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 #define VM_FAULT_READ_AHEAD 8 105 #define VM_FAULT_READ_BEHIND 7 106 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 107 108 struct faultstate { 109 vm_page_t m; 110 vm_object_t object; 111 vm_pindex_t pindex; 112 vm_prot_t prot; 113 vm_page_t first_m; 114 vm_object_t first_object; 115 vm_prot_t first_prot; 116 vm_map_t map; 117 vm_map_entry_t entry; 118 int lookup_still_valid; 119 int didlimit; 120 int hardfault; 121 int fault_flags; 122 int map_generation; 123 boolean_t wired; 124 struct vnode *vp; 125 }; 126 127 static int burst_fault = 1; 128 SYSCTL_INT(_vm, OID_AUTO, burst_fault, CTLFLAG_RW, &burst_fault, 0, ""); 129 130 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t); 131 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int); 132 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 133 static int vm_fault_ratelimit(struct vmspace *); 134 135 static __inline void 136 release_page(struct faultstate *fs) 137 { 138 vm_page_deactivate(fs->m); 139 vm_page_wakeup(fs->m); 140 fs->m = NULL; 141 } 142 143 static __inline void 144 unlock_map(struct faultstate *fs) 145 { 146 if (fs->lookup_still_valid && fs->map) { 147 vm_map_lookup_done(fs->map, fs->entry, 0); 148 fs->lookup_still_valid = FALSE; 149 } 150 } 151 152 /* 153 * Clean up after a successful call to vm_fault_object() so another call 154 * to vm_fault_object() can be made. 155 */ 156 static void 157 _cleanup_successful_fault(struct faultstate *fs, int relock) 158 { 159 if (fs->object != fs->first_object) { 160 vm_page_free(fs->first_m); 161 vm_object_pip_wakeup(fs->object); 162 fs->first_m = NULL; 163 } 164 fs->object = fs->first_object; 165 if (relock && fs->lookup_still_valid == FALSE) { 166 if (fs->map) 167 vm_map_lock_read(fs->map); 168 fs->lookup_still_valid = TRUE; 169 } 170 } 171 172 static void 173 _unlock_things(struct faultstate *fs, int dealloc) 174 { 175 vm_object_pip_wakeup(fs->first_object); 176 _cleanup_successful_fault(fs, 0); 177 if (dealloc) { 178 vm_object_deallocate(fs->first_object); 179 fs->first_object = NULL; 180 } 181 unlock_map(fs); 182 if (fs->vp != NULL) { 183 vput(fs->vp); 184 fs->vp = NULL; 185 } 186 } 187 188 #define unlock_things(fs) _unlock_things(fs, 0) 189 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 190 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 191 192 /* 193 * TRYPAGER 194 * 195 * Determine if the pager for the current object *might* contain the page. 196 * 197 * We only need to try the pager if this is not a default object (default 198 * objects are zero-fill and have no real pager), and if we are not taking 199 * a wiring fault or if the FS entry is wired. 200 */ 201 #define TRYPAGER(fs) \ 202 (fs->object->type != OBJT_DEFAULT && \ 203 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired)) 204 205 /* 206 * vm_fault: 207 * 208 * Handle a page fault occuring at the given address, requiring the given 209 * permissions, in the map specified. If successful, the page is inserted 210 * into the associated physical map. 211 * 212 * NOTE: The given address should be truncated to the proper page address. 213 * 214 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 215 * a standard error specifying why the fault is fatal is returned. 216 * 217 * The map in question must be referenced, and remains so. 218 * The caller may hold no locks. 219 */ 220 int 221 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 222 { 223 int result; 224 vm_pindex_t first_pindex; 225 struct faultstate fs; 226 227 mycpu->gd_cnt.v_vm_faults++; 228 229 fs.didlimit = 0; 230 fs.hardfault = 0; 231 fs.fault_flags = fault_flags; 232 233 RetryFault: 234 /* 235 * Find the vm_map_entry representing the backing store and resolve 236 * the top level object and page index. This may have the side 237 * effect of executing a copy-on-write on the map entry and/or 238 * creating a shadow object, but will not COW any actual VM pages. 239 * 240 * On success fs.map is left read-locked and various other fields 241 * are initialized but not otherwise referenced or locked. 242 * 243 * NOTE! vm_map_lookup will try to upgrade the fault_type to 244 * VM_FAULT_WRITE if the map entry is a virtual page table and also 245 * writable, so we can set the 'A'accessed bit in the virtual page 246 * table entry. 247 */ 248 fs.map = map; 249 result = vm_map_lookup(&fs.map, vaddr, fault_type, 250 &fs.entry, &fs.first_object, 251 &first_pindex, &fs.first_prot, &fs.wired); 252 253 /* 254 * If the lookup failed or the map protections are incompatible, 255 * the fault generally fails. However, if the caller is trying 256 * to do a user wiring we have more work to do. 257 */ 258 if (result != KERN_SUCCESS) { 259 if (result != KERN_PROTECTION_FAILURE) 260 return result; 261 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 262 return result; 263 264 /* 265 * If we are user-wiring a r/w segment, and it is COW, then 266 * we need to do the COW operation. Note that we don't 267 * currently COW RO sections now, because it is NOT desirable 268 * to COW .text. We simply keep .text from ever being COW'ed 269 * and take the heat that one cannot debug wired .text sections. 270 */ 271 result = vm_map_lookup(&fs.map, vaddr, 272 VM_PROT_READ|VM_PROT_WRITE| 273 VM_PROT_OVERRIDE_WRITE, 274 &fs.entry, &fs.first_object, 275 &first_pindex, &fs.first_prot, 276 &fs.wired); 277 if (result != KERN_SUCCESS) 278 return result; 279 280 /* 281 * If we don't COW now, on a user wire, the user will never 282 * be able to write to the mapping. If we don't make this 283 * restriction, the bookkeeping would be nearly impossible. 284 */ 285 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 286 fs.entry->max_protection &= ~VM_PROT_WRITE; 287 } 288 289 /* 290 * fs.map is read-locked 291 * 292 * Misc checks. Save the map generation number to detect races. 293 */ 294 fs.map_generation = fs.map->timestamp; 295 296 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 297 panic("vm_fault: fault on nofault entry, addr: %lx", 298 (u_long)vaddr); 299 } 300 301 /* 302 * A system map entry may return a NULL object. No object means 303 * no pager means an unrecoverable kernel fault. 304 */ 305 if (fs.first_object == NULL) { 306 panic("vm_fault: unrecoverable fault at %p in entry %p", 307 (void *)vaddr, fs.entry); 308 } 309 310 /* 311 * Make a reference to this object to prevent its disposal while we 312 * are messing with it. Once we have the reference, the map is free 313 * to be diddled. Since objects reference their shadows (and copies), 314 * they will stay around as well. 315 * 316 * Bump the paging-in-progress count to prevent size changes (e.g. 317 * truncation operations) during I/O. This must be done after 318 * obtaining the vnode lock in order to avoid possible deadlocks. 319 */ 320 vm_object_reference(fs.first_object); 321 fs.vp = vnode_pager_lock(fs.first_object); 322 vm_object_pip_add(fs.first_object, 1); 323 324 fs.lookup_still_valid = TRUE; 325 fs.first_m = NULL; 326 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 327 328 /* 329 * If the entry is wired we cannot change the page protection. 330 */ 331 if (fs.wired) 332 fault_type = fs.first_prot; 333 334 /* 335 * The page we want is at (first_object, first_pindex), but if the 336 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 337 * page table to figure out the actual pindex. 338 * 339 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 340 * ONLY 341 */ 342 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 343 result = vm_fault_vpagetable(&fs, &first_pindex, 344 fs.entry->aux.master_pde, 345 fault_type); 346 if (result == KERN_TRY_AGAIN) 347 goto RetryFault; 348 if (result != KERN_SUCCESS) 349 return (result); 350 } 351 352 /* 353 * Now we have the actual (object, pindex), fault in the page. If 354 * vm_fault_object() fails it will unlock and deallocate the FS 355 * data. If it succeeds everything remains locked and fs->object 356 * will have an additinal PIP count if it is not equal to 357 * fs->first_object 358 * 359 * vm_fault_object will set fs->prot for the pmap operation. It is 360 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 361 * page can be safely written. However, it will force a read-only 362 * mapping for a read fault if the memory is managed by a virtual 363 * page table. 364 */ 365 result = vm_fault_object(&fs, first_pindex, fault_type); 366 367 if (result == KERN_TRY_AGAIN) 368 goto RetryFault; 369 if (result != KERN_SUCCESS) 370 return (result); 371 372 /* 373 * On success vm_fault_object() does not unlock or deallocate, and fs.m 374 * will contain a busied page. 375 * 376 * Enter the page into the pmap and do pmap-related adjustments. 377 */ 378 unlock_things(&fs); 379 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired); 380 381 if (((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0) && (fs.wired == 0)) { 382 pmap_prefault(fs.map->pmap, vaddr, fs.entry); 383 } 384 385 vm_page_flag_clear(fs.m, PG_ZERO); 386 vm_page_flag_set(fs.m, PG_REFERENCED); 387 388 /* 389 * If the page is not wired down, then put it where the pageout daemon 390 * can find it. 391 */ 392 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 393 if (fs.wired) 394 vm_page_wire(fs.m); 395 else 396 vm_page_unwire(fs.m, 1); 397 } else { 398 vm_page_activate(fs.m); 399 } 400 401 if (curthread->td_lwp) { 402 if (fs.hardfault) { 403 curthread->td_lwp->lwp_ru.ru_majflt++; 404 } else { 405 curthread->td_lwp->lwp_ru.ru_minflt++; 406 } 407 } 408 409 /* 410 * Unlock everything, and return 411 */ 412 vm_page_wakeup(fs.m); 413 vm_object_deallocate(fs.first_object); 414 415 return (KERN_SUCCESS); 416 } 417 418 /* 419 * Fault in the specified virtual address in the current process map, 420 * returning a held VM page or NULL. See vm_fault_page() for more 421 * information. 422 */ 423 vm_page_t 424 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp) 425 { 426 struct lwp *lp = curthread->td_lwp; 427 vm_page_t m; 428 429 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 430 fault_type, VM_FAULT_NORMAL, errorp); 431 return(m); 432 } 433 434 /* 435 * Fault in the specified virtual address in the specified map, doing all 436 * necessary manipulation of the object store and all necessary I/O. Return 437 * a held VM page or NULL, and set *errorp. The related pmap is not 438 * updated. 439 * 440 * The returned page will be properly dirtied if VM_PROT_WRITE was specified, 441 * and marked PG_REFERENCED as well. 442 * 443 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 444 * error will be returned. 445 */ 446 vm_page_t 447 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 448 int fault_flags, int *errorp) 449 { 450 vm_pindex_t first_pindex; 451 struct faultstate fs; 452 int result; 453 vm_prot_t orig_fault_type = fault_type; 454 455 mycpu->gd_cnt.v_vm_faults++; 456 457 fs.didlimit = 0; 458 fs.hardfault = 0; 459 fs.fault_flags = fault_flags; 460 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 461 462 RetryFault: 463 /* 464 * Find the vm_map_entry representing the backing store and resolve 465 * the top level object and page index. This may have the side 466 * effect of executing a copy-on-write on the map entry and/or 467 * creating a shadow object, but will not COW any actual VM pages. 468 * 469 * On success fs.map is left read-locked and various other fields 470 * are initialized but not otherwise referenced or locked. 471 * 472 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 473 * if the map entry is a virtual page table and also writable, 474 * so we can set the 'A'accessed bit in the virtual page table entry. 475 */ 476 fs.map = map; 477 result = vm_map_lookup(&fs.map, vaddr, fault_type, 478 &fs.entry, &fs.first_object, 479 &first_pindex, &fs.first_prot, &fs.wired); 480 481 if (result != KERN_SUCCESS) { 482 *errorp = result; 483 return (NULL); 484 } 485 486 /* 487 * fs.map is read-locked 488 * 489 * Misc checks. Save the map generation number to detect races. 490 */ 491 fs.map_generation = fs.map->timestamp; 492 493 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 494 panic("vm_fault: fault on nofault entry, addr: %lx", 495 (u_long)vaddr); 496 } 497 498 /* 499 * A system map entry may return a NULL object. No object means 500 * no pager means an unrecoverable kernel fault. 501 */ 502 if (fs.first_object == NULL) { 503 panic("vm_fault: unrecoverable fault at %p in entry %p", 504 (void *)vaddr, fs.entry); 505 } 506 507 /* 508 * Make a reference to this object to prevent its disposal while we 509 * are messing with it. Once we have the reference, the map is free 510 * to be diddled. Since objects reference their shadows (and copies), 511 * they will stay around as well. 512 * 513 * Bump the paging-in-progress count to prevent size changes (e.g. 514 * truncation operations) during I/O. This must be done after 515 * obtaining the vnode lock in order to avoid possible deadlocks. 516 */ 517 vm_object_reference(fs.first_object); 518 fs.vp = vnode_pager_lock(fs.first_object); 519 vm_object_pip_add(fs.first_object, 1); 520 521 fs.lookup_still_valid = TRUE; 522 fs.first_m = NULL; 523 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 524 525 /* 526 * If the entry is wired we cannot change the page protection. 527 */ 528 if (fs.wired) 529 fault_type = fs.first_prot; 530 531 /* 532 * The page we want is at (first_object, first_pindex), but if the 533 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 534 * page table to figure out the actual pindex. 535 * 536 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 537 * ONLY 538 */ 539 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 540 result = vm_fault_vpagetable(&fs, &first_pindex, 541 fs.entry->aux.master_pde, 542 fault_type); 543 if (result == KERN_TRY_AGAIN) 544 goto RetryFault; 545 if (result != KERN_SUCCESS) { 546 *errorp = result; 547 return (NULL); 548 } 549 } 550 551 /* 552 * Now we have the actual (object, pindex), fault in the page. If 553 * vm_fault_object() fails it will unlock and deallocate the FS 554 * data. If it succeeds everything remains locked and fs->object 555 * will have an additinal PIP count if it is not equal to 556 * fs->first_object 557 */ 558 result = vm_fault_object(&fs, first_pindex, fault_type); 559 560 if (result == KERN_TRY_AGAIN) 561 goto RetryFault; 562 if (result != KERN_SUCCESS) { 563 *errorp = result; 564 return(NULL); 565 } 566 567 if ((orig_fault_type & VM_PROT_WRITE) && 568 (fs.prot & VM_PROT_WRITE) == 0) { 569 *errorp = KERN_PROTECTION_FAILURE; 570 unlock_and_deallocate(&fs); 571 return(NULL); 572 } 573 574 /* 575 * On success vm_fault_object() does not unlock or deallocate, and fs.m 576 * will contain a busied page. 577 */ 578 unlock_things(&fs); 579 580 /* 581 * Return a held page. We are not doing any pmap manipulation so do 582 * not set PG_MAPPED. However, adjust the page flags according to 583 * the fault type because the caller may not use a managed pmapping 584 * (so we don't want to lose the fact that the page will be dirtied 585 * if a write fault was specified). 586 */ 587 vm_page_hold(fs.m); 588 vm_page_flag_clear(fs.m, PG_ZERO); 589 if (fault_type & VM_PROT_WRITE) 590 vm_page_dirty(fs.m); 591 592 /* 593 * Update the pmap. We really only have to do this if a COW 594 * occured to replace the read-only page with the new page. For 595 * now just do it unconditionally. XXX 596 */ 597 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired); 598 vm_page_flag_set(fs.m, PG_REFERENCED); 599 600 /* 601 * Unbusy the page by activating it. It remains held and will not 602 * be reclaimed. 603 */ 604 vm_page_activate(fs.m); 605 606 if (curthread->td_lwp) { 607 if (fs.hardfault) { 608 curthread->td_lwp->lwp_ru.ru_majflt++; 609 } else { 610 curthread->td_lwp->lwp_ru.ru_minflt++; 611 } 612 } 613 614 /* 615 * Unlock everything, and return the held page. 616 */ 617 vm_page_wakeup(fs.m); 618 vm_object_deallocate(fs.first_object); 619 620 *errorp = 0; 621 return(fs.m); 622 } 623 624 /* 625 * Fault in the specified (object,offset), dirty the returned page as 626 * needed. If the requested fault_type cannot be done NULL and an 627 * error is returned. 628 */ 629 vm_page_t 630 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 631 vm_prot_t fault_type, int fault_flags, int *errorp) 632 { 633 int result; 634 vm_pindex_t first_pindex; 635 struct faultstate fs; 636 struct vm_map_entry entry; 637 638 bzero(&entry, sizeof(entry)); 639 entry.object.vm_object = object; 640 entry.maptype = VM_MAPTYPE_NORMAL; 641 entry.protection = entry.max_protection = fault_type; 642 643 fs.didlimit = 0; 644 fs.hardfault = 0; 645 fs.fault_flags = fault_flags; 646 fs.map = NULL; 647 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 648 649 RetryFault: 650 651 fs.first_object = object; 652 first_pindex = OFF_TO_IDX(offset); 653 fs.entry = &entry; 654 fs.first_prot = fault_type; 655 fs.wired = 0; 656 /*fs.map_generation = 0; unused */ 657 658 /* 659 * Make a reference to this object to prevent its disposal while we 660 * are messing with it. Once we have the reference, the map is free 661 * to be diddled. Since objects reference their shadows (and copies), 662 * they will stay around as well. 663 * 664 * Bump the paging-in-progress count to prevent size changes (e.g. 665 * truncation operations) during I/O. This must be done after 666 * obtaining the vnode lock in order to avoid possible deadlocks. 667 */ 668 vm_object_reference(fs.first_object); 669 fs.vp = vnode_pager_lock(fs.first_object); 670 vm_object_pip_add(fs.first_object, 1); 671 672 fs.lookup_still_valid = TRUE; 673 fs.first_m = NULL; 674 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 675 676 #if 0 677 /* XXX future - ability to operate on VM object using vpagetable */ 678 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 679 result = vm_fault_vpagetable(&fs, &first_pindex, 680 fs.entry->aux.master_pde, 681 fault_type); 682 if (result == KERN_TRY_AGAIN) 683 goto RetryFault; 684 if (result != KERN_SUCCESS) { 685 *errorp = result; 686 return (NULL); 687 } 688 } 689 #endif 690 691 /* 692 * Now we have the actual (object, pindex), fault in the page. If 693 * vm_fault_object() fails it will unlock and deallocate the FS 694 * data. If it succeeds everything remains locked and fs->object 695 * will have an additinal PIP count if it is not equal to 696 * fs->first_object 697 */ 698 result = vm_fault_object(&fs, first_pindex, fault_type); 699 700 if (result == KERN_TRY_AGAIN) 701 goto RetryFault; 702 if (result != KERN_SUCCESS) { 703 *errorp = result; 704 return(NULL); 705 } 706 707 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 708 *errorp = KERN_PROTECTION_FAILURE; 709 unlock_and_deallocate(&fs); 710 return(NULL); 711 } 712 713 /* 714 * On success vm_fault_object() does not unlock or deallocate, and fs.m 715 * will contain a busied page. 716 */ 717 unlock_things(&fs); 718 719 /* 720 * Return a held page. We are not doing any pmap manipulation so do 721 * not set PG_MAPPED. However, adjust the page flags according to 722 * the fault type because the caller may not use a managed pmapping 723 * (so we don't want to lose the fact that the page will be dirtied 724 * if a write fault was specified). 725 */ 726 vm_page_hold(fs.m); 727 vm_page_flag_clear(fs.m, PG_ZERO); 728 if (fault_type & VM_PROT_WRITE) 729 vm_page_dirty(fs.m); 730 731 /* 732 * Indicate that the page was accessed. 733 */ 734 vm_page_flag_set(fs.m, PG_REFERENCED); 735 736 /* 737 * Unbusy the page by activating it. It remains held and will not 738 * be reclaimed. 739 */ 740 vm_page_activate(fs.m); 741 742 if (curthread->td_lwp) { 743 if (fs.hardfault) { 744 mycpu->gd_cnt.v_vm_faults++; 745 curthread->td_lwp->lwp_ru.ru_majflt++; 746 } else { 747 curthread->td_lwp->lwp_ru.ru_minflt++; 748 } 749 } 750 751 /* 752 * Unlock everything, and return the held page. 753 */ 754 vm_page_wakeup(fs.m); 755 vm_object_deallocate(fs.first_object); 756 757 *errorp = 0; 758 return(fs.m); 759 } 760 761 /* 762 * Translate the virtual page number (first_pindex) that is relative 763 * to the address space into a logical page number that is relative to the 764 * backing object. Use the virtual page table pointed to by (vpte). 765 * 766 * This implements an N-level page table. Any level can terminate the 767 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 768 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 769 */ 770 static 771 int 772 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 773 vpte_t vpte, int fault_type) 774 { 775 struct sf_buf *sf; 776 int vshift = 32 - PAGE_SHIFT; /* page index bits remaining */ 777 int result = KERN_SUCCESS; 778 vpte_t *ptep; 779 780 for (;;) { 781 /* 782 * We cannot proceed if the vpte is not valid, not readable 783 * for a read fault, or not writable for a write fault. 784 */ 785 if ((vpte & VPTE_V) == 0) { 786 unlock_and_deallocate(fs); 787 return (KERN_FAILURE); 788 } 789 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) { 790 unlock_and_deallocate(fs); 791 return (KERN_FAILURE); 792 } 793 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) { 794 unlock_and_deallocate(fs); 795 return (KERN_FAILURE); 796 } 797 if ((vpte & VPTE_PS) || vshift == 0) 798 break; 799 KKASSERT(vshift >= VPTE_PAGE_BITS); 800 801 /* 802 * Get the page table page. Nominally we only read the page 803 * table, but since we are actively setting VPTE_M and VPTE_A, 804 * tell vm_fault_object() that we are writing it. 805 * 806 * There is currently no real need to optimize this. 807 */ 808 result = vm_fault_object(fs, vpte >> PAGE_SHIFT, 809 VM_PROT_READ|VM_PROT_WRITE); 810 if (result != KERN_SUCCESS) 811 return (result); 812 813 /* 814 * Process the returned fs.m and look up the page table 815 * entry in the page table page. 816 */ 817 vshift -= VPTE_PAGE_BITS; 818 sf = sf_buf_alloc(fs->m, SFB_CPUPRIVATE); 819 ptep = ((vpte_t *)sf_buf_kva(sf) + 820 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 821 vpte = *ptep; 822 823 /* 824 * Page table write-back. If the vpte is valid for the 825 * requested operation, do a write-back to the page table. 826 * 827 * XXX VPTE_M is not set properly for page directory pages. 828 * It doesn't get set in the page directory if the page table 829 * is modified during a read access. 830 */ 831 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) && 832 (vpte & VPTE_W)) { 833 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) { 834 atomic_set_int(ptep, VPTE_M|VPTE_A); 835 vm_page_dirty(fs->m); 836 } 837 } 838 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) && 839 (vpte & VPTE_R)) { 840 if ((vpte & VPTE_A) == 0) { 841 atomic_set_int(ptep, VPTE_A); 842 vm_page_dirty(fs->m); 843 } 844 } 845 sf_buf_free(sf); 846 vm_page_flag_set(fs->m, PG_REFERENCED); 847 vm_page_activate(fs->m); 848 vm_page_wakeup(fs->m); 849 cleanup_successful_fault(fs); 850 } 851 /* 852 * Combine remaining address bits with the vpte. 853 */ 854 *pindex = (vpte >> PAGE_SHIFT) + 855 (*pindex & ((1 << vshift) - 1)); 856 return (KERN_SUCCESS); 857 } 858 859 860 /* 861 * Do all operations required to fault-in (fs.first_object, pindex). Run 862 * through the shadow chain as necessary and do required COW or virtual 863 * copy operations. The caller has already fully resolved the vm_map_entry 864 * and, if appropriate, has created a copy-on-write layer. All we need to 865 * do is iterate the object chain. 866 * 867 * On failure (fs) is unlocked and deallocated and the caller may return or 868 * retry depending on the failure code. On success (fs) is NOT unlocked or 869 * deallocated, fs.m will contained a resolved, busied page, and fs.object 870 * will have an additional PIP count if it is not equal to fs.first_object. 871 */ 872 static 873 int 874 vm_fault_object(struct faultstate *fs, 875 vm_pindex_t first_pindex, vm_prot_t fault_type) 876 { 877 vm_object_t next_object; 878 vm_page_t marray[VM_FAULT_READ]; 879 vm_pindex_t pindex; 880 int faultcount; 881 882 fs->prot = fs->first_prot; 883 fs->object = fs->first_object; 884 pindex = first_pindex; 885 886 /* 887 * If a read fault occurs we try to make the page writable if 888 * possible. There are three cases where we cannot make the 889 * page mapping writable: 890 * 891 * (1) The mapping is read-only or the VM object is read-only, 892 * fs->prot above will simply not have VM_PROT_WRITE set. 893 * 894 * (2) If the mapping is a virtual page table we need to be able 895 * to detect writes so we can set VPTE_M in the virtual page 896 * table. 897 * 898 * (3) If the VM page is read-only or copy-on-write, upgrading would 899 * just result in an unnecessary COW fault. 900 * 901 * VM_PROT_VPAGED is set if faulting via a virtual page table and 902 * causes adjustments to the 'M'odify bit to also turn off write 903 * access to force a re-fault. 904 */ 905 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 906 if ((fault_type & VM_PROT_WRITE) == 0) 907 fs->prot &= ~VM_PROT_WRITE; 908 } 909 910 for (;;) { 911 /* 912 * If the object is dead, we stop here 913 */ 914 if (fs->object->flags & OBJ_DEAD) { 915 unlock_and_deallocate(fs); 916 return (KERN_PROTECTION_FAILURE); 917 } 918 919 /* 920 * See if page is resident. spl protection is required 921 * to avoid an interrupt unbusy/free race against our 922 * lookup. We must hold the protection through a page 923 * allocation or busy. 924 */ 925 crit_enter(); 926 fs->m = vm_page_lookup(fs->object, pindex); 927 if (fs->m != NULL) { 928 int queue; 929 /* 930 * Wait/Retry if the page is busy. We have to do this 931 * if the page is busy via either PG_BUSY or 932 * vm_page_t->busy because the vm_pager may be using 933 * vm_page_t->busy for pageouts ( and even pageins if 934 * it is the vnode pager ), and we could end up trying 935 * to pagein and pageout the same page simultaneously. 936 * 937 * We can theoretically allow the busy case on a read 938 * fault if the page is marked valid, but since such 939 * pages are typically already pmap'd, putting that 940 * special case in might be more effort then it is 941 * worth. We cannot under any circumstances mess 942 * around with a vm_page_t->busy page except, perhaps, 943 * to pmap it. 944 */ 945 if ((fs->m->flags & PG_BUSY) || fs->m->busy) { 946 unlock_things(fs); 947 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 948 mycpu->gd_cnt.v_intrans++; 949 vm_object_deallocate(fs->first_object); 950 fs->first_object = NULL; 951 crit_exit(); 952 return (KERN_TRY_AGAIN); 953 } 954 955 /* 956 * If reactivating a page from PQ_CACHE we may have 957 * to rate-limit. 958 */ 959 queue = fs->m->queue; 960 vm_page_unqueue_nowakeup(fs->m); 961 962 if ((queue - fs->m->pc) == PQ_CACHE && 963 vm_page_count_severe()) { 964 vm_page_activate(fs->m); 965 unlock_and_deallocate(fs); 966 vm_waitpfault(); 967 crit_exit(); 968 return (KERN_TRY_AGAIN); 969 } 970 971 /* 972 * Mark page busy for other processes, and the 973 * pagedaemon. If it still isn't completely valid 974 * (readable), jump to readrest, else we found the 975 * page and can return. 976 * 977 * We can release the spl once we have marked the 978 * page busy. 979 */ 980 vm_page_busy(fs->m); 981 crit_exit(); 982 983 if (((fs->m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 984 fs->m->object != &kernel_object) { 985 goto readrest; 986 } 987 break; /* break to PAGE HAS BEEN FOUND */ 988 } 989 990 /* 991 * Page is not resident, If this is the search termination 992 * or the pager might contain the page, allocate a new page. 993 * 994 * NOTE: We are still in a critical section. 995 */ 996 if (TRYPAGER(fs) || fs->object == fs->first_object) { 997 /* 998 * If the page is beyond the object size we fail 999 */ 1000 if (pindex >= fs->object->size) { 1001 crit_exit(); 1002 unlock_and_deallocate(fs); 1003 return (KERN_PROTECTION_FAILURE); 1004 } 1005 1006 /* 1007 * Ratelimit. 1008 */ 1009 if (fs->didlimit == 0 && curproc != NULL) { 1010 int limticks; 1011 1012 limticks = vm_fault_ratelimit(curproc->p_vmspace); 1013 if (limticks) { 1014 crit_exit(); 1015 unlock_and_deallocate(fs); 1016 tsleep(curproc, 0, "vmrate", limticks); 1017 fs->didlimit = 1; 1018 return (KERN_TRY_AGAIN); 1019 } 1020 } 1021 1022 /* 1023 * Allocate a new page for this object/offset pair. 1024 */ 1025 fs->m = NULL; 1026 if (!vm_page_count_severe()) { 1027 fs->m = vm_page_alloc(fs->object, pindex, 1028 (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO); 1029 } 1030 if (fs->m == NULL) { 1031 crit_exit(); 1032 unlock_and_deallocate(fs); 1033 vm_waitpfault(); 1034 return (KERN_TRY_AGAIN); 1035 } 1036 } 1037 crit_exit(); 1038 1039 readrest: 1040 /* 1041 * We have found a valid page or we have allocated a new page. 1042 * The page thus may not be valid or may not be entirely 1043 * valid. 1044 * 1045 * Attempt to fault-in the page if there is a chance that the 1046 * pager has it, and potentially fault in additional pages 1047 * at the same time. 1048 * 1049 * We are NOT in splvm here and if TRYPAGER is true then 1050 * fs.m will be non-NULL and will be PG_BUSY for us. 1051 */ 1052 1053 if (TRYPAGER(fs)) { 1054 int rv; 1055 int reqpage; 1056 int ahead, behind; 1057 u_char behavior = vm_map_entry_behavior(fs->entry); 1058 1059 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 1060 ahead = 0; 1061 behind = 0; 1062 } else { 1063 behind = pindex; 1064 KKASSERT(behind >= 0); 1065 if (behind > VM_FAULT_READ_BEHIND) 1066 behind = VM_FAULT_READ_BEHIND; 1067 1068 ahead = fs->object->size - pindex; 1069 if (ahead < 1) 1070 ahead = 1; 1071 if (ahead > VM_FAULT_READ_AHEAD) 1072 ahead = VM_FAULT_READ_AHEAD; 1073 } 1074 1075 if ((fs->first_object->type != OBJT_DEVICE) && 1076 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 1077 (behavior != MAP_ENTRY_BEHAV_RANDOM && 1078 pindex >= fs->entry->lastr && 1079 pindex < fs->entry->lastr + VM_FAULT_READ)) 1080 ) { 1081 vm_pindex_t firstpindex, tmppindex; 1082 1083 if (first_pindex < 2 * VM_FAULT_READ) 1084 firstpindex = 0; 1085 else 1086 firstpindex = first_pindex - 2 * VM_FAULT_READ; 1087 1088 /* 1089 * note: partially valid pages cannot be 1090 * included in the lookahead - NFS piecemeal 1091 * writes will barf on it badly. 1092 * 1093 * spl protection is required to avoid races 1094 * between the lookup and an interrupt 1095 * unbusy/free sequence occuring prior to 1096 * our busy check. 1097 */ 1098 crit_enter(); 1099 for (tmppindex = first_pindex - 1; 1100 tmppindex >= firstpindex; 1101 --tmppindex 1102 ) { 1103 vm_page_t mt; 1104 1105 mt = vm_page_lookup(fs->first_object, tmppindex); 1106 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 1107 break; 1108 if (mt->busy || 1109 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 1110 mt->hold_count || 1111 mt->wire_count) 1112 continue; 1113 if (mt->dirty == 0) 1114 vm_page_test_dirty(mt); 1115 if (mt->dirty) { 1116 vm_page_busy(mt); 1117 vm_page_protect(mt, VM_PROT_NONE); 1118 vm_page_deactivate(mt); 1119 vm_page_wakeup(mt); 1120 } else { 1121 vm_page_cache(mt); 1122 } 1123 } 1124 crit_exit(); 1125 1126 ahead += behind; 1127 behind = 0; 1128 } 1129 1130 /* 1131 * now we find out if any other pages should be paged 1132 * in at this time this routine checks to see if the 1133 * pages surrounding this fault reside in the same 1134 * object as the page for this fault. If they do, 1135 * then they are faulted in also into the object. The 1136 * array "marray" returned contains an array of 1137 * vm_page_t structs where one of them is the 1138 * vm_page_t passed to the routine. The reqpage 1139 * return value is the index into the marray for the 1140 * vm_page_t passed to the routine. 1141 * 1142 * fs.m plus the additional pages are PG_BUSY'd. 1143 */ 1144 faultcount = vm_fault_additional_pages( 1145 fs->m, behind, ahead, marray, &reqpage); 1146 1147 /* 1148 * update lastr imperfectly (we do not know how much 1149 * getpages will actually read), but good enough. 1150 */ 1151 fs->entry->lastr = pindex + faultcount - behind; 1152 1153 /* 1154 * Call the pager to retrieve the data, if any, after 1155 * releasing the lock on the map. We hold a ref on 1156 * fs.object and the pages are PG_BUSY'd. 1157 */ 1158 unlock_map(fs); 1159 1160 if (faultcount) { 1161 rv = vm_pager_get_pages(fs->object, marray, 1162 faultcount, reqpage); 1163 } else { 1164 rv = VM_PAGER_FAIL; 1165 } 1166 1167 if (rv == VM_PAGER_OK) { 1168 /* 1169 * Found the page. Leave it busy while we play 1170 * with it. 1171 */ 1172 1173 /* 1174 * Relookup in case pager changed page. Pager 1175 * is responsible for disposition of old page 1176 * if moved. 1177 * 1178 * XXX other code segments do relookups too. 1179 * It's a bad abstraction that needs to be 1180 * fixed/removed. 1181 */ 1182 fs->m = vm_page_lookup(fs->object, pindex); 1183 if (fs->m == NULL) { 1184 unlock_and_deallocate(fs); 1185 return (KERN_TRY_AGAIN); 1186 } 1187 1188 ++fs->hardfault; 1189 break; /* break to PAGE HAS BEEN FOUND */ 1190 } 1191 1192 /* 1193 * Remove the bogus page (which does not exist at this 1194 * object/offset); before doing so, we must get back 1195 * our object lock to preserve our invariant. 1196 * 1197 * Also wake up any other process that may want to bring 1198 * in this page. 1199 * 1200 * If this is the top-level object, we must leave the 1201 * busy page to prevent another process from rushing 1202 * past us, and inserting the page in that object at 1203 * the same time that we are. 1204 */ 1205 if (rv == VM_PAGER_ERROR) { 1206 if (curproc) 1207 kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); 1208 else 1209 kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm); 1210 } 1211 /* 1212 * Data outside the range of the pager or an I/O error 1213 * 1214 * The page may have been wired during the pagein, 1215 * e.g. by the buffer cache, and cannot simply be 1216 * freed. Call vnode_pager_freepag() to deal with it. 1217 */ 1218 /* 1219 * XXX - the check for kernel_map is a kludge to work 1220 * around having the machine panic on a kernel space 1221 * fault w/ I/O error. 1222 */ 1223 if (((fs->map != &kernel_map) && (rv == VM_PAGER_ERROR)) || 1224 (rv == VM_PAGER_BAD)) { 1225 vnode_pager_freepage(fs->m); 1226 fs->m = NULL; 1227 unlock_and_deallocate(fs); 1228 if (rv == VM_PAGER_ERROR) 1229 return (KERN_FAILURE); 1230 else 1231 return (KERN_PROTECTION_FAILURE); 1232 /* NOT REACHED */ 1233 } 1234 if (fs->object != fs->first_object) { 1235 vnode_pager_freepage(fs->m); 1236 fs->m = NULL; 1237 /* 1238 * XXX - we cannot just fall out at this 1239 * point, m has been freed and is invalid! 1240 */ 1241 } 1242 } 1243 1244 /* 1245 * We get here if the object has a default pager (or unwiring) 1246 * or the pager doesn't have the page. 1247 */ 1248 if (fs->object == fs->first_object) 1249 fs->first_m = fs->m; 1250 1251 /* 1252 * Move on to the next object. Lock the next object before 1253 * unlocking the current one. 1254 */ 1255 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1256 next_object = fs->object->backing_object; 1257 if (next_object == NULL) { 1258 /* 1259 * If there's no object left, fill the page in the top 1260 * object with zeros. 1261 */ 1262 if (fs->object != fs->first_object) { 1263 vm_object_pip_wakeup(fs->object); 1264 1265 fs->object = fs->first_object; 1266 pindex = first_pindex; 1267 fs->m = fs->first_m; 1268 } 1269 fs->first_m = NULL; 1270 1271 /* 1272 * Zero the page if necessary and mark it valid. 1273 */ 1274 if ((fs->m->flags & PG_ZERO) == 0) { 1275 vm_page_zero_fill(fs->m); 1276 } else { 1277 mycpu->gd_cnt.v_ozfod++; 1278 } 1279 mycpu->gd_cnt.v_zfod++; 1280 fs->m->valid = VM_PAGE_BITS_ALL; 1281 break; /* break to PAGE HAS BEEN FOUND */ 1282 } else { 1283 if (fs->object != fs->first_object) { 1284 vm_object_pip_wakeup(fs->object); 1285 } 1286 KASSERT(fs->object != next_object, ("object loop %p", next_object)); 1287 fs->object = next_object; 1288 vm_object_pip_add(fs->object, 1); 1289 } 1290 } 1291 1292 KASSERT((fs->m->flags & PG_BUSY) != 0, 1293 ("vm_fault: not busy after main loop")); 1294 1295 /* 1296 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1297 * is held.] 1298 */ 1299 1300 /* 1301 * If the page is being written, but isn't already owned by the 1302 * top-level object, we have to copy it into a new page owned by the 1303 * top-level object. 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 vm_pager_page_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 1692 /* 1693 * This routine checks around the requested page for other pages that 1694 * might be able to be faulted in. This routine brackets the viable 1695 * pages for the pages to be paged in. 1696 * 1697 * Inputs: 1698 * m, rbehind, rahead 1699 * 1700 * Outputs: 1701 * marray (array of vm_page_t), reqpage (index of requested page) 1702 * 1703 * Return value: 1704 * number of pages in marray 1705 */ 1706 static int 1707 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 1708 vm_page_t *marray, int *reqpage) 1709 { 1710 int i,j; 1711 vm_object_t object; 1712 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1713 vm_page_t rtm; 1714 int cbehind, cahead; 1715 1716 object = m->object; 1717 pindex = m->pindex; 1718 1719 /* 1720 * we don't fault-ahead for device pager 1721 */ 1722 if (object->type == OBJT_DEVICE) { 1723 *reqpage = 0; 1724 marray[0] = m; 1725 return 1; 1726 } 1727 1728 /* 1729 * if the requested page is not available, then give up now 1730 */ 1731 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1732 *reqpage = 0; /* not used by caller, fix compiler warn */ 1733 return 0; 1734 } 1735 1736 if ((cbehind == 0) && (cahead == 0)) { 1737 *reqpage = 0; 1738 marray[0] = m; 1739 return 1; 1740 } 1741 1742 if (rahead > cahead) { 1743 rahead = cahead; 1744 } 1745 1746 if (rbehind > cbehind) { 1747 rbehind = cbehind; 1748 } 1749 1750 /* 1751 * Do not do any readahead if we have insufficient free memory. 1752 * 1753 * XXX code was broken disabled before and has instability 1754 * with this conditonal fixed, so shortcut for now. 1755 */ 1756 if (burst_fault == 0 || vm_page_count_severe()) { 1757 marray[0] = m; 1758 *reqpage = 0; 1759 return 1; 1760 } 1761 1762 /* 1763 * scan backward for the read behind pages -- in memory 1764 * 1765 * Assume that if the page is not found an interrupt will not 1766 * create it. Theoretically interrupts can only remove (busy) 1767 * pages, not create new associations. 1768 */ 1769 if (pindex > 0) { 1770 if (rbehind > pindex) { 1771 rbehind = pindex; 1772 startpindex = 0; 1773 } else { 1774 startpindex = pindex - rbehind; 1775 } 1776 1777 crit_enter(); 1778 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 1779 if (vm_page_lookup(object, tpindex - 1)) 1780 break; 1781 } 1782 1783 i = 0; 1784 while (tpindex < pindex) { 1785 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM); 1786 if (rtm == NULL) { 1787 crit_exit(); 1788 for (j = 0; j < i; j++) { 1789 vm_page_free(marray[j]); 1790 } 1791 marray[0] = m; 1792 *reqpage = 0; 1793 return 1; 1794 } 1795 marray[i] = rtm; 1796 ++i; 1797 ++tpindex; 1798 } 1799 crit_exit(); 1800 } else { 1801 i = 0; 1802 } 1803 1804 /* 1805 * Assign requested page 1806 */ 1807 marray[i] = m; 1808 *reqpage = i; 1809 ++i; 1810 1811 /* 1812 * Scan forwards for read-ahead pages 1813 */ 1814 tpindex = pindex + 1; 1815 endpindex = tpindex + rahead; 1816 if (endpindex > object->size) 1817 endpindex = object->size; 1818 1819 crit_enter(); 1820 while (tpindex < endpindex) { 1821 if (vm_page_lookup(object, tpindex)) 1822 break; 1823 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM); 1824 if (rtm == NULL) 1825 break; 1826 marray[i] = rtm; 1827 ++i; 1828 ++tpindex; 1829 } 1830 crit_exit(); 1831 1832 return (i); 1833 } 1834