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 1.41 2007/01/12 22:12:53 dillon Exp $ 71 */ 72 73 /* 74 * Page fault handling module. 75 */ 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/kernel.h> 80 #include <sys/proc.h> 81 #include <sys/vnode.h> 82 #include <sys/resourcevar.h> 83 #include <sys/vmmeter.h> 84 #include <sys/vkernel.h> 85 #include <sys/sfbuf.h> 86 #include <sys/lock.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vnode_pager.h> 98 #include <vm/vm_extern.h> 99 100 #include <sys/thread2.h> 101 #include <vm/vm_page2.h> 102 103 #define VM_FAULT_READ_AHEAD 8 104 #define VM_FAULT_READ_BEHIND 7 105 #define VM_FAULT_READ (VM_FAULT_READ_AHEAD+VM_FAULT_READ_BEHIND+1) 106 107 struct faultstate { 108 vm_page_t m; 109 vm_object_t object; 110 vm_pindex_t pindex; 111 vm_prot_t prot; 112 vm_page_t first_m; 113 vm_object_t first_object; 114 vm_prot_t first_prot; 115 vm_map_t map; 116 vm_map_entry_t entry; 117 int lookup_still_valid; 118 int didlimit; 119 int hardfault; 120 int fault_flags; 121 int map_generation; 122 boolean_t wired; 123 struct vnode *vp; 124 }; 125 126 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t); 127 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int); 128 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 129 static int vm_fault_ratelimit(struct vmspace *); 130 131 static __inline void 132 release_page(struct faultstate *fs) 133 { 134 vm_page_wakeup(fs->m); 135 vm_page_deactivate(fs->m); 136 fs->m = NULL; 137 } 138 139 static __inline void 140 unlock_map(struct faultstate *fs) 141 { 142 if (fs->lookup_still_valid) { 143 vm_map_lookup_done(fs->map, fs->entry, 0); 144 fs->lookup_still_valid = FALSE; 145 } 146 } 147 148 /* 149 * Clean up after a successful call to vm_fault_object() so another call 150 * to vm_fault_object() can be made. 151 */ 152 static void 153 _cleanup_successful_fault(struct faultstate *fs, int relock) 154 { 155 if (fs->object != fs->first_object) { 156 vm_page_free(fs->first_m); 157 vm_object_pip_wakeup(fs->object); 158 fs->first_m = NULL; 159 } 160 fs->object = fs->first_object; 161 if (relock && fs->lookup_still_valid == FALSE) { 162 vm_map_lock_read(fs->map); 163 fs->lookup_still_valid = TRUE; 164 } 165 } 166 167 static void 168 _unlock_things(struct faultstate *fs, int dealloc) 169 { 170 vm_object_pip_wakeup(fs->first_object); 171 _cleanup_successful_fault(fs, 0); 172 if (dealloc) { 173 vm_object_deallocate(fs->first_object); 174 } 175 unlock_map(fs); 176 if (fs->vp != NULL) { 177 vput(fs->vp); 178 fs->vp = NULL; 179 } 180 } 181 182 #define unlock_things(fs) _unlock_things(fs, 0) 183 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 184 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 185 186 /* 187 * TRYPAGER 188 * 189 * Determine if the pager for the current object *might* contain the page. 190 * 191 * We only need to try the pager if this is not a default object (default 192 * objects are zero-fill and have no real pager), and if we are not taking 193 * a wiring fault or if the FS entry is wired. 194 */ 195 #define TRYPAGER(fs) \ 196 (fs->object->type != OBJT_DEFAULT && \ 197 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired)) 198 199 /* 200 * vm_fault: 201 * 202 * Handle a page fault occuring at the given address, requiring the given 203 * permissions, in the map specified. If successful, the page is inserted 204 * into the associated physical map. 205 * 206 * NOTE: The given address should be truncated to the proper page address. 207 * 208 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 209 * a standard error specifying why the fault is fatal is returned. 210 * 211 * The map in question must be referenced, and remains so. 212 * The caller may hold no locks. 213 */ 214 int 215 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 216 { 217 int result; 218 vm_pindex_t first_pindex; 219 struct faultstate fs; 220 221 mycpu->gd_cnt.v_vm_faults++; 222 223 fs.didlimit = 0; 224 fs.hardfault = 0; 225 fs.fault_flags = fault_flags; 226 227 RetryFault: 228 /* 229 * Find the vm_map_entry representing the backing store and resolve 230 * the top level object and page index. This may have the side 231 * effect of executing a copy-on-write on the map entry and/or 232 * creating a shadow object, but will not COW any actual VM pages. 233 * 234 * On success fs.map is left read-locked and various other fields 235 * are initialized but not otherwise referenced or locked. 236 * 237 * NOTE! vm_map_lookup will try to upgrade the fault_type to 238 * VM_FAULT_WRITE if the map entry is a virtual page table and also 239 * writable, so we can set the 'A'accessed bit in the virtual page 240 * table entry. 241 */ 242 fs.map = map; 243 result = vm_map_lookup(&fs.map, vaddr, fault_type, 244 &fs.entry, &fs.first_object, 245 &first_pindex, &fs.first_prot, &fs.wired); 246 247 /* 248 * If the lookup failed or the map protections are incompatible, 249 * the fault generally fails. However, if the caller is trying 250 * to do a user wiring we have more work to do. 251 */ 252 if (result != KERN_SUCCESS) { 253 if (result != KERN_PROTECTION_FAILURE) 254 return result; 255 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 256 return result; 257 258 /* 259 * If we are user-wiring a r/w segment, and it is COW, then 260 * we need to do the COW operation. Note that we don't 261 * currently COW RO sections now, because it is NOT desirable 262 * to COW .text. We simply keep .text from ever being COW'ed 263 * and take the heat that one cannot debug wired .text sections. 264 */ 265 result = vm_map_lookup(&fs.map, vaddr, 266 VM_PROT_READ|VM_PROT_WRITE| 267 VM_PROT_OVERRIDE_WRITE, 268 &fs.entry, &fs.first_object, 269 &first_pindex, &fs.first_prot, 270 &fs.wired); 271 if (result != KERN_SUCCESS) 272 return result; 273 274 /* 275 * If we don't COW now, on a user wire, the user will never 276 * be able to write to the mapping. If we don't make this 277 * restriction, the bookkeeping would be nearly impossible. 278 */ 279 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 280 fs.entry->max_protection &= ~VM_PROT_WRITE; 281 } 282 283 /* 284 * fs.map is read-locked 285 * 286 * Misc checks. Save the map generation number to detect races. 287 */ 288 fs.map_generation = fs.map->timestamp; 289 290 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 291 panic("vm_fault: fault on nofault entry, addr: %lx", 292 (u_long)vaddr); 293 } 294 295 /* 296 * A system map entry may return a NULL object. No object means 297 * no pager means an unrecoverable kernel fault. 298 */ 299 if (fs.first_object == NULL) { 300 panic("vm_fault: unrecoverable fault at %p in entry %p", 301 (void *)vaddr, fs.entry); 302 } 303 304 /* 305 * Make a reference to this object to prevent its disposal while we 306 * are messing with it. Once we have the reference, the map is free 307 * to be diddled. Since objects reference their shadows (and copies), 308 * they will stay around as well. 309 * 310 * Bump the paging-in-progress count to prevent size changes (e.g. 311 * truncation operations) during I/O. This must be done after 312 * obtaining the vnode lock in order to avoid possible deadlocks. 313 */ 314 vm_object_reference(fs.first_object); 315 fs.vp = vnode_pager_lock(fs.first_object); 316 vm_object_pip_add(fs.first_object, 1); 317 318 fs.lookup_still_valid = TRUE; 319 fs.first_m = NULL; 320 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 321 322 /* 323 * If the entry is wired we cannot change the page protection. 324 */ 325 if (fs.wired) 326 fault_type = fs.first_prot; 327 328 /* 329 * The page we want is at (first_object, first_pindex), but if the 330 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 331 * page table to figure out the actual pindex. 332 * 333 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 334 * ONLY 335 */ 336 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 337 result = vm_fault_vpagetable(&fs, &first_pindex, 338 fs.entry->aux.master_pde, 339 fault_type); 340 if (result == KERN_TRY_AGAIN) 341 goto RetryFault; 342 if (result != KERN_SUCCESS) 343 return (result); 344 } 345 346 /* 347 * Now we have the actual (object, pindex), fault in the page. If 348 * vm_fault_object() fails it will unlock and deallocate the FS 349 * data. If it succeeds everything remains locked and fs->object 350 * will have an additinal PIP count if it is not equal to 351 * fs->first_object 352 * 353 * vm_fault_object will set fs->prot for the pmap operation. It is 354 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 355 * page can be safely written. However, it will force a read-only 356 * mapping for a read fault if the memory is managed by a virtual 357 * page table. 358 */ 359 result = vm_fault_object(&fs, first_pindex, fault_type); 360 361 if (result == KERN_TRY_AGAIN) 362 goto RetryFault; 363 if (result != KERN_SUCCESS) 364 return (result); 365 366 /* 367 * On success vm_fault_object() does not unlock or deallocate, and fs.m 368 * will contain a busied page. 369 * 370 * Enter the page into the pmap and do pmap-related adjustments. 371 */ 372 unlock_things(&fs); 373 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired); 374 375 if (((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0) && (fs.wired == 0)) { 376 pmap_prefault(fs.map->pmap, vaddr, fs.entry); 377 } 378 379 vm_page_flag_clear(fs.m, PG_ZERO); 380 vm_page_flag_set(fs.m, PG_MAPPED|PG_REFERENCED); 381 382 /* 383 * If the page is not wired down, then put it where the pageout daemon 384 * can find it. 385 */ 386 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 387 if (fs.wired) 388 vm_page_wire(fs.m); 389 else 390 vm_page_unwire(fs.m, 1); 391 } else { 392 vm_page_activate(fs.m); 393 } 394 395 if (curthread->td_lwp) { 396 if (fs.hardfault) { 397 curthread->td_lwp->lwp_ru.ru_majflt++; 398 } else { 399 curthread->td_lwp->lwp_ru.ru_minflt++; 400 } 401 } 402 403 /* 404 * Unlock everything, and return 405 */ 406 vm_page_wakeup(fs.m); 407 vm_object_deallocate(fs.first_object); 408 409 return (KERN_SUCCESS); 410 } 411 412 /* 413 * Fault in the specified virtual address in the current process map, 414 * returning a held VM page or NULL. See vm_fault_page() for more 415 * information. 416 */ 417 vm_page_t 418 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp) 419 { 420 vm_page_t m; 421 422 m = vm_fault_page(&curproc->p_vmspace->vm_map, va, 423 fault_type, VM_FAULT_NORMAL, errorp); 424 return(m); 425 } 426 427 /* 428 * Fault in the specified virtual address in the specified map, doing all 429 * necessary manipulation of the object store and all necessary I/O. Return 430 * a held VM page or NULL, and set *errorp. The related pmap is not 431 * updated. 432 * 433 * The returned page will be properly dirtied if VM_PROT_WRITE was specified, 434 * and marked PG_REFERENCED as well. 435 */ 436 vm_page_t 437 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 438 int fault_flags, int *errorp) 439 { 440 int result; 441 vm_pindex_t first_pindex; 442 struct faultstate fs; 443 444 mycpu->gd_cnt.v_vm_faults++; 445 446 fs.didlimit = 0; 447 fs.hardfault = 0; 448 fs.fault_flags = fault_flags; 449 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 450 451 RetryFault: 452 /* 453 * Find the vm_map_entry representing the backing store and resolve 454 * the top level object and page index. This may have the side 455 * effect of executing a copy-on-write on the map entry and/or 456 * creating a shadow object, but will not COW any actual VM pages. 457 * 458 * On success fs.map is left read-locked and various other fields 459 * are initialized but not otherwise referenced or locked. 460 * 461 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 462 * if the map entry is a virtual page table and also writable, 463 * so we can set the 'A'accessed bit in the virtual page table entry. 464 */ 465 fs.map = map; 466 result = vm_map_lookup(&fs.map, vaddr, fault_type, 467 &fs.entry, &fs.first_object, 468 &first_pindex, &fs.first_prot, &fs.wired); 469 470 if (result != KERN_SUCCESS) { 471 *errorp = result; 472 return (NULL); 473 } 474 475 /* 476 * fs.map is read-locked 477 * 478 * Misc checks. Save the map generation number to detect races. 479 */ 480 fs.map_generation = fs.map->timestamp; 481 482 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 483 panic("vm_fault: fault on nofault entry, addr: %lx", 484 (u_long)vaddr); 485 } 486 487 /* 488 * A system map entry may return a NULL object. No object means 489 * no pager means an unrecoverable kernel fault. 490 */ 491 if (fs.first_object == NULL) { 492 panic("vm_fault: unrecoverable fault at %p in entry %p", 493 (void *)vaddr, fs.entry); 494 } 495 496 /* 497 * Make a reference to this object to prevent its disposal while we 498 * are messing with it. Once we have the reference, the map is free 499 * to be diddled. Since objects reference their shadows (and copies), 500 * they will stay around as well. 501 * 502 * Bump the paging-in-progress count to prevent size changes (e.g. 503 * truncation operations) during I/O. This must be done after 504 * obtaining the vnode lock in order to avoid possible deadlocks. 505 */ 506 vm_object_reference(fs.first_object); 507 fs.vp = vnode_pager_lock(fs.first_object); 508 vm_object_pip_add(fs.first_object, 1); 509 510 fs.lookup_still_valid = TRUE; 511 fs.first_m = NULL; 512 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 513 514 /* 515 * If the entry is wired we cannot change the page protection. 516 */ 517 if (fs.wired) 518 fault_type = fs.first_prot; 519 520 /* 521 * The page we want is at (first_object, first_pindex), but if the 522 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 523 * page table to figure out the actual pindex. 524 * 525 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 526 * ONLY 527 */ 528 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 529 result = vm_fault_vpagetable(&fs, &first_pindex, 530 fs.entry->aux.master_pde, 531 fault_type); 532 if (result == KERN_TRY_AGAIN) 533 goto RetryFault; 534 if (result != KERN_SUCCESS) { 535 *errorp = result; 536 return (NULL); 537 } 538 } 539 540 /* 541 * Now we have the actual (object, pindex), fault in the page. If 542 * vm_fault_object() fails it will unlock and deallocate the FS 543 * data. If it succeeds everything remains locked and fs->object 544 * will have an additinal PIP count if it is not equal to 545 * fs->first_object 546 */ 547 result = vm_fault_object(&fs, first_pindex, fault_type); 548 549 if (result == KERN_TRY_AGAIN) 550 goto RetryFault; 551 if (result != KERN_SUCCESS) { 552 *errorp = result; 553 return(NULL); 554 } 555 556 /* 557 * On success vm_fault_object() does not unlock or deallocate, and fs.m 558 * will contain a busied page. 559 */ 560 unlock_things(&fs); 561 562 /* 563 * Return a held page. We are not doing any pmap manipulation so do 564 * not set PG_MAPPED. However, adjust the page flags according to 565 * the fault type because the caller may not use a managed pmapping 566 * (so we don't want to lose the fact that the page will be dirtied 567 * if a write fault was specified). 568 */ 569 vm_page_hold(fs.m); 570 vm_page_flag_clear(fs.m, PG_ZERO); 571 if (fault_type & VM_PROT_WRITE) 572 vm_page_dirty(fs.m); 573 574 /* 575 * Update the pmap. We really only have to do this if a COW 576 * occured to replace the read-only page with the new page. For 577 * now just do it unconditionally. XXX 578 */ 579 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired); 580 vm_page_flag_set(fs.m, PG_REFERENCED|PG_MAPPED); 581 582 /* 583 * Unbusy the page by activating it. It remains held and will not 584 * be reclaimed. 585 */ 586 vm_page_activate(fs.m); 587 588 if (curthread->td_lwp) { 589 if (fs.hardfault) { 590 curthread->td_lwp->lwp_ru.ru_majflt++; 591 } else { 592 curthread->td_lwp->lwp_ru.ru_minflt++; 593 } 594 } 595 596 /* 597 * Unlock everything, and return the held page. 598 */ 599 vm_page_wakeup(fs.m); 600 vm_object_deallocate(fs.first_object); 601 602 *errorp = 0; 603 return(fs.m); 604 } 605 606 /* 607 * Translate the virtual page number (first_pindex) that is relative 608 * to the address space into a logical page number that is relative to the 609 * backing object. Use the virtual page table pointed to by (vpte). 610 * 611 * This implements an N-level page table. Any level can terminate the 612 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 613 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 614 */ 615 static 616 int 617 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 618 vpte_t vpte, int fault_type) 619 { 620 struct sf_buf *sf; 621 int vshift = 32 - PAGE_SHIFT; /* page index bits remaining */ 622 int result = KERN_SUCCESS; 623 vpte_t *ptep; 624 625 for (;;) { 626 /* 627 * We cannot proceed if the vpte is not valid, not readable 628 * for a read fault, or not writable for a write fault. 629 */ 630 if ((vpte & VPTE_V) == 0) { 631 unlock_and_deallocate(fs); 632 return (KERN_FAILURE); 633 } 634 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) { 635 unlock_and_deallocate(fs); 636 return (KERN_FAILURE); 637 } 638 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) { 639 unlock_and_deallocate(fs); 640 return (KERN_FAILURE); 641 } 642 if ((vpte & VPTE_PS) || vshift == 0) 643 break; 644 KKASSERT(vshift >= VPTE_PAGE_BITS); 645 646 /* 647 * Get the page table page. Nominally we only read the page 648 * table, but since we are actively setting VPTE_M and VPTE_A, 649 * tell vm_fault_object() that we are writing it. 650 * 651 * There is currently no real need to optimize this. 652 */ 653 result = vm_fault_object(fs, vpte >> PAGE_SHIFT, 654 VM_PROT_READ|VM_PROT_WRITE); 655 if (result != KERN_SUCCESS) 656 return (result); 657 658 /* 659 * Process the returned fs.m and look up the page table 660 * entry in the page table page. 661 */ 662 vshift -= VPTE_PAGE_BITS; 663 sf = sf_buf_alloc(fs->m, SFB_CPUPRIVATE); 664 ptep = ((vpte_t *)sf_buf_kva(sf) + 665 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 666 vpte = *ptep; 667 668 /* 669 * Page table write-back. If the vpte is valid for the 670 * requested operation, do a write-back to the page table. 671 * 672 * XXX VPTE_M is not set properly for page directory pages. 673 * It doesn't get set in the page directory if the page table 674 * is modified during a read access. 675 */ 676 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) && 677 (vpte & VPTE_W)) { 678 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) { 679 atomic_set_int(ptep, VPTE_M|VPTE_A); 680 vm_page_dirty(fs->m); 681 } 682 } 683 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) && 684 (vpte & VPTE_R)) { 685 if ((vpte & VPTE_A) == 0) { 686 atomic_set_int(ptep, VPTE_A); 687 vm_page_dirty(fs->m); 688 } 689 } 690 sf_buf_free(sf); 691 vm_page_flag_set(fs->m, PG_REFERENCED); 692 vm_page_activate(fs->m); 693 vm_page_wakeup(fs->m); 694 cleanup_successful_fault(fs); 695 } 696 /* 697 * Combine remaining address bits with the vpte. 698 */ 699 *pindex = (vpte >> PAGE_SHIFT) + 700 (*pindex & ((1 << vshift) - 1)); 701 return (KERN_SUCCESS); 702 } 703 704 705 /* 706 * Do all operations required to fault-in (fs.first_object, pindex). Run 707 * through the shadow chain as necessary and do required COW or virtual 708 * copy operations. The caller has already fully resolved the vm_map_entry 709 * and, if appropriate, has created a copy-on-write layer. All we need to 710 * do is iterate the object chain. 711 * 712 * On failure (fs) is unlocked and deallocated and the caller may return or 713 * retry depending on the failure code. On success (fs) is NOT unlocked or 714 * deallocated, fs.m will contained a resolved, busied page, and fs.object 715 * will have an additional PIP count if it is not equal to fs.first_object. 716 */ 717 static 718 int 719 vm_fault_object(struct faultstate *fs, 720 vm_pindex_t first_pindex, vm_prot_t fault_type) 721 { 722 vm_object_t next_object; 723 vm_page_t marray[VM_FAULT_READ]; 724 vm_pindex_t pindex; 725 int faultcount; 726 727 fs->prot = fs->first_prot; 728 fs->object = fs->first_object; 729 pindex = first_pindex; 730 731 /* 732 * If a read fault occurs we try to make the page writable if 733 * possible. There are three cases where we cannot make the 734 * page mapping writable: 735 * 736 * (1) The mapping is read-only or the VM object is read-only, 737 * fs->prot above will simply not have VM_PROT_WRITE set. 738 * 739 * (2) If the mapping is a virtual page table we need to be able 740 * to detect writes so we can set VPTE_M in the virtual page 741 * table. 742 * 743 * (3) If the VM page is read-only or copy-on-write, upgrading would 744 * just result in an unnecessary COW fault. 745 * 746 * VM_PROT_VPAGED is set if faulting via a virtual page table and 747 * causes adjustments to the 'M'odify bit to also turn off write 748 * access to force a re-fault. 749 */ 750 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 751 if ((fault_type & VM_PROT_WRITE) == 0) 752 fs->prot &= ~VM_PROT_WRITE; 753 } 754 755 for (;;) { 756 /* 757 * If the object is dead, we stop here 758 */ 759 if (fs->object->flags & OBJ_DEAD) { 760 unlock_and_deallocate(fs); 761 return (KERN_PROTECTION_FAILURE); 762 } 763 764 /* 765 * See if page is resident. spl protection is required 766 * to avoid an interrupt unbusy/free race against our 767 * lookup. We must hold the protection through a page 768 * allocation or busy. 769 */ 770 crit_enter(); 771 fs->m = vm_page_lookup(fs->object, pindex); 772 if (fs->m != NULL) { 773 int queue; 774 /* 775 * Wait/Retry if the page is busy. We have to do this 776 * if the page is busy via either PG_BUSY or 777 * vm_page_t->busy because the vm_pager may be using 778 * vm_page_t->busy for pageouts ( and even pageins if 779 * it is the vnode pager ), and we could end up trying 780 * to pagein and pageout the same page simultaneously. 781 * 782 * We can theoretically allow the busy case on a read 783 * fault if the page is marked valid, but since such 784 * pages are typically already pmap'd, putting that 785 * special case in might be more effort then it is 786 * worth. We cannot under any circumstances mess 787 * around with a vm_page_t->busy page except, perhaps, 788 * to pmap it. 789 */ 790 if ((fs->m->flags & PG_BUSY) || fs->m->busy) { 791 unlock_things(fs); 792 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 793 mycpu->gd_cnt.v_intrans++; 794 vm_object_deallocate(fs->first_object); 795 crit_exit(); 796 return (KERN_TRY_AGAIN); 797 } 798 799 /* 800 * If reactivating a page from PQ_CACHE we may have 801 * to rate-limit. 802 */ 803 queue = fs->m->queue; 804 vm_page_unqueue_nowakeup(fs->m); 805 806 if ((queue - fs->m->pc) == PQ_CACHE && 807 vm_page_count_severe()) { 808 vm_page_activate(fs->m); 809 unlock_and_deallocate(fs); 810 vm_waitpfault(); 811 crit_exit(); 812 return (KERN_TRY_AGAIN); 813 } 814 815 /* 816 * Mark page busy for other processes, and the 817 * pagedaemon. If it still isn't completely valid 818 * (readable), jump to readrest, else we found the 819 * page and can return. 820 * 821 * We can release the spl once we have marked the 822 * page busy. 823 */ 824 vm_page_busy(fs->m); 825 crit_exit(); 826 827 if (((fs->m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) && 828 fs->m->object != &kernel_object) { 829 goto readrest; 830 } 831 break; /* break to PAGE HAS BEEN FOUND */ 832 } 833 834 /* 835 * Page is not resident, If this is the search termination 836 * or the pager might contain the page, allocate a new page. 837 * 838 * NOTE: We are still in a critical section. 839 */ 840 if (TRYPAGER(fs) || fs->object == fs->first_object) { 841 /* 842 * If the page is beyond the object size we fail 843 */ 844 if (pindex >= fs->object->size) { 845 crit_exit(); 846 unlock_and_deallocate(fs); 847 return (KERN_PROTECTION_FAILURE); 848 } 849 850 /* 851 * Ratelimit. 852 */ 853 if (fs->didlimit == 0 && curproc != NULL) { 854 int limticks; 855 856 limticks = vm_fault_ratelimit(curproc->p_vmspace); 857 if (limticks) { 858 crit_exit(); 859 unlock_and_deallocate(fs); 860 tsleep(curproc, 0, "vmrate", limticks); 861 fs->didlimit = 1; 862 return (KERN_TRY_AGAIN); 863 } 864 } 865 866 /* 867 * Allocate a new page for this object/offset pair. 868 */ 869 fs->m = NULL; 870 if (!vm_page_count_severe()) { 871 fs->m = vm_page_alloc(fs->object, pindex, 872 (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO); 873 } 874 if (fs->m == NULL) { 875 crit_exit(); 876 unlock_and_deallocate(fs); 877 vm_waitpfault(); 878 return (KERN_TRY_AGAIN); 879 } 880 } 881 crit_exit(); 882 883 readrest: 884 /* 885 * We have found a valid page or we have allocated a new page. 886 * The page thus may not be valid or may not be entirely 887 * valid. 888 * 889 * Attempt to fault-in the page if there is a chance that the 890 * pager has it, and potentially fault in additional pages 891 * at the same time. 892 * 893 * We are NOT in splvm here and if TRYPAGER is true then 894 * fs.m will be non-NULL and will be PG_BUSY for us. 895 */ 896 897 if (TRYPAGER(fs)) { 898 int rv; 899 int reqpage; 900 int ahead, behind; 901 u_char behavior = vm_map_entry_behavior(fs->entry); 902 903 if (behavior == MAP_ENTRY_BEHAV_RANDOM) { 904 ahead = 0; 905 behind = 0; 906 } else { 907 behind = pindex; 908 if (behind > VM_FAULT_READ_BEHIND) 909 behind = VM_FAULT_READ_BEHIND; 910 911 ahead = fs->object->size - pindex; 912 if (ahead < 1) 913 ahead = 1; 914 if (ahead > VM_FAULT_READ_AHEAD) 915 ahead = VM_FAULT_READ_AHEAD; 916 } 917 918 if ((fs->first_object->type != OBJT_DEVICE) && 919 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 920 (behavior != MAP_ENTRY_BEHAV_RANDOM && 921 pindex >= fs->entry->lastr && 922 pindex < fs->entry->lastr + VM_FAULT_READ)) 923 ) { 924 vm_pindex_t firstpindex, tmppindex; 925 926 if (first_pindex < 2 * VM_FAULT_READ) 927 firstpindex = 0; 928 else 929 firstpindex = first_pindex - 2 * VM_FAULT_READ; 930 931 /* 932 * note: partially valid pages cannot be 933 * included in the lookahead - NFS piecemeal 934 * writes will barf on it badly. 935 * 936 * spl protection is required to avoid races 937 * between the lookup and an interrupt 938 * unbusy/free sequence occuring prior to 939 * our busy check. 940 */ 941 crit_enter(); 942 for (tmppindex = first_pindex - 1; 943 tmppindex >= firstpindex; 944 --tmppindex 945 ) { 946 vm_page_t mt; 947 948 mt = vm_page_lookup(fs->first_object, tmppindex); 949 if (mt == NULL || (mt->valid != VM_PAGE_BITS_ALL)) 950 break; 951 if (mt->busy || 952 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) || 953 mt->hold_count || 954 mt->wire_count) 955 continue; 956 if (mt->dirty == 0) 957 vm_page_test_dirty(mt); 958 if (mt->dirty) { 959 vm_page_protect(mt, VM_PROT_NONE); 960 vm_page_deactivate(mt); 961 } else { 962 vm_page_cache(mt); 963 } 964 } 965 crit_exit(); 966 967 ahead += behind; 968 behind = 0; 969 } 970 971 /* 972 * now we find out if any other pages should be paged 973 * in at this time this routine checks to see if the 974 * pages surrounding this fault reside in the same 975 * object as the page for this fault. If they do, 976 * then they are faulted in also into the object. The 977 * array "marray" returned contains an array of 978 * vm_page_t structs where one of them is the 979 * vm_page_t passed to the routine. The reqpage 980 * return value is the index into the marray for the 981 * vm_page_t passed to the routine. 982 * 983 * fs.m plus the additional pages are PG_BUSY'd. 984 */ 985 faultcount = vm_fault_additional_pages( 986 fs->m, behind, ahead, marray, &reqpage); 987 988 /* 989 * update lastr imperfectly (we do not know how much 990 * getpages will actually read), but good enough. 991 */ 992 fs->entry->lastr = pindex + faultcount - behind; 993 994 /* 995 * Call the pager to retrieve the data, if any, after 996 * releasing the lock on the map. We hold a ref on 997 * fs.object and the pages are PG_BUSY'd. 998 */ 999 unlock_map(fs); 1000 1001 if (faultcount) { 1002 rv = vm_pager_get_pages(fs->object, marray, 1003 faultcount, reqpage); 1004 } else { 1005 rv = VM_PAGER_FAIL; 1006 } 1007 1008 if (rv == VM_PAGER_OK) { 1009 /* 1010 * Found the page. Leave it busy while we play 1011 * with it. 1012 */ 1013 1014 /* 1015 * Relookup in case pager changed page. Pager 1016 * is responsible for disposition of old page 1017 * if moved. 1018 * 1019 * XXX other code segments do relookups too. 1020 * It's a bad abstraction that needs to be 1021 * fixed/removed. 1022 */ 1023 fs->m = vm_page_lookup(fs->object, pindex); 1024 if (fs->m == NULL) { 1025 unlock_and_deallocate(fs); 1026 return (KERN_TRY_AGAIN); 1027 } 1028 1029 ++fs->hardfault; 1030 break; /* break to PAGE HAS BEEN FOUND */ 1031 } 1032 1033 /* 1034 * Remove the bogus page (which does not exist at this 1035 * object/offset); before doing so, we must get back 1036 * our object lock to preserve our invariant. 1037 * 1038 * Also wake up any other process that may want to bring 1039 * in this page. 1040 * 1041 * If this is the top-level object, we must leave the 1042 * busy page to prevent another process from rushing 1043 * past us, and inserting the page in that object at 1044 * the same time that we are. 1045 */ 1046 if (rv == VM_PAGER_ERROR) { 1047 if (curproc) 1048 kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm); 1049 else 1050 kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm); 1051 } 1052 /* 1053 * Data outside the range of the pager or an I/O error 1054 */ 1055 /* 1056 * XXX - the check for kernel_map is a kludge to work 1057 * around having the machine panic on a kernel space 1058 * fault w/ I/O error. 1059 */ 1060 if (((fs->map != &kernel_map) && (rv == VM_PAGER_ERROR)) || 1061 (rv == VM_PAGER_BAD)) { 1062 vm_page_free(fs->m); 1063 fs->m = NULL; 1064 unlock_and_deallocate(fs); 1065 if (rv == VM_PAGER_ERROR) 1066 return (KERN_FAILURE); 1067 else 1068 return (KERN_PROTECTION_FAILURE); 1069 /* NOT REACHED */ 1070 } 1071 if (fs->object != fs->first_object) { 1072 vm_page_free(fs->m); 1073 fs->m = NULL; 1074 /* 1075 * XXX - we cannot just fall out at this 1076 * point, m has been freed and is invalid! 1077 */ 1078 } 1079 } 1080 1081 /* 1082 * We get here if the object has a default pager (or unwiring) 1083 * or the pager doesn't have the page. 1084 */ 1085 if (fs->object == fs->first_object) 1086 fs->first_m = fs->m; 1087 1088 /* 1089 * Move on to the next object. Lock the next object before 1090 * unlocking the current one. 1091 */ 1092 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1093 next_object = fs->object->backing_object; 1094 if (next_object == NULL) { 1095 /* 1096 * If there's no object left, fill the page in the top 1097 * object with zeros. 1098 */ 1099 if (fs->object != fs->first_object) { 1100 vm_object_pip_wakeup(fs->object); 1101 1102 fs->object = fs->first_object; 1103 pindex = first_pindex; 1104 fs->m = fs->first_m; 1105 } 1106 fs->first_m = NULL; 1107 1108 /* 1109 * Zero the page if necessary and mark it valid. 1110 */ 1111 if ((fs->m->flags & PG_ZERO) == 0) { 1112 vm_page_zero_fill(fs->m); 1113 } else { 1114 mycpu->gd_cnt.v_ozfod++; 1115 } 1116 mycpu->gd_cnt.v_zfod++; 1117 fs->m->valid = VM_PAGE_BITS_ALL; 1118 break; /* break to PAGE HAS BEEN FOUND */ 1119 } else { 1120 if (fs->object != fs->first_object) { 1121 vm_object_pip_wakeup(fs->object); 1122 } 1123 KASSERT(fs->object != next_object, ("object loop %p", next_object)); 1124 fs->object = next_object; 1125 vm_object_pip_add(fs->object, 1); 1126 } 1127 } 1128 1129 KASSERT((fs->m->flags & PG_BUSY) != 0, 1130 ("vm_fault: not busy after main loop")); 1131 1132 /* 1133 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1134 * is held.] 1135 */ 1136 1137 /* 1138 * If the page is being written, but isn't already owned by the 1139 * top-level object, we have to copy it into a new page owned by the 1140 * top-level object. 1141 */ 1142 if (fs->object != fs->first_object) { 1143 /* 1144 * We only really need to copy if we want to write it. 1145 */ 1146 if (fault_type & VM_PROT_WRITE) { 1147 /* 1148 * This allows pages to be virtually copied from a 1149 * backing_object into the first_object, where the 1150 * backing object has no other refs to it, and cannot 1151 * gain any more refs. Instead of a bcopy, we just 1152 * move the page from the backing object to the 1153 * first object. Note that we must mark the page 1154 * dirty in the first object so that it will go out 1155 * to swap when needed. 1156 */ 1157 if (fs->map_generation == fs->map->timestamp && 1158 /* 1159 * Only one shadow object 1160 */ 1161 (fs->object->shadow_count == 1) && 1162 /* 1163 * No COW refs, except us 1164 */ 1165 (fs->object->ref_count == 1) && 1166 /* 1167 * No one else can look this object up 1168 */ 1169 (fs->object->handle == NULL) && 1170 /* 1171 * No other ways to look the object up 1172 */ 1173 ((fs->object->type == OBJT_DEFAULT) || 1174 (fs->object->type == OBJT_SWAP)) && 1175 /* 1176 * We don't chase down the shadow chain 1177 */ 1178 (fs->object == fs->first_object->backing_object) && 1179 1180 /* 1181 * grab the lock if we need to 1182 */ 1183 (fs->lookup_still_valid || 1184 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 1185 ) { 1186 1187 fs->lookup_still_valid = 1; 1188 /* 1189 * get rid of the unnecessary page 1190 */ 1191 vm_page_protect(fs->first_m, VM_PROT_NONE); 1192 vm_page_free(fs->first_m); 1193 fs->first_m = NULL; 1194 1195 /* 1196 * grab the page and put it into the 1197 * process'es object. The page is 1198 * automatically made dirty. 1199 */ 1200 vm_page_rename(fs->m, fs->first_object, first_pindex); 1201 fs->first_m = fs->m; 1202 vm_page_busy(fs->first_m); 1203 fs->m = NULL; 1204 mycpu->gd_cnt.v_cow_optim++; 1205 } else { 1206 /* 1207 * Oh, well, lets copy it. 1208 */ 1209 vm_page_copy(fs->m, fs->first_m); 1210 } 1211 1212 if (fs->m) { 1213 /* 1214 * We no longer need the old page or object. 1215 */ 1216 release_page(fs); 1217 } 1218 1219 /* 1220 * fs->object != fs->first_object due to above 1221 * conditional 1222 */ 1223 vm_object_pip_wakeup(fs->object); 1224 1225 /* 1226 * Only use the new page below... 1227 */ 1228 1229 mycpu->gd_cnt.v_cow_faults++; 1230 fs->m = fs->first_m; 1231 fs->object = fs->first_object; 1232 pindex = first_pindex; 1233 } else { 1234 /* 1235 * If it wasn't a write fault avoid having to copy 1236 * the page by mapping it read-only. 1237 */ 1238 fs->prot &= ~VM_PROT_WRITE; 1239 } 1240 } 1241 1242 /* 1243 * We may have had to unlock a map to do I/O. If we did then 1244 * lookup_still_valid will be FALSE. If the map generation count 1245 * also changed then all sorts of things could have happened while 1246 * we were doing the I/O and we need to retry. 1247 */ 1248 1249 if (!fs->lookup_still_valid && 1250 (fs->map->timestamp != fs->map_generation)) { 1251 release_page(fs); 1252 unlock_and_deallocate(fs); 1253 return (KERN_TRY_AGAIN); 1254 } 1255 1256 /* 1257 * Put this page into the physical map. We had to do the unlock above 1258 * because pmap_enter may cause other faults. We don't put the page 1259 * back on the active queue until later so that the page-out daemon 1260 * won't find us (yet). 1261 */ 1262 if (fs->prot & VM_PROT_WRITE) { 1263 vm_page_flag_set(fs->m, PG_WRITEABLE); 1264 vm_object_set_writeable_dirty(fs->m->object); 1265 1266 /* 1267 * If the fault is a write, we know that this page is being 1268 * written NOW so dirty it explicitly to save on 1269 * pmap_is_modified() calls later. 1270 * 1271 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 1272 * if the page is already dirty to prevent data written with 1273 * the expectation of being synced from not being synced. 1274 * Likewise if this entry does not request NOSYNC then make 1275 * sure the page isn't marked NOSYNC. Applications sharing 1276 * data should use the same flags to avoid ping ponging. 1277 * 1278 * Also tell the backing pager, if any, that it should remove 1279 * any swap backing since the page is now dirty. 1280 */ 1281 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) { 1282 if (fs->m->dirty == 0) 1283 vm_page_flag_set(fs->m, PG_NOSYNC); 1284 } else { 1285 vm_page_flag_clear(fs->m, PG_NOSYNC); 1286 } 1287 if (fs->fault_flags & VM_FAULT_DIRTY) { 1288 crit_enter(); 1289 vm_page_dirty(fs->m); 1290 vm_pager_page_unswapped(fs->m); 1291 crit_exit(); 1292 } 1293 } 1294 1295 /* 1296 * Page had better still be busy. We are still locked up and 1297 * fs->object will have another PIP reference if it is not equal 1298 * to fs->first_object. 1299 */ 1300 KASSERT(fs->m->flags & PG_BUSY, 1301 ("vm_fault: page %p not busy!", fs->m)); 1302 1303 /* 1304 * Sanity check: page must be completely valid or it is not fit to 1305 * map into user space. vm_pager_get_pages() ensures this. 1306 */ 1307 if (fs->m->valid != VM_PAGE_BITS_ALL) { 1308 vm_page_zero_invalid(fs->m, TRUE); 1309 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 1310 } 1311 1312 return (KERN_SUCCESS); 1313 } 1314 1315 /* 1316 * Wire down a range of virtual addresses in a map. The entry in question 1317 * should be marked in-transition and the map must be locked. We must 1318 * release the map temporarily while faulting-in the page to avoid a 1319 * deadlock. Note that the entry may be clipped while we are blocked but 1320 * will never be freed. 1321 */ 1322 int 1323 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 1324 { 1325 boolean_t fictitious; 1326 vm_offset_t start; 1327 vm_offset_t end; 1328 vm_offset_t va; 1329 vm_paddr_t pa; 1330 pmap_t pmap; 1331 int rv; 1332 1333 pmap = vm_map_pmap(map); 1334 start = entry->start; 1335 end = entry->end; 1336 fictitious = entry->object.vm_object && 1337 (entry->object.vm_object->type == OBJT_DEVICE); 1338 1339 vm_map_unlock(map); 1340 map->timestamp++; 1341 1342 /* 1343 * We simulate a fault to get the page and enter it in the physical 1344 * map. 1345 */ 1346 for (va = start; va < end; va += PAGE_SIZE) { 1347 if (user_wire) { 1348 rv = vm_fault(map, va, VM_PROT_READ, 1349 VM_FAULT_USER_WIRE); 1350 } else { 1351 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 1352 VM_FAULT_CHANGE_WIRING); 1353 } 1354 if (rv) { 1355 while (va > start) { 1356 va -= PAGE_SIZE; 1357 if ((pa = pmap_extract(pmap, va)) == 0) 1358 continue; 1359 pmap_change_wiring(pmap, va, FALSE); 1360 if (!fictitious) 1361 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1362 } 1363 vm_map_lock(map); 1364 return (rv); 1365 } 1366 } 1367 vm_map_lock(map); 1368 return (KERN_SUCCESS); 1369 } 1370 1371 /* 1372 * Unwire a range of virtual addresses in a map. The map should be 1373 * locked. 1374 */ 1375 void 1376 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 1377 { 1378 boolean_t fictitious; 1379 vm_offset_t start; 1380 vm_offset_t end; 1381 vm_offset_t va; 1382 vm_paddr_t pa; 1383 pmap_t pmap; 1384 1385 pmap = vm_map_pmap(map); 1386 start = entry->start; 1387 end = entry->end; 1388 fictitious = entry->object.vm_object && 1389 (entry->object.vm_object->type == OBJT_DEVICE); 1390 1391 /* 1392 * Since the pages are wired down, we must be able to get their 1393 * mappings from the physical map system. 1394 */ 1395 for (va = start; va < end; va += PAGE_SIZE) { 1396 pa = pmap_extract(pmap, va); 1397 if (pa != 0) { 1398 pmap_change_wiring(pmap, va, FALSE); 1399 if (!fictitious) 1400 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1); 1401 } 1402 } 1403 } 1404 1405 /* 1406 * Reduce the rate at which memory is allocated to a process based 1407 * on the perceived load on the VM system. As the load increases 1408 * the allocation burst rate goes down and the delay increases. 1409 * 1410 * Rate limiting does not apply when faulting active or inactive 1411 * pages. When faulting 'cache' pages, rate limiting only applies 1412 * if the system currently has a severe page deficit. 1413 * 1414 * XXX vm_pagesupply should be increased when a page is freed. 1415 * 1416 * We sleep up to 1/10 of a second. 1417 */ 1418 static int 1419 vm_fault_ratelimit(struct vmspace *vmspace) 1420 { 1421 if (vm_load_enable == 0) 1422 return(0); 1423 if (vmspace->vm_pagesupply > 0) { 1424 --vmspace->vm_pagesupply; 1425 return(0); 1426 } 1427 #ifdef INVARIANTS 1428 if (vm_load_debug) { 1429 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n", 1430 vm_load, 1431 (1000 - vm_load ) / 10, vm_load * hz / 10000, 1432 curproc->p_pid, curproc->p_comm); 1433 } 1434 #endif 1435 vmspace->vm_pagesupply = (1000 - vm_load) / 10; 1436 return(vm_load * hz / 10000); 1437 } 1438 1439 /* 1440 * Routine: 1441 * vm_fault_copy_entry 1442 * Function: 1443 * Copy all of the pages from a wired-down map entry to another. 1444 * 1445 * In/out conditions: 1446 * The source and destination maps must be locked for write. 1447 * The source map entry must be wired down (or be a sharing map 1448 * entry corresponding to a main map entry that is wired down). 1449 */ 1450 1451 void 1452 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 1453 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 1454 { 1455 vm_object_t dst_object; 1456 vm_object_t src_object; 1457 vm_ooffset_t dst_offset; 1458 vm_ooffset_t src_offset; 1459 vm_prot_t prot; 1460 vm_offset_t vaddr; 1461 vm_page_t dst_m; 1462 vm_page_t src_m; 1463 1464 #ifdef lint 1465 src_map++; 1466 #endif /* lint */ 1467 1468 src_object = src_entry->object.vm_object; 1469 src_offset = src_entry->offset; 1470 1471 /* 1472 * Create the top-level object for the destination entry. (Doesn't 1473 * actually shadow anything - we copy the pages directly.) 1474 */ 1475 vm_map_entry_allocate_object(dst_entry); 1476 dst_object = dst_entry->object.vm_object; 1477 1478 prot = dst_entry->max_protection; 1479 1480 /* 1481 * Loop through all of the pages in the entry's range, copying each 1482 * one from the source object (it should be there) to the destination 1483 * object. 1484 */ 1485 for (vaddr = dst_entry->start, dst_offset = 0; 1486 vaddr < dst_entry->end; 1487 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 1488 1489 /* 1490 * Allocate a page in the destination object 1491 */ 1492 do { 1493 dst_m = vm_page_alloc(dst_object, 1494 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL); 1495 if (dst_m == NULL) { 1496 vm_wait(); 1497 } 1498 } while (dst_m == NULL); 1499 1500 /* 1501 * Find the page in the source object, and copy it in. 1502 * (Because the source is wired down, the page will be in 1503 * memory.) 1504 */ 1505 src_m = vm_page_lookup(src_object, 1506 OFF_TO_IDX(dst_offset + src_offset)); 1507 if (src_m == NULL) 1508 panic("vm_fault_copy_wired: page missing"); 1509 1510 vm_page_copy(src_m, dst_m); 1511 1512 /* 1513 * Enter it in the pmap... 1514 */ 1515 1516 vm_page_flag_clear(dst_m, PG_ZERO); 1517 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE); 1518 vm_page_flag_set(dst_m, PG_WRITEABLE|PG_MAPPED); 1519 1520 /* 1521 * Mark it no longer busy, and put it on the active list. 1522 */ 1523 vm_page_activate(dst_m); 1524 vm_page_wakeup(dst_m); 1525 } 1526 } 1527 1528 1529 /* 1530 * This routine checks around the requested page for other pages that 1531 * might be able to be faulted in. This routine brackets the viable 1532 * pages for the pages to be paged in. 1533 * 1534 * Inputs: 1535 * m, rbehind, rahead 1536 * 1537 * Outputs: 1538 * marray (array of vm_page_t), reqpage (index of requested page) 1539 * 1540 * Return value: 1541 * number of pages in marray 1542 */ 1543 static int 1544 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 1545 vm_page_t *marray, int *reqpage) 1546 { 1547 int i,j; 1548 vm_object_t object; 1549 vm_pindex_t pindex, startpindex, endpindex, tpindex; 1550 vm_page_t rtm; 1551 int cbehind, cahead; 1552 1553 object = m->object; 1554 pindex = m->pindex; 1555 1556 /* 1557 * we don't fault-ahead for device pager 1558 */ 1559 if (object->type == OBJT_DEVICE) { 1560 *reqpage = 0; 1561 marray[0] = m; 1562 return 1; 1563 } 1564 1565 /* 1566 * if the requested page is not available, then give up now 1567 */ 1568 1569 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 1570 return 0; 1571 } 1572 1573 if ((cbehind == 0) && (cahead == 0)) { 1574 *reqpage = 0; 1575 marray[0] = m; 1576 return 1; 1577 } 1578 1579 if (rahead > cahead) { 1580 rahead = cahead; 1581 } 1582 1583 if (rbehind > cbehind) { 1584 rbehind = cbehind; 1585 } 1586 1587 /* 1588 * try to do any readahead that we might have free pages for. 1589 */ 1590 if ((rahead + rbehind) > 1591 ((vmstats.v_free_count + vmstats.v_cache_count) - vmstats.v_free_reserved)) { 1592 pagedaemon_wakeup(); 1593 marray[0] = m; 1594 *reqpage = 0; 1595 return 1; 1596 } 1597 1598 /* 1599 * scan backward for the read behind pages -- in memory 1600 * 1601 * Assume that if the page is not found an interrupt will not 1602 * create it. Theoretically interrupts can only remove (busy) 1603 * pages, not create new associations. 1604 */ 1605 if (pindex > 0) { 1606 if (rbehind > pindex) { 1607 rbehind = pindex; 1608 startpindex = 0; 1609 } else { 1610 startpindex = pindex - rbehind; 1611 } 1612 1613 crit_enter(); 1614 for ( tpindex = pindex - 1; tpindex >= startpindex; tpindex -= 1) { 1615 if (vm_page_lookup( object, tpindex)) { 1616 startpindex = tpindex + 1; 1617 break; 1618 } 1619 if (tpindex == 0) 1620 break; 1621 } 1622 1623 for(i = 0, tpindex = startpindex; tpindex < pindex; i++, tpindex++) { 1624 1625 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1626 if (rtm == NULL) { 1627 crit_exit(); 1628 for (j = 0; j < i; j++) { 1629 vm_page_free(marray[j]); 1630 } 1631 marray[0] = m; 1632 *reqpage = 0; 1633 return 1; 1634 } 1635 1636 marray[i] = rtm; 1637 } 1638 crit_exit(); 1639 } else { 1640 startpindex = 0; 1641 i = 0; 1642 } 1643 1644 marray[i] = m; 1645 /* page offset of the required page */ 1646 *reqpage = i; 1647 1648 tpindex = pindex + 1; 1649 i++; 1650 1651 /* 1652 * scan forward for the read ahead pages 1653 */ 1654 endpindex = tpindex + rahead; 1655 if (endpindex > object->size) 1656 endpindex = object->size; 1657 1658 crit_enter(); 1659 for( ; tpindex < endpindex; i++, tpindex++) { 1660 1661 if (vm_page_lookup(object, tpindex)) { 1662 break; 1663 } 1664 1665 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_NORMAL); 1666 if (rtm == NULL) { 1667 break; 1668 } 1669 1670 marray[i] = rtm; 1671 } 1672 crit_exit(); 1673 1674 /* return number of bytes of pages */ 1675 return i; 1676 } 1677