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