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 1.34 2007/01/01 22:51:18 corecode 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 #include <sys/vkernel.h> 85 #include <sys/sfbuf.h> 86 #include <sys/lock.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vnode_pager.h> 98 #include <vm/vm_extern.h> 99 100 #include <sys/thread2.h> 101 #include <vm/vm_page2.h> 102 103 #define VM_FAULT_READ_AHEAD 8 104 #define VM_FAULT_READ_BEHIND 7 105 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 106 107 struct faultstate { 108 vm_page_t m; 109 vm_object_t object; 110 vm_pindex_t pindex; 111 vm_prot_t prot; 112 vm_page_t first_m; 113 vm_object_t first_object; 114 vm_prot_t first_prot; 115 vm_map_t map; 116 vm_map_entry_t entry; 117 int lookup_still_valid; 118 int didlimit; 119 int hardfault; 120 int fault_flags; 121 int map_generation; 122 boolean_t wired; 123 struct vnode *vp; 124 }; 125 126 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t); 127 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t); 128 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 129 static int vm_fault_ratelimit(struct vmspace *); 130 131 static __inline void 132 release_page(struct faultstate *fs) 133 { 134 vm_page_wakeup(fs->m); 135 vm_page_deactivate(fs->m); 136 fs->m = NULL; 137 } 138 139 static __inline void 140 unlock_map(struct faultstate *fs) 141 { 142 if (fs->lookup_still_valid) { 143 vm_map_lookup_done(fs->map, fs->entry, 0); 144 fs->lookup_still_valid = FALSE; 145 } 146 } 147 148 /* 149 * Clean up after a successful call to vm_fault_object() so another call 150 * to vm_fault_object() can be made. 151 */ 152 static void 153 _cleanup_successful_fault(struct faultstate *fs, int relock) 154 { 155 if (fs->object != fs->first_object) { 156 vm_page_free(fs->first_m); 157 vm_object_pip_wakeup(fs->object); 158 fs->first_m = NULL; 159 } 160 fs->object = fs->first_object; 161 if (relock && fs->lookup_still_valid == FALSE) { 162 vm_map_lock_read(fs->map); 163 fs->lookup_still_valid = TRUE; 164 } 165 } 166 167 static void 168 _unlock_things(struct faultstate *fs, int dealloc) 169 { 170 vm_object_pip_wakeup(fs->first_object); 171 _cleanup_successful_fault(fs, 0); 172 if (dealloc) { 173 vm_object_deallocate(fs->first_object); 174 } 175 unlock_map(fs); 176 if (fs->vp != NULL) { 177 vput(fs->vp); 178 fs->vp = NULL; 179 } 180 } 181 182 #define unlock_things(fs) _unlock_things(fs, 0) 183 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 184 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 185 186 /* 187 * TRYPAGER 188 * 189 * Determine if the pager for the current object *might* contain the page. 190 * 191 * We only need to try the pager if this is not a default object (default 192 * objects are zero-fill and have no real pager), and if we are not taking 193 * a wiring fault or if the FS entry is wired. 194 */ 195 #define TRYPAGER(fs) \ 196 (fs->object->type != OBJT_DEFAULT && \ 197 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired)) 198 199 /* 200 * vm_fault: 201 * 202 * Handle a page fault occuring at the given address, requiring the given 203 * permissions, in the map specified. If successful, the page is inserted 204 * into the associated physical map. 205 * 206 * NOTE: The given address should be truncated to the proper page address. 207 * 208 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 209 * a standard error specifying why the fault is fatal is returned. 210 * 211 * The map in question must be referenced, and remains so. 212 * The caller may hold no locks. 213 */ 214 int 215 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 216 { 217 int result; 218 vm_pindex_t first_pindex; 219 struct faultstate fs; 220 221 mycpu->gd_cnt.v_vm_faults++; 222 223 fs.didlimit = 0; 224 fs.hardfault = 0; 225 fs.fault_flags = fault_flags; 226 227 RetryFault: 228 /* 229 * Find the vm_map_entry representing the backing store and resolve 230 * the top level object and page index. This may have the side 231 * effect of executing a copy-on-write on the map entry and/or 232 * creating a shadow object, but will not COW any actual VM pages. 233 * 234 * On success fs.map is left read-locked and various other fields 235 * are initialized but not otherwise referenced or locked. 236 * 237 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 238 * if the map entry is a virtual page table and also writable, 239 * so we can set the 'A'accessed bit in the virtual page table entry. 240 */ 241 fs.map = map; 242 result = vm_map_lookup(&fs.map, vaddr, fault_type, 243 &fs.entry, &fs.first_object, 244 &first_pindex, &fs.first_prot, &fs.wired); 245 246 /* 247 * If the lookup failed or the map protections are incompatible, 248 * the fault generally fails. However, if the caller is trying 249 * to do a user wiring we have more work to do. 250 */ 251 if (result != KERN_SUCCESS) { 252 if (result != KERN_PROTECTION_FAILURE) 253 return result; 254 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 255 return result; 256 257 /* 258 * If we are user-wiring a r/w segment, and it is COW, then 259 * we need to do the COW operation. Note that we don't 260 * currently COW RO sections now, because it is NOT desirable 261 * to COW .text. We simply keep .text from ever being COW'ed 262 * and take the heat that one cannot debug wired .text sections. 263 */ 264 result = vm_map_lookup(&fs.map, vaddr, 265 VM_PROT_READ|VM_PROT_WRITE| 266 VM_PROT_OVERRIDE_WRITE, 267 &fs.entry, &fs.first_object, 268 &first_pindex, &fs.first_prot, 269 &fs.wired); 270 if (result != KERN_SUCCESS) 271 return result; 272 273 /* 274 * If we don't COW now, on a user wire, the user will never 275 * be able to write to the mapping. If we don't make this 276 * restriction, the bookkeeping would be nearly impossible. 277 */ 278 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 279 fs.entry->max_protection &= ~VM_PROT_WRITE; 280 } 281 282 /* 283 * fs.map is read-locked 284 * 285 * Misc checks. Save the map generation number to detect races. 286 */ 287 fs.map_generation = fs.map->timestamp; 288 289 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 290 panic("vm_fault: fault on nofault entry, addr: %lx", 291 (u_long)vaddr); 292 } 293 294 /* 295 * A system map entry may return a NULL object. No object means 296 * no pager means an unrecoverable kernel fault. 297 */ 298 if (fs.first_object == NULL) { 299 panic("vm_fault: unrecoverable fault at %p in entry %p", 300 (void *)vaddr, fs.entry); 301 } 302 303 /* 304 * Make a reference to this object to prevent its disposal while we 305 * are messing with it. Once we have the reference, the map is free 306 * to be diddled. Since objects reference their shadows (and copies), 307 * they will stay around as well. 308 * 309 * Bump the paging-in-progress count to prevent size changes (e.g. 310 * truncation operations) during I/O. This must be done after 311 * obtaining the vnode lock in order to avoid possible deadlocks. 312 */ 313 vm_object_reference(fs.first_object); 314 fs.vp = vnode_pager_lock(fs.first_object); 315 vm_object_pip_add(fs.first_object, 1); 316 317 fs.lookup_still_valid = TRUE; 318 fs.first_m = NULL; 319 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 320 321 /* 322 * If the entry is wired we cannot change the page protection. 323 */ 324 if (fs.wired) 325 fault_type = fs.first_prot; 326 327 /* 328 * The page we want is at (first_object, first_pindex), but if the 329 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 330 * page table to figure out the actual pindex. 331 * 332 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 333 * ONLY 334 */ 335 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 336 result = vm_fault_vpagetable(&fs, &first_pindex, 337 fs.entry->aux.master_pde); 338 if (result == KERN_TRY_AGAIN) 339 goto RetryFault; 340 if (result != KERN_SUCCESS) 341 return (result); 342 } 343 344 /* 345 * Now we have the actual (object, pindex), fault in the page. If 346 * vm_fault_object() fails it will unlock and deallocate the FS 347 * data. If it succeeds everything remains locked and fs->object 348 * will have an additinal PIP count if it is not equal to 349 * fs->first_object 350 */ 351 result = vm_fault_object(&fs, first_pindex, fault_type); 352 353 if (result == KERN_TRY_AGAIN) 354 goto RetryFault; 355 if (result != KERN_SUCCESS) 356 return (result); 357 358 /* 359 * On success vm_fault_object() does not unlock or deallocate, and fs.m 360 * will contain a busied page. 361 * 362 * Enter the page into the pmap and do pmap-related adjustments. 363 */ 364 unlock_things(&fs); 365 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired); 366 367 if (((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0) && (fs.wired == 0)) { 368 pmap_prefault(fs.map->pmap, vaddr, fs.entry); 369 } 370 371 vm_page_flag_clear(fs.m, PG_ZERO); 372 vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED); 373 if (fs.fault_flags & VM_FAULT_HOLD) 374 vm_page_hold(fs.m); 375 376 /* 377 * If the page is not wired down, then put it where the pageout daemon 378 * can find it. 379 */ 380 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 381 if (fs.wired) 382 vm_page_wire(fs.m); 383 else 384 vm_page_unwire(fs.m, 1); 385 } else { 386 vm_page_activate(fs.m); 387 } 388 389 if (curthread->td_lwp) { 390 if (fs.hardfault) { 391 curthread->td_lwp->lwp_ru.ru_majflt++; 392 } else { 393 curthread->td_lwp->lwp_ru.ru_minflt++; 394 } 395 } 396 397 /* 398 * Unlock everything, and return 399 */ 400 vm_page_wakeup(fs.m); 401 vm_object_deallocate(fs.first_object); 402 403 return (KERN_SUCCESS); 404 } 405 406 /* 407 * Translate the virtual page number (first_pindex) that is relative 408 * to the address space into a logical page number that is relative to the 409 * backing object. Use the virtual page table pointed to by (vpte). 410 * 411 * This implements an N-level page table. Any level can terminate the 412 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 413 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 414 */ 415 static 416 int 417 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, vpte_t vpte) 418 { 419 struct sf_buf *sf; 420 int vshift = 32 - PAGE_SHIFT; /* page index bits remaining */ 421 int result = KERN_SUCCESS; 422 423 for (;;) { 424 if ((vpte & VPTE_V) == 0) { 425 unlock_and_deallocate(fs); 426 return (KERN_FAILURE); 427 } 428 if ((vpte & VPTE_PS) || vshift == 0) 429 break; 430 KKASSERT(vshift >= VPTE_PAGE_BITS); 431 432 /* 433 * Get the page table page 434 */ 435 result = vm_fault_object(fs, vpte >> PAGE_SHIFT, VM_PROT_READ); 436 if (result != KERN_SUCCESS) 437 return (result); 438 439 /* 440 * Process the returned fs.m and look up the page table 441 * entry in the page table page. 442 */ 443 vshift -= VPTE_PAGE_BITS; 444 sf = sf_buf_alloc(fs->m, SFB_CPUPRIVATE); 445 vpte = *((vpte_t *)sf_buf_kva(sf) + 446 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 447 sf_buf_free(sf); 448 vm_page_flag_set(fs->m, PG_REFERENCED); 449 vm_page_activate(fs->m); 450 vm_page_wakeup(fs->m); 451 cleanup_successful_fault(fs); 452 } 453 /* 454 * Combine remaining address bits with the vpte. 455 */ 456 *pindex = (vpte >> PAGE_SHIFT) + 457 (*pindex & ((1 << vshift) - 1)); 458 return (KERN_SUCCESS); 459 } 460 461 462 /* 463 * Do all operations required to fault-in (fs.first_object, pindex). Run 464 * through the shadow chain as necessary and do required COW or virtual 465 * copy operations. The caller has already fully resolved the vm_map_entry 466 * and, if appropriate, has created a copy-on-write layer. All we need to 467 * do is iterate the object chain. 468 * 469 * On failure (fs) is unlocked and deallocated and the caller may return or 470 * retry depending on the failure code. On success (fs) is NOT unlocked or 471 * deallocated, fs.m will contained a resolved, busied page, and fs.object 472 * will have an additional PIP count if it is not equal to fs.first_object. 473 */ 474 static 475 int 476 vm_fault_object(struct faultstate *fs, 477 vm_pindex_t first_pindex, vm_prot_t fault_type) 478 { 479 vm_object_t next_object; 480 vm_page_t marray[VM_FAULT_READ]; 481 vm_pindex_t pindex; 482 int faultcount; 483 484 fs->prot = fs->first_prot; 485 fs->object = fs->first_object; 486 pindex = first_pindex; 487 488 for (;;) { 489 /* 490 * If the object is dead, we stop here 491 */ 492 if (fs->object->flags & OBJ_DEAD) { 493 unlock_and_deallocate(fs); 494 return (KERN_PROTECTION_FAILURE); 495 } 496 497 /* 498 * See if page is resident. spl protection is required 499 * to avoid an interrupt unbusy/free race against our 500 * lookup. We must hold the protection through a page 501 * allocation or busy. 502 */ 503 crit_enter(); 504 fs->m = vm_page_lookup(fs->object, pindex); 505 if (fs->m != NULL) { 506 int queue; 507 /* 508 * Wait/Retry if the page is busy. We have to do this 509 * if the page is busy via either PG_BUSY or 510 * vm_page_t->busy because the vm_pager may be using 511 * vm_page_t->busy for pageouts ( and even pageins if 512 * it is the vnode pager ), and we could end up trying 513 * to pagein and pageout the same page simultaneously. 514 * 515 * We can theoretically allow the busy case on a read 516 * fault if the page is marked valid, but since such 517 * pages are typically already pmap'd, putting that 518 * special case in might be more effort then it is 519 * worth. We cannot under any circumstances mess 520 * around with a vm_page_t->busy page except, perhaps, 521 * to pmap it. 522 */ 523 if ((fs->m->flags & PG_BUSY) || fs->m->busy) { 524 unlock_things(fs); 525 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 526 mycpu->gd_cnt.v_intrans++; 527 vm_object_deallocate(fs->first_object); 528 crit_exit(); 529 return (KERN_TRY_AGAIN); 530 } 531 532 /* 533 * If reactivating a page from PQ_CACHE we may have 534 * to rate-limit. 535 */ 536 queue = fs->m->queue; 537 vm_page_unqueue_nowakeup(fs->m); 538 539 if ((queue - fs->m->pc) == PQ_CACHE && 540 vm_page_count_severe()) { 541 vm_page_activate(fs->m); 542 unlock_and_deallocate(fs); 543 vm_waitpfault(); 544 crit_exit(); 545 return (KERN_TRY_AGAIN); 546 } 547 548 /* 549 * Mark page busy for other processes, and the 550 * pagedaemon. If it still isn't completely valid 551 * (readable), jump to readrest, else we found the 552 * page and can return. 553 * 554 * We can release the spl once we have marked the 555 * page busy. 556 */ 557 vm_page_busy(fs->m); 558 crit_exit(); 559 560 if (((fs->m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 561 fs->m->object != &kernel_object) { 562 goto readrest; 563 } 564 break; /* break to PAGE HAS BEEN FOUND */ 565 } 566 567 /* 568 * Page is not resident, If this is the search termination 569 * or the pager might contain the page, allocate a new page. 570 * 571 * NOTE: We are still in a critical section. 572 */ 573 if (TRYPAGER(fs) || fs->object == fs->first_object) { 574 /* 575 * If the page is beyond the object size we fail 576 */ 577 if (pindex >= fs->object->size) { 578 crit_exit(); 579 unlock_and_deallocate(fs); 580 return (KERN_PROTECTION_FAILURE); 581 } 582 583 /* 584 * Ratelimit. 585 */ 586 if (fs->didlimit == 0 && curproc != NULL) { 587 int limticks; 588 589 limticks = vm_fault_ratelimit(curproc->p_vmspace); 590 if (limticks) { 591 crit_exit(); 592 unlock_and_deallocate(fs); 593 tsleep(curproc, 0, "vmrate", limticks); 594 fs->didlimit = 1; 595 return (KERN_TRY_AGAIN); 596 } 597 } 598 599 /* 600 * Allocate a new page for this object/offset pair. 601 */ 602 fs->m = NULL; 603 if (!vm_page_count_severe()) { 604 fs->m = vm_page_alloc(fs->object, pindex, 605 (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO); 606 } 607 if (fs->m == NULL) { 608 crit_exit(); 609 unlock_and_deallocate(fs); 610 vm_waitpfault(); 611 return (KERN_TRY_AGAIN); 612 } 613 } 614 crit_exit(); 615 616 readrest: 617 /* 618 * We have found a valid page or we have allocated a new page. 619 * The page thus may not be valid or may not be entirely 620 * valid. 621 * 622 * Attempt to fault-in the page if there is a chance that the 623 * pager has it, and potentially fault in additional pages 624 * at the same time. 625 * 626 * We are NOT in splvm here and if TRYPAGER is true then 627 * fs.m will be non-NULL and will be PG_BUSY for us. 628 */ 629 630 if (TRYPAGER(fs)) { 631 int rv; 632 int reqpage; 633 int ahead, behind; 634 u_char behavior = vm_map_entry_behavior(fs->entry); 635 636 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 637 ahead = 0; 638 behind = 0; 639 } else { 640 behind = pindex; 641 if (behind > VM_FAULT_READ_BEHIND) 642 behind = VM_FAULT_READ_BEHIND; 643 644 ahead = fs->object->size - pindex; 645 if (ahead < 1) 646 ahead = 1; 647 if (ahead > VM_FAULT_READ_AHEAD) 648 ahead = VM_FAULT_READ_AHEAD; 649 } 650 651 if ((fs->first_object->type != OBJT_DEVICE) && 652 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 653 (behavior != MAP_ENTRY_BEHAV_RANDOM && 654 pindex >= fs->entry->lastr && 655 pindex < fs->entry->lastr + VM_FAULT_READ)) 656 ) { 657 vm_pindex_t firstpindex, tmppindex; 658 659 if (first_pindex < 2 * VM_FAULT_READ) 660 firstpindex = 0; 661 else 662 firstpindex = first_pindex - 2 * VM_FAULT_READ; 663 664 /* 665 * note: partially valid pages cannot be 666 * included in the lookahead - NFS piecemeal 667 * writes will barf on it badly. 668 * 669 * spl protection is required to avoid races 670 * between the lookup and an interrupt 671 * unbusy/free sequence occuring prior to 672 * our busy check. 673 */ 674 crit_enter(); 675 for (tmppindex = first_pindex - 1; 676 tmppindex >= firstpindex; 677 --tmppindex 678 ) { 679 vm_page_t mt; 680 681 mt = vm_page_lookup(fs->first_object, tmppindex); 682 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 683 break; 684 if (mt->busy || 685 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 686 mt->hold_count || 687 mt->wire_count) 688 continue; 689 if (mt->dirty == 0) 690 vm_page_test_dirty(mt); 691 if (mt->dirty) { 692 vm_page_protect(mt, VM_PROT_NONE); 693 vm_page_deactivate(mt); 694 } else { 695 vm_page_cache(mt); 696 } 697 } 698 crit_exit(); 699 700 ahead += behind; 701 behind = 0; 702 } 703 704 /* 705 * now we find out if any other pages should be paged 706 * in at this time this routine checks to see if the 707 * pages surrounding this fault reside in the same 708 * object as the page for this fault. If they do, 709 * then they are faulted in also into the object. The 710 * array "marray" returned contains an array of 711 * vm_page_t structs where one of them is the 712 * vm_page_t passed to the routine. The reqpage 713 * return value is the index into the marray for the 714 * vm_page_t passed to the routine. 715 * 716 * fs.m plus the additional pages are PG_BUSY'd. 717 */ 718 faultcount = vm_fault_additional_pages( 719 fs->m, behind, ahead, marray, &reqpage); 720 721 /* 722 * update lastr imperfectly (we do not know how much 723 * getpages will actually read), but good enough. 724 */ 725 fs->entry->lastr = pindex + faultcount - behind; 726 727 /* 728 * Call the pager to retrieve the data, if any, after 729 * releasing the lock on the map. We hold a ref on 730 * fs.object and the pages are PG_BUSY'd. 731 */ 732 unlock_map(fs); 733 734 if (faultcount) { 735 rv = vm_pager_get_pages(fs->object, marray, 736 faultcount, reqpage); 737 } else { 738 rv = VM_PAGER_FAIL; 739 } 740 741 if (rv == VM_PAGER_OK) { 742 /* 743 * Found the page. Leave it busy while we play 744 * with it. 745 */ 746 747 /* 748 * Relookup in case pager changed page. Pager 749 * is responsible for disposition of old page 750 * if moved. 751 * 752 * XXX other code segments do relookups too. 753 * It's a bad abstraction that needs to be 754 * fixed/removed. 755 */ 756 fs->m = vm_page_lookup(fs->object, pindex); 757 if (fs->m == NULL) { 758 unlock_and_deallocate(fs); 759 return (KERN_TRY_AGAIN); 760 } 761 762 ++fs->hardfault; 763 break; /* break to PAGE HAS BEEN FOUND */ 764 } 765 766 /* 767 * Remove the bogus page (which does not exist at this 768 * object/offset); before doing so, we must get back 769 * our object lock to preserve our invariant. 770 * 771 * Also wake up any other process that may want to bring 772 * in this page. 773 * 774 * If this is the top-level object, we must leave the 775 * busy page to prevent another process from rushing 776 * past us, and inserting the page in that object at 777 * the same time that we are. 778 */ 779 if (rv == VM_PAGER_ERROR) { 780 if (curproc) 781 kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); 782 else 783 kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm); 784 } 785 /* 786 * Data outside the range of the pager or an I/O error 787 */ 788 /* 789 * XXX - the check for kernel_map is a kludge to work 790 * around having the machine panic on a kernel space 791 * fault w/ I/O error. 792 */ 793 if (((fs->map != &kernel_map) && (rv == VM_PAGER_ERROR)) || 794 (rv == VM_PAGER_BAD)) { 795 vm_page_free(fs->m); 796 fs->m = NULL; 797 unlock_and_deallocate(fs); 798 if (rv == VM_PAGER_ERROR) 799 return (KERN_FAILURE); 800 else 801 return (KERN_PROTECTION_FAILURE); 802 /* NOT REACHED */ 803 } 804 if (fs->object != fs->first_object) { 805 vm_page_free(fs->m); 806 fs->m = NULL; 807 /* 808 * XXX - we cannot just fall out at this 809 * point, m has been freed and is invalid! 810 */ 811 } 812 } 813 814 /* 815 * We get here if the object has a default pager (or unwiring) 816 * or the pager doesn't have the page. 817 */ 818 if (fs->object == fs->first_object) 819 fs->first_m = fs->m; 820 821 /* 822 * Move on to the next object. Lock the next object before 823 * unlocking the current one. 824 */ 825 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 826 next_object = fs->object->backing_object; 827 if (next_object == NULL) { 828 /* 829 * If there's no object left, fill the page in the top 830 * object with zeros. 831 */ 832 if (fs->object != fs->first_object) { 833 vm_object_pip_wakeup(fs->object); 834 835 fs->object = fs->first_object; 836 pindex = first_pindex; 837 fs->m = fs->first_m; 838 } 839 fs->first_m = NULL; 840 841 /* 842 * Zero the page if necessary and mark it valid. 843 */ 844 if ((fs->m->flags & PG_ZERO) == 0) { 845 vm_page_zero_fill(fs->m); 846 } else { 847 mycpu->gd_cnt.v_ozfod++; 848 } 849 mycpu->gd_cnt.v_zfod++; 850 fs->m->valid = VM_PAGE_BITS_ALL; 851 break; /* break to PAGE HAS BEEN FOUND */ 852 } else { 853 if (fs->object != fs->first_object) { 854 vm_object_pip_wakeup(fs->object); 855 } 856 KASSERT(fs->object != next_object, ("object loop %p", next_object)); 857 fs->object = next_object; 858 vm_object_pip_add(fs->object, 1); 859 } 860 } 861 862 KASSERT((fs->m->flags & PG_BUSY) != 0, 863 ("vm_fault: not busy after main loop")); 864 865 /* 866 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 867 * is held.] 868 */ 869 870 /* 871 * If the page is being written, but isn't already owned by the 872 * top-level object, we have to copy it into a new page owned by the 873 * top-level object. 874 */ 875 if (fs->object != fs->first_object) { 876 /* 877 * We only really need to copy if we want to write it. 878 */ 879 if (fault_type & VM_PROT_WRITE) { 880 /* 881 * This allows pages to be virtually copied from a 882 * backing_object into the first_object, where the 883 * backing object has no other refs to it, and cannot 884 * gain any more refs. Instead of a bcopy, we just 885 * move the page from the backing object to the 886 * first object. Note that we must mark the page 887 * dirty in the first object so that it will go out 888 * to swap when needed. 889 */ 890 if (fs->map_generation == fs->map->timestamp && 891 /* 892 * Only one shadow object 893 */ 894 (fs->object->shadow_count == 1) && 895 /* 896 * No COW refs, except us 897 */ 898 (fs->object->ref_count == 1) && 899 /* 900 * No one else can look this object up 901 */ 902 (fs->object->handle == NULL) && 903 /* 904 * No other ways to look the object up 905 */ 906 ((fs->object->type == OBJT_DEFAULT) || 907 (fs->object->type == OBJT_SWAP)) && 908 /* 909 * We don't chase down the shadow chain 910 */ 911 (fs->object == fs->first_object->backing_object) && 912 913 /* 914 * grab the lock if we need to 915 */ 916 (fs->lookup_still_valid || 917 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 918 ) { 919 920 fs->lookup_still_valid = 1; 921 /* 922 * get rid of the unnecessary page 923 */ 924 vm_page_protect(fs->first_m, VM_PROT_NONE); 925 vm_page_free(fs->first_m); 926 fs->first_m = NULL; 927 928 /* 929 * grab the page and put it into the 930 * process'es object. The page is 931 * automatically made dirty. 932 */ 933 vm_page_rename(fs->m, fs->first_object, first_pindex); 934 fs->first_m = fs->m; 935 vm_page_busy(fs->first_m); 936 fs->m = NULL; 937 mycpu->gd_cnt.v_cow_optim++; 938 } else { 939 /* 940 * Oh, well, lets copy it. 941 */ 942 vm_page_copy(fs->m, fs->first_m); 943 } 944 945 if (fs->m) { 946 /* 947 * We no longer need the old page or object. 948 */ 949 release_page(fs); 950 } 951 952 /* 953 * fs->object != fs->first_object due to above 954 * conditional 955 */ 956 vm_object_pip_wakeup(fs->object); 957 958 /* 959 * Only use the new page below... 960 */ 961 962 mycpu->gd_cnt.v_cow_faults++; 963 fs->m = fs->first_m; 964 fs->object = fs->first_object; 965 pindex = first_pindex; 966 } else { 967 /* 968 * If it wasn't a write fault avoid having to copy 969 * the page by mapping it read-only. 970 */ 971 fs->prot &= ~VM_PROT_WRITE; 972 } 973 } 974 975 /* 976 * We may have had to unlock a map to do I/O. If we did then 977 * lookup_still_valid will be FALSE. If the map generation count 978 * also changed then all sorts of things could have happened while 979 * we were doing the I/O and we need to retry. 980 */ 981 982 if (!fs->lookup_still_valid && 983 (fs->map->timestamp != fs->map_generation)) { 984 release_page(fs); 985 unlock_and_deallocate(fs); 986 return (KERN_TRY_AGAIN); 987 } 988 989 /* 990 * Put this page into the physical map. We had to do the unlock above 991 * because pmap_enter may cause other faults. We don't put the page 992 * back on the active queue until later so that the page-out daemon 993 * won't find us (yet). 994 */ 995 if (fs->prot & VM_PROT_WRITE) { 996 vm_page_flag_set(fs->m, PG_WRITEABLE); 997 vm_object_set_writeable_dirty(fs->m->object); 998 999 /* 1000 * If the fault is a write, we know that this page is being 1001 * written NOW so dirty it explicitly to save on 1002 * pmap_is_modified() calls later. 1003 * 1004 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 1005 * if the page is already dirty to prevent data written with 1006 * the expectation of being synced from not being synced. 1007 * Likewise if this entry does not request NOSYNC then make 1008 * sure the page isn't marked NOSYNC. Applications sharing 1009 * data should use the same flags to avoid ping ponging. 1010 * 1011 * Also tell the backing pager, if any, that it should remove 1012 * any swap backing since the page is now dirty. 1013 */ 1014 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) { 1015 if (fs->m->dirty == 0) 1016 vm_page_flag_set(fs->m, PG_NOSYNC); 1017 } else { 1018 vm_page_flag_clear(fs->m, PG_NOSYNC); 1019 } 1020 if (fs->fault_flags & VM_FAULT_DIRTY) { 1021 crit_enter(); 1022 vm_page_dirty(fs->m); 1023 vm_pager_page_unswapped(fs->m); 1024 crit_exit(); 1025 } 1026 } 1027 1028 /* 1029 * Page had better still be busy. We are still locked up and 1030 * fs->object will have another PIP reference if it is not equal 1031 * to fs->first_object. 1032 */ 1033 KASSERT(fs->m->flags & PG_BUSY, 1034 ("vm_fault: page %p not busy!", fs->m)); 1035 1036 /* 1037 * Sanity check: page must be completely valid or it is not fit to 1038 * map into user space. vm_pager_get_pages() ensures this. 1039 */ 1040 if (fs->m->valid != VM_PAGE_BITS_ALL) { 1041 vm_page_zero_invalid(fs->m, TRUE); 1042 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 1043 } 1044 1045 return (KERN_SUCCESS); 1046 } 1047 1048 /* 1049 * quick version of vm_fault 1050 */ 1051 int 1052 vm_fault_quick(caddr_t v, int prot) 1053 { 1054 int r; 1055 1056 if (prot & VM_PROT_WRITE) 1057 r = subyte(v, fubyte(v)); 1058 else 1059 r = fubyte(v); 1060 return(r); 1061 } 1062 1063 /* 1064 * Wire down a range of virtual addresses in a map. The entry in question 1065 * should be marked in-transition and the map must be locked. We must 1066 * release the map temporarily while faulting-in the page to avoid a 1067 * deadlock. Note that the entry may be clipped while we are blocked but 1068 * will never be freed. 1069 */ 1070 int 1071 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 1072 { 1073 boolean_t fictitious; 1074 vm_offset_t start; 1075 vm_offset_t end; 1076 vm_offset_t va; 1077 vm_paddr_t pa; 1078 pmap_t pmap; 1079 int rv; 1080 1081 pmap = vm_map_pmap(map); 1082 start = entry->start; 1083 end = entry->end; 1084 fictitious = entry->object.vm_object && 1085 (entry->object.vm_object->type == OBJT_DEVICE); 1086 1087 vm_map_unlock(map); 1088 map->timestamp++; 1089 1090 /* 1091 * We simulate a fault to get the page and enter it in the physical 1092 * map. 1093 */ 1094 for (va = start; va < end; va += PAGE_SIZE) { 1095 if (user_wire) { 1096 rv = vm_fault(map, va, VM_PROT_READ, 1097 VM_FAULT_USER_WIRE); 1098 } else { 1099 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 1100 VM_FAULT_CHANGE_WIRING); 1101 } 1102 if (rv) { 1103 while (va > start) { 1104 va -= PAGE_SIZE; 1105 if ((pa = pmap_extract(pmap, va)) == 0) 1106 continue; 1107 pmap_change_wiring(pmap, va, FALSE); 1108 if (!fictitious) 1109 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1110 } 1111 vm_map_lock(map); 1112 return (rv); 1113 } 1114 } 1115 vm_map_lock(map); 1116 return (KERN_SUCCESS); 1117 } 1118 1119 /* 1120 * Unwire a range of virtual addresses in a map. The map should be 1121 * locked. 1122 */ 1123 void 1124 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 1125 { 1126 boolean_t fictitious; 1127 vm_offset_t start; 1128 vm_offset_t end; 1129 vm_offset_t va; 1130 vm_paddr_t pa; 1131 pmap_t pmap; 1132 1133 pmap = vm_map_pmap(map); 1134 start = entry->start; 1135 end = entry->end; 1136 fictitious = entry->object.vm_object && 1137 (entry->object.vm_object->type == OBJT_DEVICE); 1138 1139 /* 1140 * Since the pages are wired down, we must be able to get their 1141 * mappings from the physical map system. 1142 */ 1143 for (va = start; va < end; va += PAGE_SIZE) { 1144 pa = pmap_extract(pmap, va); 1145 if (pa != 0) { 1146 pmap_change_wiring(pmap, va, FALSE); 1147 if (!fictitious) 1148 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1149 } 1150 } 1151 } 1152 1153 /* 1154 * Reduce the rate at which memory is allocated to a process based 1155 * on the perceived load on the VM system. As the load increases 1156 * the allocation burst rate goes down and the delay increases. 1157 * 1158 * Rate limiting does not apply when faulting active or inactive 1159 * pages. When faulting 'cache' pages, rate limiting only applies 1160 * if the system currently has a severe page deficit. 1161 * 1162 * XXX vm_pagesupply should be increased when a page is freed. 1163 * 1164 * We sleep up to 1/10 of a second. 1165 */ 1166 static int 1167 vm_fault_ratelimit(struct vmspace *vmspace) 1168 { 1169 if (vm_load_enable == 0) 1170 return(0); 1171 if (vmspace->vm_pagesupply > 0) { 1172 --vmspace->vm_pagesupply; 1173 return(0); 1174 } 1175 #ifdef INVARIANTS 1176 if (vm_load_debug) { 1177 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n", 1178 vm_load, 1179 (1000 - vm_load ) / 10, vm_load * hz / 10000, 1180 curproc->p_pid, curproc->p_comm); 1181 } 1182 #endif 1183 vmspace->vm_pagesupply = (1000 - vm_load) / 10; 1184 return(vm_load * hz / 10000); 1185 } 1186 1187 /* 1188 * Routine: 1189 * vm_fault_copy_entry 1190 * Function: 1191 * Copy all of the pages from a wired-down map entry to another. 1192 * 1193 * In/out conditions: 1194 * The source and destination maps must be locked for write. 1195 * The source map entry must be wired down (or be a sharing map 1196 * entry corresponding to a main map entry that is wired down). 1197 */ 1198 1199 void 1200 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1201 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 1202 { 1203 vm_object_t dst_object; 1204 vm_object_t src_object; 1205 vm_ooffset_t dst_offset; 1206 vm_ooffset_t src_offset; 1207 vm_prot_t prot; 1208 vm_offset_t vaddr; 1209 vm_page_t dst_m; 1210 vm_page_t src_m; 1211 1212 #ifdef lint 1213 src_map++; 1214 #endif /* lint */ 1215 1216 src_object = src_entry->object.vm_object; 1217 src_offset = src_entry->offset; 1218 1219 /* 1220 * Create the top-level object for the destination entry. (Doesn't 1221 * actually shadow anything - we copy the pages directly.) 1222 */ 1223 vm_map_entry_allocate_object(dst_entry); 1224 dst_object = dst_entry->object.vm_object; 1225 1226 prot = dst_entry->max_protection; 1227 1228 /* 1229 * Loop through all of the pages in the entry's range, copying each 1230 * one from the source object (it should be there) to the destination 1231 * object. 1232 */ 1233 for (vaddr = dst_entry->start, dst_offset = 0; 1234 vaddr < dst_entry->end; 1235 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1236 1237 /* 1238 * Allocate a page in the destination object 1239 */ 1240 do { 1241 dst_m = vm_page_alloc(dst_object, 1242 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1243 if (dst_m == NULL) { 1244 vm_wait(); 1245 } 1246 } while (dst_m == NULL); 1247 1248 /* 1249 * Find the page in the source object, and copy it in. 1250 * (Because the source is wired down, the page will be in 1251 * memory.) 1252 */ 1253 src_m = vm_page_lookup(src_object, 1254 OFF_TO_IDX(dst_offset + src_offset)); 1255 if (src_m == NULL) 1256 panic("vm_fault_copy_wired: page missing"); 1257 1258 vm_page_copy(src_m, dst_m); 1259 1260 /* 1261 * Enter it in the pmap... 1262 */ 1263 1264 vm_page_flag_clear(dst_m, PG_ZERO); 1265 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1266 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED); 1267 1268 /* 1269 * Mark it no longer busy, and put it on the active list. 1270 */ 1271 vm_page_activate(dst_m); 1272 vm_page_wakeup(dst_m); 1273 } 1274 } 1275 1276 1277 /* 1278 * This routine checks around the requested page for other pages that 1279 * might be able to be faulted in. This routine brackets the viable 1280 * pages for the pages to be paged in. 1281 * 1282 * Inputs: 1283 * m, rbehind, rahead 1284 * 1285 * Outputs: 1286 * marray (array of vm_page_t), reqpage (index of requested page) 1287 * 1288 * Return value: 1289 * number of pages in marray 1290 */ 1291 static int 1292 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 1293 vm_page_t *marray, int *reqpage) 1294 { 1295 int i,j; 1296 vm_object_t object; 1297 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1298 vm_page_t rtm; 1299 int cbehind, cahead; 1300 1301 object = m->object; 1302 pindex = m->pindex; 1303 1304 /* 1305 * we don't fault-ahead for device pager 1306 */ 1307 if (object->type == OBJT_DEVICE) { 1308 *reqpage = 0; 1309 marray[0] = m; 1310 return 1; 1311 } 1312 1313 /* 1314 * if the requested page is not available, then give up now 1315 */ 1316 1317 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1318 return 0; 1319 } 1320 1321 if ((cbehind == 0) && (cahead == 0)) { 1322 *reqpage = 0; 1323 marray[0] = m; 1324 return 1; 1325 } 1326 1327 if (rahead > cahead) { 1328 rahead = cahead; 1329 } 1330 1331 if (rbehind > cbehind) { 1332 rbehind = cbehind; 1333 } 1334 1335 /* 1336 * try to do any readahead that we might have free pages for. 1337 */ 1338 if ((rahead + rbehind) > 1339 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) { 1340 pagedaemon_wakeup(); 1341 marray[0] = m; 1342 *reqpage = 0; 1343 return 1; 1344 } 1345 1346 /* 1347 * scan backward for the read behind pages -- in memory 1348 * 1349 * Assume that if the page is not found an interrupt will not 1350 * create it. Theoretically interrupts can only remove (busy) 1351 * pages, not create new associations. 1352 */ 1353 if (pindex > 0) { 1354 if (rbehind > pindex) { 1355 rbehind = pindex; 1356 startpindex = 0; 1357 } else { 1358 startpindex = pindex - rbehind; 1359 } 1360 1361 crit_enter(); 1362 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1363 if (vm_page_lookup( object, tpindex)) { 1364 startpindex = tpindex + 1; 1365 break; 1366 } 1367 if (tpindex == 0) 1368 break; 1369 } 1370 1371 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1372 1373 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1374 if (rtm == NULL) { 1375 crit_exit(); 1376 for (j = 0; j < i; j++) { 1377 vm_page_free(marray[j]); 1378 } 1379 marray[0] = m; 1380 *reqpage = 0; 1381 return 1; 1382 } 1383 1384 marray[i] = rtm; 1385 } 1386 crit_exit(); 1387 } else { 1388 startpindex = 0; 1389 i = 0; 1390 } 1391 1392 marray[i] = m; 1393 /* page offset of the required page */ 1394 *reqpage = i; 1395 1396 tpindex = pindex + 1; 1397 i++; 1398 1399 /* 1400 * scan forward for the read ahead pages 1401 */ 1402 endpindex = tpindex + rahead; 1403 if (endpindex > object->size) 1404 endpindex = object->size; 1405 1406 crit_enter(); 1407 for( ; tpindex < endpindex; i++, tpindex++) { 1408 1409 if (vm_page_lookup(object, tpindex)) { 1410 break; 1411 } 1412 1413 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1414 if (rtm == NULL) { 1415 break; 1416 } 1417 1418 marray[i] = rtm; 1419 } 1420 crit_exit(); 1421 1422 /* return number of bytes of pages */ 1423 return i; 1424 } 1425