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