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 1.47 2008/07/01 02:02:56 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 #include <sys/sysctl.h> 88 89 #include <vm/vm.h> 90 #include <vm/vm_param.h> 91 #include <vm/pmap.h> 92 #include <vm/vm_map.h> 93 #include <vm/vm_object.h> 94 #include <vm/vm_page.h> 95 #include <vm/vm_pageout.h> 96 #include <vm/vm_kern.h> 97 #include <vm/vm_pager.h> 98 #include <vm/vnode_pager.h> 99 #include <vm/vm_extern.h> 100 101 #include <sys/thread2.h> 102 #include <vm/vm_page2.h> 103 104 struct faultstate { 105 vm_page_t m; 106 vm_object_t object; 107 vm_pindex_t pindex; 108 vm_prot_t prot; 109 vm_page_t first_m; 110 vm_object_t first_object; 111 vm_prot_t first_prot; 112 vm_map_t map; 113 vm_map_entry_t entry; 114 int lookup_still_valid; 115 int didlimit; 116 int hardfault; 117 int fault_flags; 118 int map_generation; 119 boolean_t wired; 120 struct vnode *vp; 121 }; 122 123 static int vm_fast_fault = 1; 124 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, ""); 125 static int debug_cluster = 0; 126 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, ""); 127 128 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t); 129 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int); 130 #if 0 131 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 132 #endif 133 static int vm_fault_ratelimit(struct vmspace *); 134 static void vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, 135 int prot); 136 137 static __inline void 138 release_page(struct faultstate *fs) 139 { 140 vm_page_deactivate(fs->m); 141 vm_page_wakeup(fs->m); 142 fs->m = NULL; 143 } 144 145 static __inline void 146 unlock_map(struct faultstate *fs) 147 { 148 if (fs->lookup_still_valid && fs->map) { 149 vm_map_lookup_done(fs->map, fs->entry, 0); 150 fs->lookup_still_valid = FALSE; 151 } 152 } 153 154 /* 155 * Clean up after a successful call to vm_fault_object() so another call 156 * to vm_fault_object() can be made. 157 */ 158 static void 159 _cleanup_successful_fault(struct faultstate *fs, int relock) 160 { 161 if (fs->object != fs->first_object) { 162 vm_page_free(fs->first_m); 163 vm_object_pip_wakeup(fs->object); 164 fs->first_m = NULL; 165 } 166 fs->object = fs->first_object; 167 if (relock && fs->lookup_still_valid == FALSE) { 168 if (fs->map) 169 vm_map_lock_read(fs->map); 170 fs->lookup_still_valid = TRUE; 171 } 172 } 173 174 static void 175 _unlock_things(struct faultstate *fs, int dealloc) 176 { 177 vm_object_pip_wakeup(fs->first_object); 178 _cleanup_successful_fault(fs, 0); 179 if (dealloc) { 180 vm_object_deallocate(fs->first_object); 181 fs->first_object = NULL; 182 } 183 unlock_map(fs); 184 if (fs->vp != NULL) { 185 vput(fs->vp); 186 fs->vp = NULL; 187 } 188 } 189 190 #define unlock_things(fs) _unlock_things(fs, 0) 191 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 192 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 193 194 /* 195 * TRYPAGER 196 * 197 * Determine if the pager for the current object *might* contain the page. 198 * 199 * We only need to try the pager if this is not a default object (default 200 * objects are zero-fill and have no real pager), and if we are not taking 201 * a wiring fault or if the FS entry is wired. 202 */ 203 #define TRYPAGER(fs) \ 204 (fs->object->type != OBJT_DEFAULT && \ 205 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired)) 206 207 /* 208 * vm_fault: 209 * 210 * Handle a page fault occuring at the given address, requiring the given 211 * permissions, in the map specified. If successful, the page is inserted 212 * into the associated physical map. 213 * 214 * NOTE: The given address should be truncated to the proper page address. 215 * 216 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 217 * a standard error specifying why the fault is fatal is returned. 218 * 219 * The map in question must be referenced, and remains so. 220 * The caller may hold no locks. 221 */ 222 int 223 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 224 { 225 int result; 226 vm_pindex_t first_pindex; 227 struct faultstate fs; 228 229 mycpu->gd_cnt.v_vm_faults++; 230 231 fs.didlimit = 0; 232 fs.hardfault = 0; 233 fs.fault_flags = fault_flags; 234 235 RetryFault: 236 /* 237 * Find the vm_map_entry representing the backing store and resolve 238 * the top level object and page index. This may have the side 239 * effect of executing a copy-on-write on the map entry and/or 240 * creating a shadow object, but will not COW any actual VM pages. 241 * 242 * On success fs.map is left read-locked and various other fields 243 * are initialized but not otherwise referenced or locked. 244 * 245 * NOTE! vm_map_lookup will try to upgrade the fault_type to 246 * VM_FAULT_WRITE if the map entry is a virtual page table and also 247 * writable, so we can set the 'A'accessed bit in the virtual page 248 * table entry. 249 */ 250 fs.map = map; 251 result = vm_map_lookup(&fs.map, vaddr, fault_type, 252 &fs.entry, &fs.first_object, 253 &first_pindex, &fs.first_prot, &fs.wired); 254 255 /* 256 * If the lookup failed or the map protections are incompatible, 257 * the fault generally fails. However, if the caller is trying 258 * to do a user wiring we have more work to do. 259 */ 260 if (result != KERN_SUCCESS) { 261 if (result != KERN_PROTECTION_FAILURE) 262 return result; 263 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 264 return result; 265 266 /* 267 * If we are user-wiring a r/w segment, and it is COW, then 268 * we need to do the COW operation. Note that we don't 269 * currently COW RO sections now, because it is NOT desirable 270 * to COW .text. We simply keep .text from ever being COW'ed 271 * and take the heat that one cannot debug wired .text sections. 272 */ 273 result = vm_map_lookup(&fs.map, vaddr, 274 VM_PROT_READ|VM_PROT_WRITE| 275 VM_PROT_OVERRIDE_WRITE, 276 &fs.entry, &fs.first_object, 277 &first_pindex, &fs.first_prot, 278 &fs.wired); 279 if (result != KERN_SUCCESS) 280 return result; 281 282 /* 283 * If we don't COW now, on a user wire, the user will never 284 * be able to write to the mapping. If we don't make this 285 * restriction, the bookkeeping would be nearly impossible. 286 */ 287 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 288 fs.entry->max_protection &= ~VM_PROT_WRITE; 289 } 290 291 /* 292 * fs.map is read-locked 293 * 294 * Misc checks. Save the map generation number to detect races. 295 */ 296 fs.map_generation = fs.map->timestamp; 297 298 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 299 panic("vm_fault: fault on nofault entry, addr: %lx", 300 (u_long)vaddr); 301 } 302 303 /* 304 * A system map entry may return a NULL object. No object means 305 * no pager means an unrecoverable kernel fault. 306 */ 307 if (fs.first_object == NULL) { 308 panic("vm_fault: unrecoverable fault at %p in entry %p", 309 (void *)vaddr, fs.entry); 310 } 311 312 /* 313 * Make a reference to this object to prevent its disposal while we 314 * are messing with it. Once we have the reference, the map is free 315 * to be diddled. Since objects reference their shadows (and copies), 316 * they will stay around as well. 317 * 318 * Bump the paging-in-progress count to prevent size changes (e.g. 319 * truncation operations) during I/O. This must be done after 320 * obtaining the vnode lock in order to avoid possible deadlocks. 321 */ 322 vm_object_reference(fs.first_object); 323 fs.vp = vnode_pager_lock(fs.first_object); 324 vm_object_pip_add(fs.first_object, 1); 325 326 fs.lookup_still_valid = TRUE; 327 fs.first_m = NULL; 328 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 329 330 /* 331 * If the entry is wired we cannot change the page protection. 332 */ 333 if (fs.wired) 334 fault_type = fs.first_prot; 335 336 /* 337 * The page we want is at (first_object, first_pindex), but if the 338 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 339 * page table to figure out the actual pindex. 340 * 341 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 342 * ONLY 343 */ 344 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 345 result = vm_fault_vpagetable(&fs, &first_pindex, 346 fs.entry->aux.master_pde, 347 fault_type); 348 if (result == KERN_TRY_AGAIN) 349 goto RetryFault; 350 if (result != KERN_SUCCESS) 351 return (result); 352 } 353 354 /* 355 * Now we have the actual (object, pindex), fault in the page. If 356 * vm_fault_object() fails it will unlock and deallocate the FS 357 * data. If it succeeds everything remains locked and fs->object 358 * will have an additinal PIP count if it is not equal to 359 * fs->first_object 360 * 361 * vm_fault_object will set fs->prot for the pmap operation. It is 362 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 363 * page can be safely written. However, it will force a read-only 364 * mapping for a read fault if the memory is managed by a virtual 365 * page table. 366 */ 367 result = vm_fault_object(&fs, first_pindex, fault_type); 368 369 if (result == KERN_TRY_AGAIN) 370 goto RetryFault; 371 if (result != KERN_SUCCESS) 372 return (result); 373 374 /* 375 * On success vm_fault_object() does not unlock or deallocate, and fs.m 376 * will contain a busied page. 377 * 378 * Enter the page into the pmap and do pmap-related adjustments. 379 */ 380 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired); 381 382 /* 383 * Burst in a few more pages if possible. The fs.map should still 384 * be locked. 385 */ 386 if (fault_flags & VM_FAULT_BURST) { 387 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 && 388 fs.wired == 0) { 389 vm_prefault(fs.map->pmap, vaddr, fs.entry, fs.prot); 390 } 391 } 392 unlock_things(&fs); 393 394 vm_page_flag_clear(fs.m, PG_ZERO); 395 vm_page_flag_set(fs.m, PG_REFERENCED); 396 397 /* 398 * If the page is not wired down, then put it where the pageout daemon 399 * can find it. 400 */ 401 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 402 if (fs.wired) 403 vm_page_wire(fs.m); 404 else 405 vm_page_unwire(fs.m, 1); 406 } else { 407 vm_page_activate(fs.m); 408 } 409 410 if (curthread->td_lwp) { 411 if (fs.hardfault) { 412 curthread->td_lwp->lwp_ru.ru_majflt++; 413 } else { 414 curthread->td_lwp->lwp_ru.ru_minflt++; 415 } 416 } 417 418 /* 419 * Unlock everything, and return 420 */ 421 vm_page_wakeup(fs.m); 422 vm_object_deallocate(fs.first_object); 423 424 return (KERN_SUCCESS); 425 } 426 427 /* 428 * Fault in the specified virtual address in the current process map, 429 * returning a held VM page or NULL. See vm_fault_page() for more 430 * information. 431 */ 432 vm_page_t 433 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp) 434 { 435 struct lwp *lp = curthread->td_lwp; 436 vm_page_t m; 437 438 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 439 fault_type, VM_FAULT_NORMAL, errorp); 440 return(m); 441 } 442 443 /* 444 * Fault in the specified virtual address in the specified map, doing all 445 * necessary manipulation of the object store and all necessary I/O. Return 446 * a held VM page or NULL, and set *errorp. The related pmap is not 447 * updated. 448 * 449 * The returned page will be properly dirtied if VM_PROT_WRITE was specified, 450 * and marked PG_REFERENCED as well. 451 * 452 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 453 * error will be returned. 454 */ 455 vm_page_t 456 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 457 int fault_flags, int *errorp) 458 { 459 vm_pindex_t first_pindex; 460 struct faultstate fs; 461 int result; 462 vm_prot_t orig_fault_type = fault_type; 463 464 mycpu->gd_cnt.v_vm_faults++; 465 466 fs.didlimit = 0; 467 fs.hardfault = 0; 468 fs.fault_flags = fault_flags; 469 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 470 471 RetryFault: 472 /* 473 * Find the vm_map_entry representing the backing store and resolve 474 * the top level object and page index. This may have the side 475 * effect of executing a copy-on-write on the map entry and/or 476 * creating a shadow object, but will not COW any actual VM pages. 477 * 478 * On success fs.map is left read-locked and various other fields 479 * are initialized but not otherwise referenced or locked. 480 * 481 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 482 * if the map entry is a virtual page table and also writable, 483 * so we can set the 'A'accessed bit in the virtual page table entry. 484 */ 485 fs.map = map; 486 result = vm_map_lookup(&fs.map, vaddr, fault_type, 487 &fs.entry, &fs.first_object, 488 &first_pindex, &fs.first_prot, &fs.wired); 489 490 if (result != KERN_SUCCESS) { 491 *errorp = result; 492 return (NULL); 493 } 494 495 /* 496 * fs.map is read-locked 497 * 498 * Misc checks. Save the map generation number to detect races. 499 */ 500 fs.map_generation = fs.map->timestamp; 501 502 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 503 panic("vm_fault: fault on nofault entry, addr: %lx", 504 (u_long)vaddr); 505 } 506 507 /* 508 * A system map entry may return a NULL object. No object means 509 * no pager means an unrecoverable kernel fault. 510 */ 511 if (fs.first_object == NULL) { 512 panic("vm_fault: unrecoverable fault at %p in entry %p", 513 (void *)vaddr, fs.entry); 514 } 515 516 /* 517 * Make a reference to this object to prevent its disposal while we 518 * are messing with it. Once we have the reference, the map is free 519 * to be diddled. Since objects reference their shadows (and copies), 520 * they will stay around as well. 521 * 522 * Bump the paging-in-progress count to prevent size changes (e.g. 523 * truncation operations) during I/O. This must be done after 524 * obtaining the vnode lock in order to avoid possible deadlocks. 525 */ 526 vm_object_reference(fs.first_object); 527 fs.vp = vnode_pager_lock(fs.first_object); 528 vm_object_pip_add(fs.first_object, 1); 529 530 fs.lookup_still_valid = TRUE; 531 fs.first_m = NULL; 532 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 533 534 /* 535 * If the entry is wired we cannot change the page protection. 536 */ 537 if (fs.wired) 538 fault_type = fs.first_prot; 539 540 /* 541 * The page we want is at (first_object, first_pindex), but if the 542 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 543 * page table to figure out the actual pindex. 544 * 545 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 546 * ONLY 547 */ 548 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 549 result = vm_fault_vpagetable(&fs, &first_pindex, 550 fs.entry->aux.master_pde, 551 fault_type); 552 if (result == KERN_TRY_AGAIN) 553 goto RetryFault; 554 if (result != KERN_SUCCESS) { 555 *errorp = result; 556 return (NULL); 557 } 558 } 559 560 /* 561 * Now we have the actual (object, pindex), fault in the page. If 562 * vm_fault_object() fails it will unlock and deallocate the FS 563 * data. If it succeeds everything remains locked and fs->object 564 * will have an additinal PIP count if it is not equal to 565 * fs->first_object 566 */ 567 result = vm_fault_object(&fs, first_pindex, fault_type); 568 569 if (result == KERN_TRY_AGAIN) 570 goto RetryFault; 571 if (result != KERN_SUCCESS) { 572 *errorp = result; 573 return(NULL); 574 } 575 576 if ((orig_fault_type & VM_PROT_WRITE) && 577 (fs.prot & VM_PROT_WRITE) == 0) { 578 *errorp = KERN_PROTECTION_FAILURE; 579 unlock_and_deallocate(&fs); 580 return(NULL); 581 } 582 583 /* 584 * On success vm_fault_object() does not unlock or deallocate, and fs.m 585 * will contain a busied page. 586 */ 587 unlock_things(&fs); 588 589 /* 590 * Return a held page. We are not doing any pmap manipulation so do 591 * not set PG_MAPPED. However, adjust the page flags according to 592 * the fault type because the caller may not use a managed pmapping 593 * (so we don't want to lose the fact that the page will be dirtied 594 * if a write fault was specified). 595 */ 596 vm_page_hold(fs.m); 597 vm_page_flag_clear(fs.m, PG_ZERO); 598 if (fault_type & VM_PROT_WRITE) 599 vm_page_dirty(fs.m); 600 601 /* 602 * Update the pmap. We really only have to do this if a COW 603 * occured to replace the read-only page with the new page. For 604 * now just do it unconditionally. XXX 605 */ 606 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired); 607 vm_page_flag_set(fs.m, PG_REFERENCED); 608 609 /* 610 * Unbusy the page by activating it. It remains held and will not 611 * be reclaimed. 612 */ 613 vm_page_activate(fs.m); 614 615 if (curthread->td_lwp) { 616 if (fs.hardfault) { 617 curthread->td_lwp->lwp_ru.ru_majflt++; 618 } else { 619 curthread->td_lwp->lwp_ru.ru_minflt++; 620 } 621 } 622 623 /* 624 * Unlock everything, and return the held page. 625 */ 626 vm_page_wakeup(fs.m); 627 vm_object_deallocate(fs.first_object); 628 629 *errorp = 0; 630 return(fs.m); 631 } 632 633 /* 634 * Fault in the specified (object,offset), dirty the returned page as 635 * needed. If the requested fault_type cannot be done NULL and an 636 * error is returned. 637 */ 638 vm_page_t 639 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 640 vm_prot_t fault_type, int fault_flags, int *errorp) 641 { 642 int result; 643 vm_pindex_t first_pindex; 644 struct faultstate fs; 645 struct vm_map_entry entry; 646 647 bzero(&entry, sizeof(entry)); 648 entry.object.vm_object = object; 649 entry.maptype = VM_MAPTYPE_NORMAL; 650 entry.protection = entry.max_protection = fault_type; 651 652 fs.didlimit = 0; 653 fs.hardfault = 0; 654 fs.fault_flags = fault_flags; 655 fs.map = NULL; 656 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 657 658 RetryFault: 659 660 fs.first_object = object; 661 first_pindex = OFF_TO_IDX(offset); 662 fs.entry = &entry; 663 fs.first_prot = fault_type; 664 fs.wired = 0; 665 /*fs.map_generation = 0; unused */ 666 667 /* 668 * Make a reference to this object to prevent its disposal while we 669 * are messing with it. Once we have the reference, the map is free 670 * to be diddled. Since objects reference their shadows (and copies), 671 * they will stay around as well. 672 * 673 * Bump the paging-in-progress count to prevent size changes (e.g. 674 * truncation operations) during I/O. This must be done after 675 * obtaining the vnode lock in order to avoid possible deadlocks. 676 */ 677 vm_object_reference(fs.first_object); 678 fs.vp = vnode_pager_lock(fs.first_object); 679 vm_object_pip_add(fs.first_object, 1); 680 681 fs.lookup_still_valid = TRUE; 682 fs.first_m = NULL; 683 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 684 685 #if 0 686 /* XXX future - ability to operate on VM object using vpagetable */ 687 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 688 result = vm_fault_vpagetable(&fs, &first_pindex, 689 fs.entry->aux.master_pde, 690 fault_type); 691 if (result == KERN_TRY_AGAIN) 692 goto RetryFault; 693 if (result != KERN_SUCCESS) { 694 *errorp = result; 695 return (NULL); 696 } 697 } 698 #endif 699 700 /* 701 * Now we have the actual (object, pindex), fault in the page. If 702 * vm_fault_object() fails it will unlock and deallocate the FS 703 * data. If it succeeds everything remains locked and fs->object 704 * will have an additinal PIP count if it is not equal to 705 * fs->first_object 706 */ 707 result = vm_fault_object(&fs, first_pindex, fault_type); 708 709 if (result == KERN_TRY_AGAIN) 710 goto RetryFault; 711 if (result != KERN_SUCCESS) { 712 *errorp = result; 713 return(NULL); 714 } 715 716 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 717 *errorp = KERN_PROTECTION_FAILURE; 718 unlock_and_deallocate(&fs); 719 return(NULL); 720 } 721 722 /* 723 * On success vm_fault_object() does not unlock or deallocate, and fs.m 724 * will contain a busied page. 725 */ 726 unlock_things(&fs); 727 728 /* 729 * Return a held page. We are not doing any pmap manipulation so do 730 * not set PG_MAPPED. However, adjust the page flags according to 731 * the fault type because the caller may not use a managed pmapping 732 * (so we don't want to lose the fact that the page will be dirtied 733 * if a write fault was specified). 734 */ 735 vm_page_hold(fs.m); 736 vm_page_flag_clear(fs.m, PG_ZERO); 737 if (fault_type & VM_PROT_WRITE) 738 vm_page_dirty(fs.m); 739 740 /* 741 * Indicate that the page was accessed. 742 */ 743 vm_page_flag_set(fs.m, PG_REFERENCED); 744 745 /* 746 * Unbusy the page by activating it. It remains held and will not 747 * be reclaimed. 748 */ 749 vm_page_activate(fs.m); 750 751 if (curthread->td_lwp) { 752 if (fs.hardfault) { 753 mycpu->gd_cnt.v_vm_faults++; 754 curthread->td_lwp->lwp_ru.ru_majflt++; 755 } else { 756 curthread->td_lwp->lwp_ru.ru_minflt++; 757 } 758 } 759 760 /* 761 * Unlock everything, and return the held page. 762 */ 763 vm_page_wakeup(fs.m); 764 vm_object_deallocate(fs.first_object); 765 766 *errorp = 0; 767 return(fs.m); 768 } 769 770 /* 771 * Translate the virtual page number (first_pindex) that is relative 772 * to the address space into a logical page number that is relative to the 773 * backing object. Use the virtual page table pointed to by (vpte). 774 * 775 * This implements an N-level page table. Any level can terminate the 776 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 777 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 778 */ 779 static 780 int 781 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 782 vpte_t vpte, int fault_type) 783 { 784 struct sf_buf *sf; 785 int vshift = 32 - PAGE_SHIFT; /* page index bits remaining */ 786 int result = KERN_SUCCESS; 787 vpte_t *ptep; 788 789 for (;;) { 790 /* 791 * We cannot proceed if the vpte is not valid, not readable 792 * for a read fault, or not writable for a write fault. 793 */ 794 if ((vpte & VPTE_V) == 0) { 795 unlock_and_deallocate(fs); 796 return (KERN_FAILURE); 797 } 798 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) { 799 unlock_and_deallocate(fs); 800 return (KERN_FAILURE); 801 } 802 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) { 803 unlock_and_deallocate(fs); 804 return (KERN_FAILURE); 805 } 806 if ((vpte & VPTE_PS) || vshift == 0) 807 break; 808 KKASSERT(vshift >= VPTE_PAGE_BITS); 809 810 /* 811 * Get the page table page. Nominally we only read the page 812 * table, but since we are actively setting VPTE_M and VPTE_A, 813 * tell vm_fault_object() that we are writing it. 814 * 815 * There is currently no real need to optimize this. 816 */ 817 result = vm_fault_object(fs, vpte >> PAGE_SHIFT, 818 VM_PROT_READ|VM_PROT_WRITE); 819 if (result != KERN_SUCCESS) 820 return (result); 821 822 /* 823 * Process the returned fs.m and look up the page table 824 * entry in the page table page. 825 */ 826 vshift -= VPTE_PAGE_BITS; 827 sf = sf_buf_alloc(fs->m, SFB_CPUPRIVATE); 828 ptep = ((vpte_t *)sf_buf_kva(sf) + 829 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 830 vpte = *ptep; 831 832 /* 833 * Page table write-back. If the vpte is valid for the 834 * requested operation, do a write-back to the page table. 835 * 836 * XXX VPTE_M is not set properly for page directory pages. 837 * It doesn't get set in the page directory if the page table 838 * is modified during a read access. 839 */ 840 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) && 841 (vpte & VPTE_W)) { 842 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) { 843 atomic_set_int(ptep, VPTE_M|VPTE_A); 844 vm_page_dirty(fs->m); 845 } 846 } 847 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) && 848 (vpte & VPTE_R)) { 849 if ((vpte & VPTE_A) == 0) { 850 atomic_set_int(ptep, VPTE_A); 851 vm_page_dirty(fs->m); 852 } 853 } 854 sf_buf_free(sf); 855 vm_page_flag_set(fs->m, PG_REFERENCED); 856 vm_page_activate(fs->m); 857 vm_page_wakeup(fs->m); 858 cleanup_successful_fault(fs); 859 } 860 /* 861 * Combine remaining address bits with the vpte. 862 */ 863 *pindex = (vpte >> PAGE_SHIFT) + 864 (*pindex & ((1 << vshift) - 1)); 865 return (KERN_SUCCESS); 866 } 867 868 869 /* 870 * Do all operations required to fault-in (fs.first_object, pindex). Run 871 * through the shadow chain as necessary and do required COW or virtual 872 * copy operations. The caller has already fully resolved the vm_map_entry 873 * and, if appropriate, has created a copy-on-write layer. All we need to 874 * do is iterate the object chain. 875 * 876 * On failure (fs) is unlocked and deallocated and the caller may return or 877 * retry depending on the failure code. On success (fs) is NOT unlocked or 878 * deallocated, fs.m will contained a resolved, busied page, and fs.object 879 * will have an additional PIP count if it is not equal to fs.first_object. 880 */ 881 static 882 int 883 vm_fault_object(struct faultstate *fs, 884 vm_pindex_t first_pindex, vm_prot_t fault_type) 885 { 886 vm_object_t next_object; 887 vm_pindex_t pindex; 888 889 fs->prot = fs->first_prot; 890 fs->object = fs->first_object; 891 pindex = first_pindex; 892 893 /* 894 * If a read fault occurs we try to make the page writable if 895 * possible. There are three cases where we cannot make the 896 * page mapping writable: 897 * 898 * (1) The mapping is read-only or the VM object is read-only, 899 * fs->prot above will simply not have VM_PROT_WRITE set. 900 * 901 * (2) If the mapping is a virtual page table we need to be able 902 * to detect writes so we can set VPTE_M in the virtual page 903 * table. 904 * 905 * (3) If the VM page is read-only or copy-on-write, upgrading would 906 * just result in an unnecessary COW fault. 907 * 908 * VM_PROT_VPAGED is set if faulting via a virtual page table and 909 * causes adjustments to the 'M'odify bit to also turn off write 910 * access to force a re-fault. 911 */ 912 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 913 if ((fault_type & VM_PROT_WRITE) == 0) 914 fs->prot &= ~VM_PROT_WRITE; 915 } 916 917 for (;;) { 918 /* 919 * If the object is dead, we stop here 920 */ 921 if (fs->object->flags & OBJ_DEAD) { 922 unlock_and_deallocate(fs); 923 return (KERN_PROTECTION_FAILURE); 924 } 925 926 /* 927 * See if page is resident. spl protection is required 928 * to avoid an interrupt unbusy/free race against our 929 * lookup. We must hold the protection through a page 930 * allocation or busy. 931 */ 932 crit_enter(); 933 fs->m = vm_page_lookup(fs->object, pindex); 934 if (fs->m != NULL) { 935 int queue; 936 /* 937 * Wait/Retry if the page is busy. We have to do this 938 * if the page is busy via either PG_BUSY or 939 * vm_page_t->busy because the vm_pager may be using 940 * vm_page_t->busy for pageouts ( and even pageins if 941 * it is the vnode pager ), and we could end up trying 942 * to pagein and pageout the same page simultaneously. 943 * 944 * We can theoretically allow the busy case on a read 945 * fault if the page is marked valid, but since such 946 * pages are typically already pmap'd, putting that 947 * special case in might be more effort then it is 948 * worth. We cannot under any circumstances mess 949 * around with a vm_page_t->busy page except, perhaps, 950 * to pmap it. 951 */ 952 if ((fs->m->flags & PG_BUSY) || fs->m->busy) { 953 unlock_things(fs); 954 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 955 mycpu->gd_cnt.v_intrans++; 956 vm_object_deallocate(fs->first_object); 957 fs->first_object = NULL; 958 crit_exit(); 959 return (KERN_TRY_AGAIN); 960 } 961 962 /* 963 * If reactivating a page from PQ_CACHE we may have 964 * to rate-limit. 965 */ 966 queue = fs->m->queue; 967 vm_page_unqueue_nowakeup(fs->m); 968 969 if ((queue - fs->m->pc) == PQ_CACHE && 970 vm_page_count_severe()) { 971 vm_page_activate(fs->m); 972 unlock_and_deallocate(fs); 973 vm_waitpfault(); 974 crit_exit(); 975 return (KERN_TRY_AGAIN); 976 } 977 978 /* 979 * Mark page busy for other processes, and the 980 * pagedaemon. If it still isn't completely valid 981 * (readable), or if a read-ahead-mark is set on 982 * the VM page, jump to readrest, else we found the 983 * page and can return. 984 * 985 * We can release the spl once we have marked the 986 * page busy. 987 */ 988 vm_page_busy(fs->m); 989 crit_exit(); 990 991 if (fs->m->object != &kernel_object) { 992 if ((fs->m->valid & VM_PAGE_BITS_ALL) != 993 VM_PAGE_BITS_ALL) { 994 goto readrest; 995 } 996 if (fs->m->flags & PG_RAM) { 997 if (debug_cluster) 998 kprintf("R"); 999 vm_page_flag_clear(fs->m, PG_RAM); 1000 goto readrest; 1001 } 1002 } 1003 break; /* break to PAGE HAS BEEN FOUND */ 1004 } 1005 1006 /* 1007 * Page is not resident, If this is the search termination 1008 * or the pager might contain the page, allocate a new page. 1009 * 1010 * NOTE: We are still in a critical section. 1011 */ 1012 if (TRYPAGER(fs) || fs->object == fs->first_object) { 1013 /* 1014 * If the page is beyond the object size we fail 1015 */ 1016 if (pindex >= fs->object->size) { 1017 crit_exit(); 1018 unlock_and_deallocate(fs); 1019 return (KERN_PROTECTION_FAILURE); 1020 } 1021 1022 /* 1023 * Ratelimit. 1024 */ 1025 if (fs->didlimit == 0 && curproc != NULL) { 1026 int limticks; 1027 1028 limticks = vm_fault_ratelimit(curproc->p_vmspace); 1029 if (limticks) { 1030 crit_exit(); 1031 unlock_and_deallocate(fs); 1032 tsleep(curproc, 0, "vmrate", limticks); 1033 fs->didlimit = 1; 1034 return (KERN_TRY_AGAIN); 1035 } 1036 } 1037 1038 /* 1039 * Allocate a new page for this object/offset pair. 1040 */ 1041 fs->m = NULL; 1042 if (!vm_page_count_severe()) { 1043 fs->m = vm_page_alloc(fs->object, pindex, 1044 (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO); 1045 } 1046 if (fs->m == NULL) { 1047 crit_exit(); 1048 unlock_and_deallocate(fs); 1049 vm_waitpfault(); 1050 return (KERN_TRY_AGAIN); 1051 } 1052 } 1053 crit_exit(); 1054 1055 readrest: 1056 /* 1057 * We have found an invalid or partially valid page, a 1058 * page with a read-ahead mark which might be partially or 1059 * fully valid (and maybe dirty too), or we have allocated 1060 * a new page. 1061 * 1062 * Attempt to fault-in the page if there is a chance that the 1063 * pager has it, and potentially fault in additional pages 1064 * at the same time. 1065 * 1066 * We are NOT in splvm here and if TRYPAGER is true then 1067 * fs.m will be non-NULL and will be PG_BUSY for us. 1068 */ 1069 if (TRYPAGER(fs)) { 1070 int rv; 1071 int seqaccess; 1072 u_char behavior = vm_map_entry_behavior(fs->entry); 1073 1074 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 1075 seqaccess = 0; 1076 else 1077 seqaccess = -1; 1078 1079 /* 1080 * If sequential access is detected then attempt 1081 * to deactivate/cache pages behind the scan to 1082 * prevent resource hogging. 1083 * 1084 * Use of PG_RAM to detect sequential access 1085 * also simulates multi-zone sequential access 1086 * detection for free. 1087 * 1088 * NOTE: Partially valid dirty pages cannot be 1089 * deactivated without causing NFS picemeal 1090 * writes to barf. 1091 */ 1092 if ((fs->first_object->type != OBJT_DEVICE) && 1093 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 1094 (behavior != MAP_ENTRY_BEHAV_RANDOM && 1095 (fs->m->flags & PG_RAM))) 1096 ) { 1097 vm_pindex_t scan_pindex; 1098 int scan_count = 16; 1099 1100 if (first_pindex < 16) { 1101 scan_pindex = 0; 1102 scan_count = 0; 1103 } else { 1104 scan_pindex = first_pindex - 16; 1105 if (scan_pindex < 16) 1106 scan_count = scan_pindex; 1107 else 1108 scan_count = 16; 1109 } 1110 1111 crit_enter(); 1112 while (scan_count) { 1113 vm_page_t mt; 1114 1115 mt = vm_page_lookup(fs->first_object, 1116 scan_pindex); 1117 if (mt == NULL || 1118 (mt->valid != VM_PAGE_BITS_ALL)) { 1119 break; 1120 } 1121 if (mt->busy || 1122 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 1123 mt->hold_count || 1124 mt->wire_count) { 1125 goto skip; 1126 } 1127 if (mt->dirty == 0) 1128 vm_page_test_dirty(mt); 1129 if (mt->dirty) { 1130 vm_page_busy(mt); 1131 vm_page_protect(mt, 1132 VM_PROT_NONE); 1133 vm_page_deactivate(mt); 1134 vm_page_wakeup(mt); 1135 } else { 1136 vm_page_cache(mt); 1137 } 1138 skip: 1139 --scan_count; 1140 --scan_pindex; 1141 } 1142 crit_exit(); 1143 1144 seqaccess = 1; 1145 } 1146 1147 /* 1148 * Avoid deadlocking against the map when doing I/O. 1149 * fs.object and the page is PG_BUSY'd. 1150 */ 1151 unlock_map(fs); 1152 1153 /* 1154 * Acquire the page data. We still hold a ref on 1155 * fs.object and the page has been PG_BUSY's. 1156 * 1157 * The pager may replace the page (for example, in 1158 * order to enter a fictitious page into the 1159 * object). If it does so it is responsible for 1160 * cleaning up the passed page and properly setting 1161 * the new page PG_BUSY. 1162 * 1163 * If we got here through a PG_RAM read-ahead 1164 * mark the page may be partially dirty and thus 1165 * not freeable. Don't bother checking to see 1166 * if the pager has the page because we can't free 1167 * it anyway. We have to depend on the get_page 1168 * operation filling in any gaps whether there is 1169 * backing store or not. 1170 */ 1171 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess); 1172 1173 if (rv == VM_PAGER_OK) { 1174 /* 1175 * Relookup in case pager changed page. Pager 1176 * is responsible for disposition of old page 1177 * if moved. 1178 * 1179 * XXX other code segments do relookups too. 1180 * It's a bad abstraction that needs to be 1181 * fixed/removed. 1182 */ 1183 fs->m = vm_page_lookup(fs->object, pindex); 1184 if (fs->m == NULL) { 1185 unlock_and_deallocate(fs); 1186 return (KERN_TRY_AGAIN); 1187 } 1188 1189 ++fs->hardfault; 1190 break; /* break to PAGE HAS BEEN FOUND */ 1191 } 1192 1193 /* 1194 * Remove the bogus page (which does not exist at this 1195 * object/offset); before doing so, we must get back 1196 * our object lock to preserve our invariant. 1197 * 1198 * Also wake up any other process that may want to bring 1199 * in this page. 1200 * 1201 * If this is the top-level object, we must leave the 1202 * busy page to prevent another process from rushing 1203 * past us, and inserting the page in that object at 1204 * the same time that we are. 1205 */ 1206 if (rv == VM_PAGER_ERROR) { 1207 if (curproc) 1208 kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); 1209 else 1210 kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm); 1211 } 1212 1213 /* 1214 * Data outside the range of the pager or an I/O error 1215 * 1216 * The page may have been wired during the pagein, 1217 * e.g. by the buffer cache, and cannot simply be 1218 * freed. Call vnode_pager_freepage() to deal with it. 1219 */ 1220 /* 1221 * XXX - the check for kernel_map is a kludge to work 1222 * around having the machine panic on a kernel space 1223 * fault w/ I/O error. 1224 */ 1225 if (((fs->map != &kernel_map) && 1226 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 1227 vnode_pager_freepage(fs->m); 1228 fs->m = NULL; 1229 unlock_and_deallocate(fs); 1230 if (rv == VM_PAGER_ERROR) 1231 return (KERN_FAILURE); 1232 else 1233 return (KERN_PROTECTION_FAILURE); 1234 /* NOT REACHED */ 1235 } 1236 if (fs->object != fs->first_object) { 1237 vnode_pager_freepage(fs->m); 1238 fs->m = NULL; 1239 /* 1240 * XXX - we cannot just fall out at this 1241 * point, m has been freed and is invalid! 1242 */ 1243 } 1244 } 1245 1246 /* 1247 * We get here if the object has a default pager (or unwiring) 1248 * or the pager doesn't have the page. 1249 */ 1250 if (fs->object == fs->first_object) 1251 fs->first_m = fs->m; 1252 1253 /* 1254 * Move on to the next object. Lock the next object before 1255 * unlocking the current one. 1256 */ 1257 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1258 next_object = fs->object->backing_object; 1259 if (next_object == NULL) { 1260 /* 1261 * If there's no object left, fill the page in the top 1262 * object with zeros. 1263 */ 1264 if (fs->object != fs->first_object) { 1265 vm_object_pip_wakeup(fs->object); 1266 1267 fs->object = fs->first_object; 1268 pindex = first_pindex; 1269 fs->m = fs->first_m; 1270 } 1271 fs->first_m = NULL; 1272 1273 /* 1274 * Zero the page if necessary and mark it valid. 1275 */ 1276 if ((fs->m->flags & PG_ZERO) == 0) { 1277 vm_page_zero_fill(fs->m); 1278 } else { 1279 mycpu->gd_cnt.v_ozfod++; 1280 } 1281 mycpu->gd_cnt.v_zfod++; 1282 fs->m->valid = VM_PAGE_BITS_ALL; 1283 break; /* break to PAGE HAS BEEN FOUND */ 1284 } else { 1285 if (fs->object != fs->first_object) { 1286 vm_object_pip_wakeup(fs->object); 1287 } 1288 KASSERT(fs->object != next_object, ("object loop %p", next_object)); 1289 fs->object = next_object; 1290 vm_object_pip_add(fs->object, 1); 1291 } 1292 } 1293 1294 /* 1295 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1296 * is held.] 1297 * 1298 * If the page is being written, but isn't already owned by the 1299 * top-level object, we have to copy it into a new page owned by the 1300 * top-level object. 1301 */ 1302 KASSERT((fs->m->flags & PG_BUSY) != 0, 1303 ("vm_fault: not busy after main loop")); 1304 1305 if (fs->object != fs->first_object) { 1306 /* 1307 * We only really need to copy if we want to write it. 1308 */ 1309 if (fault_type & VM_PROT_WRITE) { 1310 /* 1311 * This allows pages to be virtually copied from a 1312 * backing_object into the first_object, where the 1313 * backing object has no other refs to it, and cannot 1314 * gain any more refs. Instead of a bcopy, we just 1315 * move the page from the backing object to the 1316 * first object. Note that we must mark the page 1317 * dirty in the first object so that it will go out 1318 * to swap when needed. 1319 */ 1320 if ( 1321 /* 1322 * Map, if present, has not changed 1323 */ 1324 (fs->map == NULL || 1325 fs->map_generation == fs->map->timestamp) && 1326 /* 1327 * Only one shadow object 1328 */ 1329 (fs->object->shadow_count == 1) && 1330 /* 1331 * No COW refs, except us 1332 */ 1333 (fs->object->ref_count == 1) && 1334 /* 1335 * No one else can look this object up 1336 */ 1337 (fs->object->handle == NULL) && 1338 /* 1339 * No other ways to look the object up 1340 */ 1341 ((fs->object->type == OBJT_DEFAULT) || 1342 (fs->object->type == OBJT_SWAP)) && 1343 /* 1344 * We don't chase down the shadow chain 1345 */ 1346 (fs->object == fs->first_object->backing_object) && 1347 1348 /* 1349 * grab the lock if we need to 1350 */ 1351 (fs->lookup_still_valid || 1352 fs->map == NULL || 1353 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 1354 ) { 1355 1356 fs->lookup_still_valid = 1; 1357 /* 1358 * get rid of the unnecessary page 1359 */ 1360 vm_page_protect(fs->first_m, VM_PROT_NONE); 1361 vm_page_free(fs->first_m); 1362 fs->first_m = NULL; 1363 1364 /* 1365 * grab the page and put it into the 1366 * process'es object. The page is 1367 * automatically made dirty. 1368 */ 1369 vm_page_rename(fs->m, fs->first_object, first_pindex); 1370 fs->first_m = fs->m; 1371 vm_page_busy(fs->first_m); 1372 fs->m = NULL; 1373 mycpu->gd_cnt.v_cow_optim++; 1374 } else { 1375 /* 1376 * Oh, well, lets copy it. 1377 */ 1378 vm_page_copy(fs->m, fs->first_m); 1379 vm_page_event(fs->m, VMEVENT_COW); 1380 } 1381 1382 if (fs->m) { 1383 /* 1384 * We no longer need the old page or object. 1385 */ 1386 release_page(fs); 1387 } 1388 1389 /* 1390 * fs->object != fs->first_object due to above 1391 * conditional 1392 */ 1393 vm_object_pip_wakeup(fs->object); 1394 1395 /* 1396 * Only use the new page below... 1397 */ 1398 1399 mycpu->gd_cnt.v_cow_faults++; 1400 fs->m = fs->first_m; 1401 fs->object = fs->first_object; 1402 pindex = first_pindex; 1403 } else { 1404 /* 1405 * If it wasn't a write fault avoid having to copy 1406 * the page by mapping it read-only. 1407 */ 1408 fs->prot &= ~VM_PROT_WRITE; 1409 } 1410 } 1411 1412 /* 1413 * We may have had to unlock a map to do I/O. If we did then 1414 * lookup_still_valid will be FALSE. If the map generation count 1415 * also changed then all sorts of things could have happened while 1416 * we were doing the I/O and we need to retry. 1417 */ 1418 1419 if (!fs->lookup_still_valid && 1420 fs->map != NULL && 1421 (fs->map->timestamp != fs->map_generation)) { 1422 release_page(fs); 1423 unlock_and_deallocate(fs); 1424 return (KERN_TRY_AGAIN); 1425 } 1426 1427 /* 1428 * If the fault is a write, we know that this page is being 1429 * written NOW so dirty it explicitly to save on pmap_is_modified() 1430 * calls later. 1431 * 1432 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 1433 * if the page is already dirty to prevent data written with 1434 * the expectation of being synced from not being synced. 1435 * Likewise if this entry does not request NOSYNC then make 1436 * sure the page isn't marked NOSYNC. Applications sharing 1437 * data should use the same flags to avoid ping ponging. 1438 * 1439 * Also tell the backing pager, if any, that it should remove 1440 * any swap backing since the page is now dirty. 1441 */ 1442 if (fs->prot & VM_PROT_WRITE) { 1443 vm_object_set_writeable_dirty(fs->m->object); 1444 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) { 1445 if (fs->m->dirty == 0) 1446 vm_page_flag_set(fs->m, PG_NOSYNC); 1447 } else { 1448 vm_page_flag_clear(fs->m, PG_NOSYNC); 1449 } 1450 if (fs->fault_flags & VM_FAULT_DIRTY) { 1451 crit_enter(); 1452 vm_page_dirty(fs->m); 1453 swap_pager_unswapped(fs->m); 1454 crit_exit(); 1455 } 1456 } 1457 1458 /* 1459 * Page had better still be busy. We are still locked up and 1460 * fs->object will have another PIP reference if it is not equal 1461 * to fs->first_object. 1462 */ 1463 KASSERT(fs->m->flags & PG_BUSY, 1464 ("vm_fault: page %p not busy!", fs->m)); 1465 1466 /* 1467 * Sanity check: page must be completely valid or it is not fit to 1468 * map into user space. vm_pager_get_pages() ensures this. 1469 */ 1470 if (fs->m->valid != VM_PAGE_BITS_ALL) { 1471 vm_page_zero_invalid(fs->m, TRUE); 1472 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 1473 } 1474 1475 return (KERN_SUCCESS); 1476 } 1477 1478 /* 1479 * Wire down a range of virtual addresses in a map. The entry in question 1480 * should be marked in-transition and the map must be locked. We must 1481 * release the map temporarily while faulting-in the page to avoid a 1482 * deadlock. Note that the entry may be clipped while we are blocked but 1483 * will never be freed. 1484 */ 1485 int 1486 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 1487 { 1488 boolean_t fictitious; 1489 vm_offset_t start; 1490 vm_offset_t end; 1491 vm_offset_t va; 1492 vm_paddr_t pa; 1493 pmap_t pmap; 1494 int rv; 1495 1496 pmap = vm_map_pmap(map); 1497 start = entry->start; 1498 end = entry->end; 1499 fictitious = entry->object.vm_object && 1500 (entry->object.vm_object->type == OBJT_DEVICE); 1501 1502 vm_map_unlock(map); 1503 map->timestamp++; 1504 1505 /* 1506 * We simulate a fault to get the page and enter it in the physical 1507 * map. 1508 */ 1509 for (va = start; va < end; va += PAGE_SIZE) { 1510 if (user_wire) { 1511 rv = vm_fault(map, va, VM_PROT_READ, 1512 VM_FAULT_USER_WIRE); 1513 } else { 1514 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 1515 VM_FAULT_CHANGE_WIRING); 1516 } 1517 if (rv) { 1518 while (va > start) { 1519 va -= PAGE_SIZE; 1520 if ((pa = pmap_extract(pmap, va)) == 0) 1521 continue; 1522 pmap_change_wiring(pmap, va, FALSE); 1523 if (!fictitious) 1524 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1525 } 1526 vm_map_lock(map); 1527 return (rv); 1528 } 1529 } 1530 vm_map_lock(map); 1531 return (KERN_SUCCESS); 1532 } 1533 1534 /* 1535 * Unwire a range of virtual addresses in a map. The map should be 1536 * locked. 1537 */ 1538 void 1539 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 1540 { 1541 boolean_t fictitious; 1542 vm_offset_t start; 1543 vm_offset_t end; 1544 vm_offset_t va; 1545 vm_paddr_t pa; 1546 pmap_t pmap; 1547 1548 pmap = vm_map_pmap(map); 1549 start = entry->start; 1550 end = entry->end; 1551 fictitious = entry->object.vm_object && 1552 (entry->object.vm_object->type == OBJT_DEVICE); 1553 1554 /* 1555 * Since the pages are wired down, we must be able to get their 1556 * mappings from the physical map system. 1557 */ 1558 for (va = start; va < end; va += PAGE_SIZE) { 1559 pa = pmap_extract(pmap, va); 1560 if (pa != 0) { 1561 pmap_change_wiring(pmap, va, FALSE); 1562 if (!fictitious) 1563 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1564 } 1565 } 1566 } 1567 1568 /* 1569 * Reduce the rate at which memory is allocated to a process based 1570 * on the perceived load on the VM system. As the load increases 1571 * the allocation burst rate goes down and the delay increases. 1572 * 1573 * Rate limiting does not apply when faulting active or inactive 1574 * pages. When faulting 'cache' pages, rate limiting only applies 1575 * if the system currently has a severe page deficit. 1576 * 1577 * XXX vm_pagesupply should be increased when a page is freed. 1578 * 1579 * We sleep up to 1/10 of a second. 1580 */ 1581 static int 1582 vm_fault_ratelimit(struct vmspace *vmspace) 1583 { 1584 if (vm_load_enable == 0) 1585 return(0); 1586 if (vmspace->vm_pagesupply > 0) { 1587 --vmspace->vm_pagesupply; 1588 return(0); 1589 } 1590 #ifdef INVARIANTS 1591 if (vm_load_debug) { 1592 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n", 1593 vm_load, 1594 (1000 - vm_load ) / 10, vm_load * hz / 10000, 1595 curproc->p_pid, curproc->p_comm); 1596 } 1597 #endif 1598 vmspace->vm_pagesupply = (1000 - vm_load) / 10; 1599 return(vm_load * hz / 10000); 1600 } 1601 1602 /* 1603 * Routine: 1604 * vm_fault_copy_entry 1605 * Function: 1606 * Copy all of the pages from a wired-down map entry to another. 1607 * 1608 * In/out conditions: 1609 * The source and destination maps must be locked for write. 1610 * The source map entry must be wired down (or be a sharing map 1611 * entry corresponding to a main map entry that is wired down). 1612 */ 1613 1614 void 1615 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1616 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 1617 { 1618 vm_object_t dst_object; 1619 vm_object_t src_object; 1620 vm_ooffset_t dst_offset; 1621 vm_ooffset_t src_offset; 1622 vm_prot_t prot; 1623 vm_offset_t vaddr; 1624 vm_page_t dst_m; 1625 vm_page_t src_m; 1626 1627 #ifdef lint 1628 src_map++; 1629 #endif /* lint */ 1630 1631 src_object = src_entry->object.vm_object; 1632 src_offset = src_entry->offset; 1633 1634 /* 1635 * Create the top-level object for the destination entry. (Doesn't 1636 * actually shadow anything - we copy the pages directly.) 1637 */ 1638 vm_map_entry_allocate_object(dst_entry); 1639 dst_object = dst_entry->object.vm_object; 1640 1641 prot = dst_entry->max_protection; 1642 1643 /* 1644 * Loop through all of the pages in the entry's range, copying each 1645 * one from the source object (it should be there) to the destination 1646 * object. 1647 */ 1648 for (vaddr = dst_entry->start, dst_offset = 0; 1649 vaddr < dst_entry->end; 1650 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1651 1652 /* 1653 * Allocate a page in the destination object 1654 */ 1655 do { 1656 dst_m = vm_page_alloc(dst_object, 1657 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1658 if (dst_m == NULL) { 1659 vm_wait(0); 1660 } 1661 } while (dst_m == NULL); 1662 1663 /* 1664 * Find the page in the source object, and copy it in. 1665 * (Because the source is wired down, the page will be in 1666 * memory.) 1667 */ 1668 src_m = vm_page_lookup(src_object, 1669 OFF_TO_IDX(dst_offset + src_offset)); 1670 if (src_m == NULL) 1671 panic("vm_fault_copy_wired: page missing"); 1672 1673 vm_page_copy(src_m, dst_m); 1674 vm_page_event(src_m, VMEVENT_COW); 1675 1676 /* 1677 * Enter it in the pmap... 1678 */ 1679 1680 vm_page_flag_clear(dst_m, PG_ZERO); 1681 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1682 1683 /* 1684 * Mark it no longer busy, and put it on the active list. 1685 */ 1686 vm_page_activate(dst_m); 1687 vm_page_wakeup(dst_m); 1688 } 1689 } 1690 1691 #if 0 1692 1693 /* 1694 * This routine checks around the requested page for other pages that 1695 * might be able to be faulted in. This routine brackets the viable 1696 * pages for the pages to be paged in. 1697 * 1698 * Inputs: 1699 * m, rbehind, rahead 1700 * 1701 * Outputs: 1702 * marray (array of vm_page_t), reqpage (index of requested page) 1703 * 1704 * Return value: 1705 * number of pages in marray 1706 */ 1707 static int 1708 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 1709 vm_page_t *marray, int *reqpage) 1710 { 1711 int i,j; 1712 vm_object_t object; 1713 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1714 vm_page_t rtm; 1715 int cbehind, cahead; 1716 1717 object = m->object; 1718 pindex = m->pindex; 1719 1720 /* 1721 * we don't fault-ahead for device pager 1722 */ 1723 if (object->type == OBJT_DEVICE) { 1724 *reqpage = 0; 1725 marray[0] = m; 1726 return 1; 1727 } 1728 1729 /* 1730 * if the requested page is not available, then give up now 1731 */ 1732 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1733 *reqpage = 0; /* not used by caller, fix compiler warn */ 1734 return 0; 1735 } 1736 1737 if ((cbehind == 0) && (cahead == 0)) { 1738 *reqpage = 0; 1739 marray[0] = m; 1740 return 1; 1741 } 1742 1743 if (rahead > cahead) { 1744 rahead = cahead; 1745 } 1746 1747 if (rbehind > cbehind) { 1748 rbehind = cbehind; 1749 } 1750 1751 /* 1752 * Do not do any readahead if we have insufficient free memory. 1753 * 1754 * XXX code was broken disabled before and has instability 1755 * with this conditonal fixed, so shortcut for now. 1756 */ 1757 if (burst_fault == 0 || vm_page_count_severe()) { 1758 marray[0] = m; 1759 *reqpage = 0; 1760 return 1; 1761 } 1762 1763 /* 1764 * scan backward for the read behind pages -- in memory 1765 * 1766 * Assume that if the page is not found an interrupt will not 1767 * create it. Theoretically interrupts can only remove (busy) 1768 * pages, not create new associations. 1769 */ 1770 if (pindex > 0) { 1771 if (rbehind > pindex) { 1772 rbehind = pindex; 1773 startpindex = 0; 1774 } else { 1775 startpindex = pindex - rbehind; 1776 } 1777 1778 crit_enter(); 1779 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 1780 if (vm_page_lookup(object, tpindex - 1)) 1781 break; 1782 } 1783 1784 i = 0; 1785 while (tpindex < pindex) { 1786 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM); 1787 if (rtm == NULL) { 1788 crit_exit(); 1789 for (j = 0; j < i; j++) { 1790 vm_page_free(marray[j]); 1791 } 1792 marray[0] = m; 1793 *reqpage = 0; 1794 return 1; 1795 } 1796 marray[i] = rtm; 1797 ++i; 1798 ++tpindex; 1799 } 1800 crit_exit(); 1801 } else { 1802 i = 0; 1803 } 1804 1805 /* 1806 * Assign requested page 1807 */ 1808 marray[i] = m; 1809 *reqpage = i; 1810 ++i; 1811 1812 /* 1813 * Scan forwards for read-ahead pages 1814 */ 1815 tpindex = pindex + 1; 1816 endpindex = tpindex + rahead; 1817 if (endpindex > object->size) 1818 endpindex = object->size; 1819 1820 crit_enter(); 1821 while (tpindex < endpindex) { 1822 if (vm_page_lookup(object, tpindex)) 1823 break; 1824 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM); 1825 if (rtm == NULL) 1826 break; 1827 marray[i] = rtm; 1828 ++i; 1829 ++tpindex; 1830 } 1831 crit_exit(); 1832 1833 return (i); 1834 } 1835 1836 #endif 1837 1838 /* 1839 * vm_prefault() provides a quick way of clustering pagefaults into a 1840 * processes address space. It is a "cousin" of pmap_object_init_pt, 1841 * except it runs at page fault time instead of mmap time. 1842 * 1843 * This code used to be per-platform pmap_prefault(). It is now 1844 * machine-independent and enhanced to also pre-fault zero-fill pages 1845 * (see vm.fast_fault) as well as make them writable, which greatly 1846 * reduces the number of page faults programs incur. 1847 * 1848 * Application performance when pre-faulting zero-fill pages is heavily 1849 * dependent on the application. Very tiny applications like /bin/echo 1850 * lose a little performance while applications of any appreciable size 1851 * gain performance. Prefaulting multiple pages also reduces SMP 1852 * congestion and can improve SMP performance significantly. 1853 * 1854 * NOTE! prot may allow writing but this only applies to the top level 1855 * object. If we wind up mapping a page extracted from a backing 1856 * object we have to make sure it is read-only. 1857 * 1858 * NOTE! The caller has already handled any COW operations on the 1859 * vm_map_entry via the normal fault code. Do NOT call this 1860 * shortcut unless the normal fault code has run on this entry. 1861 */ 1862 #define PFBAK 4 1863 #define PFFOR 4 1864 #define PAGEORDER_SIZE (PFBAK+PFFOR) 1865 1866 static int vm_prefault_pageorder[] = { 1867 -PAGE_SIZE, PAGE_SIZE, 1868 -2 * PAGE_SIZE, 2 * PAGE_SIZE, 1869 -3 * PAGE_SIZE, 3 * PAGE_SIZE, 1870 -4 * PAGE_SIZE, 4 * PAGE_SIZE 1871 }; 1872 1873 static void 1874 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot) 1875 { 1876 struct lwp *lp; 1877 vm_page_t m; 1878 vm_offset_t starta; 1879 vm_offset_t addr; 1880 vm_pindex_t index; 1881 vm_pindex_t pindex; 1882 vm_object_t object; 1883 int pprot; 1884 int i; 1885 1886 /* 1887 * We do not currently prefault mappings that use virtual page 1888 * tables. We do not prefault foreign pmaps. 1889 */ 1890 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) 1891 return; 1892 lp = curthread->td_lwp; 1893 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 1894 return; 1895 1896 object = entry->object.vm_object; 1897 1898 starta = addra - PFBAK * PAGE_SIZE; 1899 if (starta < entry->start) 1900 starta = entry->start; 1901 else if (starta > addra) 1902 starta = 0; 1903 1904 /* 1905 * critical section protection is required to maintain the 1906 * page/object association, interrupts can free pages and remove 1907 * them from their objects. 1908 */ 1909 crit_enter(); 1910 for (i = 0; i < PAGEORDER_SIZE; i++) { 1911 vm_object_t lobject; 1912 int allocated = 0; 1913 1914 addr = addra + vm_prefault_pageorder[i]; 1915 if (addr > addra + (PFFOR * PAGE_SIZE)) 1916 addr = 0; 1917 1918 if (addr < starta || addr >= entry->end) 1919 continue; 1920 1921 if (pmap_prefault_ok(pmap, addr) == 0) 1922 continue; 1923 1924 /* 1925 * Follow the VM object chain to obtain the page to be mapped 1926 * into the pmap. 1927 * 1928 * If we reach the terminal object without finding a page 1929 * and we determine it would be advantageous, then allocate 1930 * a zero-fill page for the base object. The base object 1931 * is guaranteed to be OBJT_DEFAULT for this case. 1932 * 1933 * In order to not have to check the pager via *haspage*() 1934 * we stop if any non-default object is encountered. e.g. 1935 * a vnode or swap object would stop the loop. 1936 */ 1937 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 1938 lobject = object; 1939 pindex = index; 1940 pprot = prot; 1941 1942 while ((m = vm_page_lookup(lobject, pindex)) == NULL) { 1943 if (lobject->type != OBJT_DEFAULT) 1944 break; 1945 if (lobject->backing_object == NULL) { 1946 if (vm_fast_fault == 0) 1947 break; 1948 if (vm_prefault_pageorder[i] < 0 || 1949 (prot & VM_PROT_WRITE) == 0 || 1950 vm_page_count_min(0)) { 1951 break; 1952 } 1953 /* note: allocate from base object */ 1954 m = vm_page_alloc(object, index, 1955 VM_ALLOC_NORMAL | VM_ALLOC_ZERO); 1956 1957 if ((m->flags & PG_ZERO) == 0) { 1958 vm_page_zero_fill(m); 1959 } else { 1960 vm_page_flag_clear(m, PG_ZERO); 1961 mycpu->gd_cnt.v_ozfod++; 1962 } 1963 mycpu->gd_cnt.v_zfod++; 1964 m->valid = VM_PAGE_BITS_ALL; 1965 allocated = 1; 1966 pprot = prot; 1967 /* lobject = object .. not needed */ 1968 break; 1969 } 1970 if (lobject->backing_object_offset & PAGE_MASK) 1971 break; 1972 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 1973 lobject = lobject->backing_object; 1974 pprot &= ~VM_PROT_WRITE; 1975 } 1976 /* 1977 * NOTE: lobject now invalid (if we did a zero-fill we didn't 1978 * bother assigning lobject = object). 1979 * 1980 * Give-up if the page is not available. 1981 */ 1982 if (m == NULL) 1983 break; 1984 1985 /* 1986 * Do not conditionalize on PG_RAM. If pages are present in 1987 * the VM system we assume optimal caching. If caching is 1988 * not optimal the I/O gravy train will be restarted when we 1989 * hit an unavailable page. We do not want to try to restart 1990 * the gravy train now because we really don't know how much 1991 * of the object has been cached. The cost for restarting 1992 * the gravy train should be low (since accesses will likely 1993 * be I/O bound anyway). 1994 * 1995 * The object must be marked dirty if we are mapping a 1996 * writable page. 1997 */ 1998 if (pprot & VM_PROT_WRITE) 1999 vm_object_set_writeable_dirty(m->object); 2000 2001 /* 2002 * Enter the page into the pmap if appropriate. If we had 2003 * allocated the page we have to place it on a queue. If not 2004 * we just have to make sure it isn't on the cache queue 2005 * (pages on the cache queue are not allowed to be mapped). 2006 */ 2007 if (allocated) { 2008 pmap_enter(pmap, addr, m, pprot, 0); 2009 vm_page_deactivate(m); 2010 vm_page_wakeup(m); 2011 } else if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2012 (m->busy == 0) && 2013 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) { 2014 2015 if ((m->queue - m->pc) == PQ_CACHE) { 2016 vm_page_deactivate(m); 2017 } 2018 vm_page_busy(m); 2019 pmap_enter(pmap, addr, m, pprot, 0); 2020 vm_page_wakeup(m); 2021 } 2022 } 2023 crit_exit(); 2024 } 2025