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