246 */ 1247 pmap_copy_page(fs.m, fs.first_m); 1248 vm_page_valid(fs.first_m); 1249 if (wired && (fault_flags & 1250 VM_FAULT_WIRE) == 0) { 1251 vm_page_wire(fs.first_m); 1252 vm_page_unwire(fs.m, PQ_INACTIVE); 1253 } 1254 /* 1255 * We no longer need the old page or object. 1256 */ 1257 release_page(&fs); 1258 } 1259 /* 1260 * fs.object != fs.first_object due to above 1261 * conditional 1262 */ 1263 vm_object_pip_wakeup(fs.object); 1264 VM_OBJECT_WUNLOCK(fs.object); 1265 1266 /* 1267 * We only try to prefault read-only mappings to the 1268 * neighboring pages when this copy-on-write fault is 1269 * a hard fault. In other cases, trying to prefault 1270 * is typically wasted effort. 1271 */ 1272 if (faultcount == 0) 1273 faultcount = 1; 1274 1275 /* 1276 * Only use the new page below... 1277 */ 1278 fs.object = fs.first_object; 1279 fs.pindex = fs.first_pindex; 1280 fs.m = fs.first_m; 1281 if (!is_first_object_locked) 1282 VM_OBJECT_WLOCK(fs.object); 1283 VM_CNT_INC(v_cow_faults); 1284 curthread->td_cow++; 1285 } else { 1286 prot &= ~VM_PROT_WRITE; 1287 } 1288 } 1289 1290 /* 1291 * We must verify that the maps have not changed since our last 1292 * lookup. 1293 */ 1294 if (!fs.lookup_still_valid) { 1295 if (!vm_map_trylock_read(fs.map)) { 1296 release_page(&fs); 1297 unlock_and_deallocate(&fs); 1298 goto RetryFault; 1299 } 1300 fs.lookup_still_valid = true; 1301 if (fs.map->timestamp != fs.map_generation) { 1302 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 1303 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 1304 1305 /* 1306 * If we don't need the page any longer, put it on the inactive 1307 * list (the easiest thing to do here). If no one needs it, 1308 * pageout will grab it eventually. 1309 */ 1310 if (result != KERN_SUCCESS) { 1311 release_page(&fs); 1312 unlock_and_deallocate(&fs); 1313 1314 /* 1315 * If retry of map lookup would have blocked then 1316 * retry fault from start. 1317 */ 1318 if (result == KERN_FAILURE) 1319 goto RetryFault; 1320 return (result); 1321 } 1322 if ((retry_object != fs.first_object) || 1323 (retry_pindex != fs.first_pindex)) { 1324 release_page(&fs); 1325 unlock_and_deallocate(&fs); 1326 goto RetryFault; 1327 } 1328 1329 /* 1330 * Check whether the protection has changed or the object has 1331 * been copied while we left the map unlocked. Changing from 1332 * read to write permission is OK - we leave the page 1333 * write-protected, and catch the write fault. Changing from 1334 * write to read permission means that we can't mark the page 1335 * write-enabled after all. 1336 */ 1337 prot &= retry_prot; 1338 fault_type &= retry_prot; 1339 if (prot == 0) { 1340 release_page(&fs); 1341 unlock_and_deallocate(&fs); 1342 goto RetryFault; 1343 } 1344 1345 /* Reassert because wired may have changed. */ 1346 KASSERT(wired || (fault_flags & VM_FAULT_WIRE) == 0, 1347 ("!wired && VM_FAULT_WIRE")); 1348 } 1349 } 1350 1351 /* 1352 * If the page was filled by a pager, save the virtual address that 1353 * should be faulted on next under a sequential access pattern to the 1354 * map entry. A read lock on the map suffices to update this address 1355 * safely. 1356 */ 1357 if (hardfault) 1358 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; 1359 1360 vm_page_assert_xbusied(fs.m); 1361 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true); 1362 1363 /* 1364 * Page must be completely valid or it is not fit to 1365 * map into user space. vm_pager_get_pages() ensures this. 1366 */ 1367 KASSERT(vm_page_all_valid(fs.m), 1368 ("vm_fault: page %p partially invalid", fs.m)); 1369 VM_OBJECT_WUNLOCK(fs.object); 1370 1371 /* 1372 * Put this page into the physical map. We had to do the unlock above 1373 * because pmap_enter() may sleep. We don't put the page 1374 * back on the active queue until later so that the pageout daemon 1375 * won't find it (yet). 1376 */ 1377 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, 1378 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); 1379 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 && 1380 wired == 0) 1381 vm_fault_prefault(&fs, vaddr, 1382 faultcount > 0 ? behind : PFBAK, 1383 faultcount > 0 ? ahead : PFFOR, false); 1384 VM_OBJECT_WLOCK(fs.object); 1385 1386 /* 1387 * If the page is not wired down, then put it where the pageout daemon 1388 * can find it. 1389 */ 1390 if ((fault_flags & VM_FAULT_WIRE) != 0) { 1391 vm_page_wire(fs.m); 1392 } else { 1393 vm_page_lock(fs.m); 1394 vm_page_activate(fs.m); 1395 vm_page_unlock(fs.m); 1396 } 1397 if (m_hold != NULL) { 1398 *m_hold = fs.m; 1399 vm_page_wire(fs.m); 1400 } 1401 vm_page_xunbusy(fs.m); 1402 1403 /* 1404 * Unlock everything, and return 1405 */ 1406 unlock_and_deallocate(&fs); 1407 if (hardfault) { 1408 VM_CNT_INC(v_io_faults); 1409 curthread->td_ru.ru_majflt++; 1410 #ifdef RACCT 1411 if (racct_enable && fs.object->type == OBJT_VNODE) { 1412 PROC_LOCK(curproc); 1413 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1414 racct_add_force(curproc, RACCT_WRITEBPS, 1415 PAGE_SIZE + behind * PAGE_SIZE); 1416 racct_add_force(curproc, RACCT_WRITEIOPS, 1); 1417 } else { 1418 racct_add_force(curproc, RACCT_READBPS, 1419 PAGE_SIZE + ahead * PAGE_SIZE); 1420 racct_add_force(curproc, RACCT_READIOPS, 1); 1421 } 1422 PROC_UNLOCK(curproc); 1423 } 1424 #endif 1425 } else 1426 curthread->td_ru.ru_minflt++; 1427 1428 return (KERN_SUCCESS); 1429 } 1430 1431 /* 1432 * Speed up the reclamation of pages that precede the faulting pindex within 1433 * the first object of the shadow chain. Essentially, perform the equivalent 1434 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes 1435 * the faulting pindex by the cluster size when the pages read by vm_fault() 1436 * cross a cluster-size boundary. The cluster size is the greater of the 1437 * smallest superpage size and VM_FAULT_DONTNEED_MIN. 1438 * 1439 * When "fs->first_object" is a shadow object, the pages in the backing object 1440 * that precede the faulting pindex are deactivated by vm_fault(). So, this 1441 * function must only be concerned with pages in the first object. 1442 */ 1443 static void 1444 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) 1445 { 1446 vm_map_entry_t entry; 1447 vm_object_t first_object, object; 1448 vm_offset_t end, start; 1449 vm_page_t m, m_next; 1450 vm_pindex_t pend, pstart; 1451 vm_size_t size; 1452 1453 object = fs->object; 1454 VM_OBJECT_ASSERT_WLOCKED(object); 1455 first_object = fs->first_object; 1456 if (first_object != object) { 1457 if (!VM_OBJECT_TRYWLOCK(first_object)) { 1458 VM_OBJECT_WUNLOCK(object); 1459 VM_OBJECT_WLOCK(first_object); 1460 VM_OBJECT_WLOCK(object); 1461 } 1462 } 1463 /* Neither fictitious nor unmanaged pages can be reclaimed. */ 1464 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 1465 size = VM_FAULT_DONTNEED_MIN; 1466 if (MAXPAGESIZES > 1 && size < pagesizes[1]) 1467 size = pagesizes[1]; 1468 end = rounddown2(vaddr, size); 1469 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && 1470 (entry = fs->entry)->start < end) { 1471 if (end - entry->start < size) 1472 start = entry->start; 1473 else 1474 start = end - size; 1475 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); 1476 pstart = OFF_TO_IDX(entry->offset) + atop(start - 1477 entry->start); 1478 m_next = vm_page_find_least(first_object, pstart); 1479 pend = OFF_TO_IDX(entry->offset) + atop(end - 1480 entry->start); 1481 while ((m = m_next) != NULL && m->pindex < pend) { 1482 m_next = TAILQ_NEXT(m, listq); 1483 if (!vm_page_all_valid(m) || 1484 vm_page_busied(m)) 1485 continue; 1486 1487 /* 1488 * Don't clear PGA_REFERENCED, since it would 1489 * likely represent a reference by a different 1490 * process. 1491 * 1492 * Typically, at this point, prefetched pages 1493 * are still in the inactive queue. Only 1494 * pages that triggered page faults are in the 1495 * active queue. 1496 */ 1497 vm_page_lock(m); 1498 if (!vm_page_inactive(m)) 1499 vm_page_deactivate(m); 1500 vm_page_unlock(m); 1501 } 1502 } 1503 } 1504 if (first_object != object) 1505 VM_OBJECT_WUNLOCK(first_object); 1506 } 1507 1508 /* 1509 * vm_fault_prefault provides a quick way of clustering 1510 * pagefaults into a processes address space. It is a "cousin" 1511 * of vm_map_pmap_enter, except it runs at page fault time instead 1512 * of mmap time. 1513 */ 1514 static void 1515 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 1516 int backward, int forward, bool obj_locked) 1517 { 1518 pmap_t pmap; 1519 vm_map_entry_t entry; 1520 vm_object_t backing_object, lobject; 1521 vm_offset_t addr, starta; 1522 vm_pindex_t pindex; 1523 vm_page_t m; 1524 int i; 1525 1526 pmap = fs->map->pmap; 1527 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1528 return; 1529 1530 entry = fs->entry; 1531 1532 if (addra < backward * PAGE_SIZE) { 1533 starta = entry->start; 1534 } else { 1535 starta = addra - backward * PAGE_SIZE; 1536 if (starta < entry->start) 1537 starta = entry->start; 1538 } 1539 1540 /* 1541 * Generate the sequence of virtual addresses that are candidates for 1542 * prefaulting in an outward spiral from the faulting virtual address, 1543 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra 1544 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... 1545 * If the candidate address doesn't have a backing physical page, then 1546 * the loop immediately terminates. 1547 */ 1548 for (i = 0; i < 2 * imax(backward, forward); i++) { 1549 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : 1550 PAGE_SIZE); 1551 if (addr > addra + forward * PAGE_SIZE) 1552 addr = 0; 1553 1554 if (addr < starta || addr >= entry->end) 1555 continue; 1556 1557 if (!pmap_is_prefaultable(pmap, addr)) 1558 continue; 1559 1560 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1561 lobject = entry->object.vm_object; 1562 if (!obj_locked) 1563 VM_OBJECT_RLOCK(lobject); 1564 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1565 lobject->type == OBJT_DEFAULT && 1566 (backing_object = lobject->backing_object) != NULL) { 1567 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1568 0, ("vm_fault_prefault: unaligned object offset")); 1569 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1570 VM_OBJECT_RLOCK(backing_object); 1571 if (!obj_locked || lobject != entry->object.vm_object) 1572 VM_OBJECT_RUNLOCK(lobject); 1573 lobject = backing_object; 1574 } 1575 if (m == NULL) { 1576 if (!obj_locked || lobject != entry->object.vm_object) 1577 VM_OBJECT_RUNLOCK(lobject); 1578 break; 1579 } 1580 if (vm_page_all_valid(m) && 1581 (m->flags & PG_FICTITIOUS) == 0) 1582 pmap_enter_quick(pmap, addr, m, entry->protection); 1583 if (!obj_locked || lobject != entry->object.vm_object) 1584 VM_OBJECT_RUNLOCK(lobject); 1585 } 1586 } 1587 1588 /* 1589 * Hold each of the physical pages that are mapped by the specified range of 1590 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1591 * and allow the specified types of access, "prot". If all of the implied 1592 * pages are successfully held, then the number of held pages is returned 1593 * together with pointers to those pages in the array "ma". However, if any 1594 * of the pages cannot be held, -1 is returned. 1595 */ 1596 int 1597 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1598 vm_prot_t prot, vm_page_t *ma, int max_count) 1599 { 1600 vm_offset_t end, va; 1601 vm_page_t *mp; 1602 int count; 1603 boolean_t pmap_failed; 1604 1605 if (len == 0) 1606 return (0); 1607 end = round_page(addr + len); 1608 addr = trunc_page(addr); 1609 1610 /* 1611 * Check for illegal addresses. 1612 */ 1613 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) 1614 return (-1); 1615 1616 if (atop(end - addr) > max_count) 1617 panic("vm_fault_quick_hold_pages: count > max_count"); 1618 count = atop(end - addr); 1619 1620 /* 1621 * Most likely, the physical pages are resident in the pmap, so it is 1622 * faster to try pmap_extract_and_hold() first. 1623 */ 1624 pmap_failed = FALSE; 1625 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1626 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1627 if (*mp == NULL) 1628 pmap_failed = TRUE; 1629 else if ((prot & VM_PROT_WRITE) != 0 && 1630 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1631 /* 1632 * Explicitly dirty the physical page. Otherwise, the 1633 * caller's changes may go unnoticed because they are 1634 * performed through an unmanaged mapping or by a DMA 1635 * operation. 1636 * 1637 * The object lock is not held here. 1638 * See vm_page_clear_dirty_mask(). 1639 */ 1640 vm_page_dirty(*mp); 1641 } 1642 } 1643 if (pmap_failed) { 1644 /* 1645 * One or more pages could not be held by the pmap. Either no 1646 * page was mapped at the specified virtual address or that 1647 * mapping had insufficient permissions. Attempt to fault in 1648 * and hold these pages. 1649 * 1650 * If vm_fault_disable_pagefaults() was called, 1651 * i.e., TDP_NOFAULTING is set, we must not sleep nor 1652 * acquire MD VM locks, which means we must not call 1653 * vm_fault(). Some (out of tree) callers mark 1654 * too wide a code area with vm_fault_disable_pagefaults() 1655 * already, use the VM_PROT_QUICK_NOFAULT flag to request 1656 * the proper behaviour explicitly. 1657 */ 1658 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && 1659 (curthread->td_pflags & TDP_NOFAULTING) != 0) 1660 goto error; 1661 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1662 if (*mp == NULL && vm_fault(map, va, prot, 1663 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1664 goto error; 1665 } 1666 return (count); 1667 error: 1668 for (mp = ma; mp < ma + count; mp++) 1669 if (*mp != NULL) 1670 vm_page_unwire(*mp, PQ_INACTIVE); 1671 return (-1); 1672 } 1673 1674 /* 1675 * Routine: 1676 * vm_fault_copy_entry 1677 * Function: 1678 * Create new shadow object backing dst_entry with private copy of 1679 * all underlying pages. When src_entry is equal to dst_entry, 1680 * function implements COW for wired-down map entry. Otherwise, 1681 * it forks wired entry into dst_map. 1682 * 1683 * In/out conditions: 1684 * The source and destination maps must be locked for write. 1685 * The source map entry must be wired down (or be a sharing map 1686 * entry corresponding to a main map entry that is wired down). 1687 */ 1688 void 1689 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1690 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1691 vm_ooffset_t *fork_charge) 1692 { 1693 vm_object_t backing_object, dst_object, object, src_object; 1694 vm_pindex_t dst_pindex, pindex, src_pindex; 1695 vm_prot_t access, prot; 1696 vm_offset_t vaddr; 1697 vm_page_t dst_m; 1698 vm_page_t src_m; 1699 boolean_t upgrade; 1700 1701 #ifdef lint 1702 src_map++; 1703 #endif /* lint */ 1704 1705 upgrade = src_entry == dst_entry; 1706 access = prot = dst_entry->protection; 1707 1708 src_object = src_entry->object.vm_object; 1709 src_pindex = OFF_TO_IDX(src_entry->offset); 1710 1711 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 1712 dst_object = src_object; 1713 vm_object_reference(dst_object); 1714 } else { 1715 /* 1716 * Create the top-level object for the destination entry. (Doesn't 1717 * actually shadow anything - we copy the pages directly.) 1718 */ 1719 dst_object = vm_object_allocate(OBJT_DEFAULT, 1720 atop(dst_entry->end - dst_entry->start)); 1721 #if VM_NRESERVLEVEL > 0 1722 dst_object->flags |= OBJ_COLORED; 1723 dst_object->pg_color = atop(dst_entry->start); 1724 #endif 1725 dst_object->domain = src_object->domain; 1726 dst_object->charge = dst_entry->end - dst_entry->start; 1727 } 1728 1729 VM_OBJECT_WLOCK(dst_object); 1730 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1731 ("vm_fault_copy_entry: vm_object not NULL")); 1732 if (src_object != dst_object) { 1733 dst_entry->object.vm_object = dst_object; 1734 dst_entry->offset = 0; 1735 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; 1736 } 1737 if (fork_charge != NULL) { 1738 KASSERT(dst_entry->cred == NULL, 1739 ("vm_fault_copy_entry: leaked swp charge")); 1740 dst_object->cred = curthread->td_ucred; 1741 crhold(dst_object->cred); 1742 *fork_charge += dst_object->charge; 1743 } else if ((dst_object->type == OBJT_DEFAULT || 1744 dst_object->type == OBJT_SWAP) && 1745 dst_object->cred == NULL) { 1746 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", 1747 dst_entry)); 1748 dst_object->cred = dst_entry->cred; 1749 dst_entry->cred = NULL; 1750 } 1751 1752 /* 1753 * If not an upgrade, then enter the mappings in the pmap as 1754 * read and/or execute accesses. Otherwise, enter them as 1755 * write accesses. 1756 * 1757 * A writeable large page mapping is only created if all of 1758 * the constituent small page mappings are modified. Marking 1759 * PTEs as modified on inception allows promotion to happen 1760 * without taking potentially large number of soft faults. 1761 */ 1762 if (!upgrade) 1763 access &= ~VM_PROT_WRITE; 1764 1765 /* 1766 * Loop through all of the virtual pages within the entry's 1767 * range, copying each page from the source object to the 1768 * destination object. Since the source is wired, those pages 1769 * must exist. In contrast, the destination is pageable. 1770 * Since the destination object doesn't share any backing storage 1771 * with the source object, all of its pages must be dirtied, 1772 * regardless of whether they can be written. 1773 */ 1774 for (vaddr = dst_entry->start, dst_pindex = 0; 1775 vaddr < dst_entry->end; 1776 vaddr += PAGE_SIZE, dst_pindex++) { 1777 again: 1778 /* 1779 * Find the page in the source object, and copy it in. 1780 * Because the source is wired down, the page will be 1781 * in memory. 1782 */ 1783 if (src_object != dst_object) 1784 VM_OBJECT_RLOCK(src_object); 1785 object = src_object; 1786 pindex = src_pindex + dst_pindex; 1787 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1788 (backing_object = object->backing_object) != NULL) { 1789 /* 1790 * Unless the source mapping is read-only or 1791 * it is presently being upgraded from 1792 * read-only, the first object in the shadow 1793 * chain should provide all of the pages. In 1794 * other words, this loop body should never be 1795 * executed when the source mapping is already 1796 * read/write. 1797 */ 1798 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || 1799 upgrade, 1800 ("vm_fault_copy_entry: main object missing page")); 1801 1802 VM_OBJECT_RLOCK(backing_object); 1803 pindex += OFF_TO_IDX(object->backing_object_offset); 1804 if (object != dst_object) 1805 VM_OBJECT_RUNLOCK(object); 1806 object = backing_object; 1807 } 1808 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); 1809 1810 if (object != dst_object) { 1811 /* 1812 * Allocate a page in the destination object. 1813 */ 1814 dst_m = vm_page_alloc(dst_object, (src_object == 1815 dst_object ? src_pindex : 0) + dst_pindex, 1816 VM_ALLOC_NORMAL); 1817 if (dst_m == NULL) { 1818 VM_OBJECT_WUNLOCK(dst_object); 1819 VM_OBJECT_RUNLOCK(object); 1820 vm_wait(dst_object); 1821 VM_OBJECT_WLOCK(dst_object); 1822 goto again; 1823 } 1824 pmap_copy_page(src_m, dst_m); 1825 VM_OBJECT_RUNLOCK(object); 1826 dst_m->dirty = dst_m->valid = src_m->valid; 1827 } else { 1828 dst_m = src_m; 1829 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0) 1830 goto again; 1831 if (dst_m->pindex >= dst_object->size) { 1832 /* 1833 * We are upgrading. Index can occur 1834 * out of bounds if the object type is 1835 * vnode and the file was truncated. 1836 */ 1837 vm_page_xunbusy(dst_m); 1838 break; 1839 } 1840 } 1841 VM_OBJECT_WUNLOCK(dst_object); 1842 1843 /* 1844 * Enter it in the pmap. If a wired, copy-on-write 1845 * mapping is being replaced by a write-enabled 1846 * mapping, then wire that new mapping. 1847 * 1848 * The page can be invalid if the user called 1849 * msync(MS_INVALIDATE) or truncated the backing vnode 1850 * or shared memory object. In this case, do not 1851 * insert it into pmap, but still do the copy so that 1852 * all copies of the wired map entry have similar 1853 * backing pages. 1854 */ 1855 if (vm_page_all_valid(dst_m)) { 1856 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, 1857 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); 1858 } 1859 1860 /* 1861 * Mark it no longer busy, and put it on the active list. 1862 */ 1863 VM_OBJECT_WLOCK(dst_object); 1864 1865 if (upgrade) { 1866 if (src_m != dst_m) { 1867 vm_page_unwire(src_m, PQ_INACTIVE); 1868 vm_page_wire(dst_m); 1869 } else { 1870 KASSERT(vm_page_wired(dst_m), 1871 ("dst_m %p is not wired", dst_m)); 1872 } 1873 } else { 1874 vm_page_lock(dst_m); 1875 vm_page_activate(dst_m); 1876 vm_page_unlock(dst_m); 1877 } 1878 vm_page_xunbusy(dst_m); 1879 } 1880 VM_OBJECT_WUNLOCK(dst_object); 1881 if (upgrade) { 1882 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1883 vm_object_deallocate(src_object); 1884 } 1885 } 1886 1887 /* 1888 * Block entry into the machine-independent layer's page fault handler by 1889 * the calling thread. Subsequent calls to vm_fault() by that thread will 1890 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 1891 * spurious page faults. 1892 */ 1893 int 1894 vm_fault_disable_pagefaults(void) 1895 { 1896 1897 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 1898 } 1899 1900 void 1901 vm_fault_enable_pagefaults(int save) 1902 { 1903 1904 curthread_pflags_restore(save); 1905 } 1906