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