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/mutex.h> 88 #include <sys/proc.h> 89 #include <sys/racct.h> 90 #include <sys/refcount.h> 91 #include <sys/resourcevar.h> 92 #include <sys/rwlock.h> 93 #include <sys/signalvar.h> 94 #include <sys/sysctl.h> 95 #include <sys/sysent.h> 96 #include <sys/vmmeter.h> 97 #include <sys/vnode.h> 98 #ifdef KTRACE 99 #include <sys/ktrace.h> 100 #endif 101 102 #include <vm/vm.h> 103 #include <vm/vm_param.h> 104 #include <vm/pmap.h> 105 #include <vm/vm_map.h> 106 #include <vm/vm_object.h> 107 #include <vm/vm_page.h> 108 #include <vm/vm_pageout.h> 109 #include <vm/vm_kern.h> 110 #include <vm/vm_pager.h> 111 #include <vm/vm_extern.h> 112 #include <vm/vm_reserv.h> 113 114 #define PFBAK 4 115 #define PFFOR 4 116 117 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) 118 119 #define VM_FAULT_DONTNEED_MIN 1048576 120 121 struct faultstate { 122 /* Fault parameters. */ 123 vm_offset_t vaddr; 124 vm_page_t *m_hold; 125 vm_prot_t fault_type; 126 vm_prot_t prot; 127 int fault_flags; 128 int oom; 129 boolean_t wired; 130 131 /* Page reference for cow. */ 132 vm_page_t m_cow; 133 134 /* Current object. */ 135 vm_object_t object; 136 vm_pindex_t pindex; 137 vm_page_t m; 138 139 /* Top-level map object. */ 140 vm_object_t first_object; 141 vm_pindex_t first_pindex; 142 vm_page_t first_m; 143 144 /* Map state. */ 145 vm_map_t map; 146 vm_map_entry_t entry; 147 int map_generation; 148 bool lookup_still_valid; 149 150 /* Vnode if locked. */ 151 struct vnode *vp; 152 }; 153 154 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, 155 int ahead); 156 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 157 int backward, int forward, bool obj_locked); 158 159 static int vm_pfault_oom_attempts = 3; 160 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_attempts, CTLFLAG_RWTUN, 161 &vm_pfault_oom_attempts, 0, 162 "Number of page allocation attempts in page fault handler before it " 163 "triggers OOM handling"); 164 165 static int vm_pfault_oom_wait = 10; 166 SYSCTL_INT(_vm, OID_AUTO, pfault_oom_wait, CTLFLAG_RWTUN, 167 &vm_pfault_oom_wait, 0, 168 "Number of seconds to wait for free pages before retrying " 169 "the page fault handler"); 170 171 static inline void 172 fault_page_release(vm_page_t *mp) 173 { 174 vm_page_t m; 175 176 m = *mp; 177 if (m != NULL) { 178 /* 179 * We are likely to loop around again and attempt to busy 180 * this page. Deactivating it leaves it available for 181 * pageout while optimizing fault restarts. 182 */ 183 vm_page_deactivate(m); 184 vm_page_xunbusy(m); 185 *mp = NULL; 186 } 187 } 188 189 static inline void 190 fault_page_free(vm_page_t *mp) 191 { 192 vm_page_t m; 193 194 m = *mp; 195 if (m != NULL) { 196 VM_OBJECT_ASSERT_WLOCKED(m->object); 197 if (!vm_page_wired(m)) 198 vm_page_free(m); 199 else 200 vm_page_xunbusy(m); 201 *mp = NULL; 202 } 203 } 204 205 static inline void 206 unlock_map(struct faultstate *fs) 207 { 208 209 if (fs->lookup_still_valid) { 210 vm_map_lookup_done(fs->map, fs->entry); 211 fs->lookup_still_valid = false; 212 } 213 } 214 215 static void 216 unlock_vp(struct faultstate *fs) 217 { 218 219 if (fs->vp != NULL) { 220 vput(fs->vp); 221 fs->vp = NULL; 222 } 223 } 224 225 static void 226 fault_deallocate(struct faultstate *fs) 227 { 228 229 fault_page_release(&fs->m_cow); 230 fault_page_release(&fs->m); 231 vm_object_pip_wakeup(fs->object); 232 if (fs->object != fs->first_object) { 233 VM_OBJECT_WLOCK(fs->first_object); 234 fault_page_free(&fs->first_m); 235 VM_OBJECT_WUNLOCK(fs->first_object); 236 vm_object_pip_wakeup(fs->first_object); 237 } 238 vm_object_deallocate(fs->first_object); 239 unlock_map(fs); 240 unlock_vp(fs); 241 } 242 243 static void 244 unlock_and_deallocate(struct faultstate *fs) 245 { 246 247 VM_OBJECT_WUNLOCK(fs->object); 248 fault_deallocate(fs); 249 } 250 251 static void 252 vm_fault_dirty(struct faultstate *fs, vm_page_t m) 253 { 254 bool need_dirty; 255 256 if (((fs->prot & VM_PROT_WRITE) == 0 && 257 (fs->fault_flags & VM_FAULT_DIRTY) == 0) || 258 (m->oflags & VPO_UNMANAGED) != 0) 259 return; 260 261 VM_PAGE_OBJECT_BUSY_ASSERT(m); 262 263 need_dirty = ((fs->fault_type & VM_PROT_WRITE) != 0 && 264 (fs->fault_flags & VM_FAULT_WIRE) == 0) || 265 (fs->fault_flags & VM_FAULT_DIRTY) != 0; 266 267 vm_object_set_writeable_dirty(m->object); 268 269 /* 270 * If the fault is a write, we know that this page is being 271 * written NOW so dirty it explicitly to save on 272 * pmap_is_modified() calls later. 273 * 274 * Also, since the page is now dirty, we can possibly tell 275 * the pager to release any swap backing the page. 276 */ 277 if (need_dirty && vm_page_set_dirty(m) == 0) { 278 /* 279 * If this is a NOSYNC mmap we do not want to set PGA_NOSYNC 280 * if the page is already dirty to prevent data written with 281 * the expectation of being synced from not being synced. 282 * Likewise if this entry does not request NOSYNC then make 283 * sure the page isn't marked NOSYNC. Applications sharing 284 * data should use the same flags to avoid ping ponging. 285 */ 286 if ((fs->entry->eflags & MAP_ENTRY_NOSYNC) != 0) 287 vm_page_aflag_set(m, PGA_NOSYNC); 288 else 289 vm_page_aflag_clear(m, PGA_NOSYNC); 290 } 291 292 } 293 294 /* 295 * Unlocks fs.first_object and fs.map on success. 296 */ 297 static int 298 vm_fault_soft_fast(struct faultstate *fs) 299 { 300 vm_page_t m, m_map; 301 #if VM_NRESERVLEVEL > 0 302 vm_page_t m_super; 303 int flags; 304 #endif 305 int psind, rv; 306 vm_offset_t vaddr; 307 308 MPASS(fs->vp == NULL); 309 vaddr = fs->vaddr; 310 vm_object_busy(fs->first_object); 311 m = vm_page_lookup(fs->first_object, fs->first_pindex); 312 /* A busy page can be mapped for read|execute access. */ 313 if (m == NULL || ((fs->prot & VM_PROT_WRITE) != 0 && 314 vm_page_busied(m)) || !vm_page_all_valid(m)) { 315 rv = KERN_FAILURE; 316 goto out; 317 } 318 m_map = m; 319 psind = 0; 320 #if VM_NRESERVLEVEL > 0 321 if ((m->flags & PG_FICTITIOUS) == 0 && 322 (m_super = vm_reserv_to_superpage(m)) != NULL && 323 rounddown2(vaddr, pagesizes[m_super->psind]) >= fs->entry->start && 324 roundup2(vaddr + 1, pagesizes[m_super->psind]) <= fs->entry->end && 325 (vaddr & (pagesizes[m_super->psind] - 1)) == (VM_PAGE_TO_PHYS(m) & 326 (pagesizes[m_super->psind] - 1)) && !fs->wired && 327 pmap_ps_enabled(fs->map->pmap)) { 328 flags = PS_ALL_VALID; 329 if ((fs->prot & VM_PROT_WRITE) != 0) { 330 /* 331 * Create a superpage mapping allowing write access 332 * only if none of the constituent pages are busy and 333 * all of them are already dirty (except possibly for 334 * the page that was faulted on). 335 */ 336 flags |= PS_NONE_BUSY; 337 if ((fs->first_object->flags & OBJ_UNMANAGED) == 0) 338 flags |= PS_ALL_DIRTY; 339 } 340 if (vm_page_ps_test(m_super, flags, m)) { 341 m_map = m_super; 342 psind = m_super->psind; 343 vaddr = rounddown2(vaddr, pagesizes[psind]); 344 /* Preset the modified bit for dirty superpages. */ 345 if ((flags & PS_ALL_DIRTY) != 0) 346 fs->fault_type |= VM_PROT_WRITE; 347 } 348 } 349 #endif 350 rv = pmap_enter(fs->map->pmap, vaddr, m_map, fs->prot, fs->fault_type | 351 PMAP_ENTER_NOSLEEP | (fs->wired ? PMAP_ENTER_WIRED : 0), psind); 352 if (rv != KERN_SUCCESS) 353 goto out; 354 if (fs->m_hold != NULL) { 355 (*fs->m_hold) = m; 356 vm_page_wire(m); 357 } 358 if (psind == 0 && !fs->wired) 359 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR, true); 360 VM_OBJECT_RUNLOCK(fs->first_object); 361 vm_fault_dirty(fs, m); 362 vm_map_lookup_done(fs->map, fs->entry); 363 curthread->td_ru.ru_minflt++; 364 365 out: 366 vm_object_unbusy(fs->first_object); 367 return (rv); 368 } 369 370 static void 371 vm_fault_restore_map_lock(struct faultstate *fs) 372 { 373 374 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 375 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0); 376 377 if (!vm_map_trylock_read(fs->map)) { 378 VM_OBJECT_WUNLOCK(fs->first_object); 379 vm_map_lock_read(fs->map); 380 VM_OBJECT_WLOCK(fs->first_object); 381 } 382 fs->lookup_still_valid = true; 383 } 384 385 static void 386 vm_fault_populate_check_page(vm_page_t m) 387 { 388 389 /* 390 * Check each page to ensure that the pager is obeying the 391 * interface: the page must be installed in the object, fully 392 * valid, and exclusively busied. 393 */ 394 MPASS(m != NULL); 395 MPASS(vm_page_all_valid(m)); 396 MPASS(vm_page_xbusied(m)); 397 } 398 399 static void 400 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, 401 vm_pindex_t last) 402 { 403 vm_page_t m; 404 vm_pindex_t pidx; 405 406 VM_OBJECT_ASSERT_WLOCKED(object); 407 MPASS(first <= last); 408 for (pidx = first, m = vm_page_lookup(object, pidx); 409 pidx <= last; pidx++, m = vm_page_next(m)) { 410 vm_fault_populate_check_page(m); 411 vm_page_deactivate(m); 412 vm_page_xunbusy(m); 413 } 414 } 415 416 static int 417 vm_fault_populate(struct faultstate *fs) 418 { 419 vm_offset_t vaddr; 420 vm_page_t m; 421 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; 422 int bdry_idx, i, npages, psind, rv; 423 424 MPASS(fs->object == fs->first_object); 425 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 426 MPASS(blockcount_read(&fs->first_object->paging_in_progress) > 0); 427 MPASS(fs->first_object->backing_object == NULL); 428 MPASS(fs->lookup_still_valid); 429 430 pager_first = OFF_TO_IDX(fs->entry->offset); 431 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; 432 unlock_map(fs); 433 unlock_vp(fs); 434 435 /* 436 * Call the pager (driver) populate() method. 437 * 438 * There is no guarantee that the method will be called again 439 * if the current fault is for read, and a future fault is 440 * for write. Report the entry's maximum allowed protection 441 * to the driver. 442 */ 443 rv = vm_pager_populate(fs->first_object, fs->first_pindex, 444 fs->fault_type, fs->entry->max_protection, &pager_first, 445 &pager_last); 446 447 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 448 if (rv == VM_PAGER_BAD) { 449 /* 450 * VM_PAGER_BAD is the backdoor for a pager to request 451 * normal fault handling. 452 */ 453 vm_fault_restore_map_lock(fs); 454 if (fs->map->timestamp != fs->map_generation) 455 return (KERN_RESTART); 456 return (KERN_NOT_RECEIVER); 457 } 458 if (rv != VM_PAGER_OK) 459 return (KERN_FAILURE); /* AKA SIGSEGV */ 460 461 /* Ensure that the driver is obeying the interface. */ 462 MPASS(pager_first <= pager_last); 463 MPASS(fs->first_pindex <= pager_last); 464 MPASS(fs->first_pindex >= pager_first); 465 MPASS(pager_last < fs->first_object->size); 466 467 vm_fault_restore_map_lock(fs); 468 bdry_idx = (fs->entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 469 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 470 if (fs->map->timestamp != fs->map_generation) { 471 if (bdry_idx == 0) { 472 vm_fault_populate_cleanup(fs->first_object, pager_first, 473 pager_last); 474 } else { 475 m = vm_page_lookup(fs->first_object, pager_first); 476 if (m != fs->m) 477 vm_page_xunbusy(m); 478 } 479 return (KERN_RESTART); 480 } 481 482 /* 483 * The map is unchanged after our last unlock. Process the fault. 484 * 485 * First, the special case of largepage mappings, where 486 * populate only busies the first page in superpage run. 487 */ 488 if (bdry_idx != 0) { 489 KASSERT(PMAP_HAS_LARGEPAGES, 490 ("missing pmap support for large pages")); 491 m = vm_page_lookup(fs->first_object, pager_first); 492 vm_fault_populate_check_page(m); 493 VM_OBJECT_WUNLOCK(fs->first_object); 494 vaddr = fs->entry->start + IDX_TO_OFF(pager_first) - 495 fs->entry->offset; 496 /* assert alignment for entry */ 497 KASSERT((vaddr & (pagesizes[bdry_idx] - 1)) == 0, 498 ("unaligned superpage start %#jx pager_first %#jx offset %#jx vaddr %#jx", 499 (uintmax_t)fs->entry->start, (uintmax_t)pager_first, 500 (uintmax_t)fs->entry->offset, (uintmax_t)vaddr)); 501 KASSERT((VM_PAGE_TO_PHYS(m) & (pagesizes[bdry_idx] - 1)) == 0, 502 ("unaligned superpage m %p %#jx", m, 503 (uintmax_t)VM_PAGE_TO_PHYS(m))); 504 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, 505 fs->fault_type | (fs->wired ? PMAP_ENTER_WIRED : 0) | 506 PMAP_ENTER_LARGEPAGE, bdry_idx); 507 VM_OBJECT_WLOCK(fs->first_object); 508 vm_page_xunbusy(m); 509 if (rv != KERN_SUCCESS) 510 goto out; 511 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) { 512 for (i = 0; i < atop(pagesizes[bdry_idx]); i++) 513 vm_page_wire(m + i); 514 } 515 if (fs->m_hold != NULL) { 516 *fs->m_hold = m + (fs->first_pindex - pager_first); 517 vm_page_wire(*fs->m_hold); 518 } 519 goto out; 520 } 521 522 /* 523 * The range [pager_first, pager_last] that is given to the 524 * pager is only a hint. The pager may populate any range 525 * within the object that includes the requested page index. 526 * In case the pager expanded the range, clip it to fit into 527 * the map entry. 528 */ 529 map_first = OFF_TO_IDX(fs->entry->offset); 530 if (map_first > pager_first) { 531 vm_fault_populate_cleanup(fs->first_object, pager_first, 532 map_first - 1); 533 pager_first = map_first; 534 } 535 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; 536 if (map_last < pager_last) { 537 vm_fault_populate_cleanup(fs->first_object, map_last + 1, 538 pager_last); 539 pager_last = map_last; 540 } 541 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); 542 pidx <= pager_last; 543 pidx += npages, m = vm_page_next(&m[npages - 1])) { 544 vaddr = fs->entry->start + IDX_TO_OFF(pidx) - fs->entry->offset; 545 546 psind = m->psind; 547 if (psind > 0 && ((vaddr & (pagesizes[psind] - 1)) != 0 || 548 pidx + OFF_TO_IDX(pagesizes[psind]) - 1 > pager_last || 549 !pmap_ps_enabled(fs->map->pmap) || fs->wired)) 550 psind = 0; 551 552 npages = atop(pagesizes[psind]); 553 for (i = 0; i < npages; i++) { 554 vm_fault_populate_check_page(&m[i]); 555 vm_fault_dirty(fs, &m[i]); 556 } 557 VM_OBJECT_WUNLOCK(fs->first_object); 558 rv = pmap_enter(fs->map->pmap, vaddr, m, fs->prot, fs->fault_type | 559 (fs->wired ? PMAP_ENTER_WIRED : 0), psind); 560 561 /* 562 * pmap_enter() may fail for a superpage mapping if additional 563 * protection policies prevent the full mapping. 564 * For example, this will happen on amd64 if the entire 565 * address range does not share the same userspace protection 566 * key. Revert to single-page mappings if this happens. 567 */ 568 MPASS(rv == KERN_SUCCESS || 569 (psind > 0 && rv == KERN_PROTECTION_FAILURE)); 570 if (__predict_false(psind > 0 && 571 rv == KERN_PROTECTION_FAILURE)) { 572 for (i = 0; i < npages; i++) { 573 rv = pmap_enter(fs->map->pmap, vaddr + ptoa(i), 574 &m[i], fs->prot, fs->fault_type | 575 (fs->wired ? PMAP_ENTER_WIRED : 0), 0); 576 MPASS(rv == KERN_SUCCESS); 577 } 578 } 579 580 VM_OBJECT_WLOCK(fs->first_object); 581 for (i = 0; i < npages; i++) { 582 if ((fs->fault_flags & VM_FAULT_WIRE) != 0) 583 vm_page_wire(&m[i]); 584 else 585 vm_page_activate(&m[i]); 586 if (fs->m_hold != NULL && m[i].pindex == fs->first_pindex) { 587 (*fs->m_hold) = &m[i]; 588 vm_page_wire(&m[i]); 589 } 590 vm_page_xunbusy(&m[i]); 591 } 592 } 593 out: 594 curthread->td_ru.ru_majflt++; 595 return (rv); 596 } 597 598 static int prot_fault_translation; 599 SYSCTL_INT(_machdep, OID_AUTO, prot_fault_translation, CTLFLAG_RWTUN, 600 &prot_fault_translation, 0, 601 "Control signal to deliver on protection fault"); 602 603 /* compat definition to keep common code for signal translation */ 604 #define UCODE_PAGEFLT 12 605 #ifdef T_PAGEFLT 606 _Static_assert(UCODE_PAGEFLT == T_PAGEFLT, "T_PAGEFLT"); 607 #endif 608 609 /* 610 * vm_fault_trap: 611 * 612 * Handle a page fault occurring at the given address, 613 * requiring the given permissions, in the map specified. 614 * If successful, the page is inserted into the 615 * associated physical map. 616 * 617 * NOTE: the given address should be truncated to the 618 * proper page address. 619 * 620 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 621 * a standard error specifying why the fault is fatal is returned. 622 * 623 * The map in question must be referenced, and remains so. 624 * Caller may hold no locks. 625 */ 626 int 627 vm_fault_trap(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 628 int fault_flags, int *signo, int *ucode) 629 { 630 int result; 631 632 MPASS(signo == NULL || ucode != NULL); 633 #ifdef KTRACE 634 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULT)) 635 ktrfault(vaddr, fault_type); 636 #endif 637 result = vm_fault(map, trunc_page(vaddr), fault_type, fault_flags, 638 NULL); 639 KASSERT(result == KERN_SUCCESS || result == KERN_FAILURE || 640 result == KERN_INVALID_ADDRESS || 641 result == KERN_RESOURCE_SHORTAGE || 642 result == KERN_PROTECTION_FAILURE || 643 result == KERN_OUT_OF_BOUNDS, 644 ("Unexpected Mach error %d from vm_fault()", result)); 645 #ifdef KTRACE 646 if (map != kernel_map && KTRPOINT(curthread, KTR_FAULTEND)) 647 ktrfaultend(result); 648 #endif 649 if (result != KERN_SUCCESS && signo != NULL) { 650 switch (result) { 651 case KERN_FAILURE: 652 case KERN_INVALID_ADDRESS: 653 *signo = SIGSEGV; 654 *ucode = SEGV_MAPERR; 655 break; 656 case KERN_RESOURCE_SHORTAGE: 657 *signo = SIGBUS; 658 *ucode = BUS_OOMERR; 659 break; 660 case KERN_OUT_OF_BOUNDS: 661 *signo = SIGBUS; 662 *ucode = BUS_OBJERR; 663 break; 664 case KERN_PROTECTION_FAILURE: 665 if (prot_fault_translation == 0) { 666 /* 667 * Autodetect. This check also covers 668 * the images without the ABI-tag ELF 669 * note. 670 */ 671 if (SV_CURPROC_ABI() == SV_ABI_FREEBSD && 672 curproc->p_osrel >= P_OSREL_SIGSEGV) { 673 *signo = SIGSEGV; 674 *ucode = SEGV_ACCERR; 675 } else { 676 *signo = SIGBUS; 677 *ucode = UCODE_PAGEFLT; 678 } 679 } else if (prot_fault_translation == 1) { 680 /* Always compat mode. */ 681 *signo = SIGBUS; 682 *ucode = UCODE_PAGEFLT; 683 } else { 684 /* Always SIGSEGV mode. */ 685 *signo = SIGSEGV; 686 *ucode = SEGV_ACCERR; 687 } 688 break; 689 default: 690 KASSERT(0, ("Unexpected Mach error %d from vm_fault()", 691 result)); 692 break; 693 } 694 } 695 return (result); 696 } 697 698 static int 699 vm_fault_lock_vnode(struct faultstate *fs, bool objlocked) 700 { 701 struct vnode *vp; 702 int error, locked; 703 704 if (fs->object->type != OBJT_VNODE) 705 return (KERN_SUCCESS); 706 vp = fs->object->handle; 707 if (vp == fs->vp) { 708 ASSERT_VOP_LOCKED(vp, "saved vnode is not locked"); 709 return (KERN_SUCCESS); 710 } 711 712 /* 713 * Perform an unlock in case the desired vnode changed while 714 * the map was unlocked during a retry. 715 */ 716 unlock_vp(fs); 717 718 locked = VOP_ISLOCKED(vp); 719 if (locked != LK_EXCLUSIVE) 720 locked = LK_SHARED; 721 722 /* 723 * We must not sleep acquiring the vnode lock while we have 724 * the page exclusive busied or the object's 725 * paging-in-progress count incremented. Otherwise, we could 726 * deadlock. 727 */ 728 error = vget(vp, locked | LK_CANRECURSE | LK_NOWAIT); 729 if (error == 0) { 730 fs->vp = vp; 731 return (KERN_SUCCESS); 732 } 733 734 vhold(vp); 735 if (objlocked) 736 unlock_and_deallocate(fs); 737 else 738 fault_deallocate(fs); 739 error = vget(vp, locked | LK_RETRY | LK_CANRECURSE); 740 vdrop(vp); 741 fs->vp = vp; 742 KASSERT(error == 0, ("vm_fault: vget failed %d", error)); 743 return (KERN_RESOURCE_SHORTAGE); 744 } 745 746 /* 747 * Calculate the desired readahead. Handle drop-behind. 748 * 749 * Returns the number of readahead blocks to pass to the pager. 750 */ 751 static int 752 vm_fault_readahead(struct faultstate *fs) 753 { 754 int era, nera; 755 u_char behavior; 756 757 KASSERT(fs->lookup_still_valid, ("map unlocked")); 758 era = fs->entry->read_ahead; 759 behavior = vm_map_entry_behavior(fs->entry); 760 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 761 nera = 0; 762 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 763 nera = VM_FAULT_READ_AHEAD_MAX; 764 if (fs->vaddr == fs->entry->next_read) 765 vm_fault_dontneed(fs, fs->vaddr, nera); 766 } else if (fs->vaddr == fs->entry->next_read) { 767 /* 768 * This is a sequential fault. Arithmetically 769 * increase the requested number of pages in 770 * the read-ahead window. The requested 771 * number of pages is "# of sequential faults 772 * x (read ahead min + 1) + read ahead min" 773 */ 774 nera = VM_FAULT_READ_AHEAD_MIN; 775 if (era > 0) { 776 nera += era + 1; 777 if (nera > VM_FAULT_READ_AHEAD_MAX) 778 nera = VM_FAULT_READ_AHEAD_MAX; 779 } 780 if (era == VM_FAULT_READ_AHEAD_MAX) 781 vm_fault_dontneed(fs, fs->vaddr, nera); 782 } else { 783 /* 784 * This is a non-sequential fault. 785 */ 786 nera = 0; 787 } 788 if (era != nera) { 789 /* 790 * A read lock on the map suffices to update 791 * the read ahead count safely. 792 */ 793 fs->entry->read_ahead = nera; 794 } 795 796 return (nera); 797 } 798 799 static int 800 vm_fault_lookup(struct faultstate *fs) 801 { 802 int result; 803 804 KASSERT(!fs->lookup_still_valid, 805 ("vm_fault_lookup: Map already locked.")); 806 result = vm_map_lookup(&fs->map, fs->vaddr, fs->fault_type | 807 VM_PROT_FAULT_LOOKUP, &fs->entry, &fs->first_object, 808 &fs->first_pindex, &fs->prot, &fs->wired); 809 if (result != KERN_SUCCESS) { 810 unlock_vp(fs); 811 return (result); 812 } 813 814 fs->map_generation = fs->map->timestamp; 815 816 if (fs->entry->eflags & MAP_ENTRY_NOFAULT) { 817 panic("%s: fault on nofault entry, addr: %#lx", 818 __func__, (u_long)fs->vaddr); 819 } 820 821 if (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION && 822 fs->entry->wiring_thread != curthread) { 823 vm_map_unlock_read(fs->map); 824 vm_map_lock(fs->map); 825 if (vm_map_lookup_entry(fs->map, fs->vaddr, &fs->entry) && 826 (fs->entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 827 unlock_vp(fs); 828 fs->entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 829 vm_map_unlock_and_wait(fs->map, 0); 830 } else 831 vm_map_unlock(fs->map); 832 return (KERN_RESOURCE_SHORTAGE); 833 } 834 835 MPASS((fs->entry->eflags & MAP_ENTRY_GUARD) == 0); 836 837 if (fs->wired) 838 fs->fault_type = fs->prot | (fs->fault_type & VM_PROT_COPY); 839 else 840 KASSERT((fs->fault_flags & VM_FAULT_WIRE) == 0, 841 ("!fs->wired && VM_FAULT_WIRE")); 842 fs->lookup_still_valid = true; 843 844 return (KERN_SUCCESS); 845 } 846 847 static int 848 vm_fault_relookup(struct faultstate *fs) 849 { 850 vm_object_t retry_object; 851 vm_pindex_t retry_pindex; 852 vm_prot_t retry_prot; 853 int result; 854 855 if (!vm_map_trylock_read(fs->map)) 856 return (KERN_RESTART); 857 858 fs->lookup_still_valid = true; 859 if (fs->map->timestamp == fs->map_generation) 860 return (KERN_SUCCESS); 861 862 result = vm_map_lookup_locked(&fs->map, fs->vaddr, fs->fault_type, 863 &fs->entry, &retry_object, &retry_pindex, &retry_prot, 864 &fs->wired); 865 if (result != KERN_SUCCESS) { 866 /* 867 * If retry of map lookup would have blocked then 868 * retry fault from start. 869 */ 870 if (result == KERN_FAILURE) 871 return (KERN_RESTART); 872 return (result); 873 } 874 if (retry_object != fs->first_object || 875 retry_pindex != fs->first_pindex) 876 return (KERN_RESTART); 877 878 /* 879 * Check whether the protection has changed or the object has 880 * been copied while we left the map unlocked. Changing from 881 * read to write permission is OK - we leave the page 882 * write-protected, and catch the write fault. Changing from 883 * write to read permission means that we can't mark the page 884 * write-enabled after all. 885 */ 886 fs->prot &= retry_prot; 887 fs->fault_type &= retry_prot; 888 if (fs->prot == 0) 889 return (KERN_RESTART); 890 891 /* Reassert because wired may have changed. */ 892 KASSERT(fs->wired || (fs->fault_flags & VM_FAULT_WIRE) == 0, 893 ("!wired && VM_FAULT_WIRE")); 894 895 return (KERN_SUCCESS); 896 } 897 898 static void 899 vm_fault_cow(struct faultstate *fs) 900 { 901 bool is_first_object_locked; 902 903 /* 904 * This allows pages to be virtually copied from a backing_object 905 * into the first_object, where the backing object has no other 906 * refs to it, and cannot gain any more refs. Instead of a bcopy, 907 * we just move the page from the backing object to the first 908 * object. Note that we must mark the page dirty in the first 909 * object so that it will go out to swap when needed. 910 */ 911 is_first_object_locked = false; 912 if ( 913 /* 914 * Only one shadow object and no other refs. 915 */ 916 fs->object->shadow_count == 1 && fs->object->ref_count == 1 && 917 /* 918 * No other ways to look the object up 919 */ 920 fs->object->handle == NULL && (fs->object->flags & OBJ_ANON) != 0 && 921 /* 922 * We don't chase down the shadow chain and we can acquire locks. 923 */ 924 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs->first_object)) && 925 fs->object == fs->first_object->backing_object && 926 VM_OBJECT_TRYWLOCK(fs->object)) { 927 /* 928 * Remove but keep xbusy for replace. fs->m is moved into 929 * fs->first_object and left busy while fs->first_m is 930 * conditionally freed. 931 */ 932 vm_page_remove_xbusy(fs->m); 933 vm_page_replace(fs->m, fs->first_object, fs->first_pindex, 934 fs->first_m); 935 vm_page_dirty(fs->m); 936 #if VM_NRESERVLEVEL > 0 937 /* 938 * Rename the reservation. 939 */ 940 vm_reserv_rename(fs->m, fs->first_object, fs->object, 941 OFF_TO_IDX(fs->first_object->backing_object_offset)); 942 #endif 943 VM_OBJECT_WUNLOCK(fs->object); 944 VM_OBJECT_WUNLOCK(fs->first_object); 945 fs->first_m = fs->m; 946 fs->m = NULL; 947 VM_CNT_INC(v_cow_optim); 948 } else { 949 if (is_first_object_locked) 950 VM_OBJECT_WUNLOCK(fs->first_object); 951 /* 952 * Oh, well, lets copy it. 953 */ 954 pmap_copy_page(fs->m, fs->first_m); 955 vm_page_valid(fs->first_m); 956 if (fs->wired && (fs->fault_flags & VM_FAULT_WIRE) == 0) { 957 vm_page_wire(fs->first_m); 958 vm_page_unwire(fs->m, PQ_INACTIVE); 959 } 960 /* 961 * Save the cow page to be released after 962 * pmap_enter is complete. 963 */ 964 fs->m_cow = fs->m; 965 fs->m = NULL; 966 } 967 /* 968 * fs->object != fs->first_object due to above 969 * conditional 970 */ 971 vm_object_pip_wakeup(fs->object); 972 973 /* 974 * Only use the new page below... 975 */ 976 fs->object = fs->first_object; 977 fs->pindex = fs->first_pindex; 978 fs->m = fs->first_m; 979 VM_CNT_INC(v_cow_faults); 980 curthread->td_cow++; 981 } 982 983 static bool 984 vm_fault_next(struct faultstate *fs) 985 { 986 vm_object_t next_object; 987 988 /* 989 * The requested page does not exist at this object/ 990 * offset. Remove the invalid page from the object, 991 * waking up anyone waiting for it, and continue on to 992 * the next object. However, if this is the top-level 993 * object, we must leave the busy page in place to 994 * prevent another process from rushing past us, and 995 * inserting the page in that object at the same time 996 * that we are. 997 */ 998 if (fs->object == fs->first_object) { 999 fs->first_m = fs->m; 1000 fs->m = NULL; 1001 } else 1002 fault_page_free(&fs->m); 1003 1004 /* 1005 * Move on to the next object. Lock the next object before 1006 * unlocking the current one. 1007 */ 1008 VM_OBJECT_ASSERT_WLOCKED(fs->object); 1009 next_object = fs->object->backing_object; 1010 if (next_object == NULL) 1011 return (false); 1012 MPASS(fs->first_m != NULL); 1013 KASSERT(fs->object != next_object, ("object loop %p", next_object)); 1014 VM_OBJECT_WLOCK(next_object); 1015 vm_object_pip_add(next_object, 1); 1016 if (fs->object != fs->first_object) 1017 vm_object_pip_wakeup(fs->object); 1018 fs->pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1019 VM_OBJECT_WUNLOCK(fs->object); 1020 fs->object = next_object; 1021 1022 return (true); 1023 } 1024 1025 static void 1026 vm_fault_zerofill(struct faultstate *fs) 1027 { 1028 1029 /* 1030 * If there's no object left, fill the page in the top 1031 * object with zeros. 1032 */ 1033 if (fs->object != fs->first_object) { 1034 vm_object_pip_wakeup(fs->object); 1035 fs->object = fs->first_object; 1036 fs->pindex = fs->first_pindex; 1037 } 1038 MPASS(fs->first_m != NULL); 1039 MPASS(fs->m == NULL); 1040 fs->m = fs->first_m; 1041 fs->first_m = NULL; 1042 1043 /* 1044 * Zero the page if necessary and mark it valid. 1045 */ 1046 if ((fs->m->flags & PG_ZERO) == 0) { 1047 pmap_zero_page(fs->m); 1048 } else { 1049 VM_CNT_INC(v_ozfod); 1050 } 1051 VM_CNT_INC(v_zfod); 1052 vm_page_valid(fs->m); 1053 } 1054 1055 /* 1056 * Allocate a page directly or via the object populate method. 1057 */ 1058 static int 1059 vm_fault_allocate(struct faultstate *fs) 1060 { 1061 struct domainset *dset; 1062 int alloc_req; 1063 int rv; 1064 1065 if ((fs->object->flags & OBJ_SIZEVNLOCK) != 0) { 1066 rv = vm_fault_lock_vnode(fs, true); 1067 MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE); 1068 if (rv == KERN_RESOURCE_SHORTAGE) 1069 return (rv); 1070 } 1071 1072 if (fs->pindex >= fs->object->size) 1073 return (KERN_OUT_OF_BOUNDS); 1074 1075 if (fs->object == fs->first_object && 1076 (fs->first_object->flags & OBJ_POPULATE) != 0 && 1077 fs->first_object->shadow_count == 0) { 1078 rv = vm_fault_populate(fs); 1079 switch (rv) { 1080 case KERN_SUCCESS: 1081 case KERN_FAILURE: 1082 case KERN_PROTECTION_FAILURE: 1083 case KERN_RESTART: 1084 return (rv); 1085 case KERN_NOT_RECEIVER: 1086 /* 1087 * Pager's populate() method 1088 * returned VM_PAGER_BAD. 1089 */ 1090 break; 1091 default: 1092 panic("inconsistent return codes"); 1093 } 1094 } 1095 1096 /* 1097 * Allocate a new page for this object/offset pair. 1098 * 1099 * Unlocked read of the p_flag is harmless. At worst, the P_KILLED 1100 * might be not observed there, and allocation can fail, causing 1101 * restart and new reading of the p_flag. 1102 */ 1103 dset = fs->object->domain.dr_policy; 1104 if (dset == NULL) 1105 dset = curthread->td_domain.dr_policy; 1106 if (!vm_page_count_severe_set(&dset->ds_mask) || P_KILLED(curproc)) { 1107 #if VM_NRESERVLEVEL > 0 1108 vm_object_color(fs->object, atop(fs->vaddr) - fs->pindex); 1109 #endif 1110 alloc_req = P_KILLED(curproc) ? 1111 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 1112 if (fs->object->type != OBJT_VNODE && 1113 fs->object->backing_object == NULL) 1114 alloc_req |= VM_ALLOC_ZERO; 1115 fs->m = vm_page_alloc(fs->object, fs->pindex, alloc_req); 1116 } 1117 if (fs->m == NULL) { 1118 unlock_and_deallocate(fs); 1119 if (vm_pfault_oom_attempts < 0 || 1120 fs->oom < vm_pfault_oom_attempts) { 1121 fs->oom++; 1122 vm_waitpfault(dset, vm_pfault_oom_wait * hz); 1123 } else { 1124 if (bootverbose) 1125 printf( 1126 "proc %d (%s) failed to alloc page on fault, starting OOM\n", 1127 curproc->p_pid, curproc->p_comm); 1128 vm_pageout_oom(VM_OOM_MEM_PF); 1129 fs->oom = 0; 1130 } 1131 return (KERN_RESOURCE_SHORTAGE); 1132 } 1133 fs->oom = 0; 1134 1135 return (KERN_NOT_RECEIVER); 1136 } 1137 1138 /* 1139 * Call the pager to retrieve the page if there is a chance 1140 * that the pager has it, and potentially retrieve additional 1141 * pages at the same time. 1142 */ 1143 static int 1144 vm_fault_getpages(struct faultstate *fs, int nera, int *behindp, int *aheadp) 1145 { 1146 vm_offset_t e_end, e_start; 1147 int ahead, behind, cluster_offset, rv; 1148 u_char behavior; 1149 1150 /* 1151 * Prepare for unlocking the map. Save the map 1152 * entry's start and end addresses, which are used to 1153 * optimize the size of the pager operation below. 1154 * Even if the map entry's addresses change after 1155 * unlocking the map, using the saved addresses is 1156 * safe. 1157 */ 1158 e_start = fs->entry->start; 1159 e_end = fs->entry->end; 1160 behavior = vm_map_entry_behavior(fs->entry); 1161 1162 /* 1163 * Release the map lock before locking the vnode or 1164 * sleeping in the pager. (If the current object has 1165 * a shadow, then an earlier iteration of this loop 1166 * may have already unlocked the map.) 1167 */ 1168 unlock_map(fs); 1169 1170 rv = vm_fault_lock_vnode(fs, false); 1171 MPASS(rv == KERN_SUCCESS || rv == KERN_RESOURCE_SHORTAGE); 1172 if (rv == KERN_RESOURCE_SHORTAGE) 1173 return (rv); 1174 KASSERT(fs->vp == NULL || !fs->map->system_map, 1175 ("vm_fault: vnode-backed object mapped by system map")); 1176 1177 /* 1178 * Page in the requested page and hint the pager, 1179 * that it may bring up surrounding pages. 1180 */ 1181 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || 1182 P_KILLED(curproc)) { 1183 behind = 0; 1184 ahead = 0; 1185 } else { 1186 /* Is this a sequential fault? */ 1187 if (nera > 0) { 1188 behind = 0; 1189 ahead = nera; 1190 } else { 1191 /* 1192 * Request a cluster of pages that is 1193 * aligned to a VM_FAULT_READ_DEFAULT 1194 * page offset boundary within the 1195 * object. Alignment to a page offset 1196 * boundary is more likely to coincide 1197 * with the underlying file system 1198 * block than alignment to a virtual 1199 * address boundary. 1200 */ 1201 cluster_offset = fs->pindex % VM_FAULT_READ_DEFAULT; 1202 behind = ulmin(cluster_offset, 1203 atop(fs->vaddr - e_start)); 1204 ahead = VM_FAULT_READ_DEFAULT - 1 - cluster_offset; 1205 } 1206 ahead = ulmin(ahead, atop(e_end - fs->vaddr) - 1); 1207 } 1208 *behindp = behind; 1209 *aheadp = ahead; 1210 rv = vm_pager_get_pages(fs->object, &fs->m, 1, behindp, aheadp); 1211 if (rv == VM_PAGER_OK) 1212 return (KERN_SUCCESS); 1213 if (rv == VM_PAGER_ERROR) 1214 printf("vm_fault: pager read error, pid %d (%s)\n", 1215 curproc->p_pid, curproc->p_comm); 1216 /* 1217 * If an I/O error occurred or the requested page was 1218 * outside the range of the pager, clean up and return 1219 * an error. 1220 */ 1221 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) 1222 return (KERN_OUT_OF_BOUNDS); 1223 return (KERN_NOT_RECEIVER); 1224 } 1225 1226 /* 1227 * Wait/Retry if the page is busy. We have to do this if the page is 1228 * either exclusive or shared busy because the vm_pager may be using 1229 * read busy for pageouts (and even pageins if it is the vnode pager), 1230 * and we could end up trying to pagein and pageout the same page 1231 * simultaneously. 1232 * 1233 * We can theoretically allow the busy case on a read fault if the page 1234 * is marked valid, but since such pages are typically already pmap'd, 1235 * putting that special case in might be more effort then it is worth. 1236 * We cannot under any circumstances mess around with a shared busied 1237 * page except, perhaps, to pmap it. 1238 */ 1239 static void 1240 vm_fault_busy_sleep(struct faultstate *fs) 1241 { 1242 /* 1243 * Reference the page before unlocking and 1244 * sleeping so that the page daemon is less 1245 * likely to reclaim it. 1246 */ 1247 vm_page_aflag_set(fs->m, PGA_REFERENCED); 1248 if (fs->object != fs->first_object) { 1249 fault_page_release(&fs->first_m); 1250 vm_object_pip_wakeup(fs->first_object); 1251 } 1252 vm_object_pip_wakeup(fs->object); 1253 unlock_map(fs); 1254 if (fs->m == vm_page_lookup(fs->object, fs->pindex)) 1255 vm_page_busy_sleep(fs->m, "vmpfw", false); 1256 else 1257 VM_OBJECT_WUNLOCK(fs->object); 1258 VM_CNT_INC(v_intrans); 1259 vm_object_deallocate(fs->first_object); 1260 } 1261 1262 int 1263 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 1264 int fault_flags, vm_page_t *m_hold) 1265 { 1266 struct faultstate fs; 1267 int ahead, behind, faultcount; 1268 int nera, result, rv; 1269 bool dead, hardfault; 1270 1271 VM_CNT_INC(v_vm_faults); 1272 1273 if ((curthread->td_pflags & TDP_NOFAULTING) != 0) 1274 return (KERN_PROTECTION_FAILURE); 1275 1276 fs.vp = NULL; 1277 fs.vaddr = vaddr; 1278 fs.m_hold = m_hold; 1279 fs.fault_flags = fault_flags; 1280 fs.map = map; 1281 fs.lookup_still_valid = false; 1282 fs.oom = 0; 1283 faultcount = 0; 1284 nera = -1; 1285 hardfault = false; 1286 1287 RetryFault: 1288 fs.fault_type = fault_type; 1289 1290 /* 1291 * Find the backing store object and offset into it to begin the 1292 * search. 1293 */ 1294 result = vm_fault_lookup(&fs); 1295 if (result != KERN_SUCCESS) { 1296 if (result == KERN_RESOURCE_SHORTAGE) 1297 goto RetryFault; 1298 return (result); 1299 } 1300 1301 /* 1302 * Try to avoid lock contention on the top-level object through 1303 * special-case handling of some types of page faults, specifically, 1304 * those that are mapping an existing page from the top-level object. 1305 * Under this condition, a read lock on the object suffices, allowing 1306 * multiple page faults of a similar type to run in parallel. 1307 */ 1308 if (fs.vp == NULL /* avoid locked vnode leak */ && 1309 (fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) == 0 && 1310 (fs.fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0) { 1311 VM_OBJECT_RLOCK(fs.first_object); 1312 rv = vm_fault_soft_fast(&fs); 1313 if (rv == KERN_SUCCESS) 1314 return (rv); 1315 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { 1316 VM_OBJECT_RUNLOCK(fs.first_object); 1317 VM_OBJECT_WLOCK(fs.first_object); 1318 } 1319 } else { 1320 VM_OBJECT_WLOCK(fs.first_object); 1321 } 1322 1323 /* 1324 * Make a reference to this object to prevent its disposal while we 1325 * are messing with it. Once we have the reference, the map is free 1326 * to be diddled. Since objects reference their shadows (and copies), 1327 * they will stay around as well. 1328 * 1329 * Bump the paging-in-progress count to prevent size changes (e.g. 1330 * truncation operations) during I/O. 1331 */ 1332 vm_object_reference_locked(fs.first_object); 1333 vm_object_pip_add(fs.first_object, 1); 1334 1335 fs.m_cow = fs.m = fs.first_m = NULL; 1336 1337 /* 1338 * Search for the page at object/offset. 1339 */ 1340 fs.object = fs.first_object; 1341 fs.pindex = fs.first_pindex; 1342 1343 if ((fs.entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) != 0) { 1344 rv = vm_fault_allocate(&fs); 1345 switch (rv) { 1346 case KERN_RESTART: 1347 unlock_and_deallocate(&fs); 1348 /* FALLTHROUGH */ 1349 case KERN_RESOURCE_SHORTAGE: 1350 goto RetryFault; 1351 case KERN_SUCCESS: 1352 case KERN_FAILURE: 1353 case KERN_PROTECTION_FAILURE: 1354 case KERN_OUT_OF_BOUNDS: 1355 unlock_and_deallocate(&fs); 1356 return (rv); 1357 case KERN_NOT_RECEIVER: 1358 break; 1359 default: 1360 panic("vm_fault: Unhandled rv %d", rv); 1361 } 1362 } 1363 1364 while (TRUE) { 1365 KASSERT(fs.m == NULL, 1366 ("page still set %p at loop start", fs.m)); 1367 /* 1368 * If the object is marked for imminent termination, 1369 * we retry here, since the collapse pass has raced 1370 * with us. Otherwise, if we see terminally dead 1371 * object, return fail. 1372 */ 1373 if ((fs.object->flags & OBJ_DEAD) != 0) { 1374 dead = fs.object->type == OBJT_DEAD; 1375 unlock_and_deallocate(&fs); 1376 if (dead) 1377 return (KERN_PROTECTION_FAILURE); 1378 pause("vmf_de", 1); 1379 goto RetryFault; 1380 } 1381 1382 /* 1383 * See if page is resident 1384 */ 1385 fs.m = vm_page_lookup(fs.object, fs.pindex); 1386 if (fs.m != NULL) { 1387 if (vm_page_tryxbusy(fs.m) == 0) { 1388 vm_fault_busy_sleep(&fs); 1389 goto RetryFault; 1390 } 1391 1392 /* 1393 * The page is marked busy for other processes and the 1394 * pagedaemon. If it still is completely valid we 1395 * are done. 1396 */ 1397 if (vm_page_all_valid(fs.m)) { 1398 VM_OBJECT_WUNLOCK(fs.object); 1399 break; /* break to PAGE HAS BEEN FOUND. */ 1400 } 1401 } 1402 VM_OBJECT_ASSERT_WLOCKED(fs.object); 1403 1404 /* 1405 * Page is not resident. If the pager might contain the page 1406 * or this is the beginning of the search, allocate a new 1407 * page. (Default objects are zero-fill, so there is no real 1408 * pager for them.) 1409 */ 1410 if (fs.m == NULL && (fs.object->type != OBJT_DEFAULT || 1411 fs.object == fs.first_object)) { 1412 rv = vm_fault_allocate(&fs); 1413 switch (rv) { 1414 case KERN_RESTART: 1415 unlock_and_deallocate(&fs); 1416 /* FALLTHROUGH */ 1417 case KERN_RESOURCE_SHORTAGE: 1418 goto RetryFault; 1419 case KERN_SUCCESS: 1420 case KERN_FAILURE: 1421 case KERN_PROTECTION_FAILURE: 1422 case KERN_OUT_OF_BOUNDS: 1423 unlock_and_deallocate(&fs); 1424 return (rv); 1425 case KERN_NOT_RECEIVER: 1426 break; 1427 default: 1428 panic("vm_fault: Unhandled rv %d", rv); 1429 } 1430 } 1431 1432 /* 1433 * Default objects have no pager so no exclusive busy exists 1434 * to protect this page in the chain. Skip to the next 1435 * object without dropping the lock to preserve atomicity of 1436 * shadow faults. 1437 */ 1438 if (fs.object->type != OBJT_DEFAULT) { 1439 /* 1440 * At this point, we have either allocated a new page 1441 * or found an existing page that is only partially 1442 * valid. 1443 * 1444 * We hold a reference on the current object and the 1445 * page is exclusive busied. The exclusive busy 1446 * prevents simultaneous faults and collapses while 1447 * the object lock is dropped. 1448 */ 1449 VM_OBJECT_WUNLOCK(fs.object); 1450 1451 /* 1452 * If the pager for the current object might have 1453 * the page, then determine the number of additional 1454 * pages to read and potentially reprioritize 1455 * previously read pages for earlier reclamation. 1456 * These operations should only be performed once per 1457 * page fault. Even if the current pager doesn't 1458 * have the page, the number of additional pages to 1459 * read will apply to subsequent objects in the 1460 * shadow chain. 1461 */ 1462 if (nera == -1 && !P_KILLED(curproc)) 1463 nera = vm_fault_readahead(&fs); 1464 1465 rv = vm_fault_getpages(&fs, nera, &behind, &ahead); 1466 if (rv == KERN_SUCCESS) { 1467 faultcount = behind + 1 + ahead; 1468 hardfault = true; 1469 break; /* break to PAGE HAS BEEN FOUND. */ 1470 } 1471 if (rv == KERN_RESOURCE_SHORTAGE) 1472 goto RetryFault; 1473 VM_OBJECT_WLOCK(fs.object); 1474 if (rv == KERN_OUT_OF_BOUNDS) { 1475 fault_page_free(&fs.m); 1476 unlock_and_deallocate(&fs); 1477 return (rv); 1478 } 1479 } 1480 1481 /* 1482 * The page was not found in the current object. Try to 1483 * traverse into a backing object or zero fill if none is 1484 * found. 1485 */ 1486 if (vm_fault_next(&fs)) 1487 continue; 1488 if ((fs.fault_flags & VM_FAULT_NOFILL) != 0) { 1489 if (fs.first_object == fs.object) 1490 fault_page_free(&fs.first_m); 1491 unlock_and_deallocate(&fs); 1492 return (KERN_OUT_OF_BOUNDS); 1493 } 1494 VM_OBJECT_WUNLOCK(fs.object); 1495 vm_fault_zerofill(&fs); 1496 /* Don't try to prefault neighboring pages. */ 1497 faultcount = 1; 1498 break; /* break to PAGE HAS BEEN FOUND. */ 1499 } 1500 1501 /* 1502 * PAGE HAS BEEN FOUND. A valid page has been found and exclusively 1503 * busied. The object lock must no longer be held. 1504 */ 1505 vm_page_assert_xbusied(fs.m); 1506 VM_OBJECT_ASSERT_UNLOCKED(fs.object); 1507 1508 /* 1509 * If the page is being written, but isn't already owned by the 1510 * top-level object, we have to copy it into a new page owned by the 1511 * top-level object. 1512 */ 1513 if (fs.object != fs.first_object) { 1514 /* 1515 * We only really need to copy if we want to write it. 1516 */ 1517 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1518 vm_fault_cow(&fs); 1519 /* 1520 * We only try to prefault read-only mappings to the 1521 * neighboring pages when this copy-on-write fault is 1522 * a hard fault. In other cases, trying to prefault 1523 * is typically wasted effort. 1524 */ 1525 if (faultcount == 0) 1526 faultcount = 1; 1527 1528 } else { 1529 fs.prot &= ~VM_PROT_WRITE; 1530 } 1531 } 1532 1533 /* 1534 * We must verify that the maps have not changed since our last 1535 * lookup. 1536 */ 1537 if (!fs.lookup_still_valid) { 1538 result = vm_fault_relookup(&fs); 1539 if (result != KERN_SUCCESS) { 1540 fault_deallocate(&fs); 1541 if (result == KERN_RESTART) 1542 goto RetryFault; 1543 return (result); 1544 } 1545 } 1546 VM_OBJECT_ASSERT_UNLOCKED(fs.object); 1547 1548 /* 1549 * If the page was filled by a pager, save the virtual address that 1550 * should be faulted on next under a sequential access pattern to the 1551 * map entry. A read lock on the map suffices to update this address 1552 * safely. 1553 */ 1554 if (hardfault) 1555 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; 1556 1557 /* 1558 * Page must be completely valid or it is not fit to 1559 * map into user space. vm_pager_get_pages() ensures this. 1560 */ 1561 vm_page_assert_xbusied(fs.m); 1562 KASSERT(vm_page_all_valid(fs.m), 1563 ("vm_fault: page %p partially invalid", fs.m)); 1564 1565 vm_fault_dirty(&fs, fs.m); 1566 1567 /* 1568 * Put this page into the physical map. We had to do the unlock above 1569 * because pmap_enter() may sleep. We don't put the page 1570 * back on the active queue until later so that the pageout daemon 1571 * won't find it (yet). 1572 */ 1573 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, 1574 fs.fault_type | (fs.wired ? PMAP_ENTER_WIRED : 0), 0); 1575 if (faultcount != 1 && (fs.fault_flags & VM_FAULT_WIRE) == 0 && 1576 fs.wired == 0) 1577 vm_fault_prefault(&fs, vaddr, 1578 faultcount > 0 ? behind : PFBAK, 1579 faultcount > 0 ? ahead : PFFOR, false); 1580 1581 /* 1582 * If the page is not wired down, then put it where the pageout daemon 1583 * can find it. 1584 */ 1585 if ((fs.fault_flags & VM_FAULT_WIRE) != 0) 1586 vm_page_wire(fs.m); 1587 else 1588 vm_page_activate(fs.m); 1589 if (fs.m_hold != NULL) { 1590 (*fs.m_hold) = fs.m; 1591 vm_page_wire(fs.m); 1592 } 1593 vm_page_xunbusy(fs.m); 1594 fs.m = NULL; 1595 1596 /* 1597 * Unlock everything, and return 1598 */ 1599 fault_deallocate(&fs); 1600 if (hardfault) { 1601 VM_CNT_INC(v_io_faults); 1602 curthread->td_ru.ru_majflt++; 1603 #ifdef RACCT 1604 if (racct_enable && fs.object->type == OBJT_VNODE) { 1605 PROC_LOCK(curproc); 1606 if ((fs.fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1607 racct_add_force(curproc, RACCT_WRITEBPS, 1608 PAGE_SIZE + behind * PAGE_SIZE); 1609 racct_add_force(curproc, RACCT_WRITEIOPS, 1); 1610 } else { 1611 racct_add_force(curproc, RACCT_READBPS, 1612 PAGE_SIZE + ahead * PAGE_SIZE); 1613 racct_add_force(curproc, RACCT_READIOPS, 1); 1614 } 1615 PROC_UNLOCK(curproc); 1616 } 1617 #endif 1618 } else 1619 curthread->td_ru.ru_minflt++; 1620 1621 return (KERN_SUCCESS); 1622 } 1623 1624 /* 1625 * Speed up the reclamation of pages that precede the faulting pindex within 1626 * the first object of the shadow chain. Essentially, perform the equivalent 1627 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes 1628 * the faulting pindex by the cluster size when the pages read by vm_fault() 1629 * cross a cluster-size boundary. The cluster size is the greater of the 1630 * smallest superpage size and VM_FAULT_DONTNEED_MIN. 1631 * 1632 * When "fs->first_object" is a shadow object, the pages in the backing object 1633 * that precede the faulting pindex are deactivated by vm_fault(). So, this 1634 * function must only be concerned with pages in the first object. 1635 */ 1636 static void 1637 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) 1638 { 1639 vm_map_entry_t entry; 1640 vm_object_t first_object, object; 1641 vm_offset_t end, start; 1642 vm_page_t m, m_next; 1643 vm_pindex_t pend, pstart; 1644 vm_size_t size; 1645 1646 object = fs->object; 1647 VM_OBJECT_ASSERT_UNLOCKED(object); 1648 first_object = fs->first_object; 1649 /* Neither fictitious nor unmanaged pages can be reclaimed. */ 1650 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 1651 VM_OBJECT_RLOCK(first_object); 1652 size = VM_FAULT_DONTNEED_MIN; 1653 if (MAXPAGESIZES > 1 && size < pagesizes[1]) 1654 size = pagesizes[1]; 1655 end = rounddown2(vaddr, size); 1656 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && 1657 (entry = fs->entry)->start < end) { 1658 if (end - entry->start < size) 1659 start = entry->start; 1660 else 1661 start = end - size; 1662 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); 1663 pstart = OFF_TO_IDX(entry->offset) + atop(start - 1664 entry->start); 1665 m_next = vm_page_find_least(first_object, pstart); 1666 pend = OFF_TO_IDX(entry->offset) + atop(end - 1667 entry->start); 1668 while ((m = m_next) != NULL && m->pindex < pend) { 1669 m_next = TAILQ_NEXT(m, listq); 1670 if (!vm_page_all_valid(m) || 1671 vm_page_busied(m)) 1672 continue; 1673 1674 /* 1675 * Don't clear PGA_REFERENCED, since it would 1676 * likely represent a reference by a different 1677 * process. 1678 * 1679 * Typically, at this point, prefetched pages 1680 * are still in the inactive queue. Only 1681 * pages that triggered page faults are in the 1682 * active queue. The test for whether the page 1683 * is in the inactive queue is racy; in the 1684 * worst case we will requeue the page 1685 * unnecessarily. 1686 */ 1687 if (!vm_page_inactive(m)) 1688 vm_page_deactivate(m); 1689 } 1690 } 1691 VM_OBJECT_RUNLOCK(first_object); 1692 } 1693 } 1694 1695 /* 1696 * vm_fault_prefault provides a quick way of clustering 1697 * pagefaults into a processes address space. It is a "cousin" 1698 * of vm_map_pmap_enter, except it runs at page fault time instead 1699 * of mmap time. 1700 */ 1701 static void 1702 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 1703 int backward, int forward, bool obj_locked) 1704 { 1705 pmap_t pmap; 1706 vm_map_entry_t entry; 1707 vm_object_t backing_object, lobject; 1708 vm_offset_t addr, starta; 1709 vm_pindex_t pindex; 1710 vm_page_t m; 1711 int i; 1712 1713 pmap = fs->map->pmap; 1714 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1715 return; 1716 1717 entry = fs->entry; 1718 1719 if (addra < backward * PAGE_SIZE) { 1720 starta = entry->start; 1721 } else { 1722 starta = addra - backward * PAGE_SIZE; 1723 if (starta < entry->start) 1724 starta = entry->start; 1725 } 1726 1727 /* 1728 * Generate the sequence of virtual addresses that are candidates for 1729 * prefaulting in an outward spiral from the faulting virtual address, 1730 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra 1731 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... 1732 * If the candidate address doesn't have a backing physical page, then 1733 * the loop immediately terminates. 1734 */ 1735 for (i = 0; i < 2 * imax(backward, forward); i++) { 1736 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : 1737 PAGE_SIZE); 1738 if (addr > addra + forward * PAGE_SIZE) 1739 addr = 0; 1740 1741 if (addr < starta || addr >= entry->end) 1742 continue; 1743 1744 if (!pmap_is_prefaultable(pmap, addr)) 1745 continue; 1746 1747 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1748 lobject = entry->object.vm_object; 1749 if (!obj_locked) 1750 VM_OBJECT_RLOCK(lobject); 1751 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1752 lobject->type == OBJT_DEFAULT && 1753 (backing_object = lobject->backing_object) != NULL) { 1754 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1755 0, ("vm_fault_prefault: unaligned object offset")); 1756 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1757 VM_OBJECT_RLOCK(backing_object); 1758 if (!obj_locked || lobject != entry->object.vm_object) 1759 VM_OBJECT_RUNLOCK(lobject); 1760 lobject = backing_object; 1761 } 1762 if (m == NULL) { 1763 if (!obj_locked || lobject != entry->object.vm_object) 1764 VM_OBJECT_RUNLOCK(lobject); 1765 break; 1766 } 1767 if (vm_page_all_valid(m) && 1768 (m->flags & PG_FICTITIOUS) == 0) 1769 pmap_enter_quick(pmap, addr, m, entry->protection); 1770 if (!obj_locked || lobject != entry->object.vm_object) 1771 VM_OBJECT_RUNLOCK(lobject); 1772 } 1773 } 1774 1775 /* 1776 * Hold each of the physical pages that are mapped by the specified range of 1777 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1778 * and allow the specified types of access, "prot". If all of the implied 1779 * pages are successfully held, then the number of held pages is returned 1780 * together with pointers to those pages in the array "ma". However, if any 1781 * of the pages cannot be held, -1 is returned. 1782 */ 1783 int 1784 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1785 vm_prot_t prot, vm_page_t *ma, int max_count) 1786 { 1787 vm_offset_t end, va; 1788 vm_page_t *mp; 1789 int count; 1790 boolean_t pmap_failed; 1791 1792 if (len == 0) 1793 return (0); 1794 end = round_page(addr + len); 1795 addr = trunc_page(addr); 1796 1797 if (!vm_map_range_valid(map, addr, end)) 1798 return (-1); 1799 1800 if (atop(end - addr) > max_count) 1801 panic("vm_fault_quick_hold_pages: count > max_count"); 1802 count = atop(end - addr); 1803 1804 /* 1805 * Most likely, the physical pages are resident in the pmap, so it is 1806 * faster to try pmap_extract_and_hold() first. 1807 */ 1808 pmap_failed = FALSE; 1809 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1810 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1811 if (*mp == NULL) 1812 pmap_failed = TRUE; 1813 else if ((prot & VM_PROT_WRITE) != 0 && 1814 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1815 /* 1816 * Explicitly dirty the physical page. Otherwise, the 1817 * caller's changes may go unnoticed because they are 1818 * performed through an unmanaged mapping or by a DMA 1819 * operation. 1820 * 1821 * The object lock is not held here. 1822 * See vm_page_clear_dirty_mask(). 1823 */ 1824 vm_page_dirty(*mp); 1825 } 1826 } 1827 if (pmap_failed) { 1828 /* 1829 * One or more pages could not be held by the pmap. Either no 1830 * page was mapped at the specified virtual address or that 1831 * mapping had insufficient permissions. Attempt to fault in 1832 * and hold these pages. 1833 * 1834 * If vm_fault_disable_pagefaults() was called, 1835 * i.e., TDP_NOFAULTING is set, we must not sleep nor 1836 * acquire MD VM locks, which means we must not call 1837 * vm_fault(). Some (out of tree) callers mark 1838 * too wide a code area with vm_fault_disable_pagefaults() 1839 * already, use the VM_PROT_QUICK_NOFAULT flag to request 1840 * the proper behaviour explicitly. 1841 */ 1842 if ((prot & VM_PROT_QUICK_NOFAULT) != 0 && 1843 (curthread->td_pflags & TDP_NOFAULTING) != 0) 1844 goto error; 1845 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1846 if (*mp == NULL && vm_fault(map, va, prot, 1847 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1848 goto error; 1849 } 1850 return (count); 1851 error: 1852 for (mp = ma; mp < ma + count; mp++) 1853 if (*mp != NULL) 1854 vm_page_unwire(*mp, PQ_INACTIVE); 1855 return (-1); 1856 } 1857 1858 /* 1859 * Routine: 1860 * vm_fault_copy_entry 1861 * Function: 1862 * Create new shadow object backing dst_entry with private copy of 1863 * all underlying pages. When src_entry is equal to dst_entry, 1864 * function implements COW for wired-down map entry. Otherwise, 1865 * it forks wired entry into dst_map. 1866 * 1867 * In/out conditions: 1868 * The source and destination maps must be locked for write. 1869 * The source map entry must be wired down (or be a sharing map 1870 * entry corresponding to a main map entry that is wired down). 1871 */ 1872 void 1873 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1874 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1875 vm_ooffset_t *fork_charge) 1876 { 1877 vm_object_t backing_object, dst_object, object, src_object; 1878 vm_pindex_t dst_pindex, pindex, src_pindex; 1879 vm_prot_t access, prot; 1880 vm_offset_t vaddr; 1881 vm_page_t dst_m; 1882 vm_page_t src_m; 1883 boolean_t upgrade; 1884 1885 #ifdef lint 1886 src_map++; 1887 #endif /* lint */ 1888 1889 upgrade = src_entry == dst_entry; 1890 access = prot = dst_entry->protection; 1891 1892 src_object = src_entry->object.vm_object; 1893 src_pindex = OFF_TO_IDX(src_entry->offset); 1894 1895 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 1896 dst_object = src_object; 1897 vm_object_reference(dst_object); 1898 } else { 1899 /* 1900 * Create the top-level object for the destination entry. 1901 * Doesn't actually shadow anything - we copy the pages 1902 * directly. 1903 */ 1904 dst_object = vm_object_allocate_anon(atop(dst_entry->end - 1905 dst_entry->start), NULL, NULL, 0); 1906 #if VM_NRESERVLEVEL > 0 1907 dst_object->flags |= OBJ_COLORED; 1908 dst_object->pg_color = atop(dst_entry->start); 1909 #endif 1910 dst_object->domain = src_object->domain; 1911 dst_object->charge = dst_entry->end - dst_entry->start; 1912 } 1913 1914 VM_OBJECT_WLOCK(dst_object); 1915 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1916 ("vm_fault_copy_entry: vm_object not NULL")); 1917 if (src_object != dst_object) { 1918 dst_entry->object.vm_object = dst_object; 1919 dst_entry->offset = 0; 1920 dst_entry->eflags &= ~MAP_ENTRY_VN_EXEC; 1921 } 1922 if (fork_charge != NULL) { 1923 KASSERT(dst_entry->cred == NULL, 1924 ("vm_fault_copy_entry: leaked swp charge")); 1925 dst_object->cred = curthread->td_ucred; 1926 crhold(dst_object->cred); 1927 *fork_charge += dst_object->charge; 1928 } else if ((dst_object->type == OBJT_DEFAULT || 1929 dst_object->type == OBJT_SWAP) && 1930 dst_object->cred == NULL) { 1931 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", 1932 dst_entry)); 1933 dst_object->cred = dst_entry->cred; 1934 dst_entry->cred = NULL; 1935 } 1936 1937 /* 1938 * If not an upgrade, then enter the mappings in the pmap as 1939 * read and/or execute accesses. Otherwise, enter them as 1940 * write accesses. 1941 * 1942 * A writeable large page mapping is only created if all of 1943 * the constituent small page mappings are modified. Marking 1944 * PTEs as modified on inception allows promotion to happen 1945 * without taking potentially large number of soft faults. 1946 */ 1947 if (!upgrade) 1948 access &= ~VM_PROT_WRITE; 1949 1950 /* 1951 * Loop through all of the virtual pages within the entry's 1952 * range, copying each page from the source object to the 1953 * destination object. Since the source is wired, those pages 1954 * must exist. In contrast, the destination is pageable. 1955 * Since the destination object doesn't share any backing storage 1956 * with the source object, all of its pages must be dirtied, 1957 * regardless of whether they can be written. 1958 */ 1959 for (vaddr = dst_entry->start, dst_pindex = 0; 1960 vaddr < dst_entry->end; 1961 vaddr += PAGE_SIZE, dst_pindex++) { 1962 again: 1963 /* 1964 * Find the page in the source object, and copy it in. 1965 * Because the source is wired down, the page will be 1966 * in memory. 1967 */ 1968 if (src_object != dst_object) 1969 VM_OBJECT_RLOCK(src_object); 1970 object = src_object; 1971 pindex = src_pindex + dst_pindex; 1972 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1973 (backing_object = object->backing_object) != NULL) { 1974 /* 1975 * Unless the source mapping is read-only or 1976 * it is presently being upgraded from 1977 * read-only, the first object in the shadow 1978 * chain should provide all of the pages. In 1979 * other words, this loop body should never be 1980 * executed when the source mapping is already 1981 * read/write. 1982 */ 1983 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || 1984 upgrade, 1985 ("vm_fault_copy_entry: main object missing page")); 1986 1987 VM_OBJECT_RLOCK(backing_object); 1988 pindex += OFF_TO_IDX(object->backing_object_offset); 1989 if (object != dst_object) 1990 VM_OBJECT_RUNLOCK(object); 1991 object = backing_object; 1992 } 1993 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); 1994 1995 if (object != dst_object) { 1996 /* 1997 * Allocate a page in the destination object. 1998 */ 1999 dst_m = vm_page_alloc(dst_object, (src_object == 2000 dst_object ? src_pindex : 0) + dst_pindex, 2001 VM_ALLOC_NORMAL); 2002 if (dst_m == NULL) { 2003 VM_OBJECT_WUNLOCK(dst_object); 2004 VM_OBJECT_RUNLOCK(object); 2005 vm_wait(dst_object); 2006 VM_OBJECT_WLOCK(dst_object); 2007 goto again; 2008 } 2009 pmap_copy_page(src_m, dst_m); 2010 VM_OBJECT_RUNLOCK(object); 2011 dst_m->dirty = dst_m->valid = src_m->valid; 2012 } else { 2013 dst_m = src_m; 2014 if (vm_page_busy_acquire(dst_m, VM_ALLOC_WAITFAIL) == 0) 2015 goto again; 2016 if (dst_m->pindex >= dst_object->size) { 2017 /* 2018 * We are upgrading. Index can occur 2019 * out of bounds if the object type is 2020 * vnode and the file was truncated. 2021 */ 2022 vm_page_xunbusy(dst_m); 2023 break; 2024 } 2025 } 2026 VM_OBJECT_WUNLOCK(dst_object); 2027 2028 /* 2029 * Enter it in the pmap. If a wired, copy-on-write 2030 * mapping is being replaced by a write-enabled 2031 * mapping, then wire that new mapping. 2032 * 2033 * The page can be invalid if the user called 2034 * msync(MS_INVALIDATE) or truncated the backing vnode 2035 * or shared memory object. In this case, do not 2036 * insert it into pmap, but still do the copy so that 2037 * all copies of the wired map entry have similar 2038 * backing pages. 2039 */ 2040 if (vm_page_all_valid(dst_m)) { 2041 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, 2042 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); 2043 } 2044 2045 /* 2046 * Mark it no longer busy, and put it on the active list. 2047 */ 2048 VM_OBJECT_WLOCK(dst_object); 2049 2050 if (upgrade) { 2051 if (src_m != dst_m) { 2052 vm_page_unwire(src_m, PQ_INACTIVE); 2053 vm_page_wire(dst_m); 2054 } else { 2055 KASSERT(vm_page_wired(dst_m), 2056 ("dst_m %p is not wired", dst_m)); 2057 } 2058 } else { 2059 vm_page_activate(dst_m); 2060 } 2061 vm_page_xunbusy(dst_m); 2062 } 2063 VM_OBJECT_WUNLOCK(dst_object); 2064 if (upgrade) { 2065 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 2066 vm_object_deallocate(src_object); 2067 } 2068 } 2069 2070 /* 2071 * Block entry into the machine-independent layer's page fault handler by 2072 * the calling thread. Subsequent calls to vm_fault() by that thread will 2073 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 2074 * spurious page faults. 2075 */ 2076 int 2077 vm_fault_disable_pagefaults(void) 2078 { 2079 2080 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 2081 } 2082 2083 void 2084 vm_fault_enable_pagefaults(int save) 2085 { 2086 2087 curthread_pflags_restore(save); 2088 } 2089