1 /* 2 * Copyright (c) 2003-2014 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * --- 35 * 36 * Copyright (c) 1991, 1993 37 * The Regents of the University of California. All rights reserved. 38 * Copyright (c) 1994 John S. Dyson 39 * All rights reserved. 40 * Copyright (c) 1994 David Greenman 41 * All rights reserved. 42 * 43 * 44 * This code is derived from software contributed to Berkeley by 45 * The Mach Operating System project at Carnegie-Mellon University. 46 * 47 * Redistribution and use in source and binary forms, with or without 48 * modification, are permitted provided that the following conditions 49 * are met: 50 * 1. Redistributions of source code must retain the above copyright 51 * notice, this list of conditions and the following disclaimer. 52 * 2. Redistributions in binary form must reproduce the above copyright 53 * notice, this list of conditions and the following disclaimer in the 54 * documentation and/or other materials provided with the distribution. 55 * 3. Neither the name of the University nor the names of its contributors 56 * may be used to endorse or promote products derived from this software 57 * without specific prior written permission. 58 * 59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 69 * SUCH DAMAGE. 70 * 71 * --- 72 * 73 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 74 * All rights reserved. 75 * 76 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 77 * 78 * Permission to use, copy, modify and distribute this software and 79 * its documentation is hereby granted, provided that both the copyright 80 * notice and this permission notice appear in all copies of the 81 * software, derivative works or modified versions, and any portions 82 * thereof, and that both notices appear in supporting documentation. 83 * 84 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 85 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 86 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 87 * 88 * Carnegie Mellon requests users of this software to return to 89 * 90 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 91 * School of Computer Science 92 * Carnegie Mellon University 93 * Pittsburgh PA 15213-3890 94 * 95 * any improvements or extensions that they make and grant Carnegie the 96 * rights to redistribute these changes. 97 */ 98 99 /* 100 * Page fault handling module. 101 */ 102 103 #include <sys/param.h> 104 #include <sys/systm.h> 105 #include <sys/kernel.h> 106 #include <sys/proc.h> 107 #include <sys/vnode.h> 108 #include <sys/resourcevar.h> 109 #include <sys/vmmeter.h> 110 #include <sys/vkernel.h> 111 #include <sys/lock.h> 112 #include <sys/sysctl.h> 113 114 #include <cpu/lwbuf.h> 115 116 #include <vm/vm.h> 117 #include <vm/vm_param.h> 118 #include <vm/pmap.h> 119 #include <vm/vm_map.h> 120 #include <vm/vm_object.h> 121 #include <vm/vm_page.h> 122 #include <vm/vm_pageout.h> 123 #include <vm/vm_kern.h> 124 #include <vm/vm_pager.h> 125 #include <vm/vnode_pager.h> 126 #include <vm/vm_extern.h> 127 128 #include <vm/vm_page2.h> 129 130 struct faultstate { 131 vm_page_t m; 132 vm_object_t object; 133 vm_pindex_t pindex; 134 vm_prot_t prot; 135 vm_page_t first_m; 136 vm_object_t first_object; 137 vm_prot_t first_prot; 138 vm_map_t map; 139 vm_map_entry_t entry; 140 int lookup_still_valid; 141 int hardfault; 142 int fault_flags; 143 int map_generation; 144 int shared; 145 int first_shared; 146 int wflags; 147 struct vnode *vp; 148 }; 149 150 static int debug_fault = 0; 151 SYSCTL_INT(_vm, OID_AUTO, debug_fault, CTLFLAG_RW, &debug_fault, 0, ""); 152 static int debug_cluster = 0; 153 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, ""); 154 static int virtual_copy_enable = 1; 155 SYSCTL_INT(_vm, OID_AUTO, virtual_copy_enable, CTLFLAG_RW, 156 &virtual_copy_enable, 0, ""); 157 int vm_shared_fault = 1; 158 TUNABLE_INT("vm.shared_fault", &vm_shared_fault); 159 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, 160 &vm_shared_fault, 0, "Allow shared token on vm_object"); 161 162 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int); 163 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, 164 vpte_t, int, int); 165 #if 0 166 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 167 #endif 168 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry); 169 static void vm_prefault(pmap_t pmap, vm_offset_t addra, 170 vm_map_entry_t entry, int prot, int fault_flags); 171 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 172 vm_map_entry_t entry, int prot, int fault_flags); 173 174 static __inline void 175 release_page(struct faultstate *fs) 176 { 177 vm_page_deactivate(fs->m); 178 vm_page_wakeup(fs->m); 179 fs->m = NULL; 180 } 181 182 /* 183 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse 184 * requires relocking and then checking the timestamp. 185 * 186 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do 187 * not have to update fs->map_generation here. 188 * 189 * NOTE: This function can fail due to a deadlock against the caller's 190 * holding of a vm_page BUSY. 191 */ 192 static __inline int 193 relock_map(struct faultstate *fs) 194 { 195 int error; 196 197 if (fs->lookup_still_valid == FALSE && fs->map) { 198 error = vm_map_lock_read_to(fs->map); 199 if (error == 0) 200 fs->lookup_still_valid = TRUE; 201 } else { 202 error = 0; 203 } 204 return error; 205 } 206 207 static __inline void 208 unlock_map(struct faultstate *fs) 209 { 210 if (fs->lookup_still_valid && fs->map) { 211 vm_map_lookup_done(fs->map, fs->entry, 0); 212 fs->lookup_still_valid = FALSE; 213 } 214 } 215 216 /* 217 * Clean up after a successful call to vm_fault_object() so another call 218 * to vm_fault_object() can be made. 219 */ 220 static void 221 _cleanup_successful_fault(struct faultstate *fs, int relock) 222 { 223 /* 224 * We allocated a junk page for a COW operation that did 225 * not occur, the page must be freed. 226 */ 227 if (fs->object != fs->first_object) { 228 KKASSERT(fs->first_shared == 0); 229 vm_page_free(fs->first_m); 230 vm_object_pip_wakeup(fs->object); 231 fs->first_m = NULL; 232 } 233 234 /* 235 * Reset fs->object. 236 */ 237 fs->object = fs->first_object; 238 if (relock && fs->lookup_still_valid == FALSE) { 239 if (fs->map) 240 vm_map_lock_read(fs->map); 241 fs->lookup_still_valid = TRUE; 242 } 243 } 244 245 static void 246 _unlock_things(struct faultstate *fs, int dealloc) 247 { 248 _cleanup_successful_fault(fs, 0); 249 if (dealloc) { 250 /*vm_object_deallocate(fs->first_object);*/ 251 /*fs->first_object = NULL; drop used later on */ 252 } 253 unlock_map(fs); 254 if (fs->vp != NULL) { 255 vput(fs->vp); 256 fs->vp = NULL; 257 } 258 } 259 260 #define unlock_things(fs) _unlock_things(fs, 0) 261 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 262 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 263 264 /* 265 * Virtual copy tests. Used by the fault code to determine if a 266 * page can be moved from an orphan vm_object into its shadow 267 * instead of copying its contents. 268 */ 269 static __inline int 270 virtual_copy_test(struct faultstate *fs) 271 { 272 /* 273 * Must be holding exclusive locks 274 */ 275 if (fs->first_shared || fs->shared || virtual_copy_enable == 0) 276 return 0; 277 278 /* 279 * Map, if present, has not changed 280 */ 281 if (fs->map && fs->map_generation != fs->map->timestamp) 282 return 0; 283 284 /* 285 * Only one shadow object 286 */ 287 if (fs->object->shadow_count != 1) 288 return 0; 289 290 /* 291 * No COW refs, except us 292 */ 293 if (fs->object->ref_count != 1) 294 return 0; 295 296 /* 297 * No one else can look this object up 298 */ 299 if (fs->object->handle != NULL) 300 return 0; 301 302 /* 303 * No other ways to look the object up 304 */ 305 if (fs->object->type != OBJT_DEFAULT && 306 fs->object->type != OBJT_SWAP) 307 return 0; 308 309 /* 310 * We don't chase down the shadow chain 311 */ 312 if (fs->object != fs->first_object->backing_object) 313 return 0; 314 315 return 1; 316 } 317 318 static __inline int 319 virtual_copy_ok(struct faultstate *fs) 320 { 321 if (virtual_copy_test(fs)) { 322 /* 323 * Grab the lock and re-test changeable items. 324 */ 325 if (fs->lookup_still_valid == FALSE && fs->map) { 326 if (lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT)) 327 return 0; 328 fs->lookup_still_valid = TRUE; 329 if (virtual_copy_test(fs)) { 330 fs->map_generation = ++fs->map->timestamp; 331 return 1; 332 } 333 fs->lookup_still_valid = FALSE; 334 lockmgr(&fs->map->lock, LK_RELEASE); 335 } 336 } 337 return 0; 338 } 339 340 /* 341 * TRYPAGER 342 * 343 * Determine if the pager for the current object *might* contain the page. 344 * 345 * We only need to try the pager if this is not a default object (default 346 * objects are zero-fill and have no real pager), and if we are not taking 347 * a wiring fault or if the FS entry is wired. 348 */ 349 #define TRYPAGER(fs) \ 350 (fs->object->type != OBJT_DEFAULT && \ 351 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || \ 352 (fs->wflags & FW_WIRED))) 353 354 /* 355 * vm_fault: 356 * 357 * Handle a page fault occuring at the given address, requiring the given 358 * permissions, in the map specified. If successful, the page is inserted 359 * into the associated physical map. 360 * 361 * NOTE: The given address should be truncated to the proper page address. 362 * 363 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 364 * a standard error specifying why the fault is fatal is returned. 365 * 366 * The map in question must be referenced, and remains so. 367 * The caller may hold no locks. 368 * No other requirements. 369 */ 370 int 371 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 372 { 373 int result; 374 vm_pindex_t first_pindex; 375 struct faultstate fs; 376 struct lwp *lp; 377 struct proc *p; 378 thread_t td; 379 struct vm_map_ilock ilock; 380 int didilock; 381 int growstack; 382 int retry = 0; 383 int inherit_prot; 384 385 inherit_prot = fault_type & VM_PROT_NOSYNC; 386 fs.hardfault = 0; 387 fs.fault_flags = fault_flags; 388 fs.vp = NULL; 389 fs.shared = vm_shared_fault; 390 fs.first_shared = vm_shared_fault; 391 growstack = 1; 392 393 /* 394 * vm_map interactions 395 */ 396 td = curthread; 397 if ((lp = td->td_lwp) != NULL) 398 lp->lwp_flags |= LWP_PAGING; 399 400 RetryFault: 401 /* 402 * Find the vm_map_entry representing the backing store and resolve 403 * the top level object and page index. This may have the side 404 * effect of executing a copy-on-write on the map entry, 405 * creating a shadow object, or splitting an anonymous entry for 406 * performance, but will not COW any actual VM pages. 407 * 408 * On success fs.map is left read-locked and various other fields 409 * are initialized but not otherwise referenced or locked. 410 * 411 * NOTE! vm_map_lookup will try to upgrade the fault_type to 412 * VM_FAULT_WRITE if the map entry is a virtual page table 413 * and also writable, so we can set the 'A'accessed bit in 414 * the virtual page table entry. 415 */ 416 fs.map = map; 417 result = vm_map_lookup(&fs.map, vaddr, fault_type, 418 &fs.entry, &fs.first_object, 419 &first_pindex, &fs.first_prot, &fs.wflags); 420 421 /* 422 * If the lookup failed or the map protections are incompatible, 423 * the fault generally fails. 424 * 425 * The failure could be due to TDF_NOFAULT if vm_map_lookup() 426 * tried to do a COW fault. 427 * 428 * If the caller is trying to do a user wiring we have more work 429 * to do. 430 */ 431 if (result != KERN_SUCCESS) { 432 if (result == KERN_FAILURE_NOFAULT) { 433 result = KERN_FAILURE; 434 goto done; 435 } 436 if (result != KERN_PROTECTION_FAILURE || 437 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 438 { 439 if (result == KERN_INVALID_ADDRESS && growstack && 440 map != &kernel_map && curproc != NULL) { 441 result = vm_map_growstack(map, vaddr); 442 if (result == KERN_SUCCESS) { 443 growstack = 0; 444 ++retry; 445 goto RetryFault; 446 } 447 result = KERN_FAILURE; 448 } 449 goto done; 450 } 451 452 /* 453 * If we are user-wiring a r/w segment, and it is COW, then 454 * we need to do the COW operation. Note that we don't 455 * currently COW RO sections now, because it is NOT desirable 456 * to COW .text. We simply keep .text from ever being COW'ed 457 * and take the heat that one cannot debug wired .text sections. 458 * 459 * XXX Try to allow the above by specifying OVERRIDE_WRITE. 460 */ 461 result = vm_map_lookup(&fs.map, vaddr, 462 VM_PROT_READ|VM_PROT_WRITE| 463 VM_PROT_OVERRIDE_WRITE, 464 &fs.entry, &fs.first_object, 465 &first_pindex, &fs.first_prot, 466 &fs.wflags); 467 if (result != KERN_SUCCESS) { 468 /* could also be KERN_FAILURE_NOFAULT */ 469 result = KERN_FAILURE; 470 goto done; 471 } 472 473 /* 474 * If we don't COW now, on a user wire, the user will never 475 * be able to write to the mapping. If we don't make this 476 * restriction, the bookkeeping would be nearly impossible. 477 * 478 * XXX We have a shared lock, this will have a MP race but 479 * I don't see how it can hurt anything. 480 */ 481 if ((fs.entry->protection & VM_PROT_WRITE) == 0) { 482 atomic_clear_char(&fs.entry->max_protection, 483 VM_PROT_WRITE); 484 } 485 } 486 487 /* 488 * fs.map is read-locked 489 * 490 * Misc checks. Save the map generation number to detect races. 491 */ 492 fs.map_generation = fs.map->timestamp; 493 fs.lookup_still_valid = TRUE; 494 fs.first_m = NULL; 495 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 496 fs.prot = fs.first_prot; /* default (used by uksmap) */ 497 498 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) { 499 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 500 panic("vm_fault: fault on nofault entry, addr: %p", 501 (void *)vaddr); 502 } 503 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) && 504 vaddr >= fs.entry->start && 505 vaddr < fs.entry->start + PAGE_SIZE) { 506 panic("vm_fault: fault on stack guard, addr: %p", 507 (void *)vaddr); 508 } 509 } 510 511 /* 512 * A user-kernel shared map has no VM object and bypasses 513 * everything. We execute the uksmap function with a temporary 514 * fictitious vm_page. The address is directly mapped with no 515 * management. 516 */ 517 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 518 struct vm_page fakem; 519 520 bzero(&fakem, sizeof(fakem)); 521 fakem.pindex = first_pindex; 522 fakem.flags = PG_FICTITIOUS | PG_UNMANAGED; 523 fakem.busy_count = PBUSY_LOCKED; 524 fakem.valid = VM_PAGE_BITS_ALL; 525 fakem.pat_mode = VM_MEMATTR_DEFAULT; 526 if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) { 527 result = KERN_FAILURE; 528 unlock_things(&fs); 529 goto done2; 530 } 531 pmap_enter(fs.map->pmap, vaddr, &fakem, fs.prot | inherit_prot, 532 (fs.wflags & FW_WIRED), fs.entry); 533 goto done_success; 534 } 535 536 /* 537 * A system map entry may return a NULL object. No object means 538 * no pager means an unrecoverable kernel fault. 539 */ 540 if (fs.first_object == NULL) { 541 panic("vm_fault: unrecoverable fault at %p in entry %p", 542 (void *)vaddr, fs.entry); 543 } 544 545 /* 546 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 547 * is set. 548 * 549 * Unfortunately a deadlock can occur if we are forced to page-in 550 * from swap, but diving all the way into the vm_pager_get_page() 551 * function to find out is too much. Just check the object type. 552 * 553 * The deadlock is a CAM deadlock on a busy VM page when trying 554 * to finish an I/O if another process gets stuck in 555 * vop_helper_read_shortcut() due to a swap fault. 556 */ 557 if ((td->td_flags & TDF_NOFAULT) && 558 (retry || 559 fs.first_object->type == OBJT_VNODE || 560 fs.first_object->type == OBJT_SWAP || 561 fs.first_object->backing_object)) { 562 result = KERN_FAILURE; 563 unlock_things(&fs); 564 goto done2; 565 } 566 567 /* 568 * If the entry is wired we cannot change the page protection. 569 */ 570 if (fs.wflags & FW_WIRED) 571 fault_type = fs.first_prot; 572 573 /* 574 * We generally want to avoid unnecessary exclusive modes on backing 575 * and terminal objects because this can seriously interfere with 576 * heavily fork()'d processes (particularly /bin/sh scripts). 577 * 578 * However, we also want to avoid unnecessary retries due to needed 579 * shared->exclusive promotion for common faults. Exclusive mode is 580 * always needed if any page insertion, rename, or free occurs in an 581 * object (and also indirectly if any I/O is done). 582 * 583 * The main issue here is going to be fs.first_shared. If the 584 * first_object has a backing object which isn't shadowed and the 585 * process is single-threaded we might as well use an exclusive 586 * lock/chain right off the bat. 587 */ 588 if (fs.first_shared && fs.first_object->backing_object && 589 LIST_EMPTY(&fs.first_object->shadow_head) && 590 td->td_proc && td->td_proc->p_nthreads == 1) { 591 fs.first_shared = 0; 592 } 593 594 /* 595 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 596 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 597 * we can try shared first. 598 */ 599 if (fault_flags & VM_FAULT_UNSWAP) { 600 fs.first_shared = 0; 601 } 602 603 /* 604 * Obtain a top-level object lock, shared or exclusive depending 605 * on fs.first_shared. If a shared lock winds up being insufficient 606 * we will retry with an exclusive lock. 607 * 608 * The vnode pager lock is always shared. 609 */ 610 if (fs.first_shared) 611 vm_object_hold_shared(fs.first_object); 612 else 613 vm_object_hold(fs.first_object); 614 if (fs.vp == NULL) 615 fs.vp = vnode_pager_lock(fs.first_object); 616 617 /* 618 * The page we want is at (first_object, first_pindex), but if the 619 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 620 * page table to figure out the actual pindex. 621 * 622 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 623 * ONLY 624 */ 625 didilock = 0; 626 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 627 vm_map_interlock(fs.map, &ilock, vaddr, vaddr + PAGE_SIZE); 628 didilock = 1; 629 result = vm_fault_vpagetable(&fs, &first_pindex, 630 fs.entry->aux.master_pde, 631 fault_type, 1); 632 if (result == KERN_TRY_AGAIN) { 633 vm_map_deinterlock(fs.map, &ilock); 634 vm_object_drop(fs.first_object); 635 ++retry; 636 goto RetryFault; 637 } 638 if (result != KERN_SUCCESS) { 639 vm_map_deinterlock(fs.map, &ilock); 640 goto done; 641 } 642 } 643 644 /* 645 * Now we have the actual (object, pindex), fault in the page. If 646 * vm_fault_object() fails it will unlock and deallocate the FS 647 * data. If it succeeds everything remains locked and fs->object 648 * will have an additional PIP count if it is not equal to 649 * fs->first_object 650 * 651 * vm_fault_object will set fs->prot for the pmap operation. It is 652 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 653 * page can be safely written. However, it will force a read-only 654 * mapping for a read fault if the memory is managed by a virtual 655 * page table. 656 * 657 * If the fault code uses the shared object lock shortcut 658 * we must not try to burst (we can't allocate VM pages). 659 */ 660 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 661 662 if (debug_fault > 0) { 663 --debug_fault; 664 kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x " 665 "fs.m=%p fs.prot=%02x fs.wflags=%02x fs.entry=%p\n", 666 result, (intmax_t)vaddr, fault_type, fault_flags, 667 fs.m, fs.prot, fs.wflags, fs.entry); 668 } 669 670 if (result == KERN_TRY_AGAIN) { 671 if (didilock) 672 vm_map_deinterlock(fs.map, &ilock); 673 vm_object_drop(fs.first_object); 674 ++retry; 675 goto RetryFault; 676 } 677 if (result != KERN_SUCCESS) { 678 if (didilock) 679 vm_map_deinterlock(fs.map, &ilock); 680 goto done; 681 } 682 683 /* 684 * On success vm_fault_object() does not unlock or deallocate, and fs.m 685 * will contain a busied page. 686 * 687 * Enter the page into the pmap and do pmap-related adjustments. 688 */ 689 KKASSERT(fs.lookup_still_valid == TRUE); 690 vm_page_flag_set(fs.m, PG_REFERENCED); 691 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot | inherit_prot, 692 fs.wflags & FW_WIRED, fs.entry); 693 694 if (didilock) 695 vm_map_deinterlock(fs.map, &ilock); 696 697 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */ 698 KKASSERT(fs.m->busy_count & PBUSY_LOCKED); 699 700 /* 701 * If the page is not wired down, then put it where the pageout daemon 702 * can find it. 703 */ 704 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 705 if (fs.wflags & FW_WIRED) 706 vm_page_wire(fs.m); 707 else 708 vm_page_unwire(fs.m, 1); 709 } else { 710 vm_page_activate(fs.m); 711 } 712 vm_page_wakeup(fs.m); 713 714 /* 715 * Burst in a few more pages if possible. The fs.map should still 716 * be locked. To avoid interlocking against a vnode->getblk 717 * operation we had to be sure to unbusy our primary vm_page above 718 * first. 719 * 720 * A normal burst can continue down backing store, only execute 721 * if we are holding an exclusive lock, otherwise the exclusive 722 * locks the burst code gets might cause excessive SMP collisions. 723 * 724 * A quick burst can be utilized when there is no backing object 725 * (i.e. a shared file mmap). 726 */ 727 if ((fault_flags & VM_FAULT_BURST) && 728 (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 && 729 (fs.wflags & FW_WIRED) == 0) { 730 if (fs.first_shared == 0 && fs.shared == 0) { 731 vm_prefault(fs.map->pmap, vaddr, 732 fs.entry, fs.prot, fault_flags); 733 } else { 734 vm_prefault_quick(fs.map->pmap, vaddr, 735 fs.entry, fs.prot, fault_flags); 736 } 737 } 738 739 done_success: 740 mycpu->gd_cnt.v_vm_faults++; 741 if (td->td_lwp) 742 ++td->td_lwp->lwp_ru.ru_minflt; 743 744 /* 745 * Unlock everything, and return 746 */ 747 unlock_things(&fs); 748 749 if (td->td_lwp) { 750 if (fs.hardfault) { 751 td->td_lwp->lwp_ru.ru_majflt++; 752 } else { 753 td->td_lwp->lwp_ru.ru_minflt++; 754 } 755 } 756 757 /*vm_object_deallocate(fs.first_object);*/ 758 /*fs.m = NULL; */ 759 /*fs.first_object = NULL; must still drop later */ 760 761 result = KERN_SUCCESS; 762 done: 763 if (fs.first_object) 764 vm_object_drop(fs.first_object); 765 done2: 766 if (lp) 767 lp->lwp_flags &= ~LWP_PAGING; 768 769 #if !defined(NO_SWAPPING) 770 /* 771 * Check the process RSS limit and force deactivation and 772 * (asynchronous) paging if necessary. This is a complex operation, 773 * only do it for direct user-mode faults, for now. 774 * 775 * To reduce overhead implement approximately a ~16MB hysteresis. 776 */ 777 p = td->td_proc; 778 if ((fault_flags & VM_FAULT_USERMODE) && lp && 779 p->p_limit && map->pmap && vm_pageout_memuse_mode >= 1 && 780 map != &kernel_map) { 781 vm_pindex_t limit; 782 vm_pindex_t size; 783 784 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 785 p->p_rlimit[RLIMIT_RSS].rlim_max)); 786 size = pmap_resident_tlnw_count(map->pmap); 787 if (limit >= 0 && size > 4096 && size - 4096 >= limit) { 788 vm_pageout_map_deactivate_pages(map, limit); 789 } 790 } 791 #endif 792 793 return (result); 794 } 795 796 /* 797 * Fault in the specified virtual address in the current process map, 798 * returning a held VM page or NULL. See vm_fault_page() for more 799 * information. 800 * 801 * No requirements. 802 */ 803 vm_page_t 804 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, 805 int *errorp, int *busyp) 806 { 807 struct lwp *lp = curthread->td_lwp; 808 vm_page_t m; 809 810 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 811 fault_type, VM_FAULT_NORMAL, 812 errorp, busyp); 813 return(m); 814 } 815 816 /* 817 * Fault in the specified virtual address in the specified map, doing all 818 * necessary manipulation of the object store and all necessary I/O. Return 819 * a held VM page or NULL, and set *errorp. The related pmap is not 820 * updated. 821 * 822 * If busyp is not NULL then *busyp will be set to TRUE if this routine 823 * decides to return a busied page (aka VM_PROT_WRITE), or FALSE if it 824 * does not (VM_PROT_WRITE not specified or busyp is NULL). If busyp is 825 * NULL the returned page is only held. 826 * 827 * If the caller has no intention of writing to the page's contents, busyp 828 * can be passed as NULL along with VM_PROT_WRITE to force a COW operation 829 * without busying the page. 830 * 831 * The returned page will also be marked PG_REFERENCED. 832 * 833 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 834 * error will be returned. 835 * 836 * No requirements. 837 */ 838 vm_page_t 839 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 840 int fault_flags, int *errorp, int *busyp) 841 { 842 vm_pindex_t first_pindex; 843 struct faultstate fs; 844 int result; 845 int retry; 846 int growstack; 847 int didcow; 848 vm_prot_t orig_fault_type = fault_type; 849 850 retry = 0; 851 didcow = 0; 852 fs.hardfault = 0; 853 fs.fault_flags = fault_flags; 854 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 855 856 /* 857 * Dive the pmap (concurrency possible). If we find the 858 * appropriate page we can terminate early and quickly. 859 * 860 * This works great for normal programs but will always return 861 * NULL for host lookups of vkernel maps in VMM mode. 862 * 863 * NOTE: pmap_fault_page_quick() might not busy the page. If 864 * VM_PROT_WRITE is set in fault_type and pmap_fault_page_quick() 865 * returns non-NULL, it will safely dirty the returned vm_page_t 866 * for us. We cannot safely dirty it here (it might not be 867 * busy). 868 */ 869 fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type, busyp); 870 if (fs.m) { 871 *errorp = 0; 872 return(fs.m); 873 } 874 875 /* 876 * Otherwise take a concurrency hit and do a formal page 877 * fault. 878 */ 879 fs.vp = NULL; 880 fs.shared = vm_shared_fault; 881 fs.first_shared = vm_shared_fault; 882 growstack = 1; 883 884 /* 885 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 886 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 887 * we can try shared first. 888 */ 889 if (fault_flags & VM_FAULT_UNSWAP) { 890 fs.first_shared = 0; 891 } 892 893 RetryFault: 894 /* 895 * Find the vm_map_entry representing the backing store and resolve 896 * the top level object and page index. This may have the side 897 * effect of executing a copy-on-write on the map entry and/or 898 * creating a shadow object, but will not COW any actual VM pages. 899 * 900 * On success fs.map is left read-locked and various other fields 901 * are initialized but not otherwise referenced or locked. 902 * 903 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 904 * if the map entry is a virtual page table and also writable, 905 * so we can set the 'A'accessed bit in the virtual page table 906 * entry. 907 */ 908 fs.map = map; 909 result = vm_map_lookup(&fs.map, vaddr, fault_type, 910 &fs.entry, &fs.first_object, 911 &first_pindex, &fs.first_prot, &fs.wflags); 912 913 if (result != KERN_SUCCESS) { 914 if (result == KERN_FAILURE_NOFAULT) { 915 *errorp = KERN_FAILURE; 916 fs.m = NULL; 917 goto done; 918 } 919 if (result != KERN_PROTECTION_FAILURE || 920 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 921 { 922 if (result == KERN_INVALID_ADDRESS && growstack && 923 map != &kernel_map && curproc != NULL) { 924 result = vm_map_growstack(map, vaddr); 925 if (result == KERN_SUCCESS) { 926 growstack = 0; 927 ++retry; 928 goto RetryFault; 929 } 930 result = KERN_FAILURE; 931 } 932 fs.m = NULL; 933 *errorp = result; 934 goto done; 935 } 936 937 /* 938 * If we are user-wiring a r/w segment, and it is COW, then 939 * we need to do the COW operation. Note that we don't 940 * currently COW RO sections now, because it is NOT desirable 941 * to COW .text. We simply keep .text from ever being COW'ed 942 * and take the heat that one cannot debug wired .text sections. 943 */ 944 result = vm_map_lookup(&fs.map, vaddr, 945 VM_PROT_READ|VM_PROT_WRITE| 946 VM_PROT_OVERRIDE_WRITE, 947 &fs.entry, &fs.first_object, 948 &first_pindex, &fs.first_prot, 949 &fs.wflags); 950 if (result != KERN_SUCCESS) { 951 /* could also be KERN_FAILURE_NOFAULT */ 952 *errorp = KERN_FAILURE; 953 fs.m = NULL; 954 goto done; 955 } 956 957 /* 958 * If we don't COW now, on a user wire, the user will never 959 * be able to write to the mapping. If we don't make this 960 * restriction, the bookkeeping would be nearly impossible. 961 * 962 * XXX We have a shared lock, this will have a MP race but 963 * I don't see how it can hurt anything. 964 */ 965 if ((fs.entry->protection & VM_PROT_WRITE) == 0) { 966 atomic_clear_char(&fs.entry->max_protection, 967 VM_PROT_WRITE); 968 } 969 } 970 971 /* 972 * fs.map is read-locked 973 * 974 * Misc checks. Save the map generation number to detect races. 975 */ 976 fs.map_generation = fs.map->timestamp; 977 fs.lookup_still_valid = TRUE; 978 fs.first_m = NULL; 979 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 980 981 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 982 panic("vm_fault: fault on nofault entry, addr: %lx", 983 (u_long)vaddr); 984 } 985 986 /* 987 * A user-kernel shared map has no VM object and bypasses 988 * everything. We execute the uksmap function with a temporary 989 * fictitious vm_page. The address is directly mapped with no 990 * management. 991 */ 992 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 993 struct vm_page fakem; 994 995 bzero(&fakem, sizeof(fakem)); 996 fakem.pindex = first_pindex; 997 fakem.flags = PG_FICTITIOUS | PG_UNMANAGED; 998 fakem.busy_count = PBUSY_LOCKED; 999 fakem.valid = VM_PAGE_BITS_ALL; 1000 fakem.pat_mode = VM_MEMATTR_DEFAULT; 1001 if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) { 1002 *errorp = KERN_FAILURE; 1003 fs.m = NULL; 1004 unlock_things(&fs); 1005 goto done2; 1006 } 1007 fs.m = PHYS_TO_VM_PAGE(fakem.phys_addr); 1008 vm_page_hold(fs.m); 1009 if (busyp) 1010 *busyp = 0; /* don't need to busy R or W */ 1011 unlock_things(&fs); 1012 *errorp = 0; 1013 goto done; 1014 } 1015 1016 1017 /* 1018 * A system map entry may return a NULL object. No object means 1019 * no pager means an unrecoverable kernel fault. 1020 */ 1021 if (fs.first_object == NULL) { 1022 panic("vm_fault: unrecoverable fault at %p in entry %p", 1023 (void *)vaddr, fs.entry); 1024 } 1025 1026 /* 1027 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 1028 * is set. 1029 * 1030 * Unfortunately a deadlock can occur if we are forced to page-in 1031 * from swap, but diving all the way into the vm_pager_get_page() 1032 * function to find out is too much. Just check the object type. 1033 */ 1034 if ((curthread->td_flags & TDF_NOFAULT) && 1035 (retry || 1036 fs.first_object->type == OBJT_VNODE || 1037 fs.first_object->type == OBJT_SWAP || 1038 fs.first_object->backing_object)) { 1039 *errorp = KERN_FAILURE; 1040 unlock_things(&fs); 1041 fs.m = NULL; 1042 goto done2; 1043 } 1044 1045 /* 1046 * If the entry is wired we cannot change the page protection. 1047 */ 1048 if (fs.wflags & FW_WIRED) 1049 fault_type = fs.first_prot; 1050 1051 /* 1052 * Make a reference to this object to prevent its disposal while we 1053 * are messing with it. Once we have the reference, the map is free 1054 * to be diddled. Since objects reference their shadows (and copies), 1055 * they will stay around as well. 1056 * 1057 * The reference should also prevent an unexpected collapse of the 1058 * parent that might move pages from the current object into the 1059 * parent unexpectedly, resulting in corruption. 1060 * 1061 * Bump the paging-in-progress count to prevent size changes (e.g. 1062 * truncation operations) during I/O. This must be done after 1063 * obtaining the vnode lock in order to avoid possible deadlocks. 1064 */ 1065 if (fs.first_shared) 1066 vm_object_hold_shared(fs.first_object); 1067 else 1068 vm_object_hold(fs.first_object); 1069 if (fs.vp == NULL) 1070 fs.vp = vnode_pager_lock(fs.first_object); /* shared */ 1071 1072 /* 1073 * The page we want is at (first_object, first_pindex), but if the 1074 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 1075 * page table to figure out the actual pindex. 1076 * 1077 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 1078 * ONLY 1079 */ 1080 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1081 result = vm_fault_vpagetable(&fs, &first_pindex, 1082 fs.entry->aux.master_pde, 1083 fault_type, 1); 1084 if (result == KERN_TRY_AGAIN) { 1085 vm_object_drop(fs.first_object); 1086 ++retry; 1087 goto RetryFault; 1088 } 1089 if (result != KERN_SUCCESS) { 1090 *errorp = result; 1091 fs.m = NULL; 1092 goto done; 1093 } 1094 } 1095 1096 /* 1097 * Now we have the actual (object, pindex), fault in the page. If 1098 * vm_fault_object() fails it will unlock and deallocate the FS 1099 * data. If it succeeds everything remains locked and fs->object 1100 * will have an additinal PIP count if it is not equal to 1101 * fs->first_object 1102 */ 1103 fs.m = NULL; 1104 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 1105 1106 if (result == KERN_TRY_AGAIN) { 1107 vm_object_drop(fs.first_object); 1108 ++retry; 1109 didcow |= fs.wflags & FW_DIDCOW; 1110 goto RetryFault; 1111 } 1112 if (result != KERN_SUCCESS) { 1113 *errorp = result; 1114 fs.m = NULL; 1115 goto done; 1116 } 1117 1118 if ((orig_fault_type & VM_PROT_WRITE) && 1119 (fs.prot & VM_PROT_WRITE) == 0) { 1120 *errorp = KERN_PROTECTION_FAILURE; 1121 unlock_and_deallocate(&fs); 1122 fs.m = NULL; 1123 goto done; 1124 } 1125 1126 /* 1127 * Generally speaking we don't want to update the pmap because 1128 * this routine can be called many times for situations that do 1129 * not require updating the pmap, not to mention the page might 1130 * already be in the pmap. 1131 * 1132 * However, if our vm_map_lookup() results in a COW, we need to 1133 * at least remove the pte from the pmap to guarantee proper 1134 * visibility of modifications made to the process. For example, 1135 * modifications made by vkernel uiocopy/related routines and 1136 * modifications made by ptrace(). 1137 */ 1138 vm_page_flag_set(fs.m, PG_REFERENCED); 1139 #if 0 1140 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, 1141 fs.wflags & FW_WIRED, NULL); 1142 mycpu->gd_cnt.v_vm_faults++; 1143 if (curthread->td_lwp) 1144 ++curthread->td_lwp->lwp_ru.ru_minflt; 1145 #endif 1146 if ((fs.wflags | didcow) | FW_DIDCOW) { 1147 pmap_remove(fs.map->pmap, 1148 vaddr & ~PAGE_MASK, 1149 (vaddr & ~PAGE_MASK) + PAGE_SIZE); 1150 } 1151 1152 /* 1153 * On success vm_fault_object() does not unlock or deallocate, and fs.m 1154 * will contain a busied page. So we must unlock here after having 1155 * messed with the pmap. 1156 */ 1157 unlock_things(&fs); 1158 1159 /* 1160 * Return a held page. We are not doing any pmap manipulation so do 1161 * not set PG_MAPPED. However, adjust the page flags according to 1162 * the fault type because the caller may not use a managed pmapping 1163 * (so we don't want to lose the fact that the page will be dirtied 1164 * if a write fault was specified). 1165 */ 1166 if (fault_type & VM_PROT_WRITE) 1167 vm_page_dirty(fs.m); 1168 vm_page_activate(fs.m); 1169 1170 if (curthread->td_lwp) { 1171 if (fs.hardfault) { 1172 curthread->td_lwp->lwp_ru.ru_majflt++; 1173 } else { 1174 curthread->td_lwp->lwp_ru.ru_minflt++; 1175 } 1176 } 1177 1178 /* 1179 * Unlock everything, and return the held or busied page. 1180 */ 1181 if (busyp) { 1182 if (fault_type & VM_PROT_WRITE) { 1183 vm_page_dirty(fs.m); 1184 *busyp = 1; 1185 } else { 1186 *busyp = 0; 1187 vm_page_hold(fs.m); 1188 vm_page_wakeup(fs.m); 1189 } 1190 } else { 1191 vm_page_hold(fs.m); 1192 vm_page_wakeup(fs.m); 1193 } 1194 /*vm_object_deallocate(fs.first_object);*/ 1195 /*fs.first_object = NULL; */ 1196 *errorp = 0; 1197 1198 done: 1199 if (fs.first_object) 1200 vm_object_drop(fs.first_object); 1201 done2: 1202 return(fs.m); 1203 } 1204 1205 /* 1206 * Fault in the specified (object,offset), dirty the returned page as 1207 * needed. If the requested fault_type cannot be done NULL and an 1208 * error is returned. 1209 * 1210 * A held (but not busied) page is returned. 1211 * 1212 * The passed in object must be held as specified by the shared 1213 * argument. 1214 */ 1215 vm_page_t 1216 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 1217 vm_prot_t fault_type, int fault_flags, 1218 int *sharedp, int *errorp) 1219 { 1220 int result; 1221 vm_pindex_t first_pindex; 1222 struct faultstate fs; 1223 struct vm_map_entry entry; 1224 1225 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1226 bzero(&entry, sizeof(entry)); 1227 entry.object.vm_object = object; 1228 entry.maptype = VM_MAPTYPE_NORMAL; 1229 entry.protection = entry.max_protection = fault_type; 1230 1231 fs.hardfault = 0; 1232 fs.fault_flags = fault_flags; 1233 fs.map = NULL; 1234 fs.shared = vm_shared_fault; 1235 fs.first_shared = *sharedp; 1236 fs.vp = NULL; 1237 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 1238 1239 /* 1240 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 1241 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 1242 * we can try shared first. 1243 */ 1244 if (fs.first_shared && (fault_flags & VM_FAULT_UNSWAP)) { 1245 fs.first_shared = 0; 1246 vm_object_upgrade(object); 1247 } 1248 1249 /* 1250 * Retry loop as needed (typically for shared->exclusive transitions) 1251 */ 1252 RetryFault: 1253 *sharedp = fs.first_shared; 1254 first_pindex = OFF_TO_IDX(offset); 1255 fs.first_object = object; 1256 fs.entry = &entry; 1257 fs.first_prot = fault_type; 1258 fs.wflags = 0; 1259 /*fs.map_generation = 0; unused */ 1260 1261 /* 1262 * Make a reference to this object to prevent its disposal while we 1263 * are messing with it. Once we have the reference, the map is free 1264 * to be diddled. Since objects reference their shadows (and copies), 1265 * they will stay around as well. 1266 * 1267 * The reference should also prevent an unexpected collapse of the 1268 * parent that might move pages from the current object into the 1269 * parent unexpectedly, resulting in corruption. 1270 * 1271 * Bump the paging-in-progress count to prevent size changes (e.g. 1272 * truncation operations) during I/O. This must be done after 1273 * obtaining the vnode lock in order to avoid possible deadlocks. 1274 */ 1275 if (fs.vp == NULL) 1276 fs.vp = vnode_pager_lock(fs.first_object); 1277 1278 fs.lookup_still_valid = TRUE; 1279 fs.first_m = NULL; 1280 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 1281 1282 #if 0 1283 /* XXX future - ability to operate on VM object using vpagetable */ 1284 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1285 result = vm_fault_vpagetable(&fs, &first_pindex, 1286 fs.entry->aux.master_pde, 1287 fault_type, 0); 1288 if (result == KERN_TRY_AGAIN) { 1289 if (fs.first_shared == 0 && *sharedp) 1290 vm_object_upgrade(object); 1291 goto RetryFault; 1292 } 1293 if (result != KERN_SUCCESS) { 1294 *errorp = result; 1295 return (NULL); 1296 } 1297 } 1298 #endif 1299 1300 /* 1301 * Now we have the actual (object, pindex), fault in the page. If 1302 * vm_fault_object() fails it will unlock and deallocate the FS 1303 * data. If it succeeds everything remains locked and fs->object 1304 * will have an additinal PIP count if it is not equal to 1305 * fs->first_object 1306 * 1307 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact. 1308 * We may have to upgrade its lock to handle the requested fault. 1309 */ 1310 result = vm_fault_object(&fs, first_pindex, fault_type, 0); 1311 1312 if (result == KERN_TRY_AGAIN) { 1313 if (fs.first_shared == 0 && *sharedp) 1314 vm_object_upgrade(object); 1315 goto RetryFault; 1316 } 1317 if (result != KERN_SUCCESS) { 1318 *errorp = result; 1319 return(NULL); 1320 } 1321 1322 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 1323 *errorp = KERN_PROTECTION_FAILURE; 1324 unlock_and_deallocate(&fs); 1325 return(NULL); 1326 } 1327 1328 /* 1329 * On success vm_fault_object() does not unlock or deallocate, so we 1330 * do it here. Note that the returned fs.m will be busied. 1331 */ 1332 unlock_things(&fs); 1333 1334 /* 1335 * Return a held page. We are not doing any pmap manipulation so do 1336 * not set PG_MAPPED. However, adjust the page flags according to 1337 * the fault type because the caller may not use a managed pmapping 1338 * (so we don't want to lose the fact that the page will be dirtied 1339 * if a write fault was specified). 1340 */ 1341 vm_page_hold(fs.m); 1342 vm_page_activate(fs.m); 1343 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY)) 1344 vm_page_dirty(fs.m); 1345 if (fault_flags & VM_FAULT_UNSWAP) 1346 swap_pager_unswapped(fs.m); 1347 1348 /* 1349 * Indicate that the page was accessed. 1350 */ 1351 vm_page_flag_set(fs.m, PG_REFERENCED); 1352 1353 if (curthread->td_lwp) { 1354 if (fs.hardfault) { 1355 curthread->td_lwp->lwp_ru.ru_majflt++; 1356 } else { 1357 curthread->td_lwp->lwp_ru.ru_minflt++; 1358 } 1359 } 1360 1361 /* 1362 * Unlock everything, and return the held page. 1363 */ 1364 vm_page_wakeup(fs.m); 1365 /*vm_object_deallocate(fs.first_object);*/ 1366 /*fs.first_object = NULL; */ 1367 1368 *errorp = 0; 1369 return(fs.m); 1370 } 1371 1372 /* 1373 * Translate the virtual page number (first_pindex) that is relative 1374 * to the address space into a logical page number that is relative to the 1375 * backing object. Use the virtual page table pointed to by (vpte). 1376 * 1377 * Possibly downgrade the protection based on the vpte bits. 1378 * 1379 * This implements an N-level page table. Any level can terminate the 1380 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 1381 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 1382 */ 1383 static 1384 int 1385 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 1386 vpte_t vpte, int fault_type, int allow_nofault) 1387 { 1388 struct lwbuf *lwb; 1389 struct lwbuf lwb_cache; 1390 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */ 1391 int result; 1392 vpte_t *ptep; 1393 1394 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1395 for (;;) { 1396 /* 1397 * We cannot proceed if the vpte is not valid, not readable 1398 * for a read fault, not writable for a write fault, or 1399 * not executable for an instruction execution fault. 1400 */ 1401 if ((vpte & VPTE_V) == 0) { 1402 unlock_and_deallocate(fs); 1403 return (KERN_FAILURE); 1404 } 1405 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) { 1406 unlock_and_deallocate(fs); 1407 return (KERN_FAILURE); 1408 } 1409 if ((fault_type & VM_PROT_EXECUTE) && (vpte & VPTE_NX)) { 1410 unlock_and_deallocate(fs); 1411 return (KERN_FAILURE); 1412 } 1413 if ((vpte & VPTE_PS) || vshift == 0) 1414 break; 1415 1416 /* 1417 * Get the page table page. Nominally we only read the page 1418 * table, but since we are actively setting VPTE_M and VPTE_A, 1419 * tell vm_fault_object() that we are writing it. 1420 * 1421 * There is currently no real need to optimize this. 1422 */ 1423 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT, 1424 VM_PROT_READ|VM_PROT_WRITE, 1425 allow_nofault); 1426 if (result != KERN_SUCCESS) 1427 return (result); 1428 1429 /* 1430 * Process the returned fs.m and look up the page table 1431 * entry in the page table page. 1432 */ 1433 vshift -= VPTE_PAGE_BITS; 1434 lwb = lwbuf_alloc(fs->m, &lwb_cache); 1435 ptep = ((vpte_t *)lwbuf_kva(lwb) + 1436 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 1437 vm_page_activate(fs->m); 1438 1439 /* 1440 * Page table write-back - entire operation including 1441 * validation of the pte must be atomic to avoid races 1442 * against the vkernel changing the pte. 1443 * 1444 * If the vpte is valid for the* requested operation, do 1445 * a write-back to the page table. 1446 * 1447 * XXX VPTE_M is not set properly for page directory pages. 1448 * It doesn't get set in the page directory if the page table 1449 * is modified during a read access. 1450 */ 1451 for (;;) { 1452 vpte_t nvpte; 1453 1454 /* 1455 * Reload for the cmpset, but make sure the pte is 1456 * still valid. 1457 */ 1458 vpte = *ptep; 1459 cpu_ccfence(); 1460 nvpte = vpte; 1461 1462 if ((vpte & VPTE_V) == 0) 1463 break; 1464 1465 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW)) 1466 nvpte |= VPTE_M | VPTE_A; 1467 if (fault_type & (VM_PROT_READ | VM_PROT_EXECUTE)) 1468 nvpte |= VPTE_A; 1469 if (vpte == nvpte) 1470 break; 1471 if (atomic_cmpset_long(ptep, vpte, nvpte)) { 1472 vm_page_dirty(fs->m); 1473 break; 1474 } 1475 } 1476 lwbuf_free(lwb); 1477 vm_page_flag_set(fs->m, PG_REFERENCED); 1478 vm_page_wakeup(fs->m); 1479 fs->m = NULL; 1480 cleanup_successful_fault(fs); 1481 } 1482 1483 /* 1484 * When the vkernel sets VPTE_RW it expects the real kernel to 1485 * reflect VPTE_M back when the page is modified via the mapping. 1486 * In order to accomplish this the real kernel must map the page 1487 * read-only for read faults and use write faults to reflect VPTE_M 1488 * back. 1489 * 1490 * Once VPTE_M has been set, the real kernel's pte allows writing. 1491 * If the vkernel clears VPTE_M the vkernel must be sure to 1492 * MADV_INVAL the real kernel's mappings to force the real kernel 1493 * to re-fault on the next write so oit can set VPTE_M again. 1494 */ 1495 if ((fault_type & VM_PROT_WRITE) == 0 && 1496 (vpte & (VPTE_RW | VPTE_M)) != (VPTE_RW | VPTE_M)) { 1497 fs->first_prot &= ~VM_PROT_WRITE; 1498 } 1499 1500 /* 1501 * Disable EXECUTE perms if NX bit is set. 1502 */ 1503 if (vpte & VPTE_NX) 1504 fs->first_prot &= ~VM_PROT_EXECUTE; 1505 1506 /* 1507 * Combine remaining address bits with the vpte. 1508 */ 1509 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) + 1510 (*pindex & ((1L << vshift) - 1)); 1511 return (KERN_SUCCESS); 1512 } 1513 1514 1515 /* 1516 * This is the core of the vm_fault code. 1517 * 1518 * Do all operations required to fault-in (fs.first_object, pindex). Run 1519 * through the shadow chain as necessary and do required COW or virtual 1520 * copy operations. The caller has already fully resolved the vm_map_entry 1521 * and, if appropriate, has created a copy-on-write layer. All we need to 1522 * do is iterate the object chain. 1523 * 1524 * On failure (fs) is unlocked and deallocated and the caller may return or 1525 * retry depending on the failure code. On success (fs) is NOT unlocked or 1526 * deallocated, fs.m will contained a resolved, busied page, and fs.object 1527 * will have an additional PIP count if it is not equal to fs.first_object. 1528 * 1529 * If locks based on fs->first_shared or fs->shared are insufficient, 1530 * clear the appropriate field(s) and return RETRY. COWs require that 1531 * first_shared be 0, while page allocations (or frees) require that 1532 * shared be 0. Renames require that both be 0. 1533 * 1534 * NOTE! fs->[first_]shared might be set with VM_FAULT_DIRTY also set. 1535 * we will have to retry with it exclusive if the vm_page is 1536 * PG_SWAPPED. 1537 * 1538 * fs->first_object must be held on call. 1539 */ 1540 static 1541 int 1542 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex, 1543 vm_prot_t fault_type, int allow_nofault) 1544 { 1545 vm_object_t next_object; 1546 vm_pindex_t pindex; 1547 int error; 1548 1549 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1550 fs->prot = fs->first_prot; 1551 fs->object = fs->first_object; 1552 pindex = first_pindex; 1553 1554 vm_object_chain_acquire(fs->first_object, fs->shared); 1555 vm_object_pip_add(fs->first_object, 1); 1556 1557 /* 1558 * If a read fault occurs we try to upgrade the page protection 1559 * and make it also writable if possible. There are three cases 1560 * where we cannot make the page mapping writable: 1561 * 1562 * (1) The mapping is read-only or the VM object is read-only, 1563 * fs->prot above will simply not have VM_PROT_WRITE set. 1564 * 1565 * (2) If the mapping is a virtual page table fs->first_prot will 1566 * have already been properly adjusted by vm_fault_vpagetable(). 1567 * to detect writes so we can set VPTE_M in the virtual page 1568 * table. Used by vkernels. 1569 * 1570 * (3) If the VM page is read-only or copy-on-write, upgrading would 1571 * just result in an unnecessary COW fault. 1572 * 1573 * (4) If the pmap specifically requests A/M bit emulation, downgrade 1574 * here. 1575 */ 1576 #if 0 1577 /* see vpagetable code */ 1578 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1579 if ((fault_type & VM_PROT_WRITE) == 0) 1580 fs->prot &= ~VM_PROT_WRITE; 1581 } 1582 #endif 1583 1584 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 1585 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 1586 if ((fault_type & VM_PROT_WRITE) == 0) 1587 fs->prot &= ~VM_PROT_WRITE; 1588 } 1589 1590 /* vm_object_hold(fs->object); implied b/c object == first_object */ 1591 1592 for (;;) { 1593 /* 1594 * The entire backing chain from first_object to object 1595 * inclusive is chainlocked. 1596 * 1597 * If the object is dead, we stop here 1598 */ 1599 if (fs->object->flags & OBJ_DEAD) { 1600 vm_object_pip_wakeup(fs->first_object); 1601 vm_object_chain_release_all(fs->first_object, 1602 fs->object); 1603 if (fs->object != fs->first_object) 1604 vm_object_drop(fs->object); 1605 unlock_and_deallocate(fs); 1606 return (KERN_PROTECTION_FAILURE); 1607 } 1608 1609 /* 1610 * See if the page is resident. Wait/Retry if the page is 1611 * busy (lots of stuff may have changed so we can't continue 1612 * in that case). 1613 * 1614 * We can theoretically allow the soft-busy case on a read 1615 * fault if the page is marked valid, but since such 1616 * pages are typically already pmap'd, putting that 1617 * special case in might be more effort then it is 1618 * worth. We cannot under any circumstances mess 1619 * around with a vm_page_t->busy page except, perhaps, 1620 * to pmap it. 1621 */ 1622 fs->m = vm_page_lookup_busy_try(fs->object, pindex, 1623 TRUE, &error); 1624 if (error) { 1625 vm_object_pip_wakeup(fs->first_object); 1626 vm_object_chain_release_all(fs->first_object, 1627 fs->object); 1628 if (fs->object != fs->first_object) 1629 vm_object_drop(fs->object); 1630 unlock_things(fs); 1631 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 1632 mycpu->gd_cnt.v_intrans++; 1633 /*vm_object_deallocate(fs->first_object);*/ 1634 /*fs->first_object = NULL;*/ 1635 fs->m = NULL; 1636 return (KERN_TRY_AGAIN); 1637 } 1638 if (fs->m) { 1639 /* 1640 * The page is busied for us. 1641 * 1642 * If reactivating a page from PQ_CACHE we may have 1643 * to rate-limit. 1644 */ 1645 int queue = fs->m->queue; 1646 vm_page_unqueue_nowakeup(fs->m); 1647 1648 if ((queue - fs->m->pc) == PQ_CACHE && 1649 vm_page_count_severe()) { 1650 vm_page_activate(fs->m); 1651 vm_page_wakeup(fs->m); 1652 fs->m = NULL; 1653 vm_object_pip_wakeup(fs->first_object); 1654 vm_object_chain_release_all(fs->first_object, 1655 fs->object); 1656 if (fs->object != fs->first_object) 1657 vm_object_drop(fs->object); 1658 unlock_and_deallocate(fs); 1659 if (allow_nofault == 0 || 1660 (curthread->td_flags & TDF_NOFAULT) == 0) { 1661 thread_t td; 1662 1663 vm_wait_pfault(); 1664 td = curthread; 1665 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1666 return (KERN_PROTECTION_FAILURE); 1667 } 1668 return (KERN_TRY_AGAIN); 1669 } 1670 1671 /* 1672 * If it still isn't completely valid (readable), 1673 * or if a read-ahead-mark is set on the VM page, 1674 * jump to readrest, else we found the page and 1675 * can return. 1676 * 1677 * We can release the spl once we have marked the 1678 * page busy. 1679 */ 1680 if (fs->m->object != &kernel_object) { 1681 if ((fs->m->valid & VM_PAGE_BITS_ALL) != 1682 VM_PAGE_BITS_ALL) { 1683 goto readrest; 1684 } 1685 if (fs->m->flags & PG_RAM) { 1686 if (debug_cluster) 1687 kprintf("R"); 1688 vm_page_flag_clear(fs->m, PG_RAM); 1689 goto readrest; 1690 } 1691 } 1692 break; /* break to PAGE HAS BEEN FOUND */ 1693 } 1694 1695 /* 1696 * Page is not resident, If this is the search termination 1697 * or the pager might contain the page, allocate a new page. 1698 */ 1699 if (TRYPAGER(fs) || fs->object == fs->first_object) { 1700 /* 1701 * Allocating, must be exclusive. 1702 */ 1703 if (fs->object == fs->first_object && 1704 fs->first_shared) { 1705 fs->first_shared = 0; 1706 vm_object_pip_wakeup(fs->first_object); 1707 vm_object_chain_release_all(fs->first_object, 1708 fs->object); 1709 if (fs->object != fs->first_object) 1710 vm_object_drop(fs->object); 1711 unlock_and_deallocate(fs); 1712 return (KERN_TRY_AGAIN); 1713 } 1714 if (fs->object != fs->first_object && 1715 fs->shared) { 1716 fs->first_shared = 0; 1717 fs->shared = 0; 1718 vm_object_pip_wakeup(fs->first_object); 1719 vm_object_chain_release_all(fs->first_object, 1720 fs->object); 1721 if (fs->object != fs->first_object) 1722 vm_object_drop(fs->object); 1723 unlock_and_deallocate(fs); 1724 return (KERN_TRY_AGAIN); 1725 } 1726 1727 /* 1728 * If the page is beyond the object size we fail 1729 */ 1730 if (pindex >= fs->object->size) { 1731 vm_object_pip_wakeup(fs->first_object); 1732 vm_object_chain_release_all(fs->first_object, 1733 fs->object); 1734 if (fs->object != fs->first_object) 1735 vm_object_drop(fs->object); 1736 unlock_and_deallocate(fs); 1737 return (KERN_PROTECTION_FAILURE); 1738 } 1739 1740 /* 1741 * Allocate a new page for this object/offset pair. 1742 * 1743 * It is possible for the allocation to race, so 1744 * handle the case. 1745 */ 1746 fs->m = NULL; 1747 if (!vm_page_count_severe()) { 1748 fs->m = vm_page_alloc(fs->object, pindex, 1749 ((fs->vp || fs->object->backing_object) ? 1750 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1751 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1752 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1753 } 1754 if (fs->m == NULL) { 1755 vm_object_pip_wakeup(fs->first_object); 1756 vm_object_chain_release_all(fs->first_object, 1757 fs->object); 1758 if (fs->object != fs->first_object) 1759 vm_object_drop(fs->object); 1760 unlock_and_deallocate(fs); 1761 if (allow_nofault == 0 || 1762 (curthread->td_flags & TDF_NOFAULT) == 0) { 1763 thread_t td; 1764 1765 vm_wait_pfault(); 1766 td = curthread; 1767 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1768 return (KERN_PROTECTION_FAILURE); 1769 } 1770 return (KERN_TRY_AGAIN); 1771 } 1772 1773 /* 1774 * Fall through to readrest. We have a new page which 1775 * will have to be paged (since m->valid will be 0). 1776 */ 1777 } 1778 1779 readrest: 1780 /* 1781 * We have found an invalid or partially valid page, a 1782 * page with a read-ahead mark which might be partially or 1783 * fully valid (and maybe dirty too), or we have allocated 1784 * a new page. 1785 * 1786 * Attempt to fault-in the page if there is a chance that the 1787 * pager has it, and potentially fault in additional pages 1788 * at the same time. 1789 * 1790 * If TRYPAGER is true then fs.m will be non-NULL and busied 1791 * for us. 1792 */ 1793 if (TRYPAGER(fs)) { 1794 int rv; 1795 int seqaccess; 1796 u_char behavior = vm_map_entry_behavior(fs->entry); 1797 1798 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 1799 seqaccess = 0; 1800 else 1801 seqaccess = -1; 1802 1803 /* 1804 * Doing I/O may synchronously insert additional 1805 * pages so we can't be shared at this point either. 1806 * 1807 * NOTE: We can't free fs->m here in the allocated 1808 * case (fs->object != fs->first_object) as 1809 * this would require an exclusively locked 1810 * VM object. 1811 */ 1812 if (fs->object == fs->first_object && 1813 fs->first_shared) { 1814 vm_page_deactivate(fs->m); 1815 vm_page_wakeup(fs->m); 1816 fs->m = NULL; 1817 fs->first_shared = 0; 1818 vm_object_pip_wakeup(fs->first_object); 1819 vm_object_chain_release_all(fs->first_object, 1820 fs->object); 1821 if (fs->object != fs->first_object) 1822 vm_object_drop(fs->object); 1823 unlock_and_deallocate(fs); 1824 return (KERN_TRY_AGAIN); 1825 } 1826 if (fs->object != fs->first_object && 1827 fs->shared) { 1828 vm_page_deactivate(fs->m); 1829 vm_page_wakeup(fs->m); 1830 fs->m = NULL; 1831 fs->first_shared = 0; 1832 fs->shared = 0; 1833 vm_object_pip_wakeup(fs->first_object); 1834 vm_object_chain_release_all(fs->first_object, 1835 fs->object); 1836 if (fs->object != fs->first_object) 1837 vm_object_drop(fs->object); 1838 unlock_and_deallocate(fs); 1839 return (KERN_TRY_AGAIN); 1840 } 1841 1842 /* 1843 * Avoid deadlocking against the map when doing I/O. 1844 * fs.object and the page is BUSY'd. 1845 * 1846 * NOTE: Once unlocked, fs->entry can become stale 1847 * so this will NULL it out. 1848 * 1849 * NOTE: fs->entry is invalid until we relock the 1850 * map and verify that the timestamp has not 1851 * changed. 1852 */ 1853 unlock_map(fs); 1854 1855 /* 1856 * Acquire the page data. We still hold a ref on 1857 * fs.object and the page has been BUSY's. 1858 * 1859 * The pager may replace the page (for example, in 1860 * order to enter a fictitious page into the 1861 * object). If it does so it is responsible for 1862 * cleaning up the passed page and properly setting 1863 * the new page BUSY. 1864 * 1865 * If we got here through a PG_RAM read-ahead 1866 * mark the page may be partially dirty and thus 1867 * not freeable. Don't bother checking to see 1868 * if the pager has the page because we can't free 1869 * it anyway. We have to depend on the get_page 1870 * operation filling in any gaps whether there is 1871 * backing store or not. 1872 */ 1873 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess); 1874 1875 if (rv == VM_PAGER_OK) { 1876 /* 1877 * Relookup in case pager changed page. Pager 1878 * is responsible for disposition of old page 1879 * if moved. 1880 * 1881 * XXX other code segments do relookups too. 1882 * It's a bad abstraction that needs to be 1883 * fixed/removed. 1884 */ 1885 fs->m = vm_page_lookup(fs->object, pindex); 1886 if (fs->m == NULL) { 1887 vm_object_pip_wakeup(fs->first_object); 1888 vm_object_chain_release_all( 1889 fs->first_object, fs->object); 1890 if (fs->object != fs->first_object) 1891 vm_object_drop(fs->object); 1892 unlock_and_deallocate(fs); 1893 return (KERN_TRY_AGAIN); 1894 } 1895 ++fs->hardfault; 1896 break; /* break to PAGE HAS BEEN FOUND */ 1897 } 1898 1899 /* 1900 * Remove the bogus page (which does not exist at this 1901 * object/offset); before doing so, we must get back 1902 * our object lock to preserve our invariant. 1903 * 1904 * Also wake up any other process that may want to bring 1905 * in this page. 1906 * 1907 * If this is the top-level object, we must leave the 1908 * busy page to prevent another process from rushing 1909 * past us, and inserting the page in that object at 1910 * the same time that we are. 1911 */ 1912 if (rv == VM_PAGER_ERROR) { 1913 if (curproc) { 1914 kprintf("vm_fault: pager read error, " 1915 "pid %d (%s)\n", 1916 curproc->p_pid, 1917 curproc->p_comm); 1918 } else { 1919 kprintf("vm_fault: pager read error, " 1920 "thread %p (%s)\n", 1921 curthread, 1922 curthread->td_comm); 1923 } 1924 } 1925 1926 /* 1927 * Data outside the range of the pager or an I/O error 1928 * 1929 * The page may have been wired during the pagein, 1930 * e.g. by the buffer cache, and cannot simply be 1931 * freed. Call vnode_pager_freepage() to deal with it. 1932 * 1933 * Also note that we cannot free the page if we are 1934 * holding the related object shared. XXX not sure 1935 * what to do in that case. 1936 */ 1937 if (fs->object != fs->first_object) { 1938 /* 1939 * Scrap the page. Check to see if the 1940 * vm_pager_get_page() call has already 1941 * dealt with it. 1942 */ 1943 if (fs->m) { 1944 vnode_pager_freepage(fs->m); 1945 fs->m = NULL; 1946 } 1947 1948 /* 1949 * XXX - we cannot just fall out at this 1950 * point, m has been freed and is invalid! 1951 */ 1952 } 1953 /* 1954 * XXX - the check for kernel_map is a kludge to work 1955 * around having the machine panic on a kernel space 1956 * fault w/ I/O error. 1957 */ 1958 if (((fs->map != &kernel_map) && 1959 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 1960 if (fs->m) { 1961 if (fs->first_shared) { 1962 vm_page_deactivate(fs->m); 1963 vm_page_wakeup(fs->m); 1964 } else { 1965 vnode_pager_freepage(fs->m); 1966 } 1967 fs->m = NULL; 1968 } 1969 vm_object_pip_wakeup(fs->first_object); 1970 vm_object_chain_release_all(fs->first_object, 1971 fs->object); 1972 if (fs->object != fs->first_object) 1973 vm_object_drop(fs->object); 1974 unlock_and_deallocate(fs); 1975 if (rv == VM_PAGER_ERROR) 1976 return (KERN_FAILURE); 1977 else 1978 return (KERN_PROTECTION_FAILURE); 1979 /* NOT REACHED */ 1980 } 1981 } 1982 1983 /* 1984 * We get here if the object has a default pager (or unwiring) 1985 * or the pager doesn't have the page. 1986 * 1987 * fs->first_m will be used for the COW unless we find a 1988 * deeper page to be mapped read-only, in which case the 1989 * unlock*(fs) will free first_m. 1990 */ 1991 if (fs->object == fs->first_object) 1992 fs->first_m = fs->m; 1993 1994 /* 1995 * Move on to the next object. The chain lock should prevent 1996 * the backing_object from getting ripped out from under us. 1997 * 1998 * The object lock for the next object is governed by 1999 * fs->shared. 2000 */ 2001 if ((next_object = fs->object->backing_object) != NULL) { 2002 if (fs->shared) 2003 vm_object_hold_shared(next_object); 2004 else 2005 vm_object_hold(next_object); 2006 vm_object_chain_acquire(next_object, fs->shared); 2007 KKASSERT(next_object == fs->object->backing_object); 2008 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 2009 } 2010 2011 if (next_object == NULL) { 2012 /* 2013 * If there's no object left, fill the page in the top 2014 * object with zeros. 2015 */ 2016 if (fs->object != fs->first_object) { 2017 #if 0 2018 if (fs->first_object->backing_object != 2019 fs->object) { 2020 vm_object_hold(fs->first_object->backing_object); 2021 } 2022 #endif 2023 vm_object_chain_release_all( 2024 fs->first_object->backing_object, 2025 fs->object); 2026 #if 0 2027 if (fs->first_object->backing_object != 2028 fs->object) { 2029 vm_object_drop(fs->first_object->backing_object); 2030 } 2031 #endif 2032 vm_object_pip_wakeup(fs->object); 2033 vm_object_drop(fs->object); 2034 fs->object = fs->first_object; 2035 pindex = first_pindex; 2036 fs->m = fs->first_m; 2037 } 2038 fs->first_m = NULL; 2039 2040 /* 2041 * Zero the page and mark it valid. 2042 */ 2043 vm_page_zero_fill(fs->m); 2044 mycpu->gd_cnt.v_zfod++; 2045 fs->m->valid = VM_PAGE_BITS_ALL; 2046 break; /* break to PAGE HAS BEEN FOUND */ 2047 } 2048 if (fs->object != fs->first_object) { 2049 vm_object_pip_wakeup(fs->object); 2050 vm_object_lock_swap(); 2051 vm_object_drop(fs->object); 2052 } 2053 KASSERT(fs->object != next_object, 2054 ("object loop %p", next_object)); 2055 fs->object = next_object; 2056 vm_object_pip_add(fs->object, 1); 2057 } 2058 2059 /* 2060 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 2061 * is held.] 2062 * 2063 * object still held. 2064 * vm_map may not be locked (determined by fs->lookup_still_valid) 2065 * 2066 * local shared variable may be different from fs->shared. 2067 * 2068 * If the page is being written, but isn't already owned by the 2069 * top-level object, we have to copy it into a new page owned by the 2070 * top-level object. 2071 */ 2072 KASSERT((fs->m->busy_count & PBUSY_LOCKED) != 0, 2073 ("vm_fault: not busy after main loop")); 2074 2075 if (fs->object != fs->first_object) { 2076 /* 2077 * We only really need to copy if we want to write it. 2078 */ 2079 if (fault_type & VM_PROT_WRITE) { 2080 /* 2081 * This allows pages to be virtually copied from a 2082 * backing_object into the first_object, where the 2083 * backing object has no other refs to it, and cannot 2084 * gain any more refs. Instead of a bcopy, we just 2085 * move the page from the backing object to the 2086 * first object. Note that we must mark the page 2087 * dirty in the first object so that it will go out 2088 * to swap when needed. 2089 */ 2090 if (virtual_copy_ok(fs)) { 2091 /* 2092 * (first_m) and (m) are both busied. We have 2093 * move (m) into (first_m)'s object/pindex 2094 * in an atomic fashion, then free (first_m). 2095 * 2096 * first_object is held so second remove 2097 * followed by the rename should wind 2098 * up being atomic. vm_page_free() might 2099 * block so we don't do it until after the 2100 * rename. 2101 */ 2102 vm_page_protect(fs->first_m, VM_PROT_NONE); 2103 vm_page_remove(fs->first_m); 2104 vm_page_rename(fs->m, fs->first_object, 2105 first_pindex); 2106 vm_page_free(fs->first_m); 2107 fs->first_m = fs->m; 2108 fs->m = NULL; 2109 mycpu->gd_cnt.v_cow_optim++; 2110 } else { 2111 /* 2112 * Oh, well, lets copy it. 2113 * 2114 * Why are we unmapping the original page 2115 * here? Well, in short, not all accessors 2116 * of user memory go through the pmap. The 2117 * procfs code doesn't have access user memory 2118 * via a local pmap, so vm_fault_page*() 2119 * can't call pmap_enter(). And the umtx*() 2120 * code may modify the COW'd page via a DMAP 2121 * or kernel mapping and not via the pmap, 2122 * leaving the original page still mapped 2123 * read-only into the pmap. 2124 * 2125 * So we have to remove the page from at 2126 * least the current pmap if it is in it. 2127 * 2128 * We used to just remove it from all pmaps 2129 * but that creates inefficiencies on SMP, 2130 * particularly for COW program & library 2131 * mappings that are concurrently exec'd. 2132 * Only remove the page from the current 2133 * pmap. 2134 */ 2135 KKASSERT(fs->first_shared == 0); 2136 vm_page_copy(fs->m, fs->first_m); 2137 /*vm_page_protect(fs->m, VM_PROT_NONE);*/ 2138 pmap_remove_specific( 2139 &curthread->td_lwp->lwp_vmspace->vm_pmap, 2140 fs->m); 2141 } 2142 2143 /* 2144 * We no longer need the old page or object. 2145 */ 2146 if (fs->m) 2147 release_page(fs); 2148 2149 /* 2150 * We intend to revert to first_object, undo the 2151 * chain lock through to that. 2152 */ 2153 #if 0 2154 if (fs->first_object->backing_object != fs->object) 2155 vm_object_hold(fs->first_object->backing_object); 2156 #endif 2157 vm_object_chain_release_all( 2158 fs->first_object->backing_object, 2159 fs->object); 2160 #if 0 2161 if (fs->first_object->backing_object != fs->object) 2162 vm_object_drop(fs->first_object->backing_object); 2163 #endif 2164 2165 /* 2166 * fs->object != fs->first_object due to above 2167 * conditional 2168 */ 2169 vm_object_pip_wakeup(fs->object); 2170 vm_object_drop(fs->object); 2171 2172 /* 2173 * Only use the new page below... 2174 */ 2175 mycpu->gd_cnt.v_cow_faults++; 2176 fs->m = fs->first_m; 2177 fs->object = fs->first_object; 2178 pindex = first_pindex; 2179 } else { 2180 /* 2181 * If it wasn't a write fault avoid having to copy 2182 * the page by mapping it read-only. 2183 */ 2184 fs->prot &= ~VM_PROT_WRITE; 2185 } 2186 } 2187 2188 /* 2189 * Relock the map if necessary, then check the generation count. 2190 * relock_map() will update fs->timestamp to account for the 2191 * relocking if necessary. 2192 * 2193 * If the count has changed after relocking then all sorts of 2194 * crap may have happened and we have to retry. 2195 * 2196 * NOTE: The relock_map() can fail due to a deadlock against 2197 * the vm_page we are holding BUSY. 2198 */ 2199 if (fs->lookup_still_valid == FALSE && fs->map) { 2200 if (relock_map(fs) || 2201 fs->map->timestamp != fs->map_generation) { 2202 release_page(fs); 2203 vm_object_pip_wakeup(fs->first_object); 2204 vm_object_chain_release_all(fs->first_object, 2205 fs->object); 2206 if (fs->object != fs->first_object) 2207 vm_object_drop(fs->object); 2208 unlock_and_deallocate(fs); 2209 return (KERN_TRY_AGAIN); 2210 } 2211 } 2212 2213 /* 2214 * If the fault is a write, we know that this page is being 2215 * written NOW so dirty it explicitly to save on pmap_is_modified() 2216 * calls later. 2217 * 2218 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 2219 * if the page is already dirty to prevent data written with 2220 * the expectation of being synced from not being synced. 2221 * Likewise if this entry does not request NOSYNC then make 2222 * sure the page isn't marked NOSYNC. Applications sharing 2223 * data should use the same flags to avoid ping ponging. 2224 * 2225 * Also tell the backing pager, if any, that it should remove 2226 * any swap backing since the page is now dirty. 2227 */ 2228 vm_page_activate(fs->m); 2229 if (fs->prot & VM_PROT_WRITE) { 2230 vm_object_set_writeable_dirty(fs->m->object); 2231 vm_set_nosync(fs->m, fs->entry); 2232 if (fs->fault_flags & VM_FAULT_DIRTY) { 2233 vm_page_dirty(fs->m); 2234 if (fs->m->flags & PG_SWAPPED) { 2235 /* 2236 * If the page is swapped out we have to call 2237 * swap_pager_unswapped() which requires an 2238 * exclusive object lock. If we are shared, 2239 * we must clear the shared flag and retry. 2240 */ 2241 if ((fs->object == fs->first_object && 2242 fs->first_shared) || 2243 (fs->object != fs->first_object && 2244 fs->shared)) { 2245 vm_page_wakeup(fs->m); 2246 fs->m = NULL; 2247 if (fs->object == fs->first_object) 2248 fs->first_shared = 0; 2249 else 2250 fs->shared = 0; 2251 vm_object_pip_wakeup(fs->first_object); 2252 vm_object_chain_release_all( 2253 fs->first_object, fs->object); 2254 if (fs->object != fs->first_object) 2255 vm_object_drop(fs->object); 2256 unlock_and_deallocate(fs); 2257 return (KERN_TRY_AGAIN); 2258 } 2259 swap_pager_unswapped(fs->m); 2260 } 2261 } 2262 } 2263 2264 vm_object_pip_wakeup(fs->first_object); 2265 vm_object_chain_release_all(fs->first_object, fs->object); 2266 if (fs->object != fs->first_object) 2267 vm_object_drop(fs->object); 2268 2269 /* 2270 * Page had better still be busy. We are still locked up and 2271 * fs->object will have another PIP reference if it is not equal 2272 * to fs->first_object. 2273 */ 2274 KASSERT(fs->m->busy_count & PBUSY_LOCKED, 2275 ("vm_fault: page %p not busy!", fs->m)); 2276 2277 /* 2278 * Sanity check: page must be completely valid or it is not fit to 2279 * map into user space. vm_pager_get_pages() ensures this. 2280 */ 2281 if (fs->m->valid != VM_PAGE_BITS_ALL) { 2282 vm_page_zero_invalid(fs->m, TRUE); 2283 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 2284 } 2285 2286 return (KERN_SUCCESS); 2287 } 2288 2289 /* 2290 * Wire down a range of virtual addresses in a map. The entry in question 2291 * should be marked in-transition and the map must be locked. We must 2292 * release the map temporarily while faulting-in the page to avoid a 2293 * deadlock. Note that the entry may be clipped while we are blocked but 2294 * will never be freed. 2295 * 2296 * No requirements. 2297 */ 2298 int 2299 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, 2300 boolean_t user_wire, int kmflags) 2301 { 2302 boolean_t fictitious; 2303 vm_offset_t start; 2304 vm_offset_t end; 2305 vm_offset_t va; 2306 pmap_t pmap; 2307 int rv; 2308 int wire_prot; 2309 int fault_flags; 2310 vm_page_t m; 2311 2312 if (user_wire) { 2313 wire_prot = VM_PROT_READ; 2314 fault_flags = VM_FAULT_USER_WIRE; 2315 } else { 2316 wire_prot = VM_PROT_READ | VM_PROT_WRITE; 2317 fault_flags = VM_FAULT_CHANGE_WIRING; 2318 } 2319 if (kmflags & KM_NOTLBSYNC) 2320 wire_prot |= VM_PROT_NOSYNC; 2321 2322 pmap = vm_map_pmap(map); 2323 start = entry->start; 2324 end = entry->end; 2325 2326 switch(entry->maptype) { 2327 case VM_MAPTYPE_NORMAL: 2328 case VM_MAPTYPE_VPAGETABLE: 2329 fictitious = entry->object.vm_object && 2330 ((entry->object.vm_object->type == OBJT_DEVICE) || 2331 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2332 break; 2333 case VM_MAPTYPE_UKSMAP: 2334 fictitious = TRUE; 2335 break; 2336 default: 2337 fictitious = FALSE; 2338 break; 2339 } 2340 2341 if (entry->eflags & MAP_ENTRY_KSTACK) 2342 start += PAGE_SIZE; 2343 map->timestamp++; 2344 vm_map_unlock(map); 2345 2346 /* 2347 * We simulate a fault to get the page and enter it in the physical 2348 * map. 2349 */ 2350 for (va = start; va < end; va += PAGE_SIZE) { 2351 rv = vm_fault(map, va, wire_prot, fault_flags); 2352 if (rv) { 2353 while (va > start) { 2354 va -= PAGE_SIZE; 2355 m = pmap_unwire(pmap, va); 2356 if (m && !fictitious) { 2357 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2358 vm_page_unwire(m, 1); 2359 vm_page_wakeup(m); 2360 } 2361 } 2362 goto done; 2363 } 2364 } 2365 rv = KERN_SUCCESS; 2366 done: 2367 vm_map_lock(map); 2368 2369 return (rv); 2370 } 2371 2372 /* 2373 * Unwire a range of virtual addresses in a map. The map should be 2374 * locked. 2375 */ 2376 void 2377 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 2378 { 2379 boolean_t fictitious; 2380 vm_offset_t start; 2381 vm_offset_t end; 2382 vm_offset_t va; 2383 pmap_t pmap; 2384 vm_page_t m; 2385 2386 pmap = vm_map_pmap(map); 2387 start = entry->start; 2388 end = entry->end; 2389 fictitious = entry->object.vm_object && 2390 ((entry->object.vm_object->type == OBJT_DEVICE) || 2391 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2392 if (entry->eflags & MAP_ENTRY_KSTACK) 2393 start += PAGE_SIZE; 2394 2395 /* 2396 * Since the pages are wired down, we must be able to get their 2397 * mappings from the physical map system. 2398 */ 2399 for (va = start; va < end; va += PAGE_SIZE) { 2400 m = pmap_unwire(pmap, va); 2401 if (m && !fictitious) { 2402 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2403 vm_page_unwire(m, 1); 2404 vm_page_wakeup(m); 2405 } 2406 } 2407 } 2408 2409 /* 2410 * Copy all of the pages from a wired-down map entry to another. 2411 * 2412 * The source and destination maps must be locked for write. 2413 * The source and destination maps token must be held 2414 * The source map entry must be wired down (or be a sharing map 2415 * entry corresponding to a main map entry that is wired down). 2416 * 2417 * No other requirements. 2418 * 2419 * XXX do segment optimization 2420 */ 2421 void 2422 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2423 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2424 { 2425 vm_object_t dst_object; 2426 vm_object_t src_object; 2427 vm_ooffset_t dst_offset; 2428 vm_ooffset_t src_offset; 2429 vm_prot_t prot; 2430 vm_offset_t vaddr; 2431 vm_page_t dst_m; 2432 vm_page_t src_m; 2433 2434 src_object = src_entry->object.vm_object; 2435 src_offset = src_entry->offset; 2436 2437 /* 2438 * Create the top-level object for the destination entry. (Doesn't 2439 * actually shadow anything - we copy the pages directly.) 2440 */ 2441 vm_map_entry_allocate_object(dst_entry); 2442 dst_object = dst_entry->object.vm_object; 2443 2444 prot = dst_entry->max_protection; 2445 2446 /* 2447 * Loop through all of the pages in the entry's range, copying each 2448 * one from the source object (it should be there) to the destination 2449 * object. 2450 */ 2451 vm_object_hold(src_object); 2452 vm_object_hold(dst_object); 2453 2454 for (vaddr = dst_entry->start, dst_offset = 0; 2455 vaddr < dst_entry->end; 2456 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2457 2458 /* 2459 * Allocate a page in the destination object 2460 */ 2461 do { 2462 dst_m = vm_page_alloc(dst_object, 2463 OFF_TO_IDX(dst_offset), 2464 VM_ALLOC_NORMAL); 2465 if (dst_m == NULL) { 2466 vm_wait(0); 2467 } 2468 } while (dst_m == NULL); 2469 2470 /* 2471 * Find the page in the source object, and copy it in. 2472 * (Because the source is wired down, the page will be in 2473 * memory.) 2474 */ 2475 src_m = vm_page_lookup(src_object, 2476 OFF_TO_IDX(dst_offset + src_offset)); 2477 if (src_m == NULL) 2478 panic("vm_fault_copy_wired: page missing"); 2479 2480 vm_page_copy(src_m, dst_m); 2481 2482 /* 2483 * Enter it in the pmap... 2484 */ 2485 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2486 2487 /* 2488 * Mark it no longer busy, and put it on the active list. 2489 */ 2490 vm_page_activate(dst_m); 2491 vm_page_wakeup(dst_m); 2492 } 2493 vm_object_drop(dst_object); 2494 vm_object_drop(src_object); 2495 } 2496 2497 #if 0 2498 2499 /* 2500 * This routine checks around the requested page for other pages that 2501 * might be able to be faulted in. This routine brackets the viable 2502 * pages for the pages to be paged in. 2503 * 2504 * Inputs: 2505 * m, rbehind, rahead 2506 * 2507 * Outputs: 2508 * marray (array of vm_page_t), reqpage (index of requested page) 2509 * 2510 * Return value: 2511 * number of pages in marray 2512 */ 2513 static int 2514 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2515 vm_page_t *marray, int *reqpage) 2516 { 2517 int i,j; 2518 vm_object_t object; 2519 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2520 vm_page_t rtm; 2521 int cbehind, cahead; 2522 2523 object = m->object; 2524 pindex = m->pindex; 2525 2526 /* 2527 * we don't fault-ahead for device pager 2528 */ 2529 if ((object->type == OBJT_DEVICE) || 2530 (object->type == OBJT_MGTDEVICE)) { 2531 *reqpage = 0; 2532 marray[0] = m; 2533 return 1; 2534 } 2535 2536 /* 2537 * if the requested page is not available, then give up now 2538 */ 2539 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2540 *reqpage = 0; /* not used by caller, fix compiler warn */ 2541 return 0; 2542 } 2543 2544 if ((cbehind == 0) && (cahead == 0)) { 2545 *reqpage = 0; 2546 marray[0] = m; 2547 return 1; 2548 } 2549 2550 if (rahead > cahead) { 2551 rahead = cahead; 2552 } 2553 2554 if (rbehind > cbehind) { 2555 rbehind = cbehind; 2556 } 2557 2558 /* 2559 * Do not do any readahead if we have insufficient free memory. 2560 * 2561 * XXX code was broken disabled before and has instability 2562 * with this conditonal fixed, so shortcut for now. 2563 */ 2564 if (burst_fault == 0 || vm_page_count_severe()) { 2565 marray[0] = m; 2566 *reqpage = 0; 2567 return 1; 2568 } 2569 2570 /* 2571 * scan backward for the read behind pages -- in memory 2572 * 2573 * Assume that if the page is not found an interrupt will not 2574 * create it. Theoretically interrupts can only remove (busy) 2575 * pages, not create new associations. 2576 */ 2577 if (pindex > 0) { 2578 if (rbehind > pindex) { 2579 rbehind = pindex; 2580 startpindex = 0; 2581 } else { 2582 startpindex = pindex - rbehind; 2583 } 2584 2585 vm_object_hold(object); 2586 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2587 if (vm_page_lookup(object, tpindex - 1)) 2588 break; 2589 } 2590 2591 i = 0; 2592 while (tpindex < pindex) { 2593 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2594 VM_ALLOC_NULL_OK); 2595 if (rtm == NULL) { 2596 for (j = 0; j < i; j++) { 2597 vm_page_free(marray[j]); 2598 } 2599 vm_object_drop(object); 2600 marray[0] = m; 2601 *reqpage = 0; 2602 return 1; 2603 } 2604 marray[i] = rtm; 2605 ++i; 2606 ++tpindex; 2607 } 2608 vm_object_drop(object); 2609 } else { 2610 i = 0; 2611 } 2612 2613 /* 2614 * Assign requested page 2615 */ 2616 marray[i] = m; 2617 *reqpage = i; 2618 ++i; 2619 2620 /* 2621 * Scan forwards for read-ahead pages 2622 */ 2623 tpindex = pindex + 1; 2624 endpindex = tpindex + rahead; 2625 if (endpindex > object->size) 2626 endpindex = object->size; 2627 2628 vm_object_hold(object); 2629 while (tpindex < endpindex) { 2630 if (vm_page_lookup(object, tpindex)) 2631 break; 2632 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2633 VM_ALLOC_NULL_OK); 2634 if (rtm == NULL) 2635 break; 2636 marray[i] = rtm; 2637 ++i; 2638 ++tpindex; 2639 } 2640 vm_object_drop(object); 2641 2642 return (i); 2643 } 2644 2645 #endif 2646 2647 /* 2648 * vm_prefault() provides a quick way of clustering pagefaults into a 2649 * processes address space. It is a "cousin" of pmap_object_init_pt, 2650 * except it runs at page fault time instead of mmap time. 2651 * 2652 * vm.fast_fault Enables pre-faulting zero-fill pages 2653 * 2654 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2655 * prefault. Scan stops in either direction when 2656 * a page is found to already exist. 2657 * 2658 * This code used to be per-platform pmap_prefault(). It is now 2659 * machine-independent and enhanced to also pre-fault zero-fill pages 2660 * (see vm.fast_fault) as well as make them writable, which greatly 2661 * reduces the number of page faults programs incur. 2662 * 2663 * Application performance when pre-faulting zero-fill pages is heavily 2664 * dependent on the application. Very tiny applications like /bin/echo 2665 * lose a little performance while applications of any appreciable size 2666 * gain performance. Prefaulting multiple pages also reduces SMP 2667 * congestion and can improve SMP performance significantly. 2668 * 2669 * NOTE! prot may allow writing but this only applies to the top level 2670 * object. If we wind up mapping a page extracted from a backing 2671 * object we have to make sure it is read-only. 2672 * 2673 * NOTE! The caller has already handled any COW operations on the 2674 * vm_map_entry via the normal fault code. Do NOT call this 2675 * shortcut unless the normal fault code has run on this entry. 2676 * 2677 * The related map must be locked. 2678 * No other requirements. 2679 */ 2680 static int vm_prefault_pages = 8; 2681 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2682 "Maximum number of pages to pre-fault"); 2683 static int vm_fast_fault = 1; 2684 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2685 "Burst fault zero-fill regions"); 2686 2687 /* 2688 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2689 * is not already dirty by other means. This will prevent passive 2690 * filesystem syncing as well as 'sync' from writing out the page. 2691 */ 2692 static void 2693 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2694 { 2695 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2696 if (m->dirty == 0) 2697 vm_page_flag_set(m, PG_NOSYNC); 2698 } else { 2699 vm_page_flag_clear(m, PG_NOSYNC); 2700 } 2701 } 2702 2703 static void 2704 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 2705 int fault_flags) 2706 { 2707 struct lwp *lp; 2708 vm_page_t m; 2709 vm_offset_t addr; 2710 vm_pindex_t index; 2711 vm_pindex_t pindex; 2712 vm_object_t object; 2713 int pprot; 2714 int i; 2715 int noneg; 2716 int nopos; 2717 int maxpages; 2718 2719 /* 2720 * Get stable max count value, disabled if set to 0 2721 */ 2722 maxpages = vm_prefault_pages; 2723 cpu_ccfence(); 2724 if (maxpages <= 0) 2725 return; 2726 2727 /* 2728 * We do not currently prefault mappings that use virtual page 2729 * tables. We do not prefault foreign pmaps. 2730 */ 2731 if (entry->maptype != VM_MAPTYPE_NORMAL) 2732 return; 2733 lp = curthread->td_lwp; 2734 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2735 return; 2736 2737 /* 2738 * Limit pre-fault count to 1024 pages. 2739 */ 2740 if (maxpages > 1024) 2741 maxpages = 1024; 2742 2743 object = entry->object.vm_object; 2744 KKASSERT(object != NULL); 2745 KKASSERT(object == entry->object.vm_object); 2746 2747 /* 2748 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively 2749 * now (or do something more complex XXX). 2750 */ 2751 vm_object_hold(object); 2752 vm_object_chain_acquire(object, 0); 2753 2754 noneg = 0; 2755 nopos = 0; 2756 for (i = 0; i < maxpages; ++i) { 2757 vm_object_t lobject; 2758 vm_object_t nobject; 2759 int allocated = 0; 2760 int error; 2761 2762 /* 2763 * This can eat a lot of time on a heavily contended 2764 * machine so yield on the tick if needed. 2765 */ 2766 if ((i & 7) == 7) 2767 lwkt_yield(); 2768 2769 /* 2770 * Calculate the page to pre-fault, stopping the scan in 2771 * each direction separately if the limit is reached. 2772 */ 2773 if (i & 1) { 2774 if (noneg) 2775 continue; 2776 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2777 } else { 2778 if (nopos) 2779 continue; 2780 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2781 } 2782 if (addr < entry->start) { 2783 noneg = 1; 2784 if (noneg && nopos) 2785 break; 2786 continue; 2787 } 2788 if (addr >= entry->end) { 2789 nopos = 1; 2790 if (noneg && nopos) 2791 break; 2792 continue; 2793 } 2794 2795 /* 2796 * Skip pages already mapped, and stop scanning in that 2797 * direction. When the scan terminates in both directions 2798 * we are done. 2799 */ 2800 if (pmap_prefault_ok(pmap, addr) == 0) { 2801 if (i & 1) 2802 noneg = 1; 2803 else 2804 nopos = 1; 2805 if (noneg && nopos) 2806 break; 2807 continue; 2808 } 2809 2810 /* 2811 * Follow the VM object chain to obtain the page to be mapped 2812 * into the pmap. 2813 * 2814 * If we reach the terminal object without finding a page 2815 * and we determine it would be advantageous, then allocate 2816 * a zero-fill page for the base object. The base object 2817 * is guaranteed to be OBJT_DEFAULT for this case. 2818 * 2819 * In order to not have to check the pager via *haspage*() 2820 * we stop if any non-default object is encountered. e.g. 2821 * a vnode or swap object would stop the loop. 2822 */ 2823 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2824 lobject = object; 2825 pindex = index; 2826 pprot = prot; 2827 2828 KKASSERT(lobject == entry->object.vm_object); 2829 /*vm_object_hold(lobject); implied */ 2830 2831 while ((m = vm_page_lookup_busy_try(lobject, pindex, 2832 TRUE, &error)) == NULL) { 2833 if (lobject->type != OBJT_DEFAULT) 2834 break; 2835 if (lobject->backing_object == NULL) { 2836 if (vm_fast_fault == 0) 2837 break; 2838 if ((prot & VM_PROT_WRITE) == 0 || 2839 vm_page_count_min(0)) { 2840 break; 2841 } 2842 2843 /* 2844 * NOTE: Allocated from base object 2845 */ 2846 m = vm_page_alloc(object, index, 2847 VM_ALLOC_NORMAL | 2848 VM_ALLOC_ZERO | 2849 VM_ALLOC_USE_GD | 2850 VM_ALLOC_NULL_OK); 2851 if (m == NULL) 2852 break; 2853 allocated = 1; 2854 pprot = prot; 2855 /* lobject = object .. not needed */ 2856 break; 2857 } 2858 if (lobject->backing_object_offset & PAGE_MASK) 2859 break; 2860 nobject = lobject->backing_object; 2861 vm_object_hold(nobject); 2862 KKASSERT(nobject == lobject->backing_object); 2863 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 2864 if (lobject != object) { 2865 vm_object_lock_swap(); 2866 vm_object_drop(lobject); 2867 } 2868 lobject = nobject; 2869 pprot &= ~VM_PROT_WRITE; 2870 vm_object_chain_acquire(lobject, 0); 2871 } 2872 2873 /* 2874 * NOTE: A non-NULL (m) will be associated with lobject if 2875 * it was found there, otherwise it is probably a 2876 * zero-fill page associated with the base object. 2877 * 2878 * Give-up if no page is available. 2879 */ 2880 if (m == NULL) { 2881 if (lobject != object) { 2882 #if 0 2883 if (object->backing_object != lobject) 2884 vm_object_hold(object->backing_object); 2885 #endif 2886 vm_object_chain_release_all( 2887 object->backing_object, lobject); 2888 #if 0 2889 if (object->backing_object != lobject) 2890 vm_object_drop(object->backing_object); 2891 #endif 2892 vm_object_drop(lobject); 2893 } 2894 break; 2895 } 2896 2897 /* 2898 * The object must be marked dirty if we are mapping a 2899 * writable page. m->object is either lobject or object, 2900 * both of which are still held. Do this before we 2901 * potentially drop the object. 2902 */ 2903 if (pprot & VM_PROT_WRITE) 2904 vm_object_set_writeable_dirty(m->object); 2905 2906 /* 2907 * Do not conditionalize on PG_RAM. If pages are present in 2908 * the VM system we assume optimal caching. If caching is 2909 * not optimal the I/O gravy train will be restarted when we 2910 * hit an unavailable page. We do not want to try to restart 2911 * the gravy train now because we really don't know how much 2912 * of the object has been cached. The cost for restarting 2913 * the gravy train should be low (since accesses will likely 2914 * be I/O bound anyway). 2915 */ 2916 if (lobject != object) { 2917 #if 0 2918 if (object->backing_object != lobject) 2919 vm_object_hold(object->backing_object); 2920 #endif 2921 vm_object_chain_release_all(object->backing_object, 2922 lobject); 2923 #if 0 2924 if (object->backing_object != lobject) 2925 vm_object_drop(object->backing_object); 2926 #endif 2927 vm_object_drop(lobject); 2928 } 2929 2930 /* 2931 * Enter the page into the pmap if appropriate. If we had 2932 * allocated the page we have to place it on a queue. If not 2933 * we just have to make sure it isn't on the cache queue 2934 * (pages on the cache queue are not allowed to be mapped). 2935 */ 2936 if (allocated) { 2937 /* 2938 * Page must be zerod. 2939 */ 2940 vm_page_zero_fill(m); 2941 mycpu->gd_cnt.v_zfod++; 2942 m->valid = VM_PAGE_BITS_ALL; 2943 2944 /* 2945 * Handle dirty page case 2946 */ 2947 if (pprot & VM_PROT_WRITE) 2948 vm_set_nosync(m, entry); 2949 pmap_enter(pmap, addr, m, pprot, 0, entry); 2950 mycpu->gd_cnt.v_vm_faults++; 2951 if (curthread->td_lwp) 2952 ++curthread->td_lwp->lwp_ru.ru_minflt; 2953 vm_page_deactivate(m); 2954 if (pprot & VM_PROT_WRITE) { 2955 /*vm_object_set_writeable_dirty(m->object);*/ 2956 vm_set_nosync(m, entry); 2957 if (fault_flags & VM_FAULT_DIRTY) { 2958 vm_page_dirty(m); 2959 /*XXX*/ 2960 swap_pager_unswapped(m); 2961 } 2962 } 2963 vm_page_wakeup(m); 2964 } else if (error) { 2965 /* couldn't busy page, no wakeup */ 2966 } else if ( 2967 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2968 (m->flags & PG_FICTITIOUS) == 0) { 2969 /* 2970 * A fully valid page not undergoing soft I/O can 2971 * be immediately entered into the pmap. 2972 */ 2973 if ((m->queue - m->pc) == PQ_CACHE) 2974 vm_page_deactivate(m); 2975 if (pprot & VM_PROT_WRITE) { 2976 /*vm_object_set_writeable_dirty(m->object);*/ 2977 vm_set_nosync(m, entry); 2978 if (fault_flags & VM_FAULT_DIRTY) { 2979 vm_page_dirty(m); 2980 /*XXX*/ 2981 swap_pager_unswapped(m); 2982 } 2983 } 2984 if (pprot & VM_PROT_WRITE) 2985 vm_set_nosync(m, entry); 2986 pmap_enter(pmap, addr, m, pprot, 0, entry); 2987 mycpu->gd_cnt.v_vm_faults++; 2988 if (curthread->td_lwp) 2989 ++curthread->td_lwp->lwp_ru.ru_minflt; 2990 vm_page_wakeup(m); 2991 } else { 2992 vm_page_wakeup(m); 2993 } 2994 } 2995 vm_object_chain_release(object); 2996 vm_object_drop(object); 2997 } 2998 2999 /* 3000 * Object can be held shared 3001 */ 3002 static void 3003 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 3004 vm_map_entry_t entry, int prot, int fault_flags) 3005 { 3006 struct lwp *lp; 3007 vm_page_t m; 3008 vm_offset_t addr; 3009 vm_pindex_t pindex; 3010 vm_object_t object; 3011 int i; 3012 int noneg; 3013 int nopos; 3014 int maxpages; 3015 3016 /* 3017 * Get stable max count value, disabled if set to 0 3018 */ 3019 maxpages = vm_prefault_pages; 3020 cpu_ccfence(); 3021 if (maxpages <= 0) 3022 return; 3023 3024 /* 3025 * We do not currently prefault mappings that use virtual page 3026 * tables. We do not prefault foreign pmaps. 3027 */ 3028 if (entry->maptype != VM_MAPTYPE_NORMAL) 3029 return; 3030 lp = curthread->td_lwp; 3031 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 3032 return; 3033 object = entry->object.vm_object; 3034 if (object->backing_object != NULL) 3035 return; 3036 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 3037 3038 /* 3039 * Limit pre-fault count to 1024 pages. 3040 */ 3041 if (maxpages > 1024) 3042 maxpages = 1024; 3043 3044 noneg = 0; 3045 nopos = 0; 3046 for (i = 0; i < maxpages; ++i) { 3047 int error; 3048 3049 /* 3050 * Calculate the page to pre-fault, stopping the scan in 3051 * each direction separately if the limit is reached. 3052 */ 3053 if (i & 1) { 3054 if (noneg) 3055 continue; 3056 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 3057 } else { 3058 if (nopos) 3059 continue; 3060 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 3061 } 3062 if (addr < entry->start) { 3063 noneg = 1; 3064 if (noneg && nopos) 3065 break; 3066 continue; 3067 } 3068 if (addr >= entry->end) { 3069 nopos = 1; 3070 if (noneg && nopos) 3071 break; 3072 continue; 3073 } 3074 3075 /* 3076 * Follow the VM object chain to obtain the page to be mapped 3077 * into the pmap. This version of the prefault code only 3078 * works with terminal objects. 3079 * 3080 * The page must already exist. If we encounter a problem 3081 * we stop here. 3082 * 3083 * WARNING! We cannot call swap_pager_unswapped() or insert 3084 * a new vm_page with a shared token. 3085 */ 3086 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 3087 3088 /* 3089 * Skip pages already mapped, and stop scanning in that 3090 * direction. When the scan terminates in both directions 3091 * we are done. 3092 */ 3093 if (pmap_prefault_ok(pmap, addr) == 0) { 3094 if (i & 1) 3095 noneg = 1; 3096 else 3097 nopos = 1; 3098 if (noneg && nopos) 3099 break; 3100 continue; 3101 } 3102 3103 /* 3104 * Shortcut the read-only mapping case using the far more 3105 * efficient vm_page_lookup_sbusy_try() function. This 3106 * allows us to acquire the page soft-busied only which 3107 * is especially nice for concurrent execs of the same 3108 * program. 3109 * 3110 * The lookup function also validates page suitability 3111 * (all valid bits set, and not fictitious). 3112 * 3113 * If the page is in PQ_CACHE we have to fall-through 3114 * and hard-busy it so we can move it out of PQ_CACHE. 3115 */ 3116 if ((prot & VM_PROT_WRITE) == 0) { 3117 m = vm_page_lookup_sbusy_try(object, pindex, 3118 0, PAGE_SIZE); 3119 if (m == NULL) 3120 break; 3121 if ((m->queue - m->pc) != PQ_CACHE) { 3122 pmap_enter(pmap, addr, m, prot, 0, entry); 3123 mycpu->gd_cnt.v_vm_faults++; 3124 if (curthread->td_lwp) 3125 ++curthread->td_lwp->lwp_ru.ru_minflt; 3126 vm_page_sbusy_drop(m); 3127 continue; 3128 } 3129 vm_page_sbusy_drop(m); 3130 } 3131 3132 /* 3133 * Fallback to normal vm_page lookup code. This code 3134 * hard-busies the page. Not only that, but the page 3135 * can remain in that state for a significant period 3136 * time due to pmap_enter()'s overhead. 3137 */ 3138 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3139 if (m == NULL || error) 3140 break; 3141 3142 /* 3143 * Stop if the page cannot be trivially entered into the 3144 * pmap. 3145 */ 3146 if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) || 3147 (m->flags & PG_FICTITIOUS) || 3148 ((m->flags & PG_SWAPPED) && 3149 (prot & VM_PROT_WRITE) && 3150 (fault_flags & VM_FAULT_DIRTY))) { 3151 vm_page_wakeup(m); 3152 break; 3153 } 3154 3155 /* 3156 * Enter the page into the pmap. The object might be held 3157 * shared so we can't do any (serious) modifying operation 3158 * on it. 3159 */ 3160 if ((m->queue - m->pc) == PQ_CACHE) 3161 vm_page_deactivate(m); 3162 if (prot & VM_PROT_WRITE) { 3163 vm_object_set_writeable_dirty(m->object); 3164 vm_set_nosync(m, entry); 3165 if (fault_flags & VM_FAULT_DIRTY) { 3166 vm_page_dirty(m); 3167 /* can't happeen due to conditional above */ 3168 /* swap_pager_unswapped(m); */ 3169 } 3170 } 3171 pmap_enter(pmap, addr, m, prot, 0, entry); 3172 mycpu->gd_cnt.v_vm_faults++; 3173 if (curthread->td_lwp) 3174 ++curthread->td_lwp->lwp_ru.ru_minflt; 3175 vm_page_wakeup(m); 3176 } 3177 } 3178