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