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 1.30 2006/09/13 22:25:00 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 #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 (curproc && (curproc->p_flag & P_SWAPPEDOUT) == 0 && 390 curproc->p_stats) { 391 if (fs.hardfault) { 392 curproc->p_stats->p_ru.ru_majflt++; 393 } else { 394 curproc->p_stats->p_ru.ru_minflt++; 395 } 396 } 397 398 /* 399 * Unlock everything, and return 400 */ 401 vm_page_wakeup(fs.m); 402 vm_object_deallocate(fs.first_object); 403 404 return (KERN_SUCCESS); 405 } 406 407 /* 408 * Translate the virtual page number (first_pindex) that is relative 409 * to the address space into a logical page number that is relative to the 410 * backing object. Use the virtual page table pointed to by (vpte). 411 * 412 * This implements an N-level page table. Any level can terminate the 413 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 414 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 415 */ 416 static 417 int 418 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, vpte_t vpte) 419 { 420 struct sf_buf *sf; 421 int vshift = 32 - PAGE_SHIFT; /* page index bits remaining */ 422 int result = KERN_SUCCESS; 423 424 for (;;) { 425 if ((vpte & VPTE_V) == 0) { 426 unlock_and_deallocate(fs); 427 return (KERN_FAILURE); 428 } 429 if ((vpte & VPTE_PS) || vshift == 0) 430 break; 431 KKASSERT(vshift >= VPTE_PAGE_BITS); 432 433 /* 434 * Get the page table page 435 */ 436 result = vm_fault_object(fs, vpte >> PAGE_SHIFT, VM_PROT_READ); 437 if (result != KERN_SUCCESS) 438 return (result); 439 440 /* 441 * Process the returned fs.m and look up the page table 442 * entry in the page table page. 443 */ 444 vshift -= VPTE_PAGE_BITS; 445 sf = sf_buf_alloc(fs->m, SFB_CPUPRIVATE); 446 vpte = *((vpte_t *)sf_buf_kva(sf) + 447 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 448 sf_buf_free(sf); 449 vm_page_flag_set(fs->m, PG_REFERENCED); 450 vm_page_activate(fs->m); 451 vm_page_wakeup(fs->m); 452 cleanup_successful_fault(fs); 453 } 454 /* 455 * Combine remaining address bits with the vpte. 456 */ 457 *pindex = (vpte >> PAGE_SHIFT) + 458 (*pindex & ((1 << vshift) - 1)); 459 return (KERN_SUCCESS); 460 } 461 462 463 /* 464 * Do all operations required to fault-in (fs.first_object, pindex). Run 465 * through the shadow chain as necessary and do required COW or virtual 466 * copy operations. The caller has already fully resolved the vm_map_entry 467 * and, if appropriate, has created a copy-on-write layer. All we need to 468 * do is iterate the object chain. 469 * 470 * On failure (fs) is unlocked and deallocated and the caller may return or 471 * retry depending on the failure code. On success (fs) is NOT unlocked or 472 * deallocated, fs.m will contained a resolved, busied page, and fs.object 473 * will have an additional PIP count if it is not equal to fs.first_object. 474 */ 475 static 476 int 477 vm_fault_object(struct faultstate *fs, 478 vm_pindex_t first_pindex, vm_prot_t fault_type) 479 { 480 vm_object_t next_object; 481 vm_page_t marray[VM_FAULT_READ]; 482 vm_pindex_t pindex; 483 int faultcount; 484 485 fs->prot = fs->first_prot; 486 fs->object = fs->first_object; 487 pindex = first_pindex; 488 489 for (;;) { 490 /* 491 * If the object is dead, we stop here 492 */ 493 if (fs->object->flags & OBJ_DEAD) { 494 unlock_and_deallocate(fs); 495 return (KERN_PROTECTION_FAILURE); 496 } 497 498 /* 499 * See if page is resident. spl protection is required 500 * to avoid an interrupt unbusy/free race against our 501 * lookup. We must hold the protection through a page 502 * allocation or busy. 503 */ 504 crit_enter(); 505 fs->m = vm_page_lookup(fs->object, pindex); 506 if (fs->m != NULL) { 507 int queue; 508 /* 509 * Wait/Retry if the page is busy. We have to do this 510 * if the page is busy via either PG_BUSY or 511 * vm_page_t->busy because the vm_pager may be using 512 * vm_page_t->busy for pageouts ( and even pageins if 513 * it is the vnode pager ), and we could end up trying 514 * to pagein and pageout the same page simultaneously. 515 * 516 * We can theoretically allow the busy case on a read 517 * fault if the page is marked valid, but since such 518 * pages are typically already pmap'd, putting that 519 * special case in might be more effort then it is 520 * worth. We cannot under any circumstances mess 521 * around with a vm_page_t->busy page except, perhaps, 522 * to pmap it. 523 */ 524 if ((fs->m->flags & PG_BUSY) || fs->m->busy) { 525 unlock_things(fs); 526 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 527 mycpu->gd_cnt.v_intrans++; 528 vm_object_deallocate(fs->first_object); 529 crit_exit(); 530 return (KERN_TRY_AGAIN); 531 } 532 533 /* 534 * If reactivating a page from PQ_CACHE we may have 535 * to rate-limit. 536 */ 537 queue = fs->m->queue; 538 vm_page_unqueue_nowakeup(fs->m); 539 540 if ((queue - fs->m->pc) == PQ_CACHE && 541 vm_page_count_severe()) { 542 vm_page_activate(fs->m); 543 unlock_and_deallocate(fs); 544 vm_waitpfault(); 545 crit_exit(); 546 return (KERN_TRY_AGAIN); 547 } 548 549 /* 550 * Mark page busy for other processes, and the 551 * pagedaemon. If it still isn't completely valid 552 * (readable), jump to readrest, else we found the 553 * page and can return. 554 * 555 * We can release the spl once we have marked the 556 * page busy. 557 */ 558 vm_page_busy(fs->m); 559 crit_exit(); 560 561 if (((fs->m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 562 fs->m->object != kernel_object && 563 fs->m->object != kmem_object) { 564 goto readrest; 565 } 566 break; /* break to PAGE HAS BEEN FOUND */ 567 } 568 569 /* 570 * Page is not resident, If this is the search termination 571 * or the pager might contain the page, allocate a new page. 572 * 573 * NOTE: We are still in a critical section. 574 */ 575 if (TRYPAGER(fs) || fs->object == fs->first_object) { 576 /* 577 * If the page is beyond the object size we fail 578 */ 579 if (pindex >= fs->object->size) { 580 crit_exit(); 581 unlock_and_deallocate(fs); 582 return (KERN_PROTECTION_FAILURE); 583 } 584 585 /* 586 * Ratelimit. 587 */ 588 if (fs->didlimit == 0 && curproc != NULL) { 589 int limticks; 590 591 limticks = vm_fault_ratelimit(curproc->p_vmspace); 592 if (limticks) { 593 crit_exit(); 594 unlock_and_deallocate(fs); 595 tsleep(curproc, 0, "vmrate", limticks); 596 fs->didlimit = 1; 597 return (KERN_TRY_AGAIN); 598 } 599 } 600 601 /* 602 * Allocate a new page for this object/offset pair. 603 */ 604 fs->m = NULL; 605 if (!vm_page_count_severe()) { 606 fs->m = vm_page_alloc(fs->object, pindex, 607 (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO); 608 } 609 if (fs->m == NULL) { 610 crit_exit(); 611 unlock_and_deallocate(fs); 612 vm_waitpfault(); 613 return (KERN_TRY_AGAIN); 614 } 615 } 616 crit_exit(); 617 618 readrest: 619 /* 620 * We have found a valid page or we have allocated a new page. 621 * The page thus may not be valid or may not be entirely 622 * valid. 623 * 624 * Attempt to fault-in the page if there is a chance that the 625 * pager has it, and potentially fault in additional pages 626 * at the same time. 627 * 628 * We are NOT in splvm here and if TRYPAGER is true then 629 * fs.m will be non-NULL and will be PG_BUSY for us. 630 */ 631 632 if (TRYPAGER(fs)) { 633 int rv; 634 int reqpage; 635 int ahead, behind; 636 u_char behavior = vm_map_entry_behavior(fs->entry); 637 638 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 639 ahead = 0; 640 behind = 0; 641 } else { 642 behind = pindex; 643 if (behind > VM_FAULT_READ_BEHIND) 644 behind = VM_FAULT_READ_BEHIND; 645 646 ahead = fs->object->size - pindex; 647 if (ahead < 1) 648 ahead = 1; 649 if (ahead > VM_FAULT_READ_AHEAD) 650 ahead = VM_FAULT_READ_AHEAD; 651 } 652 653 if ((fs->first_object->type != OBJT_DEVICE) && 654 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 655 (behavior != MAP_ENTRY_BEHAV_RANDOM && 656 pindex >= fs->entry->lastr && 657 pindex < fs->entry->lastr + VM_FAULT_READ)) 658 ) { 659 vm_pindex_t firstpindex, tmppindex; 660 661 if (first_pindex < 2 * VM_FAULT_READ) 662 firstpindex = 0; 663 else 664 firstpindex = first_pindex - 2 * VM_FAULT_READ; 665 666 /* 667 * note: partially valid pages cannot be 668 * included in the lookahead - NFS piecemeal 669 * writes will barf on it badly. 670 * 671 * spl protection is required to avoid races 672 * between the lookup and an interrupt 673 * unbusy/free sequence occuring prior to 674 * our busy check. 675 */ 676 crit_enter(); 677 for (tmppindex = first_pindex - 1; 678 tmppindex >= firstpindex; 679 --tmppindex 680 ) { 681 vm_page_t mt; 682 683 mt = vm_page_lookup(fs->first_object, tmppindex); 684 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 685 break; 686 if (mt->busy || 687 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 688 mt->hold_count || 689 mt->wire_count) 690 continue; 691 if (mt->dirty == 0) 692 vm_page_test_dirty(mt); 693 if (mt->dirty) { 694 vm_page_protect(mt, VM_PROT_NONE); 695 vm_page_deactivate(mt); 696 } else { 697 vm_page_cache(mt); 698 } 699 } 700 crit_exit(); 701 702 ahead += behind; 703 behind = 0; 704 } 705 706 /* 707 * now we find out if any other pages should be paged 708 * in at this time this routine checks to see if the 709 * pages surrounding this fault reside in the same 710 * object as the page for this fault. If they do, 711 * then they are faulted in also into the object. The 712 * array "marray" returned contains an array of 713 * vm_page_t structs where one of them is the 714 * vm_page_t passed to the routine. The reqpage 715 * return value is the index into the marray for the 716 * vm_page_t passed to the routine. 717 * 718 * fs.m plus the additional pages are PG_BUSY'd. 719 */ 720 faultcount = vm_fault_additional_pages( 721 fs->m, behind, ahead, marray, &reqpage); 722 723 /* 724 * update lastr imperfectly (we do not know how much 725 * getpages will actually read), but good enough. 726 */ 727 fs->entry->lastr = pindex + faultcount - behind; 728 729 /* 730 * Call the pager to retrieve the data, if any, after 731 * releasing the lock on the map. We hold a ref on 732 * fs.object and the pages are PG_BUSY'd. 733 */ 734 unlock_map(fs); 735 736 if (faultcount) { 737 rv = vm_pager_get_pages(fs->object, marray, 738 faultcount, reqpage); 739 } else { 740 rv = VM_PAGER_FAIL; 741 } 742 743 if (rv == VM_PAGER_OK) { 744 /* 745 * Found the page. Leave it busy while we play 746 * with it. 747 */ 748 749 /* 750 * Relookup in case pager changed page. Pager 751 * is responsible for disposition of old page 752 * if moved. 753 * 754 * XXX other code segments do relookups too. 755 * It's a bad abstraction that needs to be 756 * fixed/removed. 757 */ 758 fs->m = vm_page_lookup(fs->object, pindex); 759 if (fs->m == NULL) { 760 unlock_and_deallocate(fs); 761 return (KERN_TRY_AGAIN); 762 } 763 764 ++fs->hardfault; 765 break; /* break to PAGE HAS BEEN FOUND */ 766 } 767 768 /* 769 * Remove the bogus page (which does not exist at this 770 * object/offset); before doing so, we must get back 771 * our object lock to preserve our invariant. 772 * 773 * Also wake up any other process that may want to bring 774 * in this page. 775 * 776 * If this is the top-level object, we must leave the 777 * busy page to prevent another process from rushing 778 * past us, and inserting the page in that object at 779 * the same time that we are. 780 */ 781 if (rv == VM_PAGER_ERROR) { 782 if (curproc) 783 printf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); 784 else 785 printf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm); 786 } 787 /* 788 * Data outside the range of the pager or an I/O error 789 */ 790 /* 791 * XXX - the check for kernel_map is a kludge to work 792 * around having the machine panic on a kernel space 793 * fault w/ I/O error. 794 */ 795 if (((fs->map != kernel_map) && (rv == VM_PAGER_ERROR)) || 796 (rv == VM_PAGER_BAD)) { 797 vm_page_free(fs->m); 798 fs->m = NULL; 799 unlock_and_deallocate(fs); 800 if (rv == VM_PAGER_ERROR) 801 return (KERN_FAILURE); 802 else 803 return (KERN_PROTECTION_FAILURE); 804 /* NOT REACHED */ 805 } 806 if (fs->object != fs->first_object) { 807 vm_page_free(fs->m); 808 fs->m = NULL; 809 /* 810 * XXX - we cannot just fall out at this 811 * point, m has been freed and is invalid! 812 */ 813 } 814 } 815 816 /* 817 * We get here if the object has a default pager (or unwiring) 818 * or the pager doesn't have the page. 819 */ 820 if (fs->object == fs->first_object) 821 fs->first_m = fs->m; 822 823 /* 824 * Move on to the next object. Lock the next object before 825 * unlocking the current one. 826 */ 827 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 828 next_object = fs->object->backing_object; 829 if (next_object == NULL) { 830 /* 831 * If there's no object left, fill the page in the top 832 * object with zeros. 833 */ 834 if (fs->object != fs->first_object) { 835 vm_object_pip_wakeup(fs->object); 836 837 fs->object = fs->first_object; 838 pindex = first_pindex; 839 fs->m = fs->first_m; 840 } 841 fs->first_m = NULL; 842 843 /* 844 * Zero the page if necessary and mark it valid. 845 */ 846 if ((fs->m->flags & PG_ZERO) == 0) { 847 vm_page_zero_fill(fs->m); 848 } else { 849 mycpu->gd_cnt.v_ozfod++; 850 } 851 mycpu->gd_cnt.v_zfod++; 852 fs->m->valid = VM_PAGE_BITS_ALL; 853 break; /* break to PAGE HAS BEEN FOUND */ 854 } else { 855 if (fs->object != fs->first_object) { 856 vm_object_pip_wakeup(fs->object); 857 } 858 KASSERT(fs->object != next_object, ("object loop %p", next_object)); 859 fs->object = next_object; 860 vm_object_pip_add(fs->object, 1); 861 } 862 } 863 864 KASSERT((fs->m->flags & PG_BUSY) != 0, 865 ("vm_fault: not busy after main loop")); 866 867 /* 868 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 869 * is held.] 870 */ 871 872 /* 873 * If the page is being written, but isn't already owned by the 874 * top-level object, we have to copy it into a new page owned by the 875 * top-level object. 876 */ 877 if (fs->object != fs->first_object) { 878 /* 879 * We only really need to copy if we want to write it. 880 */ 881 if (fault_type & VM_PROT_WRITE) { 882 /* 883 * This allows pages to be virtually copied from a 884 * backing_object into the first_object, where the 885 * backing object has no other refs to it, and cannot 886 * gain any more refs. Instead of a bcopy, we just 887 * move the page from the backing object to the 888 * first object. Note that we must mark the page 889 * dirty in the first object so that it will go out 890 * to swap when needed. 891 */ 892 if (fs->map_generation == fs->map->timestamp && 893 /* 894 * Only one shadow object 895 */ 896 (fs->object->shadow_count == 1) && 897 /* 898 * No COW refs, except us 899 */ 900 (fs->object->ref_count == 1) && 901 /* 902 * No one else can look this object up 903 */ 904 (fs->object->handle == NULL) && 905 /* 906 * No other ways to look the object up 907 */ 908 ((fs->object->type == OBJT_DEFAULT) || 909 (fs->object->type == OBJT_SWAP)) && 910 /* 911 * We don't chase down the shadow chain 912 */ 913 (fs->object == fs->first_object->backing_object) && 914 915 /* 916 * grab the lock if we need to 917 */ 918 (fs->lookup_still_valid || 919 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 920 ) { 921 922 fs->lookup_still_valid = 1; 923 /* 924 * get rid of the unnecessary page 925 */ 926 vm_page_protect(fs->first_m, VM_PROT_NONE); 927 vm_page_free(fs->first_m); 928 fs->first_m = NULL; 929 930 /* 931 * grab the page and put it into the 932 * process'es object. The page is 933 * automatically made dirty. 934 */ 935 vm_page_rename(fs->m, fs->first_object, first_pindex); 936 fs->first_m = fs->m; 937 vm_page_busy(fs->first_m); 938 fs->m = NULL; 939 mycpu->gd_cnt.v_cow_optim++; 940 } else { 941 /* 942 * Oh, well, lets copy it. 943 */ 944 vm_page_copy(fs->m, fs->first_m); 945 } 946 947 if (fs->m) { 948 /* 949 * We no longer need the old page or object. 950 */ 951 release_page(fs); 952 } 953 954 /* 955 * fs->object != fs->first_object due to above 956 * conditional 957 */ 958 vm_object_pip_wakeup(fs->object); 959 960 /* 961 * Only use the new page below... 962 */ 963 964 mycpu->gd_cnt.v_cow_faults++; 965 fs->m = fs->first_m; 966 fs->object = fs->first_object; 967 pindex = first_pindex; 968 } else { 969 /* 970 * If it wasn't a write fault avoid having to copy 971 * the page by mapping it read-only. 972 */ 973 fs->prot &= ~VM_PROT_WRITE; 974 } 975 } 976 977 /* 978 * We may have had to unlock a map to do I/O. If we did then 979 * lookup_still_valid will be FALSE. If the map generation count 980 * also changed then all sorts of things could have happened while 981 * we were doing the I/O and we need to retry. 982 */ 983 984 if (!fs->lookup_still_valid && 985 (fs->map->timestamp != fs->map_generation)) { 986 release_page(fs); 987 unlock_and_deallocate(fs); 988 return (KERN_TRY_AGAIN); 989 } 990 991 /* 992 * Put this page into the physical map. We had to do the unlock above 993 * because pmap_enter may cause other faults. We don't put the page 994 * back on the active queue until later so that the page-out daemon 995 * won't find us (yet). 996 */ 997 if (fs->prot & VM_PROT_WRITE) { 998 vm_page_flag_set(fs->m, PG_WRITEABLE); 999 vm_object_set_writeable_dirty(fs->m->object); 1000 1001 /* 1002 * If the fault is a write, we know that this page is being 1003 * written NOW so dirty it explicitly to save on 1004 * pmap_is_modified() calls later. 1005 * 1006 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 1007 * if the page is already dirty to prevent data written with 1008 * the expectation of being synced from not being synced. 1009 * Likewise if this entry does not request NOSYNC then make 1010 * sure the page isn't marked NOSYNC. Applications sharing 1011 * data should use the same flags to avoid ping ponging. 1012 * 1013 * Also tell the backing pager, if any, that it should remove 1014 * any swap backing since the page is now dirty. 1015 */ 1016 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) { 1017 if (fs->m->dirty == 0) 1018 vm_page_flag_set(fs->m, PG_NOSYNC); 1019 } else { 1020 vm_page_flag_clear(fs->m, PG_NOSYNC); 1021 } 1022 if (fs->fault_flags & VM_FAULT_DIRTY) { 1023 crit_enter(); 1024 vm_page_dirty(fs->m); 1025 vm_pager_page_unswapped(fs->m); 1026 crit_exit(); 1027 } 1028 } 1029 1030 /* 1031 * Page had better still be busy. We are still locked up and 1032 * fs->object will have another PIP reference if it is not equal 1033 * to fs->first_object. 1034 */ 1035 KASSERT(fs->m->flags & PG_BUSY, 1036 ("vm_fault: page %p not busy!", fs->m)); 1037 1038 /* 1039 * Sanity check: page must be completely valid or it is not fit to 1040 * map into user space. vm_pager_get_pages() ensures this. 1041 */ 1042 if (fs->m->valid != VM_PAGE_BITS_ALL) { 1043 vm_page_zero_invalid(fs->m, TRUE); 1044 printf("Warning: page %p partially invalid on fault\n", fs->m); 1045 } 1046 1047 return (KERN_SUCCESS); 1048 } 1049 1050 /* 1051 * quick version of vm_fault 1052 */ 1053 int 1054 vm_fault_quick(caddr_t v, int prot) 1055 { 1056 int r; 1057 1058 if (prot & VM_PROT_WRITE) 1059 r = subyte(v, fubyte(v)); 1060 else 1061 r = fubyte(v); 1062 return(r); 1063 } 1064 1065 /* 1066 * Wire down a range of virtual addresses in a map. The entry in question 1067 * should be marked in-transition and the map must be locked. We must 1068 * release the map temporarily while faulting-in the page to avoid a 1069 * deadlock. Note that the entry may be clipped while we are blocked but 1070 * will never be freed. 1071 */ 1072 int 1073 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 1074 { 1075 boolean_t fictitious; 1076 vm_offset_t start; 1077 vm_offset_t end; 1078 vm_offset_t va; 1079 vm_paddr_t pa; 1080 pmap_t pmap; 1081 int rv; 1082 1083 pmap = vm_map_pmap(map); 1084 start = entry->start; 1085 end = entry->end; 1086 fictitious = entry->object.vm_object && 1087 (entry->object.vm_object->type == OBJT_DEVICE); 1088 1089 vm_map_unlock(map); 1090 map->timestamp++; 1091 1092 /* 1093 * We simulate a fault to get the page and enter it in the physical 1094 * map. 1095 */ 1096 for (va = start; va < end; va += PAGE_SIZE) { 1097 if (user_wire) { 1098 rv = vm_fault(map, va, VM_PROT_READ, 1099 VM_FAULT_USER_WIRE); 1100 } else { 1101 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 1102 VM_FAULT_CHANGE_WIRING); 1103 } 1104 if (rv) { 1105 while (va > start) { 1106 va -= PAGE_SIZE; 1107 if ((pa = pmap_extract(pmap, va)) == 0) 1108 continue; 1109 pmap_change_wiring(pmap, va, FALSE); 1110 if (!fictitious) 1111 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1112 } 1113 vm_map_lock(map); 1114 return (rv); 1115 } 1116 } 1117 vm_map_lock(map); 1118 return (KERN_SUCCESS); 1119 } 1120 1121 /* 1122 * Unwire a range of virtual addresses in a map. The map should be 1123 * locked. 1124 */ 1125 void 1126 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 1127 { 1128 boolean_t fictitious; 1129 vm_offset_t start; 1130 vm_offset_t end; 1131 vm_offset_t va; 1132 vm_paddr_t pa; 1133 pmap_t pmap; 1134 1135 pmap = vm_map_pmap(map); 1136 start = entry->start; 1137 end = entry->end; 1138 fictitious = entry->object.vm_object && 1139 (entry->object.vm_object->type == OBJT_DEVICE); 1140 1141 /* 1142 * Since the pages are wired down, we must be able to get their 1143 * mappings from the physical map system. 1144 */ 1145 for (va = start; va < end; va += PAGE_SIZE) { 1146 pa = pmap_extract(pmap, va); 1147 if (pa != 0) { 1148 pmap_change_wiring(pmap, va, FALSE); 1149 if (!fictitious) 1150 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1151 } 1152 } 1153 } 1154 1155 /* 1156 * Reduce the rate at which memory is allocated to a process based 1157 * on the perceived load on the VM system. As the load increases 1158 * the allocation burst rate goes down and the delay increases. 1159 * 1160 * Rate limiting does not apply when faulting active or inactive 1161 * pages. When faulting 'cache' pages, rate limiting only applies 1162 * if the system currently has a severe page deficit. 1163 * 1164 * XXX vm_pagesupply should be increased when a page is freed. 1165 * 1166 * We sleep up to 1/10 of a second. 1167 */ 1168 static int 1169 vm_fault_ratelimit(struct vmspace *vmspace) 1170 { 1171 if (vm_load_enable == 0) 1172 return(0); 1173 if (vmspace->vm_pagesupply > 0) { 1174 --vmspace->vm_pagesupply; 1175 return(0); 1176 } 1177 #ifdef INVARIANTS 1178 if (vm_load_debug) { 1179 printf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n", 1180 vm_load, 1181 (1000 - vm_load ) / 10, vm_load * hz / 10000, 1182 curproc->p_pid, curproc->p_comm); 1183 } 1184 #endif 1185 vmspace->vm_pagesupply = (1000 - vm_load) / 10; 1186 return(vm_load * hz / 10000); 1187 } 1188 1189 /* 1190 * Routine: 1191 * vm_fault_copy_entry 1192 * Function: 1193 * Copy all of the pages from a wired-down map entry to another. 1194 * 1195 * In/out conditions: 1196 * The source and destination maps must be locked for write. 1197 * The source map entry must be wired down (or be a sharing map 1198 * entry corresponding to a main map entry that is wired down). 1199 */ 1200 1201 void 1202 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1203 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 1204 { 1205 vm_object_t dst_object; 1206 vm_object_t src_object; 1207 vm_ooffset_t dst_offset; 1208 vm_ooffset_t src_offset; 1209 vm_prot_t prot; 1210 vm_offset_t vaddr; 1211 vm_page_t dst_m; 1212 vm_page_t src_m; 1213 1214 #ifdef lint 1215 src_map++; 1216 #endif /* lint */ 1217 1218 src_object = src_entry->object.vm_object; 1219 src_offset = src_entry->offset; 1220 1221 /* 1222 * Create the top-level object for the destination entry. (Doesn't 1223 * actually shadow anything - we copy the pages directly.) 1224 */ 1225 vm_map_entry_allocate_object(dst_entry); 1226 dst_object = dst_entry->object.vm_object; 1227 1228 prot = dst_entry->max_protection; 1229 1230 /* 1231 * Loop through all of the pages in the entry's range, copying each 1232 * one from the source object (it should be there) to the destination 1233 * object. 1234 */ 1235 for (vaddr = dst_entry->start, dst_offset = 0; 1236 vaddr < dst_entry->end; 1237 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1238 1239 /* 1240 * Allocate a page in the destination object 1241 */ 1242 do { 1243 dst_m = vm_page_alloc(dst_object, 1244 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1245 if (dst_m == NULL) { 1246 vm_wait(); 1247 } 1248 } while (dst_m == NULL); 1249 1250 /* 1251 * Find the page in the source object, and copy it in. 1252 * (Because the source is wired down, the page will be in 1253 * memory.) 1254 */ 1255 src_m = vm_page_lookup(src_object, 1256 OFF_TO_IDX(dst_offset + src_offset)); 1257 if (src_m == NULL) 1258 panic("vm_fault_copy_wired: page missing"); 1259 1260 vm_page_copy(src_m, dst_m); 1261 1262 /* 1263 * Enter it in the pmap... 1264 */ 1265 1266 vm_page_flag_clear(dst_m, PG_ZERO); 1267 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1268 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED); 1269 1270 /* 1271 * Mark it no longer busy, and put it on the active list. 1272 */ 1273 vm_page_activate(dst_m); 1274 vm_page_wakeup(dst_m); 1275 } 1276 } 1277 1278 1279 /* 1280 * This routine checks around the requested page for other pages that 1281 * might be able to be faulted in. This routine brackets the viable 1282 * pages for the pages to be paged in. 1283 * 1284 * Inputs: 1285 * m, rbehind, rahead 1286 * 1287 * Outputs: 1288 * marray (array of vm_page_t), reqpage (index of requested page) 1289 * 1290 * Return value: 1291 * number of pages in marray 1292 */ 1293 static int 1294 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 1295 vm_page_t *marray, int *reqpage) 1296 { 1297 int i,j; 1298 vm_object_t object; 1299 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1300 vm_page_t rtm; 1301 int cbehind, cahead; 1302 1303 object = m->object; 1304 pindex = m->pindex; 1305 1306 /* 1307 * we don't fault-ahead for device pager 1308 */ 1309 if (object->type == OBJT_DEVICE) { 1310 *reqpage = 0; 1311 marray[0] = m; 1312 return 1; 1313 } 1314 1315 /* 1316 * if the requested page is not available, then give up now 1317 */ 1318 1319 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1320 return 0; 1321 } 1322 1323 if ((cbehind == 0) && (cahead == 0)) { 1324 *reqpage = 0; 1325 marray[0] = m; 1326 return 1; 1327 } 1328 1329 if (rahead > cahead) { 1330 rahead = cahead; 1331 } 1332 1333 if (rbehind > cbehind) { 1334 rbehind = cbehind; 1335 } 1336 1337 /* 1338 * try to do any readahead that we might have free pages for. 1339 */ 1340 if ((rahead + rbehind) > 1341 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) { 1342 pagedaemon_wakeup(); 1343 marray[0] = m; 1344 *reqpage = 0; 1345 return 1; 1346 } 1347 1348 /* 1349 * scan backward for the read behind pages -- in memory 1350 * 1351 * Assume that if the page is not found an interrupt will not 1352 * create it. Theoretically interrupts can only remove (busy) 1353 * pages, not create new associations. 1354 */ 1355 if (pindex > 0) { 1356 if (rbehind > pindex) { 1357 rbehind = pindex; 1358 startpindex = 0; 1359 } else { 1360 startpindex = pindex - rbehind; 1361 } 1362 1363 crit_enter(); 1364 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1365 if (vm_page_lookup( object, tpindex)) { 1366 startpindex = tpindex + 1; 1367 break; 1368 } 1369 if (tpindex == 0) 1370 break; 1371 } 1372 1373 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1374 1375 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1376 if (rtm == NULL) { 1377 crit_exit(); 1378 for (j = 0; j < i; j++) { 1379 vm_page_free(marray[j]); 1380 } 1381 marray[0] = m; 1382 *reqpage = 0; 1383 return 1; 1384 } 1385 1386 marray[i] = rtm; 1387 } 1388 crit_exit(); 1389 } else { 1390 startpindex = 0; 1391 i = 0; 1392 } 1393 1394 marray[i] = m; 1395 /* page offset of the required page */ 1396 *reqpage = i; 1397 1398 tpindex = pindex + 1; 1399 i++; 1400 1401 /* 1402 * scan forward for the read ahead pages 1403 */ 1404 endpindex = tpindex + rahead; 1405 if (endpindex > object->size) 1406 endpindex = object->size; 1407 1408 crit_enter(); 1409 for( ; tpindex < endpindex; i++, tpindex++) { 1410 1411 if (vm_page_lookup(object, tpindex)) { 1412 break; 1413 } 1414 1415 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1416 if (rtm == NULL) { 1417 break; 1418 } 1419 1420 marray[i] = rtm; 1421 } 1422 crit_exit(); 1423 1424 /* return number of bytes of pages */ 1425 return i; 1426 } 1427