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