1 /*- 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * 9 * 10 * This code is derived from software contributed to Berkeley by 11 * The Mach Operating System project at Carnegie-Mellon University. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by the University of 24 * California, Berkeley and its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 42 * 43 * 44 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 45 * All rights reserved. 46 * 47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 48 * 49 * Permission to use, copy, modify and distribute this software and 50 * its documentation is hereby granted, provided that both the copyright 51 * notice and this permission notice appear in all copies of the 52 * software, derivative works or modified versions, and any portions 53 * thereof, and that both notices appear in supporting documentation. 54 * 55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 58 * 59 * Carnegie Mellon requests users of this software to return to 60 * 61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 62 * School of Computer Science 63 * Carnegie Mellon University 64 * Pittsburgh PA 15213-3890 65 * 66 * any improvements or extensions that they make and grant Carnegie the 67 * rights to redistribute these changes. 68 */ 69 70 /* 71 * Page fault handling module. 72 */ 73 74 #include <sys/cdefs.h> 75 __FBSDID("$FreeBSD$"); 76 77 #include "opt_ktrace.h" 78 #include "opt_vm.h" 79 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/kernel.h> 83 #include <sys/lock.h> 84 #include <sys/mman.h> 85 #include <sys/proc.h> 86 #include <sys/racct.h> 87 #include <sys/resourcevar.h> 88 #include <sys/rwlock.h> 89 #include <sys/sysctl.h> 90 #include <sys/vmmeter.h> 91 #include <sys/vnode.h> 92 #ifdef KTRACE 93 #include <sys/ktrace.h> 94 #endif 95 96 #include <vm/vm.h> 97 #include <vm/vm_param.h> 98 #include <vm/pmap.h> 99 #include <vm/vm_map.h> 100 #include <vm/vm_object.h> 101 #include <vm/vm_page.h> 102 #include <vm/vm_pageout.h> 103 #include <vm/vm_kern.h> 104 #include <vm/vm_pager.h> 105 #include <vm/vm_extern.h> 106 #include <vm/vm_reserv.h> 107 108 #define PFBAK 4 109 #define PFFOR 4 110 111 #define VM_FAULT_READ_DEFAULT (1 + VM_FAULT_READ_AHEAD_INIT) 112 #define VM_FAULT_READ_MAX (1 + VM_FAULT_READ_AHEAD_MAX) 113 114 #define VM_FAULT_DONTNEED_MIN 1048576 115 116 struct faultstate { 117 vm_page_t m; 118 vm_object_t object; 119 vm_pindex_t pindex; 120 vm_page_t first_m; 121 vm_object_t first_object; 122 vm_pindex_t first_pindex; 123 vm_map_t map; 124 vm_map_entry_t entry; 125 int map_generation; 126 bool lookup_still_valid; 127 struct vnode *vp; 128 }; 129 130 static void vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, 131 int ahead); 132 static void vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 133 int backward, int forward); 134 135 static inline void 136 release_page(struct faultstate *fs) 137 { 138 139 vm_page_xunbusy(fs->m); 140 vm_page_lock(fs->m); 141 vm_page_deactivate(fs->m); 142 vm_page_unlock(fs->m); 143 fs->m = NULL; 144 } 145 146 static inline void 147 unlock_map(struct faultstate *fs) 148 { 149 150 if (fs->lookup_still_valid) { 151 vm_map_lookup_done(fs->map, fs->entry); 152 fs->lookup_still_valid = false; 153 } 154 } 155 156 static void 157 unlock_vp(struct faultstate *fs) 158 { 159 160 if (fs->vp != NULL) { 161 vput(fs->vp); 162 fs->vp = NULL; 163 } 164 } 165 166 static void 167 unlock_and_deallocate(struct faultstate *fs) 168 { 169 170 vm_object_pip_wakeup(fs->object); 171 VM_OBJECT_WUNLOCK(fs->object); 172 if (fs->object != fs->first_object) { 173 VM_OBJECT_WLOCK(fs->first_object); 174 vm_page_lock(fs->first_m); 175 vm_page_free(fs->first_m); 176 vm_page_unlock(fs->first_m); 177 vm_object_pip_wakeup(fs->first_object); 178 VM_OBJECT_WUNLOCK(fs->first_object); 179 fs->first_m = NULL; 180 } 181 vm_object_deallocate(fs->first_object); 182 unlock_map(fs); 183 unlock_vp(fs); 184 } 185 186 static void 187 vm_fault_dirty(vm_map_entry_t entry, vm_page_t m, vm_prot_t prot, 188 vm_prot_t fault_type, int fault_flags, bool set_wd) 189 { 190 bool need_dirty; 191 192 if (((prot & VM_PROT_WRITE) == 0 && 193 (fault_flags & VM_FAULT_DIRTY) == 0) || 194 (m->oflags & VPO_UNMANAGED) != 0) 195 return; 196 197 VM_OBJECT_ASSERT_LOCKED(m->object); 198 199 need_dirty = ((fault_type & VM_PROT_WRITE) != 0 && 200 (fault_flags & VM_FAULT_WIRE) == 0) || 201 (fault_flags & VM_FAULT_DIRTY) != 0; 202 203 if (set_wd) 204 vm_object_set_writeable_dirty(m->object); 205 else 206 /* 207 * If two callers of vm_fault_dirty() with set_wd == 208 * FALSE, one for the map entry with MAP_ENTRY_NOSYNC 209 * flag set, other with flag clear, race, it is 210 * possible for the no-NOSYNC thread to see m->dirty 211 * != 0 and not clear VPO_NOSYNC. Take vm_page lock 212 * around manipulation of VPO_NOSYNC and 213 * vm_page_dirty() call, to avoid the race and keep 214 * m->oflags consistent. 215 */ 216 vm_page_lock(m); 217 218 /* 219 * If this is a NOSYNC mmap we do not want to set VPO_NOSYNC 220 * if the page is already dirty to prevent data written with 221 * the expectation of being synced from not being synced. 222 * Likewise if this entry does not request NOSYNC then make 223 * sure the page isn't marked NOSYNC. Applications sharing 224 * data should use the same flags to avoid ping ponging. 225 */ 226 if ((entry->eflags & MAP_ENTRY_NOSYNC) != 0) { 227 if (m->dirty == 0) { 228 m->oflags |= VPO_NOSYNC; 229 } 230 } else { 231 m->oflags &= ~VPO_NOSYNC; 232 } 233 234 /* 235 * If the fault is a write, we know that this page is being 236 * written NOW so dirty it explicitly to save on 237 * pmap_is_modified() calls later. 238 * 239 * Also tell the backing pager, if any, that it should remove 240 * any swap backing since the page is now dirty. 241 */ 242 if (need_dirty) 243 vm_page_dirty(m); 244 if (!set_wd) 245 vm_page_unlock(m); 246 if (need_dirty) 247 vm_pager_page_unswapped(m); 248 } 249 250 static void 251 vm_fault_fill_hold(vm_page_t *m_hold, vm_page_t m) 252 { 253 254 if (m_hold != NULL) { 255 *m_hold = m; 256 vm_page_lock(m); 257 vm_page_hold(m); 258 vm_page_unlock(m); 259 } 260 } 261 262 /* 263 * Unlocks fs.first_object and fs.map on success. 264 */ 265 static int 266 vm_fault_soft_fast(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot, 267 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) 268 { 269 vm_page_t m; 270 int rv; 271 272 MPASS(fs->vp == NULL); 273 m = vm_page_lookup(fs->first_object, fs->first_pindex); 274 /* A busy page can be mapped for read|execute access. */ 275 if (m == NULL || ((prot & VM_PROT_WRITE) != 0 && 276 vm_page_busied(m)) || m->valid != VM_PAGE_BITS_ALL) 277 return (KERN_FAILURE); 278 rv = pmap_enter(fs->map->pmap, vaddr, m, prot, fault_type | 279 PMAP_ENTER_NOSLEEP | (wired ? PMAP_ENTER_WIRED : 0), 0); 280 if (rv != KERN_SUCCESS) 281 return (rv); 282 vm_fault_fill_hold(m_hold, m); 283 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, false); 284 VM_OBJECT_RUNLOCK(fs->first_object); 285 if (!wired) 286 vm_fault_prefault(fs, vaddr, PFBAK, PFFOR); 287 vm_map_lookup_done(fs->map, fs->entry); 288 curthread->td_ru.ru_minflt++; 289 return (KERN_SUCCESS); 290 } 291 292 static void 293 vm_fault_restore_map_lock(struct faultstate *fs) 294 { 295 296 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 297 MPASS(fs->first_object->paging_in_progress > 0); 298 299 if (!vm_map_trylock_read(fs->map)) { 300 VM_OBJECT_WUNLOCK(fs->first_object); 301 vm_map_lock_read(fs->map); 302 VM_OBJECT_WLOCK(fs->first_object); 303 } 304 fs->lookup_still_valid = true; 305 } 306 307 static void 308 vm_fault_populate_check_page(vm_page_t m) 309 { 310 311 /* 312 * Check each page to ensure that the pager is obeying the 313 * interface: the page must be installed in the object, fully 314 * valid, and exclusively busied. 315 */ 316 MPASS(m != NULL); 317 MPASS(m->valid == VM_PAGE_BITS_ALL); 318 MPASS(vm_page_xbusied(m)); 319 } 320 321 static void 322 vm_fault_populate_cleanup(vm_object_t object, vm_pindex_t first, 323 vm_pindex_t last) 324 { 325 vm_page_t m; 326 vm_pindex_t pidx; 327 328 VM_OBJECT_ASSERT_WLOCKED(object); 329 MPASS(first <= last); 330 for (pidx = first, m = vm_page_lookup(object, pidx); 331 pidx <= last; pidx++, m = vm_page_next(m)) { 332 vm_fault_populate_check_page(m); 333 vm_page_lock(m); 334 vm_page_deactivate(m); 335 vm_page_unlock(m); 336 vm_page_xunbusy(m); 337 } 338 } 339 340 static int 341 vm_fault_populate(struct faultstate *fs, vm_offset_t vaddr, vm_prot_t prot, 342 int fault_type, int fault_flags, boolean_t wired, vm_page_t *m_hold) 343 { 344 vm_page_t m; 345 vm_pindex_t map_first, map_last, pager_first, pager_last, pidx; 346 int rv; 347 348 MPASS(fs->object == fs->first_object); 349 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 350 MPASS(fs->first_object->paging_in_progress > 0); 351 MPASS(fs->first_object->backing_object == NULL); 352 MPASS(fs->lookup_still_valid); 353 354 pager_first = OFF_TO_IDX(fs->entry->offset); 355 pager_last = pager_first + atop(fs->entry->end - fs->entry->start) - 1; 356 unlock_map(fs); 357 unlock_vp(fs); 358 359 /* 360 * Call the pager (driver) populate() method. 361 * 362 * There is no guarantee that the method will be called again 363 * if the current fault is for read, and a future fault is 364 * for write. Report the entry's maximum allowed protection 365 * to the driver. 366 */ 367 rv = vm_pager_populate(fs->first_object, fs->first_pindex, 368 fault_type, fs->entry->max_protection, &pager_first, &pager_last); 369 370 VM_OBJECT_ASSERT_WLOCKED(fs->first_object); 371 if (rv == VM_PAGER_BAD) { 372 /* 373 * VM_PAGER_BAD is the backdoor for a pager to request 374 * normal fault handling. 375 */ 376 vm_fault_restore_map_lock(fs); 377 if (fs->map->timestamp != fs->map_generation) 378 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ 379 return (KERN_NOT_RECEIVER); 380 } 381 if (rv != VM_PAGER_OK) 382 return (KERN_FAILURE); /* AKA SIGSEGV */ 383 384 /* Ensure that the driver is obeying the interface. */ 385 MPASS(pager_first <= pager_last); 386 MPASS(fs->first_pindex <= pager_last); 387 MPASS(fs->first_pindex >= pager_first); 388 MPASS(pager_last < fs->first_object->size); 389 390 vm_fault_restore_map_lock(fs); 391 if (fs->map->timestamp != fs->map_generation) { 392 vm_fault_populate_cleanup(fs->first_object, pager_first, 393 pager_last); 394 return (KERN_RESOURCE_SHORTAGE); /* RetryFault */ 395 } 396 397 /* 398 * The map is unchanged after our last unlock. Process the fault. 399 * 400 * The range [pager_first, pager_last] that is given to the 401 * pager is only a hint. The pager may populate any range 402 * within the object that includes the requested page index. 403 * In case the pager expanded the range, clip it to fit into 404 * the map entry. 405 */ 406 map_first = OFF_TO_IDX(fs->entry->offset); 407 if (map_first > pager_first) { 408 vm_fault_populate_cleanup(fs->first_object, pager_first, 409 map_first - 1); 410 pager_first = map_first; 411 } 412 map_last = map_first + atop(fs->entry->end - fs->entry->start) - 1; 413 if (map_last < pager_last) { 414 vm_fault_populate_cleanup(fs->first_object, map_last + 1, 415 pager_last); 416 pager_last = map_last; 417 } 418 for (pidx = pager_first, m = vm_page_lookup(fs->first_object, pidx); 419 pidx <= pager_last; pidx++, m = vm_page_next(m)) { 420 vm_fault_populate_check_page(m); 421 vm_fault_dirty(fs->entry, m, prot, fault_type, fault_flags, 422 true); 423 VM_OBJECT_WUNLOCK(fs->first_object); 424 pmap_enter(fs->map->pmap, fs->entry->start + IDX_TO_OFF(pidx) - 425 fs->entry->offset, m, prot, fault_type | (wired ? 426 PMAP_ENTER_WIRED : 0), 0); 427 VM_OBJECT_WLOCK(fs->first_object); 428 if (pidx == fs->first_pindex) 429 vm_fault_fill_hold(m_hold, m); 430 vm_page_lock(m); 431 if ((fault_flags & VM_FAULT_WIRE) != 0) { 432 KASSERT(wired, ("VM_FAULT_WIRE && !wired")); 433 vm_page_wire(m); 434 } else { 435 vm_page_activate(m); 436 } 437 vm_page_unlock(m); 438 vm_page_xunbusy(m); 439 } 440 curthread->td_ru.ru_majflt++; 441 return (KERN_SUCCESS); 442 } 443 444 /* 445 * vm_fault: 446 * 447 * Handle a page fault occurring at the given address, 448 * requiring the given permissions, in the map specified. 449 * If successful, the page is inserted into the 450 * associated physical map. 451 * 452 * NOTE: the given address should be truncated to the 453 * proper page address. 454 * 455 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 456 * a standard error specifying why the fault is fatal is returned. 457 * 458 * The map in question must be referenced, and remains so. 459 * Caller may hold no locks. 460 */ 461 int 462 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 463 int fault_flags) 464 { 465 struct thread *td; 466 int result; 467 468 td = curthread; 469 if ((td->td_pflags & TDP_NOFAULTING) != 0) 470 return (KERN_PROTECTION_FAILURE); 471 #ifdef KTRACE 472 if (map != kernel_map && KTRPOINT(td, KTR_FAULT)) 473 ktrfault(vaddr, fault_type); 474 #endif 475 result = vm_fault_hold(map, trunc_page(vaddr), fault_type, fault_flags, 476 NULL); 477 #ifdef KTRACE 478 if (map != kernel_map && KTRPOINT(td, KTR_FAULTEND)) 479 ktrfaultend(result); 480 #endif 481 return (result); 482 } 483 484 int 485 vm_fault_hold(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 486 int fault_flags, vm_page_t *m_hold) 487 { 488 struct faultstate fs; 489 struct vnode *vp; 490 vm_object_t next_object, retry_object; 491 vm_offset_t e_end, e_start; 492 vm_pindex_t retry_pindex; 493 vm_prot_t prot, retry_prot; 494 int ahead, alloc_req, behind, cluster_offset, error, era, faultcount; 495 int locked, nera, result, rv; 496 u_char behavior; 497 boolean_t wired; /* Passed by reference. */ 498 bool dead, hardfault, is_first_object_locked; 499 500 VM_CNT_INC(v_vm_faults); 501 fs.vp = NULL; 502 faultcount = 0; 503 nera = -1; 504 hardfault = false; 505 506 RetryFault:; 507 508 /* 509 * Find the backing store object and offset into it to begin the 510 * search. 511 */ 512 fs.map = map; 513 result = vm_map_lookup(&fs.map, vaddr, fault_type | 514 VM_PROT_FAULT_LOOKUP, &fs.entry, &fs.first_object, 515 &fs.first_pindex, &prot, &wired); 516 if (result != KERN_SUCCESS) { 517 unlock_vp(&fs); 518 return (result); 519 } 520 521 fs.map_generation = fs.map->timestamp; 522 523 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 524 panic("%s: fault on nofault entry, addr: %#lx", 525 __func__, (u_long)vaddr); 526 } 527 528 if (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION && 529 fs.entry->wiring_thread != curthread) { 530 vm_map_unlock_read(fs.map); 531 vm_map_lock(fs.map); 532 if (vm_map_lookup_entry(fs.map, vaddr, &fs.entry) && 533 (fs.entry->eflags & MAP_ENTRY_IN_TRANSITION)) { 534 unlock_vp(&fs); 535 fs.entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 536 vm_map_unlock_and_wait(fs.map, 0); 537 } else 538 vm_map_unlock(fs.map); 539 goto RetryFault; 540 } 541 542 MPASS((fs.entry->eflags & MAP_ENTRY_GUARD) == 0); 543 544 if (wired) 545 fault_type = prot | (fault_type & VM_PROT_COPY); 546 else 547 KASSERT((fault_flags & VM_FAULT_WIRE) == 0, 548 ("!wired && VM_FAULT_WIRE")); 549 550 /* 551 * Try to avoid lock contention on the top-level object through 552 * special-case handling of some types of page faults, specifically, 553 * those that are both (1) mapping an existing page from the top- 554 * level object and (2) not having to mark that object as containing 555 * dirty pages. Under these conditions, a read lock on the top-level 556 * object suffices, allowing multiple page faults of a similar type to 557 * run in parallel on the same top-level object. 558 */ 559 if (fs.vp == NULL /* avoid locked vnode leak */ && 560 (fault_flags & (VM_FAULT_WIRE | VM_FAULT_DIRTY)) == 0 && 561 /* avoid calling vm_object_set_writeable_dirty() */ 562 ((prot & VM_PROT_WRITE) == 0 || 563 (fs.first_object->type != OBJT_VNODE && 564 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 565 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0)) { 566 VM_OBJECT_RLOCK(fs.first_object); 567 if ((prot & VM_PROT_WRITE) == 0 || 568 (fs.first_object->type != OBJT_VNODE && 569 (fs.first_object->flags & OBJ_TMPFS_NODE) == 0) || 570 (fs.first_object->flags & OBJ_MIGHTBEDIRTY) != 0) { 571 rv = vm_fault_soft_fast(&fs, vaddr, prot, fault_type, 572 fault_flags, wired, m_hold); 573 if (rv == KERN_SUCCESS) 574 return (rv); 575 } 576 if (!VM_OBJECT_TRYUPGRADE(fs.first_object)) { 577 VM_OBJECT_RUNLOCK(fs.first_object); 578 VM_OBJECT_WLOCK(fs.first_object); 579 } 580 } else { 581 VM_OBJECT_WLOCK(fs.first_object); 582 } 583 584 /* 585 * Make a reference to this object to prevent its disposal while we 586 * are messing with it. Once we have the reference, the map is free 587 * to be diddled. Since objects reference their shadows (and copies), 588 * they will stay around as well. 589 * 590 * Bump the paging-in-progress count to prevent size changes (e.g. 591 * truncation operations) during I/O. 592 */ 593 vm_object_reference_locked(fs.first_object); 594 vm_object_pip_add(fs.first_object, 1); 595 596 fs.lookup_still_valid = true; 597 598 fs.first_m = NULL; 599 600 /* 601 * Search for the page at object/offset. 602 */ 603 fs.object = fs.first_object; 604 fs.pindex = fs.first_pindex; 605 while (TRUE) { 606 /* 607 * If the object is marked for imminent termination, 608 * we retry here, since the collapse pass has raced 609 * with us. Otherwise, if we see terminally dead 610 * object, return fail. 611 */ 612 if ((fs.object->flags & OBJ_DEAD) != 0) { 613 dead = fs.object->type == OBJT_DEAD; 614 unlock_and_deallocate(&fs); 615 if (dead) 616 return (KERN_PROTECTION_FAILURE); 617 pause("vmf_de", 1); 618 goto RetryFault; 619 } 620 621 /* 622 * See if page is resident 623 */ 624 fs.m = vm_page_lookup(fs.object, fs.pindex); 625 if (fs.m != NULL) { 626 /* 627 * Wait/Retry if the page is busy. We have to do this 628 * if the page is either exclusive or shared busy 629 * because the vm_pager may be using read busy for 630 * pageouts (and even pageins if it is the vnode 631 * pager), and we could end up trying to pagein and 632 * pageout the same page simultaneously. 633 * 634 * We can theoretically allow the busy case on a read 635 * fault if the page is marked valid, but since such 636 * pages are typically already pmap'd, putting that 637 * special case in might be more effort then it is 638 * worth. We cannot under any circumstances mess 639 * around with a shared busied page except, perhaps, 640 * to pmap it. 641 */ 642 if (vm_page_busied(fs.m)) { 643 /* 644 * Reference the page before unlocking and 645 * sleeping so that the page daemon is less 646 * likely to reclaim it. 647 */ 648 vm_page_aflag_set(fs.m, PGA_REFERENCED); 649 if (fs.object != fs.first_object) { 650 if (!VM_OBJECT_TRYWLOCK( 651 fs.first_object)) { 652 VM_OBJECT_WUNLOCK(fs.object); 653 VM_OBJECT_WLOCK(fs.first_object); 654 VM_OBJECT_WLOCK(fs.object); 655 } 656 vm_page_lock(fs.first_m); 657 vm_page_free(fs.first_m); 658 vm_page_unlock(fs.first_m); 659 vm_object_pip_wakeup(fs.first_object); 660 VM_OBJECT_WUNLOCK(fs.first_object); 661 fs.first_m = NULL; 662 } 663 unlock_map(&fs); 664 if (fs.m == vm_page_lookup(fs.object, 665 fs.pindex)) { 666 vm_page_sleep_if_busy(fs.m, "vmpfw"); 667 } 668 vm_object_pip_wakeup(fs.object); 669 VM_OBJECT_WUNLOCK(fs.object); 670 VM_CNT_INC(v_intrans); 671 vm_object_deallocate(fs.first_object); 672 goto RetryFault; 673 } 674 vm_page_lock(fs.m); 675 vm_page_remque(fs.m); 676 vm_page_unlock(fs.m); 677 678 /* 679 * Mark page busy for other processes, and the 680 * pagedaemon. If it still isn't completely valid 681 * (readable), jump to readrest, else break-out ( we 682 * found the page ). 683 */ 684 vm_page_xbusy(fs.m); 685 if (fs.m->valid != VM_PAGE_BITS_ALL) 686 goto readrest; 687 break; 688 } 689 KASSERT(fs.m == NULL, ("fs.m should be NULL, not %p", fs.m)); 690 691 /* 692 * Page is not resident. If the pager might contain the page 693 * or this is the beginning of the search, allocate a new 694 * page. (Default objects are zero-fill, so there is no real 695 * pager for them.) 696 */ 697 if (fs.object->type != OBJT_DEFAULT || 698 fs.object == fs.first_object) { 699 if (fs.pindex >= fs.object->size) { 700 unlock_and_deallocate(&fs); 701 return (KERN_PROTECTION_FAILURE); 702 } 703 704 if (fs.object == fs.first_object && 705 (fs.first_object->flags & OBJ_POPULATE) != 0 && 706 fs.first_object->shadow_count == 0) { 707 rv = vm_fault_populate(&fs, vaddr, prot, 708 fault_type, fault_flags, wired, m_hold); 709 switch (rv) { 710 case KERN_SUCCESS: 711 case KERN_FAILURE: 712 unlock_and_deallocate(&fs); 713 return (rv); 714 case KERN_RESOURCE_SHORTAGE: 715 unlock_and_deallocate(&fs); 716 goto RetryFault; 717 case KERN_NOT_RECEIVER: 718 /* 719 * Pager's populate() method 720 * returned VM_PAGER_BAD. 721 */ 722 break; 723 default: 724 panic("inconsistent return codes"); 725 } 726 } 727 728 /* 729 * Allocate a new page for this object/offset pair. 730 * 731 * Unlocked read of the p_flag is harmless. At 732 * worst, the P_KILLED might be not observed 733 * there, and allocation can fail, causing 734 * restart and new reading of the p_flag. 735 */ 736 if (!vm_page_count_severe() || P_KILLED(curproc)) { 737 #if VM_NRESERVLEVEL > 0 738 vm_object_color(fs.object, atop(vaddr) - 739 fs.pindex); 740 #endif 741 alloc_req = P_KILLED(curproc) ? 742 VM_ALLOC_SYSTEM : VM_ALLOC_NORMAL; 743 if (fs.object->type != OBJT_VNODE && 744 fs.object->backing_object == NULL) 745 alloc_req |= VM_ALLOC_ZERO; 746 fs.m = vm_page_alloc(fs.object, fs.pindex, 747 alloc_req); 748 } 749 if (fs.m == NULL) { 750 unlock_and_deallocate(&fs); 751 VM_WAITPFAULT; 752 goto RetryFault; 753 } 754 } 755 756 readrest: 757 /* 758 * At this point, we have either allocated a new page or found 759 * an existing page that is only partially valid. 760 * 761 * We hold a reference on the current object and the page is 762 * exclusive busied. 763 */ 764 765 /* 766 * If the pager for the current object might have the page, 767 * then determine the number of additional pages to read and 768 * potentially reprioritize previously read pages for earlier 769 * reclamation. These operations should only be performed 770 * once per page fault. Even if the current pager doesn't 771 * have the page, the number of additional pages to read will 772 * apply to subsequent objects in the shadow chain. 773 */ 774 if (fs.object->type != OBJT_DEFAULT && nera == -1 && 775 !P_KILLED(curproc)) { 776 KASSERT(fs.lookup_still_valid, ("map unlocked")); 777 era = fs.entry->read_ahead; 778 behavior = vm_map_entry_behavior(fs.entry); 779 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 780 nera = 0; 781 } else if (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL) { 782 nera = VM_FAULT_READ_AHEAD_MAX; 783 if (vaddr == fs.entry->next_read) 784 vm_fault_dontneed(&fs, vaddr, nera); 785 } else if (vaddr == fs.entry->next_read) { 786 /* 787 * This is a sequential fault. Arithmetically 788 * increase the requested number of pages in 789 * the read-ahead window. The requested 790 * number of pages is "# of sequential faults 791 * x (read ahead min + 1) + read ahead min" 792 */ 793 nera = VM_FAULT_READ_AHEAD_MIN; 794 if (era > 0) { 795 nera += era + 1; 796 if (nera > VM_FAULT_READ_AHEAD_MAX) 797 nera = VM_FAULT_READ_AHEAD_MAX; 798 } 799 if (era == VM_FAULT_READ_AHEAD_MAX) 800 vm_fault_dontneed(&fs, vaddr, nera); 801 } else { 802 /* 803 * This is a non-sequential fault. 804 */ 805 nera = 0; 806 } 807 if (era != nera) { 808 /* 809 * A read lock on the map suffices to update 810 * the read ahead count safely. 811 */ 812 fs.entry->read_ahead = nera; 813 } 814 815 /* 816 * Prepare for unlocking the map. Save the map 817 * entry's start and end addresses, which are used to 818 * optimize the size of the pager operation below. 819 * Even if the map entry's addresses change after 820 * unlocking the map, using the saved addresses is 821 * safe. 822 */ 823 e_start = fs.entry->start; 824 e_end = fs.entry->end; 825 } 826 827 /* 828 * Call the pager to retrieve the page if there is a chance 829 * that the pager has it, and potentially retrieve additional 830 * pages at the same time. 831 */ 832 if (fs.object->type != OBJT_DEFAULT) { 833 /* 834 * Release the map lock before locking the vnode or 835 * sleeping in the pager. (If the current object has 836 * a shadow, then an earlier iteration of this loop 837 * may have already unlocked the map.) 838 */ 839 unlock_map(&fs); 840 841 if (fs.object->type == OBJT_VNODE && 842 (vp = fs.object->handle) != fs.vp) { 843 /* 844 * Perform an unlock in case the desired vnode 845 * changed while the map was unlocked during a 846 * retry. 847 */ 848 unlock_vp(&fs); 849 850 locked = VOP_ISLOCKED(vp); 851 if (locked != LK_EXCLUSIVE) 852 locked = LK_SHARED; 853 854 /* 855 * We must not sleep acquiring the vnode lock 856 * while we have the page exclusive busied or 857 * the object's paging-in-progress count 858 * incremented. Otherwise, we could deadlock. 859 */ 860 error = vget(vp, locked | LK_CANRECURSE | 861 LK_NOWAIT, curthread); 862 if (error != 0) { 863 vhold(vp); 864 release_page(&fs); 865 unlock_and_deallocate(&fs); 866 error = vget(vp, locked | LK_RETRY | 867 LK_CANRECURSE, curthread); 868 vdrop(vp); 869 fs.vp = vp; 870 KASSERT(error == 0, 871 ("vm_fault: vget failed")); 872 goto RetryFault; 873 } 874 fs.vp = vp; 875 } 876 KASSERT(fs.vp == NULL || !fs.map->system_map, 877 ("vm_fault: vnode-backed object mapped by system map")); 878 879 /* 880 * Page in the requested page and hint the pager, 881 * that it may bring up surrounding pages. 882 */ 883 if (nera == -1 || behavior == MAP_ENTRY_BEHAV_RANDOM || 884 P_KILLED(curproc)) { 885 behind = 0; 886 ahead = 0; 887 } else { 888 /* Is this a sequential fault? */ 889 if (nera > 0) { 890 behind = 0; 891 ahead = nera; 892 } else { 893 /* 894 * Request a cluster of pages that is 895 * aligned to a VM_FAULT_READ_DEFAULT 896 * page offset boundary within the 897 * object. Alignment to a page offset 898 * boundary is more likely to coincide 899 * with the underlying file system 900 * block than alignment to a virtual 901 * address boundary. 902 */ 903 cluster_offset = fs.pindex % 904 VM_FAULT_READ_DEFAULT; 905 behind = ulmin(cluster_offset, 906 atop(vaddr - e_start)); 907 ahead = VM_FAULT_READ_DEFAULT - 1 - 908 cluster_offset; 909 } 910 ahead = ulmin(ahead, atop(e_end - vaddr) - 1); 911 } 912 rv = vm_pager_get_pages(fs.object, &fs.m, 1, 913 &behind, &ahead); 914 if (rv == VM_PAGER_OK) { 915 faultcount = behind + 1 + ahead; 916 hardfault = true; 917 break; /* break to PAGE HAS BEEN FOUND */ 918 } 919 if (rv == VM_PAGER_ERROR) 920 printf("vm_fault: pager read error, pid %d (%s)\n", 921 curproc->p_pid, curproc->p_comm); 922 923 /* 924 * If an I/O error occurred or the requested page was 925 * outside the range of the pager, clean up and return 926 * an error. 927 */ 928 if (rv == VM_PAGER_ERROR || rv == VM_PAGER_BAD) { 929 vm_page_lock(fs.m); 930 if (fs.m->wire_count == 0) 931 vm_page_free(fs.m); 932 else 933 vm_page_xunbusy_maybelocked(fs.m); 934 vm_page_unlock(fs.m); 935 fs.m = NULL; 936 unlock_and_deallocate(&fs); 937 return (rv == VM_PAGER_ERROR ? KERN_FAILURE : 938 KERN_PROTECTION_FAILURE); 939 } 940 941 /* 942 * The requested page does not exist at this object/ 943 * offset. Remove the invalid page from the object, 944 * waking up anyone waiting for it, and continue on to 945 * the next object. However, if this is the top-level 946 * object, we must leave the busy page in place to 947 * prevent another process from rushing past us, and 948 * inserting the page in that object at the same time 949 * that we are. 950 */ 951 if (fs.object != fs.first_object) { 952 vm_page_lock(fs.m); 953 if (fs.m->wire_count == 0) 954 vm_page_free(fs.m); 955 else 956 vm_page_xunbusy_maybelocked(fs.m); 957 vm_page_unlock(fs.m); 958 fs.m = NULL; 959 } 960 } 961 962 /* 963 * We get here if the object has default pager (or unwiring) 964 * or the pager doesn't have the page. 965 */ 966 if (fs.object == fs.first_object) 967 fs.first_m = fs.m; 968 969 /* 970 * Move on to the next object. Lock the next object before 971 * unlocking the current one. 972 */ 973 next_object = fs.object->backing_object; 974 if (next_object == NULL) { 975 /* 976 * If there's no object left, fill the page in the top 977 * object with zeros. 978 */ 979 if (fs.object != fs.first_object) { 980 vm_object_pip_wakeup(fs.object); 981 VM_OBJECT_WUNLOCK(fs.object); 982 983 fs.object = fs.first_object; 984 fs.pindex = fs.first_pindex; 985 fs.m = fs.first_m; 986 VM_OBJECT_WLOCK(fs.object); 987 } 988 fs.first_m = NULL; 989 990 /* 991 * Zero the page if necessary and mark it valid. 992 */ 993 if ((fs.m->flags & PG_ZERO) == 0) { 994 pmap_zero_page(fs.m); 995 } else { 996 VM_CNT_INC(v_ozfod); 997 } 998 VM_CNT_INC(v_zfod); 999 fs.m->valid = VM_PAGE_BITS_ALL; 1000 /* Don't try to prefault neighboring pages. */ 1001 faultcount = 1; 1002 break; /* break to PAGE HAS BEEN FOUND */ 1003 } else { 1004 KASSERT(fs.object != next_object, 1005 ("object loop %p", next_object)); 1006 VM_OBJECT_WLOCK(next_object); 1007 vm_object_pip_add(next_object, 1); 1008 if (fs.object != fs.first_object) 1009 vm_object_pip_wakeup(fs.object); 1010 fs.pindex += 1011 OFF_TO_IDX(fs.object->backing_object_offset); 1012 VM_OBJECT_WUNLOCK(fs.object); 1013 fs.object = next_object; 1014 } 1015 } 1016 1017 vm_page_assert_xbusied(fs.m); 1018 1019 /* 1020 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1021 * is held.] 1022 */ 1023 1024 /* 1025 * If the page is being written, but isn't already owned by the 1026 * top-level object, we have to copy it into a new page owned by the 1027 * top-level object. 1028 */ 1029 if (fs.object != fs.first_object) { 1030 /* 1031 * We only really need to copy if we want to write it. 1032 */ 1033 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1034 /* 1035 * This allows pages to be virtually copied from a 1036 * backing_object into the first_object, where the 1037 * backing object has no other refs to it, and cannot 1038 * gain any more refs. Instead of a bcopy, we just 1039 * move the page from the backing object to the 1040 * first object. Note that we must mark the page 1041 * dirty in the first object so that it will go out 1042 * to swap when needed. 1043 */ 1044 is_first_object_locked = false; 1045 if ( 1046 /* 1047 * Only one shadow object 1048 */ 1049 (fs.object->shadow_count == 1) && 1050 /* 1051 * No COW refs, except us 1052 */ 1053 (fs.object->ref_count == 1) && 1054 /* 1055 * No one else can look this object up 1056 */ 1057 (fs.object->handle == NULL) && 1058 /* 1059 * No other ways to look the object up 1060 */ 1061 ((fs.object->type == OBJT_DEFAULT) || 1062 (fs.object->type == OBJT_SWAP)) && 1063 (is_first_object_locked = VM_OBJECT_TRYWLOCK(fs.first_object)) && 1064 /* 1065 * We don't chase down the shadow chain 1066 */ 1067 fs.object == fs.first_object->backing_object) { 1068 vm_page_lock(fs.m); 1069 vm_page_remove(fs.m); 1070 vm_page_unlock(fs.m); 1071 vm_page_lock(fs.first_m); 1072 vm_page_replace_checked(fs.m, fs.first_object, 1073 fs.first_pindex, fs.first_m); 1074 vm_page_free(fs.first_m); 1075 vm_page_unlock(fs.first_m); 1076 vm_page_dirty(fs.m); 1077 #if VM_NRESERVLEVEL > 0 1078 /* 1079 * Rename the reservation. 1080 */ 1081 vm_reserv_rename(fs.m, fs.first_object, 1082 fs.object, OFF_TO_IDX( 1083 fs.first_object->backing_object_offset)); 1084 #endif 1085 /* 1086 * Removing the page from the backing object 1087 * unbusied it. 1088 */ 1089 vm_page_xbusy(fs.m); 1090 fs.first_m = fs.m; 1091 fs.m = NULL; 1092 VM_CNT_INC(v_cow_optim); 1093 } else { 1094 /* 1095 * Oh, well, lets copy it. 1096 */ 1097 pmap_copy_page(fs.m, fs.first_m); 1098 fs.first_m->valid = VM_PAGE_BITS_ALL; 1099 if (wired && (fault_flags & 1100 VM_FAULT_WIRE) == 0) { 1101 vm_page_lock(fs.first_m); 1102 vm_page_wire(fs.first_m); 1103 vm_page_unlock(fs.first_m); 1104 1105 vm_page_lock(fs.m); 1106 vm_page_unwire(fs.m, PQ_INACTIVE); 1107 vm_page_unlock(fs.m); 1108 } 1109 /* 1110 * We no longer need the old page or object. 1111 */ 1112 release_page(&fs); 1113 } 1114 /* 1115 * fs.object != fs.first_object due to above 1116 * conditional 1117 */ 1118 vm_object_pip_wakeup(fs.object); 1119 VM_OBJECT_WUNLOCK(fs.object); 1120 /* 1121 * Only use the new page below... 1122 */ 1123 fs.object = fs.first_object; 1124 fs.pindex = fs.first_pindex; 1125 fs.m = fs.first_m; 1126 if (!is_first_object_locked) 1127 VM_OBJECT_WLOCK(fs.object); 1128 VM_CNT_INC(v_cow_faults); 1129 curthread->td_cow++; 1130 } else { 1131 prot &= ~VM_PROT_WRITE; 1132 } 1133 } 1134 1135 /* 1136 * We must verify that the maps have not changed since our last 1137 * lookup. 1138 */ 1139 if (!fs.lookup_still_valid) { 1140 if (!vm_map_trylock_read(fs.map)) { 1141 release_page(&fs); 1142 unlock_and_deallocate(&fs); 1143 goto RetryFault; 1144 } 1145 fs.lookup_still_valid = true; 1146 if (fs.map->timestamp != fs.map_generation) { 1147 result = vm_map_lookup_locked(&fs.map, vaddr, fault_type, 1148 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 1149 1150 /* 1151 * If we don't need the page any longer, put it on the inactive 1152 * list (the easiest thing to do here). If no one needs it, 1153 * pageout will grab it eventually. 1154 */ 1155 if (result != KERN_SUCCESS) { 1156 release_page(&fs); 1157 unlock_and_deallocate(&fs); 1158 1159 /* 1160 * If retry of map lookup would have blocked then 1161 * retry fault from start. 1162 */ 1163 if (result == KERN_FAILURE) 1164 goto RetryFault; 1165 return (result); 1166 } 1167 if ((retry_object != fs.first_object) || 1168 (retry_pindex != fs.first_pindex)) { 1169 release_page(&fs); 1170 unlock_and_deallocate(&fs); 1171 goto RetryFault; 1172 } 1173 1174 /* 1175 * Check whether the protection has changed or the object has 1176 * been copied while we left the map unlocked. Changing from 1177 * read to write permission is OK - we leave the page 1178 * write-protected, and catch the write fault. Changing from 1179 * write to read permission means that we can't mark the page 1180 * write-enabled after all. 1181 */ 1182 prot &= retry_prot; 1183 } 1184 } 1185 1186 /* 1187 * If the page was filled by a pager, save the virtual address that 1188 * should be faulted on next under a sequential access pattern to the 1189 * map entry. A read lock on the map suffices to update this address 1190 * safely. 1191 */ 1192 if (hardfault) 1193 fs.entry->next_read = vaddr + ptoa(ahead) + PAGE_SIZE; 1194 1195 vm_fault_dirty(fs.entry, fs.m, prot, fault_type, fault_flags, true); 1196 vm_page_assert_xbusied(fs.m); 1197 1198 /* 1199 * Page must be completely valid or it is not fit to 1200 * map into user space. vm_pager_get_pages() ensures this. 1201 */ 1202 KASSERT(fs.m->valid == VM_PAGE_BITS_ALL, 1203 ("vm_fault: page %p partially invalid", fs.m)); 1204 VM_OBJECT_WUNLOCK(fs.object); 1205 1206 /* 1207 * Put this page into the physical map. We had to do the unlock above 1208 * because pmap_enter() may sleep. We don't put the page 1209 * back on the active queue until later so that the pageout daemon 1210 * won't find it (yet). 1211 */ 1212 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, 1213 fault_type | (wired ? PMAP_ENTER_WIRED : 0), 0); 1214 if (faultcount != 1 && (fault_flags & VM_FAULT_WIRE) == 0 && 1215 wired == 0) 1216 vm_fault_prefault(&fs, vaddr, 1217 faultcount > 0 ? behind : PFBAK, 1218 faultcount > 0 ? ahead : PFFOR); 1219 VM_OBJECT_WLOCK(fs.object); 1220 vm_page_lock(fs.m); 1221 1222 /* 1223 * If the page is not wired down, then put it where the pageout daemon 1224 * can find it. 1225 */ 1226 if ((fault_flags & VM_FAULT_WIRE) != 0) { 1227 KASSERT(wired, ("VM_FAULT_WIRE && !wired")); 1228 vm_page_wire(fs.m); 1229 } else 1230 vm_page_activate(fs.m); 1231 if (m_hold != NULL) { 1232 *m_hold = fs.m; 1233 vm_page_hold(fs.m); 1234 } 1235 vm_page_unlock(fs.m); 1236 vm_page_xunbusy(fs.m); 1237 1238 /* 1239 * Unlock everything, and return 1240 */ 1241 unlock_and_deallocate(&fs); 1242 if (hardfault) { 1243 VM_CNT_INC(v_io_faults); 1244 curthread->td_ru.ru_majflt++; 1245 #ifdef RACCT 1246 if (racct_enable && fs.object->type == OBJT_VNODE) { 1247 PROC_LOCK(curproc); 1248 if ((fault_type & (VM_PROT_COPY | VM_PROT_WRITE)) != 0) { 1249 racct_add_force(curproc, RACCT_WRITEBPS, 1250 PAGE_SIZE + behind * PAGE_SIZE); 1251 racct_add_force(curproc, RACCT_WRITEIOPS, 1); 1252 } else { 1253 racct_add_force(curproc, RACCT_READBPS, 1254 PAGE_SIZE + ahead * PAGE_SIZE); 1255 racct_add_force(curproc, RACCT_READIOPS, 1); 1256 } 1257 PROC_UNLOCK(curproc); 1258 } 1259 #endif 1260 } else 1261 curthread->td_ru.ru_minflt++; 1262 1263 return (KERN_SUCCESS); 1264 } 1265 1266 /* 1267 * Speed up the reclamation of pages that precede the faulting pindex within 1268 * the first object of the shadow chain. Essentially, perform the equivalent 1269 * to madvise(..., MADV_DONTNEED) on a large cluster of pages that precedes 1270 * the faulting pindex by the cluster size when the pages read by vm_fault() 1271 * cross a cluster-size boundary. The cluster size is the greater of the 1272 * smallest superpage size and VM_FAULT_DONTNEED_MIN. 1273 * 1274 * When "fs->first_object" is a shadow object, the pages in the backing object 1275 * that precede the faulting pindex are deactivated by vm_fault(). So, this 1276 * function must only be concerned with pages in the first object. 1277 */ 1278 static void 1279 vm_fault_dontneed(const struct faultstate *fs, vm_offset_t vaddr, int ahead) 1280 { 1281 vm_map_entry_t entry; 1282 vm_object_t first_object, object; 1283 vm_offset_t end, start; 1284 vm_page_t m, m_next; 1285 vm_pindex_t pend, pstart; 1286 vm_size_t size; 1287 1288 object = fs->object; 1289 VM_OBJECT_ASSERT_WLOCKED(object); 1290 first_object = fs->first_object; 1291 if (first_object != object) { 1292 if (!VM_OBJECT_TRYWLOCK(first_object)) { 1293 VM_OBJECT_WUNLOCK(object); 1294 VM_OBJECT_WLOCK(first_object); 1295 VM_OBJECT_WLOCK(object); 1296 } 1297 } 1298 /* Neither fictitious nor unmanaged pages can be reclaimed. */ 1299 if ((first_object->flags & (OBJ_FICTITIOUS | OBJ_UNMANAGED)) == 0) { 1300 size = VM_FAULT_DONTNEED_MIN; 1301 if (MAXPAGESIZES > 1 && size < pagesizes[1]) 1302 size = pagesizes[1]; 1303 end = rounddown2(vaddr, size); 1304 if (vaddr - end >= size - PAGE_SIZE - ptoa(ahead) && 1305 (entry = fs->entry)->start < end) { 1306 if (end - entry->start < size) 1307 start = entry->start; 1308 else 1309 start = end - size; 1310 pmap_advise(fs->map->pmap, start, end, MADV_DONTNEED); 1311 pstart = OFF_TO_IDX(entry->offset) + atop(start - 1312 entry->start); 1313 m_next = vm_page_find_least(first_object, pstart); 1314 pend = OFF_TO_IDX(entry->offset) + atop(end - 1315 entry->start); 1316 while ((m = m_next) != NULL && m->pindex < pend) { 1317 m_next = TAILQ_NEXT(m, listq); 1318 if (m->valid != VM_PAGE_BITS_ALL || 1319 vm_page_busied(m)) 1320 continue; 1321 1322 /* 1323 * Don't clear PGA_REFERENCED, since it would 1324 * likely represent a reference by a different 1325 * process. 1326 * 1327 * Typically, at this point, prefetched pages 1328 * are still in the inactive queue. Only 1329 * pages that triggered page faults are in the 1330 * active queue. 1331 */ 1332 vm_page_lock(m); 1333 vm_page_deactivate(m); 1334 vm_page_unlock(m); 1335 } 1336 } 1337 } 1338 if (first_object != object) 1339 VM_OBJECT_WUNLOCK(first_object); 1340 } 1341 1342 /* 1343 * vm_fault_prefault provides a quick way of clustering 1344 * pagefaults into a processes address space. It is a "cousin" 1345 * of vm_map_pmap_enter, except it runs at page fault time instead 1346 * of mmap time. 1347 */ 1348 static void 1349 vm_fault_prefault(const struct faultstate *fs, vm_offset_t addra, 1350 int backward, int forward) 1351 { 1352 pmap_t pmap; 1353 vm_map_entry_t entry; 1354 vm_object_t backing_object, lobject; 1355 vm_offset_t addr, starta; 1356 vm_pindex_t pindex; 1357 vm_page_t m; 1358 int i; 1359 1360 pmap = fs->map->pmap; 1361 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) 1362 return; 1363 1364 entry = fs->entry; 1365 1366 if (addra < backward * PAGE_SIZE) { 1367 starta = entry->start; 1368 } else { 1369 starta = addra - backward * PAGE_SIZE; 1370 if (starta < entry->start) 1371 starta = entry->start; 1372 } 1373 1374 /* 1375 * Generate the sequence of virtual addresses that are candidates for 1376 * prefaulting in an outward spiral from the faulting virtual address, 1377 * "addra". Specifically, the sequence is "addra - PAGE_SIZE", "addra 1378 * + PAGE_SIZE", "addra - 2 * PAGE_SIZE", "addra + 2 * PAGE_SIZE", ... 1379 * If the candidate address doesn't have a backing physical page, then 1380 * the loop immediately terminates. 1381 */ 1382 for (i = 0; i < 2 * imax(backward, forward); i++) { 1383 addr = addra + ((i >> 1) + 1) * ((i & 1) == 0 ? -PAGE_SIZE : 1384 PAGE_SIZE); 1385 if (addr > addra + forward * PAGE_SIZE) 1386 addr = 0; 1387 1388 if (addr < starta || addr >= entry->end) 1389 continue; 1390 1391 if (!pmap_is_prefaultable(pmap, addr)) 1392 continue; 1393 1394 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1395 lobject = entry->object.vm_object; 1396 VM_OBJECT_RLOCK(lobject); 1397 while ((m = vm_page_lookup(lobject, pindex)) == NULL && 1398 lobject->type == OBJT_DEFAULT && 1399 (backing_object = lobject->backing_object) != NULL) { 1400 KASSERT((lobject->backing_object_offset & PAGE_MASK) == 1401 0, ("vm_fault_prefault: unaligned object offset")); 1402 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1403 VM_OBJECT_RLOCK(backing_object); 1404 VM_OBJECT_RUNLOCK(lobject); 1405 lobject = backing_object; 1406 } 1407 if (m == NULL) { 1408 VM_OBJECT_RUNLOCK(lobject); 1409 break; 1410 } 1411 if (m->valid == VM_PAGE_BITS_ALL && 1412 (m->flags & PG_FICTITIOUS) == 0) 1413 pmap_enter_quick(pmap, addr, m, entry->protection); 1414 VM_OBJECT_RUNLOCK(lobject); 1415 } 1416 } 1417 1418 /* 1419 * Hold each of the physical pages that are mapped by the specified range of 1420 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1421 * and allow the specified types of access, "prot". If all of the implied 1422 * pages are successfully held, then the number of held pages is returned 1423 * together with pointers to those pages in the array "ma". However, if any 1424 * of the pages cannot be held, -1 is returned. 1425 */ 1426 int 1427 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1428 vm_prot_t prot, vm_page_t *ma, int max_count) 1429 { 1430 vm_offset_t end, va; 1431 vm_page_t *mp; 1432 int count; 1433 boolean_t pmap_failed; 1434 1435 if (len == 0) 1436 return (0); 1437 end = round_page(addr + len); 1438 addr = trunc_page(addr); 1439 1440 /* 1441 * Check for illegal addresses. 1442 */ 1443 if (addr < vm_map_min(map) || addr > end || end > vm_map_max(map)) 1444 return (-1); 1445 1446 if (atop(end - addr) > max_count) 1447 panic("vm_fault_quick_hold_pages: count > max_count"); 1448 count = atop(end - addr); 1449 1450 /* 1451 * Most likely, the physical pages are resident in the pmap, so it is 1452 * faster to try pmap_extract_and_hold() first. 1453 */ 1454 pmap_failed = FALSE; 1455 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) { 1456 *mp = pmap_extract_and_hold(map->pmap, va, prot); 1457 if (*mp == NULL) 1458 pmap_failed = TRUE; 1459 else if ((prot & VM_PROT_WRITE) != 0 && 1460 (*mp)->dirty != VM_PAGE_BITS_ALL) { 1461 /* 1462 * Explicitly dirty the physical page. Otherwise, the 1463 * caller's changes may go unnoticed because they are 1464 * performed through an unmanaged mapping or by a DMA 1465 * operation. 1466 * 1467 * The object lock is not held here. 1468 * See vm_page_clear_dirty_mask(). 1469 */ 1470 vm_page_dirty(*mp); 1471 } 1472 } 1473 if (pmap_failed) { 1474 /* 1475 * One or more pages could not be held by the pmap. Either no 1476 * page was mapped at the specified virtual address or that 1477 * mapping had insufficient permissions. Attempt to fault in 1478 * and hold these pages. 1479 */ 1480 for (mp = ma, va = addr; va < end; mp++, va += PAGE_SIZE) 1481 if (*mp == NULL && vm_fault_hold(map, va, prot, 1482 VM_FAULT_NORMAL, mp) != KERN_SUCCESS) 1483 goto error; 1484 } 1485 return (count); 1486 error: 1487 for (mp = ma; mp < ma + count; mp++) 1488 if (*mp != NULL) { 1489 vm_page_lock(*mp); 1490 vm_page_unhold(*mp); 1491 vm_page_unlock(*mp); 1492 } 1493 return (-1); 1494 } 1495 1496 /* 1497 * Routine: 1498 * vm_fault_copy_entry 1499 * Function: 1500 * Create new shadow object backing dst_entry with private copy of 1501 * all underlying pages. When src_entry is equal to dst_entry, 1502 * function implements COW for wired-down map entry. Otherwise, 1503 * it forks wired entry into dst_map. 1504 * 1505 * In/out conditions: 1506 * The source and destination maps must be locked for write. 1507 * The source map entry must be wired down (or be a sharing map 1508 * entry corresponding to a main map entry that is wired down). 1509 */ 1510 void 1511 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1512 vm_map_entry_t dst_entry, vm_map_entry_t src_entry, 1513 vm_ooffset_t *fork_charge) 1514 { 1515 vm_object_t backing_object, dst_object, object, src_object; 1516 vm_pindex_t dst_pindex, pindex, src_pindex; 1517 vm_prot_t access, prot; 1518 vm_offset_t vaddr; 1519 vm_page_t dst_m; 1520 vm_page_t src_m; 1521 boolean_t upgrade; 1522 1523 #ifdef lint 1524 src_map++; 1525 #endif /* lint */ 1526 1527 upgrade = src_entry == dst_entry; 1528 access = prot = dst_entry->protection; 1529 1530 src_object = src_entry->object.vm_object; 1531 src_pindex = OFF_TO_IDX(src_entry->offset); 1532 1533 if (upgrade && (dst_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 1534 dst_object = src_object; 1535 vm_object_reference(dst_object); 1536 } else { 1537 /* 1538 * Create the top-level object for the destination entry. (Doesn't 1539 * actually shadow anything - we copy the pages directly.) 1540 */ 1541 dst_object = vm_object_allocate(OBJT_DEFAULT, 1542 atop(dst_entry->end - dst_entry->start)); 1543 #if VM_NRESERVLEVEL > 0 1544 dst_object->flags |= OBJ_COLORED; 1545 dst_object->pg_color = atop(dst_entry->start); 1546 #endif 1547 } 1548 1549 VM_OBJECT_WLOCK(dst_object); 1550 KASSERT(upgrade || dst_entry->object.vm_object == NULL, 1551 ("vm_fault_copy_entry: vm_object not NULL")); 1552 if (src_object != dst_object) { 1553 dst_entry->object.vm_object = dst_object; 1554 dst_entry->offset = 0; 1555 dst_object->charge = dst_entry->end - dst_entry->start; 1556 } 1557 if (fork_charge != NULL) { 1558 KASSERT(dst_entry->cred == NULL, 1559 ("vm_fault_copy_entry: leaked swp charge")); 1560 dst_object->cred = curthread->td_ucred; 1561 crhold(dst_object->cred); 1562 *fork_charge += dst_object->charge; 1563 } else if (dst_object->cred == NULL) { 1564 KASSERT(dst_entry->cred != NULL, ("no cred for entry %p", 1565 dst_entry)); 1566 dst_object->cred = dst_entry->cred; 1567 dst_entry->cred = NULL; 1568 } 1569 1570 /* 1571 * If not an upgrade, then enter the mappings in the pmap as 1572 * read and/or execute accesses. Otherwise, enter them as 1573 * write accesses. 1574 * 1575 * A writeable large page mapping is only created if all of 1576 * the constituent small page mappings are modified. Marking 1577 * PTEs as modified on inception allows promotion to happen 1578 * without taking potentially large number of soft faults. 1579 */ 1580 if (!upgrade) 1581 access &= ~VM_PROT_WRITE; 1582 1583 /* 1584 * Loop through all of the virtual pages within the entry's 1585 * range, copying each page from the source object to the 1586 * destination object. Since the source is wired, those pages 1587 * must exist. In contrast, the destination is pageable. 1588 * Since the destination object does share any backing storage 1589 * with the source object, all of its pages must be dirtied, 1590 * regardless of whether they can be written. 1591 */ 1592 for (vaddr = dst_entry->start, dst_pindex = 0; 1593 vaddr < dst_entry->end; 1594 vaddr += PAGE_SIZE, dst_pindex++) { 1595 again: 1596 /* 1597 * Find the page in the source object, and copy it in. 1598 * Because the source is wired down, the page will be 1599 * in memory. 1600 */ 1601 if (src_object != dst_object) 1602 VM_OBJECT_RLOCK(src_object); 1603 object = src_object; 1604 pindex = src_pindex + dst_pindex; 1605 while ((src_m = vm_page_lookup(object, pindex)) == NULL && 1606 (backing_object = object->backing_object) != NULL) { 1607 /* 1608 * Unless the source mapping is read-only or 1609 * it is presently being upgraded from 1610 * read-only, the first object in the shadow 1611 * chain should provide all of the pages. In 1612 * other words, this loop body should never be 1613 * executed when the source mapping is already 1614 * read/write. 1615 */ 1616 KASSERT((src_entry->protection & VM_PROT_WRITE) == 0 || 1617 upgrade, 1618 ("vm_fault_copy_entry: main object missing page")); 1619 1620 VM_OBJECT_RLOCK(backing_object); 1621 pindex += OFF_TO_IDX(object->backing_object_offset); 1622 if (object != dst_object) 1623 VM_OBJECT_RUNLOCK(object); 1624 object = backing_object; 1625 } 1626 KASSERT(src_m != NULL, ("vm_fault_copy_entry: page missing")); 1627 1628 if (object != dst_object) { 1629 /* 1630 * Allocate a page in the destination object. 1631 */ 1632 dst_m = vm_page_alloc(dst_object, (src_object == 1633 dst_object ? src_pindex : 0) + dst_pindex, 1634 VM_ALLOC_NORMAL); 1635 if (dst_m == NULL) { 1636 VM_OBJECT_WUNLOCK(dst_object); 1637 VM_OBJECT_RUNLOCK(object); 1638 VM_WAIT; 1639 VM_OBJECT_WLOCK(dst_object); 1640 goto again; 1641 } 1642 pmap_copy_page(src_m, dst_m); 1643 VM_OBJECT_RUNLOCK(object); 1644 dst_m->valid = VM_PAGE_BITS_ALL; 1645 dst_m->dirty = VM_PAGE_BITS_ALL; 1646 } else { 1647 dst_m = src_m; 1648 if (vm_page_sleep_if_busy(dst_m, "fltupg")) 1649 goto again; 1650 vm_page_xbusy(dst_m); 1651 KASSERT(dst_m->valid == VM_PAGE_BITS_ALL, 1652 ("invalid dst page %p", dst_m)); 1653 } 1654 VM_OBJECT_WUNLOCK(dst_object); 1655 1656 /* 1657 * Enter it in the pmap. If a wired, copy-on-write 1658 * mapping is being replaced by a write-enabled 1659 * mapping, then wire that new mapping. 1660 */ 1661 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, 1662 access | (upgrade ? PMAP_ENTER_WIRED : 0), 0); 1663 1664 /* 1665 * Mark it no longer busy, and put it on the active list. 1666 */ 1667 VM_OBJECT_WLOCK(dst_object); 1668 1669 if (upgrade) { 1670 if (src_m != dst_m) { 1671 vm_page_lock(src_m); 1672 vm_page_unwire(src_m, PQ_INACTIVE); 1673 vm_page_unlock(src_m); 1674 vm_page_lock(dst_m); 1675 vm_page_wire(dst_m); 1676 vm_page_unlock(dst_m); 1677 } else { 1678 KASSERT(dst_m->wire_count > 0, 1679 ("dst_m %p is not wired", dst_m)); 1680 } 1681 } else { 1682 vm_page_lock(dst_m); 1683 vm_page_activate(dst_m); 1684 vm_page_unlock(dst_m); 1685 } 1686 vm_page_xunbusy(dst_m); 1687 } 1688 VM_OBJECT_WUNLOCK(dst_object); 1689 if (upgrade) { 1690 dst_entry->eflags &= ~(MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY); 1691 vm_object_deallocate(src_object); 1692 } 1693 } 1694 1695 /* 1696 * Block entry into the machine-independent layer's page fault handler by 1697 * the calling thread. Subsequent calls to vm_fault() by that thread will 1698 * return KERN_PROTECTION_FAILURE. Enable machine-dependent handling of 1699 * spurious page faults. 1700 */ 1701 int 1702 vm_fault_disable_pagefaults(void) 1703 { 1704 1705 return (curthread_pflags_set(TDP_NOFAULTING | TDP_RESETSPUR)); 1706 } 1707 1708 void 1709 vm_fault_enable_pagefaults(int save) 1710 { 1711 1712 curthread_pflags_restore(save); 1713 } 1714