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 * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $ 70 * $DragonFly: src/sys/vm/vm_fault.c,v 1.25 2006/05/17 17:47:58 dillon Exp $ 71 */ 72 73 /* 74 * Page fault handling module. 75 */ 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/kernel.h> 80 #include <sys/proc.h> 81 #include <sys/vnode.h> 82 #include <sys/resourcevar.h> 83 #include <sys/vmmeter.h> 84 85 #include <vm/vm.h> 86 #include <vm/vm_param.h> 87 #include <sys/lock.h> 88 #include <vm/pmap.h> 89 #include <vm/vm_map.h> 90 #include <vm/vm_object.h> 91 #include <vm/vm_page.h> 92 #include <vm/vm_pageout.h> 93 #include <vm/vm_kern.h> 94 #include <vm/vm_pager.h> 95 #include <vm/vnode_pager.h> 96 #include <vm/vm_extern.h> 97 98 #include <sys/thread2.h> 99 #include <vm/vm_page2.h> 100 101 static int vm_fault_additional_pages (vm_page_t, int, 102 int, vm_page_t *, int *); 103 static int vm_fault_ratelimit(struct vmspace *vmspace); 104 105 #define VM_FAULT_READ_AHEAD 8 106 #define VM_FAULT_READ_BEHIND 7 107 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 108 109 struct faultstate { 110 vm_page_t m; 111 vm_object_t object; 112 vm_pindex_t pindex; 113 vm_page_t first_m; 114 vm_object_t first_object; 115 vm_pindex_t first_pindex; 116 vm_map_t map; 117 vm_map_entry_t entry; 118 int lookup_still_valid; 119 struct vnode *vp; 120 }; 121 122 static __inline void 123 release_page(struct faultstate *fs) 124 { 125 vm_page_wakeup(fs->m); 126 vm_page_deactivate(fs->m); 127 fs->m = NULL; 128 } 129 130 static __inline void 131 unlock_map(struct faultstate *fs) 132 { 133 if (fs->lookup_still_valid) { 134 vm_map_lookup_done(fs->map, fs->entry, 0); 135 fs->lookup_still_valid = FALSE; 136 } 137 } 138 139 static void 140 _unlock_things(struct faultstate *fs, int dealloc) 141 { 142 vm_object_pip_wakeup(fs->object); 143 if (fs->object != fs->first_object) { 144 vm_page_free(fs->first_m); 145 vm_object_pip_wakeup(fs->first_object); 146 fs->first_m = NULL; 147 } 148 if (dealloc) { 149 vm_object_deallocate(fs->first_object); 150 } 151 unlock_map(fs); 152 if (fs->vp != NULL) { 153 vput(fs->vp); 154 fs->vp = NULL; 155 } 156 } 157 158 #define unlock_things(fs) _unlock_things(fs, 0) 159 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 160 161 /* 162 * TRYPAGER - used by vm_fault to calculate whether the pager for the 163 * current object *might* contain the page. 164 * 165 * default objects are zero-fill, there is no real pager. 166 */ 167 168 #define TRYPAGER (fs.object->type != OBJT_DEFAULT && \ 169 (((fault_flags & VM_FAULT_WIRE_MASK) == 0) || wired)) 170 171 /* 172 * vm_fault: 173 * 174 * Handle a page fault occurring at the given address, 175 * requiring the given permissions, in the map specified. 176 * If successful, the page is inserted into the 177 * associated physical map. 178 * 179 * NOTE: the given address should be truncated to the 180 * proper page address. 181 * 182 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 183 * a standard error specifying why the fault is fatal is returned. 184 * 185 * 186 * The map in question must be referenced, and remains so. 187 * Caller may hold no locks. 188 */ 189 int 190 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 191 { 192 vm_prot_t prot; 193 int result; 194 boolean_t wired; 195 int map_generation; 196 vm_object_t next_object; 197 vm_page_t marray[VM_FAULT_READ]; 198 int hardfault; 199 int faultcount; 200 int limticks; 201 int didlimit = 0; 202 struct faultstate fs; 203 204 mycpu->gd_cnt.v_vm_faults++; 205 hardfault = 0; 206 207 RetryFault: 208 /* 209 * Find the backing store object and offset into it to begin the 210 * search. 211 */ 212 fs.map = map; 213 if ((result = vm_map_lookup(&fs.map, vaddr, 214 fault_type, &fs.entry, &fs.first_object, 215 &fs.first_pindex, &prot, &wired)) != KERN_SUCCESS) { 216 if ((result != KERN_PROTECTION_FAILURE) || 217 ((fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)) { 218 return result; 219 } 220 221 /* 222 * If we are user-wiring a r/w segment, and it is COW, then 223 * we need to do the COW operation. Note that we don't COW 224 * currently RO sections now, because it is NOT desirable 225 * to COW .text. We simply keep .text from ever being COW'ed 226 * and take the heat that one cannot debug wired .text sections. 227 */ 228 result = vm_map_lookup(&fs.map, vaddr, 229 VM_PROT_READ|VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE, 230 &fs.entry, &fs.first_object, &fs.first_pindex, &prot, &wired); 231 if (result != KERN_SUCCESS) { 232 return result; 233 } 234 235 /* 236 * If we don't COW now, on a user wire, the user will never 237 * be able to write to the mapping. If we don't make this 238 * restriction, the bookkeeping would be nearly impossible. 239 */ 240 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 241 fs.entry->max_protection &= ~VM_PROT_WRITE; 242 } 243 244 map_generation = fs.map->timestamp; 245 246 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 247 panic("vm_fault: fault on nofault entry, addr: %lx", 248 (u_long)vaddr); 249 } 250 251 /* 252 * A system map entry may return a NULL object. No object means 253 * no pager means an unrecoverable kernel fault. 254 */ 255 if (fs.first_object == NULL) { 256 panic("vm_fault: unrecoverable fault at %p in entry %p", 257 (void *)vaddr, fs.entry); 258 } 259 260 /* 261 * Make a reference to this object to prevent its disposal while we 262 * are messing with it. Once we have the reference, the map is free 263 * to be diddled. Since objects reference their shadows (and copies), 264 * they will stay around as well. 265 * 266 * Bump the paging-in-progress count to prevent size changes (e.g. 267 * truncation operations) during I/O. This must be done after 268 * obtaining the vnode lock in order to avoid possible deadlocks. 269 */ 270 vm_object_reference(fs.first_object); 271 fs.vp = vnode_pager_lock(fs.first_object); 272 vm_object_pip_add(fs.first_object, 1); 273 274 fs.lookup_still_valid = TRUE; 275 276 if (wired) 277 fault_type = prot; 278 279 fs.first_m = NULL; 280 281 /* 282 * Search for the page at object/offset. 283 */ 284 285 fs.object = fs.first_object; 286 fs.pindex = fs.first_pindex; 287 288 while (TRUE) { 289 /* 290 * If the object is dead, we stop here 291 */ 292 293 if (fs.object->flags & OBJ_DEAD) { 294 unlock_and_deallocate(&fs); 295 return (KERN_PROTECTION_FAILURE); 296 } 297 298 /* 299 * See if page is resident. spl protection is required 300 * to avoid an interrupt unbusy/free race against our 301 * lookup. We must hold the protection through a page 302 * allocation or busy. 303 */ 304 crit_enter(); 305 fs.m = vm_page_lookup(fs.object, fs.pindex); 306 if (fs.m != NULL) { 307 int queue; 308 /* 309 * Wait/Retry if the page is busy. We have to do this 310 * if the page is busy via either PG_BUSY or 311 * vm_page_t->busy because the vm_pager may be using 312 * vm_page_t->busy for pageouts ( and even pageins if 313 * it is the vnode pager ), and we could end up trying 314 * to pagein and pageout the same page simultaneously. 315 * 316 * We can theoretically allow the busy case on a read 317 * fault if the page is marked valid, but since such 318 * pages are typically already pmap'd, putting that 319 * special case in might be more effort then it is 320 * worth. We cannot under any circumstances mess 321 * around with a vm_page_t->busy page except, perhaps, 322 * to pmap it. 323 */ 324 if ((fs.m->flags & PG_BUSY) || fs.m->busy) { 325 unlock_things(&fs); 326 vm_page_sleep_busy(fs.m, TRUE, "vmpfw"); 327 mycpu->gd_cnt.v_intrans++; 328 vm_object_deallocate(fs.first_object); 329 crit_exit(); 330 goto RetryFault; 331 } 332 333 queue = fs.m->queue; 334 vm_page_unqueue_nowakeup(fs.m); 335 336 if ((queue - fs.m->pc) == PQ_CACHE && vm_page_count_severe()) { 337 vm_page_activate(fs.m); 338 unlock_and_deallocate(&fs); 339 vm_waitpfault(); 340 crit_exit(); 341 goto RetryFault; 342 } 343 344 /* 345 * Mark page busy for other processes, and the 346 * pagedaemon. If it still isn't completely valid 347 * (readable), jump to readrest, else break-out ( we 348 * found the page ). 349 * 350 * We can release the spl once we have marked the 351 * page busy. 352 */ 353 354 vm_page_busy(fs.m); 355 crit_exit(); 356 357 if (((fs.m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 358 fs.m->object != kernel_object && fs.m->object != kmem_object) { 359 goto readrest; 360 } 361 362 break; 363 } 364 365 /* 366 * Page is not resident, If this is the search termination 367 * or the pager might contain the page, allocate a new page. 368 * 369 * note: we are still in splvm(). 370 */ 371 372 if (TRYPAGER || fs.object == fs.first_object) { 373 if (fs.pindex >= fs.object->size) { 374 crit_exit(); 375 unlock_and_deallocate(&fs); 376 return (KERN_PROTECTION_FAILURE); 377 } 378 379 /* 380 * Ratelimit. 381 */ 382 if (didlimit == 0 && curproc != NULL) { 383 limticks = 384 vm_fault_ratelimit(curproc->p_vmspace); 385 if (limticks) { 386 crit_exit(); 387 unlock_and_deallocate(&fs); 388 tsleep(curproc, 0, "vmrate", limticks); 389 didlimit = 1; 390 goto RetryFault; 391 } 392 } 393 394 /* 395 * Allocate a new page for this object/offset pair. 396 */ 397 fs.m = NULL; 398 if (!vm_page_count_severe()) { 399 fs.m = vm_page_alloc(fs.object, fs.pindex, 400 (fs.vp || fs.object->backing_object)? VM_ALLOC_NORMAL: VM_ALLOC_NORMAL | VM_ALLOC_ZERO); 401 } 402 if (fs.m == NULL) { 403 crit_exit(); 404 unlock_and_deallocate(&fs); 405 vm_waitpfault(); 406 goto RetryFault; 407 } 408 } 409 crit_exit(); 410 411 readrest: 412 /* 413 * We have found a valid page or we have allocated a new page. 414 * The page thus may not be valid or may not be entirely 415 * valid. 416 * 417 * Attempt to fault-in the page if there is a chance that the 418 * pager has it, and potentially fault in additional pages 419 * at the same time. 420 * 421 * We are NOT in splvm here and if TRYPAGER is true then 422 * fs.m will be non-NULL and will be PG_BUSY for us. 423 */ 424 425 if (TRYPAGER) { 426 int rv; 427 int reqpage; 428 int ahead, behind; 429 u_char behavior = vm_map_entry_behavior(fs.entry); 430 431 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 432 ahead = 0; 433 behind = 0; 434 } else { 435 behind = (vaddr - fs.entry->start) >> PAGE_SHIFT; 436 if (behind > VM_FAULT_READ_BEHIND) 437 behind = VM_FAULT_READ_BEHIND; 438 439 ahead = ((fs.entry->end - vaddr) >> PAGE_SHIFT) - 1; 440 if (ahead > VM_FAULT_READ_AHEAD) 441 ahead = VM_FAULT_READ_AHEAD; 442 } 443 444 if ((fs.first_object->type != OBJT_DEVICE) && 445 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 446 (behavior != MAP_ENTRY_BEHAV_RANDOM && 447 fs.pindex >= fs.entry->lastr && 448 fs.pindex < fs.entry->lastr + VM_FAULT_READ)) 449 ) { 450 vm_pindex_t firstpindex, tmppindex; 451 452 if (fs.first_pindex < 2 * VM_FAULT_READ) 453 firstpindex = 0; 454 else 455 firstpindex = fs.first_pindex - 2 * VM_FAULT_READ; 456 457 /* 458 * note: partially valid pages cannot be 459 * included in the lookahead - NFS piecemeal 460 * writes will barf on it badly. 461 * 462 * spl protection is required to avoid races 463 * between the lookup and an interrupt 464 * unbusy/free sequence occuring prior to 465 * our busy check. 466 */ 467 crit_enter(); 468 for (tmppindex = fs.first_pindex - 1; 469 tmppindex >= firstpindex; 470 --tmppindex 471 ) { 472 vm_page_t mt; 473 mt = vm_page_lookup( fs.first_object, tmppindex); 474 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 475 break; 476 if (mt->busy || 477 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 478 mt->hold_count || 479 mt->wire_count) 480 continue; 481 if (mt->dirty == 0) 482 vm_page_test_dirty(mt); 483 if (mt->dirty) { 484 vm_page_protect(mt, VM_PROT_NONE); 485 vm_page_deactivate(mt); 486 } else { 487 vm_page_cache(mt); 488 } 489 } 490 crit_exit(); 491 492 ahead += behind; 493 behind = 0; 494 } 495 496 /* 497 * now we find out if any other pages should be paged 498 * in at this time this routine checks to see if the 499 * pages surrounding this fault reside in the same 500 * object as the page for this fault. If they do, 501 * then they are faulted in also into the object. The 502 * array "marray" returned contains an array of 503 * vm_page_t structs where one of them is the 504 * vm_page_t passed to the routine. The reqpage 505 * return value is the index into the marray for the 506 * vm_page_t passed to the routine. 507 * 508 * fs.m plus the additional pages are PG_BUSY'd. 509 */ 510 faultcount = vm_fault_additional_pages( 511 fs.m, behind, ahead, marray, &reqpage); 512 513 /* 514 * update lastr imperfectly (we do not know how much 515 * getpages will actually read), but good enough. 516 */ 517 fs.entry->lastr = fs.pindex + faultcount - behind; 518 519 /* 520 * Call the pager to retrieve the data, if any, after 521 * releasing the lock on the map. We hold a ref on 522 * fs.object and the pages are PG_BUSY'd. 523 */ 524 unlock_map(&fs); 525 526 rv = faultcount ? 527 vm_pager_get_pages(fs.object, marray, faultcount, 528 reqpage) : VM_PAGER_FAIL; 529 530 if (rv == VM_PAGER_OK) { 531 /* 532 * Found the page. Leave it busy while we play 533 * with it. 534 */ 535 536 /* 537 * Relookup in case pager changed page. Pager 538 * is responsible for disposition of old page 539 * if moved. 540 * 541 * XXX other code segments do relookups too. 542 * It's a bad abstraction that needs to be 543 * fixed/removed. 544 */ 545 fs.m = vm_page_lookup(fs.object, fs.pindex); 546 if (fs.m == NULL) { 547 unlock_and_deallocate(&fs); 548 goto RetryFault; 549 } 550 551 hardfault++; 552 break; /* break to PAGE HAS BEEN FOUND */ 553 } 554 /* 555 * Remove the bogus page (which does not exist at this 556 * object/offset); before doing so, we must get back 557 * our object lock to preserve our invariant. 558 * 559 * Also wake up any other process that may want to bring 560 * in this page. 561 * 562 * If this is the top-level object, we must leave the 563 * busy page to prevent another process from rushing 564 * past us, and inserting the page in that object at 565 * the same time that we are. 566 */ 567 568 if (rv == VM_PAGER_ERROR) { 569 if (curproc) 570 printf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); 571 else 572 printf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm); 573 } 574 /* 575 * Data outside the range of the pager or an I/O error 576 */ 577 /* 578 * XXX - the check for kernel_map is a kludge to work 579 * around having the machine panic on a kernel space 580 * fault w/ I/O error. 581 */ 582 if (((fs.map != kernel_map) && (rv == VM_PAGER_ERROR)) || 583 (rv == VM_PAGER_BAD)) { 584 vm_page_free(fs.m); 585 fs.m = NULL; 586 unlock_and_deallocate(&fs); 587 return ((rv == VM_PAGER_ERROR) ? KERN_FAILURE : KERN_PROTECTION_FAILURE); 588 } 589 if (fs.object != fs.first_object) { 590 vm_page_free(fs.m); 591 fs.m = NULL; 592 /* 593 * XXX - we cannot just fall out at this 594 * point, m has been freed and is invalid! 595 */ 596 } 597 } 598 599 /* 600 * We get here if the object has default pager (or unwiring) 601 * or the pager doesn't have the page. 602 */ 603 if (fs.object == fs.first_object) 604 fs.first_m = fs.m; 605 606 /* 607 * Move on to the next object. Lock the next object before 608 * unlocking the current one. 609 */ 610 611 fs.pindex += OFF_TO_IDX(fs.object->backing_object_offset); 612 next_object = fs.object->backing_object; 613 if (next_object == NULL) { 614 /* 615 * If there's no object left, fill the page in the top 616 * object with zeros. 617 */ 618 if (fs.object != fs.first_object) { 619 vm_object_pip_wakeup(fs.object); 620 621 fs.object = fs.first_object; 622 fs.pindex = fs.first_pindex; 623 fs.m = fs.first_m; 624 } 625 fs.first_m = NULL; 626 627 /* 628 * Zero the page if necessary and mark it valid. 629 */ 630 if ((fs.m->flags & PG_ZERO) == 0) { 631 vm_page_zero_fill(fs.m); 632 } else { 633 mycpu->gd_cnt.v_ozfod++; 634 } 635 mycpu->gd_cnt.v_zfod++; 636 fs.m->valid = VM_PAGE_BITS_ALL; 637 break; /* break to PAGE HAS BEEN FOUND */ 638 } else { 639 if (fs.object != fs.first_object) { 640 vm_object_pip_wakeup(fs.object); 641 } 642 KASSERT(fs.object != next_object, ("object loop %p", next_object)); 643 fs.object = next_object; 644 vm_object_pip_add(fs.object, 1); 645 } 646 } 647 648 KASSERT((fs.m->flags & PG_BUSY) != 0, 649 ("vm_fault: not busy after main loop")); 650 651 /* 652 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 653 * is held.] 654 */ 655 656 /* 657 * If the page is being written, but isn't already owned by the 658 * top-level object, we have to copy it into a new page owned by the 659 * top-level object. 660 */ 661 662 if (fs.object != fs.first_object) { 663 /* 664 * We only really need to copy if we want to write it. 665 */ 666 667 if (fault_type & VM_PROT_WRITE) { 668 /* 669 * This allows pages to be virtually copied from a 670 * backing_object into the first_object, where the 671 * backing object has no other refs to it, and cannot 672 * gain any more refs. Instead of a bcopy, we just 673 * move the page from the backing object to the 674 * first object. Note that we must mark the page 675 * dirty in the first object so that it will go out 676 * to swap when needed. 677 */ 678 if (map_generation == fs.map->timestamp && 679 /* 680 * Only one shadow object 681 */ 682 (fs.object->shadow_count == 1) && 683 /* 684 * No COW refs, except us 685 */ 686 (fs.object->ref_count == 1) && 687 /* 688 * No one else can look this object up 689 */ 690 (fs.object->handle == NULL) && 691 /* 692 * No other ways to look the object up 693 */ 694 ((fs.object->type == OBJT_DEFAULT) || 695 (fs.object->type == OBJT_SWAP)) && 696 /* 697 * We don't chase down the shadow chain 698 */ 699 (fs.object == fs.first_object->backing_object) && 700 701 /* 702 * grab the lock if we need to 703 */ 704 (fs.lookup_still_valid || 705 lockmgr(&fs.map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 706 ) { 707 708 fs.lookup_still_valid = 1; 709 /* 710 * get rid of the unnecessary page 711 */ 712 vm_page_protect(fs.first_m, VM_PROT_NONE); 713 vm_page_free(fs.first_m); 714 fs.first_m = NULL; 715 716 /* 717 * grab the page and put it into the 718 * process'es object. The page is 719 * automatically made dirty. 720 */ 721 vm_page_rename(fs.m, fs.first_object, fs.first_pindex); 722 fs.first_m = fs.m; 723 vm_page_busy(fs.first_m); 724 fs.m = NULL; 725 mycpu->gd_cnt.v_cow_optim++; 726 } else { 727 /* 728 * Oh, well, lets copy it. 729 */ 730 vm_page_copy(fs.m, fs.first_m); 731 } 732 733 if (fs.m) { 734 /* 735 * We no longer need the old page or object. 736 */ 737 release_page(&fs); 738 } 739 740 /* 741 * fs.object != fs.first_object due to above 742 * conditional 743 */ 744 745 vm_object_pip_wakeup(fs.object); 746 747 /* 748 * Only use the new page below... 749 */ 750 751 mycpu->gd_cnt.v_cow_faults++; 752 fs.m = fs.first_m; 753 fs.object = fs.first_object; 754 fs.pindex = fs.first_pindex; 755 756 } else { 757 prot &= ~VM_PROT_WRITE; 758 } 759 } 760 761 /* 762 * We must verify that the maps have not changed since our last 763 * lookup. 764 */ 765 766 if (!fs.lookup_still_valid && 767 (fs.map->timestamp != map_generation)) { 768 vm_object_t retry_object; 769 vm_pindex_t retry_pindex; 770 vm_prot_t retry_prot; 771 772 /* 773 * Since map entries may be pageable, make sure we can take a 774 * page fault on them. 775 */ 776 777 /* 778 * Unlock vnode before the lookup to avoid deadlock. E.G. 779 * avoid a deadlock between the inode and exec_map that can 780 * occur due to locks being obtained in different orders. 781 */ 782 783 if (fs.vp != NULL) { 784 vput(fs.vp); 785 fs.vp = NULL; 786 } 787 788 if (fs.map->infork) { 789 release_page(&fs); 790 unlock_and_deallocate(&fs); 791 goto RetryFault; 792 } 793 794 /* 795 * To avoid trying to write_lock the map while another process 796 * has it read_locked (in vm_map_wire), we do not try for 797 * write permission. If the page is still writable, we will 798 * get write permission. If it is not, or has been marked 799 * needs_copy, we enter the mapping without write permission, 800 * and will merely take another fault. 801 */ 802 result = vm_map_lookup(&fs.map, vaddr, fault_type & ~VM_PROT_WRITE, 803 &fs.entry, &retry_object, &retry_pindex, &retry_prot, &wired); 804 map_generation = fs.map->timestamp; 805 806 /* 807 * If we don't need the page any longer, put it on the active 808 * list (the easiest thing to do here). If no one needs it, 809 * pageout will grab it eventually. 810 */ 811 812 if (result != KERN_SUCCESS) { 813 release_page(&fs); 814 unlock_and_deallocate(&fs); 815 return (result); 816 } 817 fs.lookup_still_valid = TRUE; 818 819 if ((retry_object != fs.first_object) || 820 (retry_pindex != fs.first_pindex)) { 821 release_page(&fs); 822 unlock_and_deallocate(&fs); 823 goto RetryFault; 824 } 825 /* 826 * Check whether the protection has changed or the object has 827 * been copied while we left the map unlocked. Changing from 828 * read to write permission is OK - we leave the page 829 * write-protected, and catch the write fault. Changing from 830 * write to read permission means that we can't mark the page 831 * write-enabled after all. 832 */ 833 prot &= retry_prot; 834 } 835 836 /* 837 * Put this page into the physical map. We had to do the unlock above 838 * because pmap_enter may cause other faults. We don't put the page 839 * back on the active queue until later so that the page-out daemon 840 * won't find us (yet). 841 */ 842 843 if (prot & VM_PROT_WRITE) { 844 vm_page_flag_set(fs.m, PG_WRITEABLE); 845 vm_object_set_writeable_dirty(fs.m->object); 846 847 /* 848 * If the fault is a write, we know that this page is being 849 * written NOW so dirty it explicitly to save on 850 * pmap_is_modified() calls later. 851 * 852 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 853 * if the page is already dirty to prevent data written with 854 * the expectation of being synced from not being synced. 855 * Likewise if this entry does not request NOSYNC then make 856 * sure the page isn't marked NOSYNC. Applications sharing 857 * data should use the same flags to avoid ping ponging. 858 * 859 * Also tell the backing pager, if any, that it should remove 860 * any swap backing since the page is now dirty. 861 */ 862 if (fs.entry->eflags & MAP_ENTRY_NOSYNC) { 863 if (fs.m->dirty == 0) 864 vm_page_flag_set(fs.m, PG_NOSYNC); 865 } else { 866 vm_page_flag_clear(fs.m, PG_NOSYNC); 867 } 868 if (fault_flags & VM_FAULT_DIRTY) { 869 crit_enter(); 870 vm_page_dirty(fs.m); 871 vm_pager_page_unswapped(fs.m); 872 crit_exit(); 873 } 874 } 875 876 /* 877 * Page had better still be busy 878 */ 879 880 KASSERT(fs.m->flags & PG_BUSY, 881 ("vm_fault: page %p not busy!", fs.m)); 882 883 unlock_things(&fs); 884 885 /* 886 * Sanity check: page must be completely valid or it is not fit to 887 * map into user space. vm_pager_get_pages() ensures this. 888 */ 889 890 if (fs.m->valid != VM_PAGE_BITS_ALL) { 891 vm_page_zero_invalid(fs.m, TRUE); 892 printf("Warning: page %p partially invalid on fault\n", fs.m); 893 } 894 895 pmap_enter(fs.map->pmap, vaddr, fs.m, prot, wired); 896 897 if (((fault_flags & VM_FAULT_WIRE_MASK) == 0) && (wired == 0)) { 898 pmap_prefault(fs.map->pmap, vaddr, fs.entry); 899 } 900 901 vm_page_flag_clear(fs.m, PG_ZERO); 902 vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED); 903 if (fault_flags & VM_FAULT_HOLD) 904 vm_page_hold(fs.m); 905 906 /* 907 * If the page is not wired down, then put it where the pageout daemon 908 * can find it. 909 */ 910 911 if (fault_flags & VM_FAULT_WIRE_MASK) { 912 if (wired) 913 vm_page_wire(fs.m); 914 else 915 vm_page_unwire(fs.m, 1); 916 } else { 917 vm_page_activate(fs.m); 918 } 919 920 if (curproc && (curproc->p_flag & P_SWAPPEDOUT) == 0 && 921 curproc->p_stats) { 922 if (hardfault) { 923 curproc->p_stats->p_ru.ru_majflt++; 924 } else { 925 curproc->p_stats->p_ru.ru_minflt++; 926 } 927 } 928 929 /* 930 * Unlock everything, and return 931 */ 932 933 vm_page_wakeup(fs.m); 934 vm_object_deallocate(fs.first_object); 935 936 return (KERN_SUCCESS); 937 938 } 939 940 /* 941 * quick version of vm_fault 942 */ 943 int 944 vm_fault_quick(caddr_t v, int prot) 945 { 946 int r; 947 948 if (prot & VM_PROT_WRITE) 949 r = subyte(v, fubyte(v)); 950 else 951 r = fubyte(v); 952 return(r); 953 } 954 955 /* 956 * Wire down a range of virtual addresses in a map. The entry in question 957 * should be marked in-transition and the map must be locked. We must 958 * release the map temporarily while faulting-in the page to avoid a 959 * deadlock. Note that the entry may be clipped while we are blocked but 960 * will never be freed. 961 */ 962 int 963 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 964 { 965 boolean_t fictitious; 966 vm_offset_t start; 967 vm_offset_t end; 968 vm_offset_t va; 969 vm_paddr_t pa; 970 pmap_t pmap; 971 int rv; 972 973 pmap = vm_map_pmap(map); 974 start = entry->start; 975 end = entry->end; 976 fictitious = entry->object.vm_object && 977 (entry->object.vm_object->type == OBJT_DEVICE); 978 979 vm_map_unlock(map); 980 map->timestamp++; 981 982 /* 983 * We simulate a fault to get the page and enter it in the physical 984 * map. 985 */ 986 for (va = start; va < end; va += PAGE_SIZE) { 987 if (user_wire) { 988 rv = vm_fault(map, va, VM_PROT_READ, 989 VM_FAULT_USER_WIRE); 990 } else { 991 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 992 VM_FAULT_CHANGE_WIRING); 993 } 994 if (rv) { 995 while (va > start) { 996 va -= PAGE_SIZE; 997 if ((pa = pmap_extract(pmap, va)) == 0) 998 continue; 999 pmap_change_wiring(pmap, va, FALSE); 1000 if (!fictitious) 1001 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1002 } 1003 vm_map_lock(map); 1004 return (rv); 1005 } 1006 } 1007 vm_map_lock(map); 1008 return (KERN_SUCCESS); 1009 } 1010 1011 /* 1012 * Unwire a range of virtual addresses in a map. The map should be 1013 * locked. 1014 */ 1015 void 1016 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 1017 { 1018 boolean_t fictitious; 1019 vm_offset_t start; 1020 vm_offset_t end; 1021 vm_offset_t va; 1022 vm_paddr_t pa; 1023 pmap_t pmap; 1024 1025 pmap = vm_map_pmap(map); 1026 start = entry->start; 1027 end = entry->end; 1028 fictitious = entry->object.vm_object && 1029 (entry->object.vm_object->type == OBJT_DEVICE); 1030 1031 /* 1032 * Since the pages are wired down, we must be able to get their 1033 * mappings from the physical map system. 1034 */ 1035 for (va = start; va < end; va += PAGE_SIZE) { 1036 pa = pmap_extract(pmap, va); 1037 if (pa != 0) { 1038 pmap_change_wiring(pmap, va, FALSE); 1039 if (!fictitious) 1040 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1041 } 1042 } 1043 } 1044 1045 /* 1046 * Reduce the rate at which memory is allocated to a process based 1047 * on the perceived load on the VM system. As the load increases 1048 * the allocation burst rate goes down and the delay increases. 1049 * 1050 * Rate limiting does not apply when faulting active or inactive 1051 * pages. When faulting 'cache' pages, rate limiting only applies 1052 * if the system currently has a severe page deficit. 1053 * 1054 * XXX vm_pagesupply should be increased when a page is freed. 1055 * 1056 * We sleep up to 1/10 of a second. 1057 */ 1058 static int 1059 vm_fault_ratelimit(struct vmspace *vmspace) 1060 { 1061 if (vm_load_enable == 0) 1062 return(0); 1063 if (vmspace->vm_pagesupply > 0) { 1064 --vmspace->vm_pagesupply; 1065 return(0); 1066 } 1067 #ifdef INVARIANTS 1068 if (vm_load_debug) { 1069 printf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n", 1070 vm_load, 1071 (1000 - vm_load ) / 10, vm_load * hz / 10000, 1072 curproc->p_pid, curproc->p_comm); 1073 } 1074 #endif 1075 vmspace->vm_pagesupply = (1000 - vm_load) / 10; 1076 return(vm_load * hz / 10000); 1077 } 1078 1079 /* 1080 * Routine: 1081 * vm_fault_copy_entry 1082 * Function: 1083 * Copy all of the pages from a wired-down map entry to another. 1084 * 1085 * In/out conditions: 1086 * The source and destination maps must be locked for write. 1087 * The source map entry must be wired down (or be a sharing map 1088 * entry corresponding to a main map entry that is wired down). 1089 */ 1090 1091 void 1092 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1093 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 1094 { 1095 vm_object_t dst_object; 1096 vm_object_t src_object; 1097 vm_ooffset_t dst_offset; 1098 vm_ooffset_t src_offset; 1099 vm_prot_t prot; 1100 vm_offset_t vaddr; 1101 vm_page_t dst_m; 1102 vm_page_t src_m; 1103 1104 #ifdef lint 1105 src_map++; 1106 #endif /* lint */ 1107 1108 src_object = src_entry->object.vm_object; 1109 src_offset = src_entry->offset; 1110 1111 /* 1112 * Create the top-level object for the destination entry. (Doesn't 1113 * actually shadow anything - we copy the pages directly.) 1114 */ 1115 dst_object = vm_object_allocate(OBJT_DEFAULT, 1116 (vm_size_t) OFF_TO_IDX(dst_entry->end - dst_entry->start)); 1117 1118 dst_entry->object.vm_object = dst_object; 1119 dst_entry->offset = 0; 1120 1121 prot = dst_entry->max_protection; 1122 1123 /* 1124 * Loop through all of the pages in the entry's range, copying each 1125 * one from the source object (it should be there) to the destination 1126 * object. 1127 */ 1128 for (vaddr = dst_entry->start, dst_offset = 0; 1129 vaddr < dst_entry->end; 1130 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1131 1132 /* 1133 * Allocate a page in the destination object 1134 */ 1135 do { 1136 dst_m = vm_page_alloc(dst_object, 1137 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1138 if (dst_m == NULL) { 1139 vm_wait(); 1140 } 1141 } while (dst_m == NULL); 1142 1143 /* 1144 * Find the page in the source object, and copy it in. 1145 * (Because the source is wired down, the page will be in 1146 * memory.) 1147 */ 1148 src_m = vm_page_lookup(src_object, 1149 OFF_TO_IDX(dst_offset + src_offset)); 1150 if (src_m == NULL) 1151 panic("vm_fault_copy_wired: page missing"); 1152 1153 vm_page_copy(src_m, dst_m); 1154 1155 /* 1156 * Enter it in the pmap... 1157 */ 1158 1159 vm_page_flag_clear(dst_m, PG_ZERO); 1160 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1161 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED); 1162 1163 /* 1164 * Mark it no longer busy, and put it on the active list. 1165 */ 1166 vm_page_activate(dst_m); 1167 vm_page_wakeup(dst_m); 1168 } 1169 } 1170 1171 1172 /* 1173 * This routine checks around the requested page for other pages that 1174 * might be able to be faulted in. This routine brackets the viable 1175 * pages for the pages to be paged in. 1176 * 1177 * Inputs: 1178 * m, rbehind, rahead 1179 * 1180 * Outputs: 1181 * marray (array of vm_page_t), reqpage (index of requested page) 1182 * 1183 * Return value: 1184 * number of pages in marray 1185 */ 1186 static int 1187 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 1188 vm_page_t *marray, int *reqpage) 1189 { 1190 int i,j; 1191 vm_object_t object; 1192 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1193 vm_page_t rtm; 1194 int cbehind, cahead; 1195 1196 object = m->object; 1197 pindex = m->pindex; 1198 1199 /* 1200 * we don't fault-ahead for device pager 1201 */ 1202 if (object->type == OBJT_DEVICE) { 1203 *reqpage = 0; 1204 marray[0] = m; 1205 return 1; 1206 } 1207 1208 /* 1209 * if the requested page is not available, then give up now 1210 */ 1211 1212 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1213 return 0; 1214 } 1215 1216 if ((cbehind == 0) && (cahead == 0)) { 1217 *reqpage = 0; 1218 marray[0] = m; 1219 return 1; 1220 } 1221 1222 if (rahead > cahead) { 1223 rahead = cahead; 1224 } 1225 1226 if (rbehind > cbehind) { 1227 rbehind = cbehind; 1228 } 1229 1230 /* 1231 * try to do any readahead that we might have free pages for. 1232 */ 1233 if ((rahead + rbehind) > 1234 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) { 1235 pagedaemon_wakeup(); 1236 marray[0] = m; 1237 *reqpage = 0; 1238 return 1; 1239 } 1240 1241 /* 1242 * scan backward for the read behind pages -- in memory 1243 * 1244 * Assume that if the page is not found an interrupt will not 1245 * create it. Theoretically interrupts can only remove (busy) 1246 * pages, not create new associations. 1247 */ 1248 if (pindex > 0) { 1249 if (rbehind > pindex) { 1250 rbehind = pindex; 1251 startpindex = 0; 1252 } else { 1253 startpindex = pindex - rbehind; 1254 } 1255 1256 crit_enter(); 1257 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1258 if (vm_page_lookup( object, tpindex)) { 1259 startpindex = tpindex + 1; 1260 break; 1261 } 1262 if (tpindex == 0) 1263 break; 1264 } 1265 1266 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1267 1268 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1269 if (rtm == NULL) { 1270 crit_exit(); 1271 for (j = 0; j < i; j++) { 1272 vm_page_free(marray[j]); 1273 } 1274 marray[0] = m; 1275 *reqpage = 0; 1276 return 1; 1277 } 1278 1279 marray[i] = rtm; 1280 } 1281 crit_exit(); 1282 } else { 1283 startpindex = 0; 1284 i = 0; 1285 } 1286 1287 marray[i] = m; 1288 /* page offset of the required page */ 1289 *reqpage = i; 1290 1291 tpindex = pindex + 1; 1292 i++; 1293 1294 /* 1295 * scan forward for the read ahead pages 1296 */ 1297 endpindex = tpindex + rahead; 1298 if (endpindex > object->size) 1299 endpindex = object->size; 1300 1301 crit_enter(); 1302 for( ; tpindex < endpindex; i++, tpindex++) { 1303 1304 if (vm_page_lookup(object, tpindex)) { 1305 break; 1306 } 1307 1308 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1309 if (rtm == NULL) { 1310 break; 1311 } 1312 1313 marray[i] = rtm; 1314 } 1315 crit_exit(); 1316 1317 /* return number of bytes of pages */ 1318 return i; 1319 } 1320