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