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 <sys/thread2.h> 129 #include <vm/vm_page2.h> 130 131 struct faultstate { 132 vm_page_t m; 133 vm_object_t object; 134 vm_pindex_t pindex; 135 vm_prot_t prot; 136 vm_page_t first_m; 137 vm_object_t first_object; 138 vm_prot_t first_prot; 139 vm_map_t map; 140 vm_map_entry_t entry; 141 int lookup_still_valid; 142 int hardfault; 143 int fault_flags; 144 int map_generation; 145 int shared; 146 int first_shared; 147 boolean_t wired; 148 struct vnode *vp; 149 }; 150 151 static int debug_fault = 0; 152 SYSCTL_INT(_vm, OID_AUTO, debug_fault, CTLFLAG_RW, &debug_fault, 0, ""); 153 static int debug_cluster = 0; 154 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, ""); 155 static int virtual_copy_enable = 1; 156 SYSCTL_INT(_vm, OID_AUTO, virtual_copy_enable, CTLFLAG_RW, 157 &virtual_copy_enable, 0, ""); 158 int vm_shared_fault = 1; 159 TUNABLE_INT("vm.shared_fault", &vm_shared_fault); 160 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, 161 &vm_shared_fault, 0, "Allow shared token on vm_object"); 162 163 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int); 164 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, 165 vpte_t, int, int); 166 #if 0 167 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 168 #endif 169 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry); 170 static void vm_prefault(pmap_t pmap, vm_offset_t addra, 171 vm_map_entry_t entry, int prot, int fault_flags); 172 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 173 vm_map_entry_t entry, int prot, int fault_flags); 174 175 static __inline void 176 release_page(struct faultstate *fs) 177 { 178 vm_page_deactivate(fs->m); 179 vm_page_wakeup(fs->m); 180 fs->m = NULL; 181 } 182 183 /* 184 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse 185 * requires relocking and then checking the timestamp. 186 * 187 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do 188 * not have to update fs->map_generation here. 189 * 190 * NOTE: This function can fail due to a deadlock against the caller's 191 * holding of a vm_page BUSY. 192 */ 193 static __inline int 194 relock_map(struct faultstate *fs) 195 { 196 int error; 197 198 if (fs->lookup_still_valid == FALSE && fs->map) { 199 error = vm_map_lock_read_to(fs->map); 200 if (error == 0) 201 fs->lookup_still_valid = TRUE; 202 } else { 203 error = 0; 204 } 205 return error; 206 } 207 208 static __inline void 209 unlock_map(struct faultstate *fs) 210 { 211 if (fs->lookup_still_valid && fs->map) { 212 vm_map_lookup_done(fs->map, fs->entry, 0); 213 fs->lookup_still_valid = FALSE; 214 } 215 } 216 217 /* 218 * Clean up after a successful call to vm_fault_object() so another call 219 * to vm_fault_object() can be made. 220 */ 221 static void 222 _cleanup_successful_fault(struct faultstate *fs, int relock) 223 { 224 /* 225 * We allocated a junk page for a COW operation that did 226 * not occur, the page must be freed. 227 */ 228 if (fs->object != fs->first_object) { 229 KKASSERT(fs->first_shared == 0); 230 vm_page_free(fs->first_m); 231 vm_object_pip_wakeup(fs->object); 232 fs->first_m = NULL; 233 } 234 235 /* 236 * Reset fs->object. 237 */ 238 fs->object = fs->first_object; 239 if (relock && fs->lookup_still_valid == FALSE) { 240 if (fs->map) 241 vm_map_lock_read(fs->map); 242 fs->lookup_still_valid = TRUE; 243 } 244 } 245 246 static void 247 _unlock_things(struct faultstate *fs, int dealloc) 248 { 249 _cleanup_successful_fault(fs, 0); 250 if (dealloc) { 251 /*vm_object_deallocate(fs->first_object);*/ 252 /*fs->first_object = NULL; drop used later on */ 253 } 254 unlock_map(fs); 255 if (fs->vp != NULL) { 256 vput(fs->vp); 257 fs->vp = NULL; 258 } 259 } 260 261 #define unlock_things(fs) _unlock_things(fs, 0) 262 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 263 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 264 265 /* 266 * Virtual copy tests. Used by the fault code to determine if a 267 * page can be moved from an orphan vm_object into its shadow 268 * instead of copying its contents. 269 */ 270 static __inline int 271 virtual_copy_test(struct faultstate *fs) 272 { 273 /* 274 * Must be holding exclusive locks 275 */ 276 if (fs->first_shared || fs->shared || virtual_copy_enable == 0) 277 return 0; 278 279 /* 280 * Map, if present, has not changed 281 */ 282 if (fs->map && fs->map_generation != fs->map->timestamp) 283 return 0; 284 285 /* 286 * Only one shadow object 287 */ 288 if (fs->object->shadow_count != 1) 289 return 0; 290 291 /* 292 * No COW refs, except us 293 */ 294 if (fs->object->ref_count != 1) 295 return 0; 296 297 /* 298 * No one else can look this object up 299 */ 300 if (fs->object->handle != NULL) 301 return 0; 302 303 /* 304 * No other ways to look the object up 305 */ 306 if (fs->object->type != OBJT_DEFAULT && 307 fs->object->type != OBJT_SWAP) 308 return 0; 309 310 /* 311 * We don't chase down the shadow chain 312 */ 313 if (fs->object != fs->first_object->backing_object) 314 return 0; 315 316 return 1; 317 } 318 319 static __inline int 320 virtual_copy_ok(struct faultstate *fs) 321 { 322 if (virtual_copy_test(fs)) { 323 /* 324 * Grab the lock and re-test changeable items. 325 */ 326 if (fs->lookup_still_valid == FALSE && fs->map) { 327 if (lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT)) 328 return 0; 329 fs->lookup_still_valid = TRUE; 330 if (virtual_copy_test(fs)) { 331 fs->map_generation = ++fs->map->timestamp; 332 return 1; 333 } 334 fs->lookup_still_valid = FALSE; 335 lockmgr(&fs->map->lock, LK_RELEASE); 336 } 337 } 338 return 0; 339 } 340 341 /* 342 * TRYPAGER 343 * 344 * Determine if the pager for the current object *might* contain the page. 345 * 346 * We only need to try the pager if this is not a default object (default 347 * objects are zero-fill and have no real pager), and if we are not taking 348 * a wiring fault or if the FS entry is wired. 349 */ 350 #define TRYPAGER(fs) \ 351 (fs->object->type != OBJT_DEFAULT && \ 352 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->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.wired); 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 result = vm_map_lookup(&fs.map, vaddr, 460 VM_PROT_READ|VM_PROT_WRITE| 461 VM_PROT_OVERRIDE_WRITE, 462 &fs.entry, &fs.first_object, 463 &first_pindex, &fs.first_prot, 464 &fs.wired); 465 if (result != KERN_SUCCESS) { 466 /* could also be KERN_FAILURE_NOFAULT */ 467 result = KERN_FAILURE; 468 goto done; 469 } 470 471 /* 472 * If we don't COW now, on a user wire, the user will never 473 * be able to write to the mapping. If we don't make this 474 * restriction, the bookkeeping would be nearly impossible. 475 * 476 * XXX We have a shared lock, this will have a MP race but 477 * I don't see how it can hurt anything. 478 */ 479 if ((fs.entry->protection & VM_PROT_WRITE) == 0) { 480 atomic_clear_char(&fs.entry->max_protection, 481 VM_PROT_WRITE); 482 } 483 } 484 485 /* 486 * fs.map is read-locked 487 * 488 * Misc checks. Save the map generation number to detect races. 489 */ 490 fs.map_generation = fs.map->timestamp; 491 fs.lookup_still_valid = TRUE; 492 fs.first_m = NULL; 493 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 494 fs.prot = fs.first_prot; /* default (used by uksmap) */ 495 496 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) { 497 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 498 panic("vm_fault: fault on nofault entry, addr: %p", 499 (void *)vaddr); 500 } 501 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) && 502 vaddr >= fs.entry->start && 503 vaddr < fs.entry->start + PAGE_SIZE) { 504 panic("vm_fault: fault on stack guard, addr: %p", 505 (void *)vaddr); 506 } 507 } 508 509 /* 510 * A user-kernel shared map has no VM object and bypasses 511 * everything. We execute the uksmap function with a temporary 512 * fictitious vm_page. The address is directly mapped with no 513 * management. 514 */ 515 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 516 struct vm_page fakem; 517 518 bzero(&fakem, sizeof(fakem)); 519 fakem.pindex = first_pindex; 520 fakem.flags = PG_FICTITIOUS | PG_UNMANAGED; 521 fakem.busy_count = PBUSY_LOCKED; 522 fakem.valid = VM_PAGE_BITS_ALL; 523 fakem.pat_mode = VM_MEMATTR_DEFAULT; 524 if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) { 525 result = KERN_FAILURE; 526 unlock_things(&fs); 527 goto done2; 528 } 529 pmap_enter(fs.map->pmap, vaddr, &fakem, fs.prot | inherit_prot, 530 fs.wired, fs.entry); 531 goto done_success; 532 } 533 534 /* 535 * A system map entry may return a NULL object. No object means 536 * no pager means an unrecoverable kernel fault. 537 */ 538 if (fs.first_object == NULL) { 539 panic("vm_fault: unrecoverable fault at %p in entry %p", 540 (void *)vaddr, fs.entry); 541 } 542 543 /* 544 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 545 * is set. 546 * 547 * Unfortunately a deadlock can occur if we are forced to page-in 548 * from swap, but diving all the way into the vm_pager_get_page() 549 * function to find out is too much. Just check the object type. 550 * 551 * The deadlock is a CAM deadlock on a busy VM page when trying 552 * to finish an I/O if another process gets stuck in 553 * vop_helper_read_shortcut() due to a swap fault. 554 */ 555 if ((td->td_flags & TDF_NOFAULT) && 556 (retry || 557 fs.first_object->type == OBJT_VNODE || 558 fs.first_object->type == OBJT_SWAP || 559 fs.first_object->backing_object)) { 560 result = KERN_FAILURE; 561 unlock_things(&fs); 562 goto done2; 563 } 564 565 /* 566 * If the entry is wired we cannot change the page protection. 567 */ 568 if (fs.wired) 569 fault_type = fs.first_prot; 570 571 /* 572 * We generally want to avoid unnecessary exclusive modes on backing 573 * and terminal objects because this can seriously interfere with 574 * heavily fork()'d processes (particularly /bin/sh scripts). 575 * 576 * However, we also want to avoid unnecessary retries due to needed 577 * shared->exclusive promotion for common faults. Exclusive mode is 578 * always needed if any page insertion, rename, or free occurs in an 579 * object (and also indirectly if any I/O is done). 580 * 581 * The main issue here is going to be fs.first_shared. If the 582 * first_object has a backing object which isn't shadowed and the 583 * process is single-threaded we might as well use an exclusive 584 * lock/chain right off the bat. 585 */ 586 if (fs.first_shared && fs.first_object->backing_object && 587 LIST_EMPTY(&fs.first_object->shadow_head) && 588 td->td_proc && td->td_proc->p_nthreads == 1) { 589 fs.first_shared = 0; 590 } 591 592 /* 593 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 594 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 595 * we can try shared first. 596 */ 597 if (fault_flags & VM_FAULT_UNSWAP) { 598 fs.first_shared = 0; 599 } 600 601 /* 602 * Obtain a top-level object lock, shared or exclusive depending 603 * on fs.first_shared. If a shared lock winds up being insufficient 604 * we will retry with an exclusive lock. 605 * 606 * The vnode pager lock is always shared. 607 */ 608 if (fs.first_shared) 609 vm_object_hold_shared(fs.first_object); 610 else 611 vm_object_hold(fs.first_object); 612 if (fs.vp == NULL) 613 fs.vp = vnode_pager_lock(fs.first_object); 614 615 /* 616 * The page we want is at (first_object, first_pindex), but if the 617 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 618 * page table to figure out the actual pindex. 619 * 620 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 621 * ONLY 622 */ 623 didilock = 0; 624 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 625 vm_map_interlock(fs.map, &ilock, vaddr, vaddr + PAGE_SIZE); 626 didilock = 1; 627 result = vm_fault_vpagetable(&fs, &first_pindex, 628 fs.entry->aux.master_pde, 629 fault_type, 1); 630 if (result == KERN_TRY_AGAIN) { 631 vm_map_deinterlock(fs.map, &ilock); 632 vm_object_drop(fs.first_object); 633 ++retry; 634 goto RetryFault; 635 } 636 if (result != KERN_SUCCESS) { 637 vm_map_deinterlock(fs.map, &ilock); 638 goto done; 639 } 640 } 641 642 /* 643 * Now we have the actual (object, pindex), fault in the page. If 644 * vm_fault_object() fails it will unlock and deallocate the FS 645 * data. If it succeeds everything remains locked and fs->object 646 * will have an additional PIP count if it is not equal to 647 * fs->first_object 648 * 649 * vm_fault_object will set fs->prot for the pmap operation. It is 650 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 651 * page can be safely written. However, it will force a read-only 652 * mapping for a read fault if the memory is managed by a virtual 653 * page table. 654 * 655 * If the fault code uses the shared object lock shortcut 656 * we must not try to burst (we can't allocate VM pages). 657 */ 658 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 659 660 if (debug_fault > 0) { 661 --debug_fault; 662 kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x " 663 "fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n", 664 result, (intmax_t)vaddr, fault_type, fault_flags, 665 fs.m, fs.prot, fs.wired, fs.entry); 666 } 667 668 if (result == KERN_TRY_AGAIN) { 669 if (didilock) 670 vm_map_deinterlock(fs.map, &ilock); 671 vm_object_drop(fs.first_object); 672 ++retry; 673 goto RetryFault; 674 } 675 if (result != KERN_SUCCESS) { 676 if (didilock) 677 vm_map_deinterlock(fs.map, &ilock); 678 goto done; 679 } 680 681 /* 682 * On success vm_fault_object() does not unlock or deallocate, and fs.m 683 * will contain a busied page. 684 * 685 * Enter the page into the pmap and do pmap-related adjustments. 686 */ 687 KKASSERT(fs.lookup_still_valid == TRUE); 688 vm_page_flag_set(fs.m, PG_REFERENCED); 689 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot | inherit_prot, 690 fs.wired, fs.entry); 691 692 if (didilock) 693 vm_map_deinterlock(fs.map, &ilock); 694 695 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */ 696 KKASSERT(fs.m->busy_count & PBUSY_LOCKED); 697 698 /* 699 * If the page is not wired down, then put it where the pageout daemon 700 * can find it. 701 */ 702 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 703 if (fs.wired) 704 vm_page_wire(fs.m); 705 else 706 vm_page_unwire(fs.m, 1); 707 } else { 708 vm_page_activate(fs.m); 709 } 710 vm_page_wakeup(fs.m); 711 712 /* 713 * Burst in a few more pages if possible. The fs.map should still 714 * be locked. To avoid interlocking against a vnode->getblk 715 * operation we had to be sure to unbusy our primary vm_page above 716 * first. 717 * 718 * A normal burst can continue down backing store, only execute 719 * if we are holding an exclusive lock, otherwise the exclusive 720 * locks the burst code gets might cause excessive SMP collisions. 721 * 722 * A quick burst can be utilized when there is no backing object 723 * (i.e. a shared file mmap). 724 */ 725 if ((fault_flags & VM_FAULT_BURST) && 726 (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 && 727 fs.wired == 0) { 728 if (fs.first_shared == 0 && fs.shared == 0) { 729 vm_prefault(fs.map->pmap, vaddr, 730 fs.entry, fs.prot, fault_flags); 731 } else { 732 vm_prefault_quick(fs.map->pmap, vaddr, 733 fs.entry, fs.prot, fault_flags); 734 } 735 } 736 737 done_success: 738 mycpu->gd_cnt.v_vm_faults++; 739 if (td->td_lwp) 740 ++td->td_lwp->lwp_ru.ru_minflt; 741 742 /* 743 * Unlock everything, and return 744 */ 745 unlock_things(&fs); 746 747 if (td->td_lwp) { 748 if (fs.hardfault) { 749 td->td_lwp->lwp_ru.ru_majflt++; 750 } else { 751 td->td_lwp->lwp_ru.ru_minflt++; 752 } 753 } 754 755 /*vm_object_deallocate(fs.first_object);*/ 756 /*fs.m = NULL; */ 757 /*fs.first_object = NULL; must still drop later */ 758 759 result = KERN_SUCCESS; 760 done: 761 if (fs.first_object) 762 vm_object_drop(fs.first_object); 763 done2: 764 if (lp) 765 lp->lwp_flags &= ~LWP_PAGING; 766 767 #if !defined(NO_SWAPPING) 768 /* 769 * Check the process RSS limit and force deactivation and 770 * (asynchronous) paging if necessary. This is a complex operation, 771 * only do it for direct user-mode faults, for now. 772 * 773 * To reduce overhead implement approximately a ~16MB hysteresis. 774 */ 775 p = td->td_proc; 776 if ((fault_flags & VM_FAULT_USERMODE) && lp && 777 p->p_limit && map->pmap && vm_pageout_memuse_mode >= 1 && 778 map != &kernel_map) { 779 vm_pindex_t limit; 780 vm_pindex_t size; 781 782 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 783 p->p_rlimit[RLIMIT_RSS].rlim_max)); 784 size = pmap_resident_tlnw_count(map->pmap); 785 if (limit >= 0 && size > 4096 && size - 4096 >= limit) { 786 vm_pageout_map_deactivate_pages(map, limit); 787 } 788 } 789 #endif 790 791 return (result); 792 } 793 794 /* 795 * Fault in the specified virtual address in the current process map, 796 * returning a held VM page or NULL. See vm_fault_page() for more 797 * information. 798 * 799 * No requirements. 800 */ 801 vm_page_t 802 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, 803 int *errorp, int *busyp) 804 { 805 struct lwp *lp = curthread->td_lwp; 806 vm_page_t m; 807 808 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 809 fault_type, VM_FAULT_NORMAL, 810 errorp, busyp); 811 return(m); 812 } 813 814 /* 815 * Fault in the specified virtual address in the specified map, doing all 816 * necessary manipulation of the object store and all necessary I/O. Return 817 * a held VM page or NULL, and set *errorp. The related pmap is not 818 * updated. 819 * 820 * If busyp is not NULL then *busyp will be set to TRUE if this routine 821 * decides to return a busied page (aka VM_PROT_WRITE), or FALSE if it 822 * does not (VM_PROT_WRITE not specified or busyp is NULL). If busyp is 823 * NULL the returned page is only held. 824 * 825 * If the caller has no intention of writing to the page's contents, busyp 826 * can be passed as NULL along with VM_PROT_WRITE to force a COW operation 827 * without busying the page. 828 * 829 * The returned page will also be marked PG_REFERENCED. 830 * 831 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 832 * error will be returned. 833 * 834 * No requirements. 835 */ 836 vm_page_t 837 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 838 int fault_flags, int *errorp, int *busyp) 839 { 840 vm_pindex_t first_pindex; 841 struct faultstate fs; 842 int result; 843 int retry; 844 int growstack; 845 vm_prot_t orig_fault_type = fault_type; 846 847 retry = 0; 848 fs.hardfault = 0; 849 fs.fault_flags = fault_flags; 850 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 851 852 /* 853 * Dive the pmap (concurrency possible). If we find the 854 * appropriate page we can terminate early and quickly. 855 * 856 * This works great for normal programs but will always return 857 * NULL for host lookups of vkernel maps in VMM mode. 858 * 859 * NOTE: pmap_fault_page_quick() might not busy the page. If 860 * VM_PROT_WRITE or VM_PROT_OVERRIDE_WRITE is set in 861 * fault_type and pmap_fault_page_quick() returns non-NULL, 862 * it will safely dirty the returned vm_page_t for us. We 863 * cannot safely dirty it here (it might not be busy). 864 */ 865 fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type, busyp); 866 if (fs.m) { 867 *errorp = 0; 868 return(fs.m); 869 } 870 871 /* 872 * Otherwise take a concurrency hit and do a formal page 873 * fault. 874 */ 875 fs.vp = NULL; 876 fs.shared = vm_shared_fault; 877 fs.first_shared = vm_shared_fault; 878 growstack = 1; 879 880 /* 881 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 882 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 883 * we can try shared first. 884 */ 885 if (fault_flags & VM_FAULT_UNSWAP) { 886 fs.first_shared = 0; 887 } 888 889 RetryFault: 890 /* 891 * Find the vm_map_entry representing the backing store and resolve 892 * the top level object and page index. This may have the side 893 * effect of executing a copy-on-write on the map entry and/or 894 * creating a shadow object, but will not COW any actual VM pages. 895 * 896 * On success fs.map is left read-locked and various other fields 897 * are initialized but not otherwise referenced or locked. 898 * 899 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 900 * if the map entry is a virtual page table and also writable, 901 * so we can set the 'A'accessed bit in the virtual page table 902 * entry. 903 */ 904 fs.map = map; 905 result = vm_map_lookup(&fs.map, vaddr, fault_type, 906 &fs.entry, &fs.first_object, 907 &first_pindex, &fs.first_prot, &fs.wired); 908 909 if (result != KERN_SUCCESS) { 910 if (result == KERN_FAILURE_NOFAULT) { 911 *errorp = KERN_FAILURE; 912 fs.m = NULL; 913 goto done; 914 } 915 if (result != KERN_PROTECTION_FAILURE || 916 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 917 { 918 if (result == KERN_INVALID_ADDRESS && growstack && 919 map != &kernel_map && curproc != NULL) { 920 result = vm_map_growstack(map, vaddr); 921 if (result == KERN_SUCCESS) { 922 growstack = 0; 923 ++retry; 924 goto RetryFault; 925 } 926 result = KERN_FAILURE; 927 } 928 fs.m = NULL; 929 *errorp = result; 930 goto done; 931 } 932 933 /* 934 * If we are user-wiring a r/w segment, and it is COW, then 935 * we need to do the COW operation. Note that we don't 936 * currently COW RO sections now, because it is NOT desirable 937 * to COW .text. We simply keep .text from ever being COW'ed 938 * and take the heat that one cannot debug wired .text sections. 939 */ 940 result = vm_map_lookup(&fs.map, vaddr, 941 VM_PROT_READ|VM_PROT_WRITE| 942 VM_PROT_OVERRIDE_WRITE, 943 &fs.entry, &fs.first_object, 944 &first_pindex, &fs.first_prot, 945 &fs.wired); 946 if (result != KERN_SUCCESS) { 947 /* could also be KERN_FAILURE_NOFAULT */ 948 *errorp = KERN_FAILURE; 949 fs.m = NULL; 950 goto done; 951 } 952 953 /* 954 * If we don't COW now, on a user wire, the user will never 955 * be able to write to the mapping. If we don't make this 956 * restriction, the bookkeeping would be nearly impossible. 957 * 958 * XXX We have a shared lock, this will have a MP race but 959 * I don't see how it can hurt anything. 960 */ 961 if ((fs.entry->protection & VM_PROT_WRITE) == 0) { 962 atomic_clear_char(&fs.entry->max_protection, 963 VM_PROT_WRITE); 964 } 965 } 966 967 /* 968 * fs.map is read-locked 969 * 970 * Misc checks. Save the map generation number to detect races. 971 */ 972 fs.map_generation = fs.map->timestamp; 973 fs.lookup_still_valid = TRUE; 974 fs.first_m = NULL; 975 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 976 977 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 978 panic("vm_fault: fault on nofault entry, addr: %lx", 979 (u_long)vaddr); 980 } 981 982 /* 983 * A user-kernel shared map has no VM object and bypasses 984 * everything. We execute the uksmap function with a temporary 985 * fictitious vm_page. The address is directly mapped with no 986 * management. 987 */ 988 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 989 struct vm_page fakem; 990 991 bzero(&fakem, sizeof(fakem)); 992 fakem.pindex = first_pindex; 993 fakem.flags = PG_FICTITIOUS | PG_UNMANAGED; 994 fakem.busy_count = PBUSY_LOCKED; 995 fakem.valid = VM_PAGE_BITS_ALL; 996 fakem.pat_mode = VM_MEMATTR_DEFAULT; 997 if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) { 998 *errorp = KERN_FAILURE; 999 fs.m = NULL; 1000 unlock_things(&fs); 1001 goto done2; 1002 } 1003 fs.m = PHYS_TO_VM_PAGE(fakem.phys_addr); 1004 vm_page_hold(fs.m); 1005 if (busyp) 1006 *busyp = 0; /* don't need to busy R or W */ 1007 unlock_things(&fs); 1008 *errorp = 0; 1009 goto done; 1010 } 1011 1012 1013 /* 1014 * A system map entry may return a NULL object. No object means 1015 * no pager means an unrecoverable kernel fault. 1016 */ 1017 if (fs.first_object == NULL) { 1018 panic("vm_fault: unrecoverable fault at %p in entry %p", 1019 (void *)vaddr, fs.entry); 1020 } 1021 1022 /* 1023 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 1024 * is set. 1025 * 1026 * Unfortunately a deadlock can occur if we are forced to page-in 1027 * from swap, but diving all the way into the vm_pager_get_page() 1028 * function to find out is too much. Just check the object type. 1029 */ 1030 if ((curthread->td_flags & TDF_NOFAULT) && 1031 (retry || 1032 fs.first_object->type == OBJT_VNODE || 1033 fs.first_object->type == OBJT_SWAP || 1034 fs.first_object->backing_object)) { 1035 *errorp = KERN_FAILURE; 1036 unlock_things(&fs); 1037 fs.m = NULL; 1038 goto done2; 1039 } 1040 1041 /* 1042 * If the entry is wired we cannot change the page protection. 1043 */ 1044 if (fs.wired) 1045 fault_type = fs.first_prot; 1046 1047 /* 1048 * Make a reference to this object to prevent its disposal while we 1049 * are messing with it. Once we have the reference, the map is free 1050 * to be diddled. Since objects reference their shadows (and copies), 1051 * they will stay around as well. 1052 * 1053 * The reference should also prevent an unexpected collapse of the 1054 * parent that might move pages from the current object into the 1055 * parent unexpectedly, resulting in corruption. 1056 * 1057 * Bump the paging-in-progress count to prevent size changes (e.g. 1058 * truncation operations) during I/O. This must be done after 1059 * obtaining the vnode lock in order to avoid possible deadlocks. 1060 */ 1061 if (fs.first_shared) 1062 vm_object_hold_shared(fs.first_object); 1063 else 1064 vm_object_hold(fs.first_object); 1065 if (fs.vp == NULL) 1066 fs.vp = vnode_pager_lock(fs.first_object); /* shared */ 1067 1068 /* 1069 * The page we want is at (first_object, first_pindex), but if the 1070 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 1071 * page table to figure out the actual pindex. 1072 * 1073 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 1074 * ONLY 1075 */ 1076 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1077 result = vm_fault_vpagetable(&fs, &first_pindex, 1078 fs.entry->aux.master_pde, 1079 fault_type, 1); 1080 if (result == KERN_TRY_AGAIN) { 1081 vm_object_drop(fs.first_object); 1082 ++retry; 1083 goto RetryFault; 1084 } 1085 if (result != KERN_SUCCESS) { 1086 *errorp = result; 1087 fs.m = NULL; 1088 goto done; 1089 } 1090 } 1091 1092 /* 1093 * Now we have the actual (object, pindex), fault in the page. If 1094 * vm_fault_object() fails it will unlock and deallocate the FS 1095 * data. If it succeeds everything remains locked and fs->object 1096 * will have an additinal PIP count if it is not equal to 1097 * fs->first_object 1098 */ 1099 fs.m = NULL; 1100 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 1101 1102 if (result == KERN_TRY_AGAIN) { 1103 vm_object_drop(fs.first_object); 1104 ++retry; 1105 goto RetryFault; 1106 } 1107 if (result != KERN_SUCCESS) { 1108 *errorp = result; 1109 fs.m = NULL; 1110 goto done; 1111 } 1112 1113 if ((orig_fault_type & VM_PROT_WRITE) && 1114 (fs.prot & VM_PROT_WRITE) == 0) { 1115 *errorp = KERN_PROTECTION_FAILURE; 1116 unlock_and_deallocate(&fs); 1117 fs.m = NULL; 1118 goto done; 1119 } 1120 1121 /* 1122 * DO NOT UPDATE THE PMAP!!! This function may be called for 1123 * a pmap unrelated to the current process pmap, in which case 1124 * the current cpu core will not be listed in the pmap's pm_active 1125 * mask. Thus invalidation interlocks will fail to work properly. 1126 * 1127 * (for example, 'ps' uses procfs to read program arguments from 1128 * each process's stack). 1129 * 1130 * In addition to the above this function will be called to acquire 1131 * a page that might already be faulted in, re-faulting it 1132 * continuously is a waste of time. 1133 * 1134 * XXX could this have been the cause of our random seg-fault 1135 * issues? procfs accesses user stacks. 1136 */ 1137 vm_page_flag_set(fs.m, PG_REFERENCED); 1138 #if 0 1139 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL); 1140 mycpu->gd_cnt.v_vm_faults++; 1141 if (curthread->td_lwp) 1142 ++curthread->td_lwp->lwp_ru.ru_minflt; 1143 #endif 1144 1145 /* 1146 * On success vm_fault_object() does not unlock or deallocate, and fs.m 1147 * will contain a busied page. So we must unlock here after having 1148 * messed with the pmap. 1149 */ 1150 unlock_things(&fs); 1151 1152 /* 1153 * Return a held page. We are not doing any pmap manipulation so do 1154 * not set PG_MAPPED. However, adjust the page flags according to 1155 * the fault type because the caller may not use a managed pmapping 1156 * (so we don't want to lose the fact that the page will be dirtied 1157 * if a write fault was specified). 1158 */ 1159 if (fault_type & VM_PROT_WRITE) 1160 vm_page_dirty(fs.m); 1161 vm_page_activate(fs.m); 1162 1163 if (curthread->td_lwp) { 1164 if (fs.hardfault) { 1165 curthread->td_lwp->lwp_ru.ru_majflt++; 1166 } else { 1167 curthread->td_lwp->lwp_ru.ru_minflt++; 1168 } 1169 } 1170 1171 /* 1172 * Unlock everything, and return the held or busied page. 1173 */ 1174 if (busyp) { 1175 if (fault_type & (VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE)) { 1176 vm_page_dirty(fs.m); 1177 *busyp = 1; 1178 } else { 1179 *busyp = 0; 1180 vm_page_hold(fs.m); 1181 vm_page_wakeup(fs.m); 1182 } 1183 } else { 1184 vm_page_hold(fs.m); 1185 vm_page_wakeup(fs.m); 1186 } 1187 /*vm_object_deallocate(fs.first_object);*/ 1188 /*fs.first_object = NULL; */ 1189 *errorp = 0; 1190 1191 done: 1192 if (fs.first_object) 1193 vm_object_drop(fs.first_object); 1194 done2: 1195 return(fs.m); 1196 } 1197 1198 /* 1199 * Fault in the specified (object,offset), dirty the returned page as 1200 * needed. If the requested fault_type cannot be done NULL and an 1201 * error is returned. 1202 * 1203 * A held (but not busied) page is returned. 1204 * 1205 * The passed in object must be held as specified by the shared 1206 * argument. 1207 */ 1208 vm_page_t 1209 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 1210 vm_prot_t fault_type, int fault_flags, 1211 int *sharedp, int *errorp) 1212 { 1213 int result; 1214 vm_pindex_t first_pindex; 1215 struct faultstate fs; 1216 struct vm_map_entry entry; 1217 1218 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1219 bzero(&entry, sizeof(entry)); 1220 entry.object.vm_object = object; 1221 entry.maptype = VM_MAPTYPE_NORMAL; 1222 entry.protection = entry.max_protection = fault_type; 1223 1224 fs.hardfault = 0; 1225 fs.fault_flags = fault_flags; 1226 fs.map = NULL; 1227 fs.shared = vm_shared_fault; 1228 fs.first_shared = *sharedp; 1229 fs.vp = NULL; 1230 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 1231 1232 /* 1233 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 1234 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 1235 * we can try shared first. 1236 */ 1237 if (fs.first_shared && (fault_flags & VM_FAULT_UNSWAP)) { 1238 fs.first_shared = 0; 1239 vm_object_upgrade(object); 1240 } 1241 1242 /* 1243 * Retry loop as needed (typically for shared->exclusive transitions) 1244 */ 1245 RetryFault: 1246 *sharedp = fs.first_shared; 1247 first_pindex = OFF_TO_IDX(offset); 1248 fs.first_object = object; 1249 fs.entry = &entry; 1250 fs.first_prot = fault_type; 1251 fs.wired = 0; 1252 /*fs.map_generation = 0; unused */ 1253 1254 /* 1255 * Make a reference to this object to prevent its disposal while we 1256 * are messing with it. Once we have the reference, the map is free 1257 * to be diddled. Since objects reference their shadows (and copies), 1258 * they will stay around as well. 1259 * 1260 * The reference should also prevent an unexpected collapse of the 1261 * parent that might move pages from the current object into the 1262 * parent unexpectedly, resulting in corruption. 1263 * 1264 * Bump the paging-in-progress count to prevent size changes (e.g. 1265 * truncation operations) during I/O. This must be done after 1266 * obtaining the vnode lock in order to avoid possible deadlocks. 1267 */ 1268 if (fs.vp == NULL) 1269 fs.vp = vnode_pager_lock(fs.first_object); 1270 1271 fs.lookup_still_valid = TRUE; 1272 fs.first_m = NULL; 1273 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 1274 1275 #if 0 1276 /* XXX future - ability to operate on VM object using vpagetable */ 1277 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1278 result = vm_fault_vpagetable(&fs, &first_pindex, 1279 fs.entry->aux.master_pde, 1280 fault_type, 0); 1281 if (result == KERN_TRY_AGAIN) { 1282 if (fs.first_shared == 0 && *sharedp) 1283 vm_object_upgrade(object); 1284 goto RetryFault; 1285 } 1286 if (result != KERN_SUCCESS) { 1287 *errorp = result; 1288 return (NULL); 1289 } 1290 } 1291 #endif 1292 1293 /* 1294 * Now we have the actual (object, pindex), fault in the page. If 1295 * vm_fault_object() fails it will unlock and deallocate the FS 1296 * data. If it succeeds everything remains locked and fs->object 1297 * will have an additinal PIP count if it is not equal to 1298 * fs->first_object 1299 * 1300 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact. 1301 * We may have to upgrade its lock to handle the requested fault. 1302 */ 1303 result = vm_fault_object(&fs, first_pindex, fault_type, 0); 1304 1305 if (result == KERN_TRY_AGAIN) { 1306 if (fs.first_shared == 0 && *sharedp) 1307 vm_object_upgrade(object); 1308 goto RetryFault; 1309 } 1310 if (result != KERN_SUCCESS) { 1311 *errorp = result; 1312 return(NULL); 1313 } 1314 1315 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 1316 *errorp = KERN_PROTECTION_FAILURE; 1317 unlock_and_deallocate(&fs); 1318 return(NULL); 1319 } 1320 1321 /* 1322 * On success vm_fault_object() does not unlock or deallocate, so we 1323 * do it here. Note that the returned fs.m will be busied. 1324 */ 1325 unlock_things(&fs); 1326 1327 /* 1328 * Return a held page. We are not doing any pmap manipulation so do 1329 * not set PG_MAPPED. However, adjust the page flags according to 1330 * the fault type because the caller may not use a managed pmapping 1331 * (so we don't want to lose the fact that the page will be dirtied 1332 * if a write fault was specified). 1333 */ 1334 vm_page_hold(fs.m); 1335 vm_page_activate(fs.m); 1336 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY)) 1337 vm_page_dirty(fs.m); 1338 if (fault_flags & VM_FAULT_UNSWAP) 1339 swap_pager_unswapped(fs.m); 1340 1341 /* 1342 * Indicate that the page was accessed. 1343 */ 1344 vm_page_flag_set(fs.m, PG_REFERENCED); 1345 1346 if (curthread->td_lwp) { 1347 if (fs.hardfault) { 1348 curthread->td_lwp->lwp_ru.ru_majflt++; 1349 } else { 1350 curthread->td_lwp->lwp_ru.ru_minflt++; 1351 } 1352 } 1353 1354 /* 1355 * Unlock everything, and return the held page. 1356 */ 1357 vm_page_wakeup(fs.m); 1358 /*vm_object_deallocate(fs.first_object);*/ 1359 /*fs.first_object = NULL; */ 1360 1361 *errorp = 0; 1362 return(fs.m); 1363 } 1364 1365 /* 1366 * Translate the virtual page number (first_pindex) that is relative 1367 * to the address space into a logical page number that is relative to the 1368 * backing object. Use the virtual page table pointed to by (vpte). 1369 * 1370 * Possibly downgrade the protection based on the vpte bits. 1371 * 1372 * This implements an N-level page table. Any level can terminate the 1373 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 1374 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 1375 */ 1376 static 1377 int 1378 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 1379 vpte_t vpte, int fault_type, int allow_nofault) 1380 { 1381 struct lwbuf *lwb; 1382 struct lwbuf lwb_cache; 1383 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */ 1384 int result; 1385 vpte_t *ptep; 1386 1387 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1388 for (;;) { 1389 /* 1390 * We cannot proceed if the vpte is not valid, not readable 1391 * for a read fault, not writable for a write fault, or 1392 * not executable for an instruction execution fault. 1393 */ 1394 if ((vpte & VPTE_V) == 0) { 1395 unlock_and_deallocate(fs); 1396 return (KERN_FAILURE); 1397 } 1398 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) { 1399 unlock_and_deallocate(fs); 1400 return (KERN_FAILURE); 1401 } 1402 if ((fault_type & VM_PROT_EXECUTE) && (vpte & VPTE_NX)) { 1403 unlock_and_deallocate(fs); 1404 return (KERN_FAILURE); 1405 } 1406 if ((vpte & VPTE_PS) || vshift == 0) 1407 break; 1408 1409 /* 1410 * Get the page table page. Nominally we only read the page 1411 * table, but since we are actively setting VPTE_M and VPTE_A, 1412 * tell vm_fault_object() that we are writing it. 1413 * 1414 * There is currently no real need to optimize this. 1415 */ 1416 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT, 1417 VM_PROT_READ|VM_PROT_WRITE, 1418 allow_nofault); 1419 if (result != KERN_SUCCESS) 1420 return (result); 1421 1422 /* 1423 * Process the returned fs.m and look up the page table 1424 * entry in the page table page. 1425 */ 1426 vshift -= VPTE_PAGE_BITS; 1427 lwb = lwbuf_alloc(fs->m, &lwb_cache); 1428 ptep = ((vpte_t *)lwbuf_kva(lwb) + 1429 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 1430 vm_page_activate(fs->m); 1431 1432 /* 1433 * Page table write-back - entire operation including 1434 * validation of the pte must be atomic to avoid races 1435 * against the vkernel changing the pte. 1436 * 1437 * If the vpte is valid for the* requested operation, do 1438 * a write-back to the page table. 1439 * 1440 * XXX VPTE_M is not set properly for page directory pages. 1441 * It doesn't get set in the page directory if the page table 1442 * is modified during a read access. 1443 */ 1444 for (;;) { 1445 vpte_t nvpte; 1446 1447 /* 1448 * Reload for the cmpset, but make sure the pte is 1449 * still valid. 1450 */ 1451 vpte = *ptep; 1452 cpu_ccfence(); 1453 nvpte = vpte; 1454 1455 if ((vpte & VPTE_V) == 0) 1456 break; 1457 1458 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW)) 1459 nvpte |= VPTE_M | VPTE_A; 1460 if (fault_type & (VM_PROT_READ | VM_PROT_EXECUTE)) 1461 nvpte |= VPTE_A; 1462 if (vpte == nvpte) 1463 break; 1464 if (atomic_cmpset_long(ptep, vpte, nvpte)) { 1465 vm_page_dirty(fs->m); 1466 break; 1467 } 1468 } 1469 lwbuf_free(lwb); 1470 vm_page_flag_set(fs->m, PG_REFERENCED); 1471 vm_page_wakeup(fs->m); 1472 fs->m = NULL; 1473 cleanup_successful_fault(fs); 1474 } 1475 1476 /* 1477 * When the vkernel sets VPTE_RW it expects the real kernel to 1478 * reflect VPTE_M back when the page is modified via the mapping. 1479 * In order to accomplish this the real kernel must map the page 1480 * read-only for read faults and use write faults to reflect VPTE_M 1481 * back. 1482 * 1483 * Once VPTE_M has been set, the real kernel's pte allows writing. 1484 * If the vkernel clears VPTE_M the vkernel must be sure to 1485 * MADV_INVAL the real kernel's mappings to force the real kernel 1486 * to re-fault on the next write so oit can set VPTE_M again. 1487 */ 1488 if ((fault_type & VM_PROT_WRITE) == 0 && 1489 (vpte & (VPTE_RW | VPTE_M)) != (VPTE_RW | VPTE_M)) { 1490 fs->first_prot &= ~VM_PROT_WRITE; 1491 } 1492 1493 /* 1494 * Disable EXECUTE perms if NX bit is set. 1495 */ 1496 if (vpte & VPTE_NX) 1497 fs->first_prot &= ~VM_PROT_EXECUTE; 1498 1499 /* 1500 * Combine remaining address bits with the vpte. 1501 */ 1502 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) + 1503 (*pindex & ((1L << vshift) - 1)); 1504 return (KERN_SUCCESS); 1505 } 1506 1507 1508 /* 1509 * This is the core of the vm_fault code. 1510 * 1511 * Do all operations required to fault-in (fs.first_object, pindex). Run 1512 * through the shadow chain as necessary and do required COW or virtual 1513 * copy operations. The caller has already fully resolved the vm_map_entry 1514 * and, if appropriate, has created a copy-on-write layer. All we need to 1515 * do is iterate the object chain. 1516 * 1517 * On failure (fs) is unlocked and deallocated and the caller may return or 1518 * retry depending on the failure code. On success (fs) is NOT unlocked or 1519 * deallocated, fs.m will contained a resolved, busied page, and fs.object 1520 * will have an additional PIP count if it is not equal to fs.first_object. 1521 * 1522 * If locks based on fs->first_shared or fs->shared are insufficient, 1523 * clear the appropriate field(s) and return RETRY. COWs require that 1524 * first_shared be 0, while page allocations (or frees) require that 1525 * shared be 0. Renames require that both be 0. 1526 * 1527 * NOTE! fs->[first_]shared might be set with VM_FAULT_DIRTY also set. 1528 * we will have to retry with it exclusive if the vm_page is 1529 * PG_SWAPPED. 1530 * 1531 * fs->first_object must be held on call. 1532 */ 1533 static 1534 int 1535 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex, 1536 vm_prot_t fault_type, int allow_nofault) 1537 { 1538 vm_object_t next_object; 1539 vm_pindex_t pindex; 1540 int error; 1541 1542 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1543 fs->prot = fs->first_prot; 1544 fs->object = fs->first_object; 1545 pindex = first_pindex; 1546 1547 vm_object_chain_acquire(fs->first_object, fs->shared); 1548 vm_object_pip_add(fs->first_object, 1); 1549 1550 /* 1551 * If a read fault occurs we try to upgrade the page protection 1552 * and make it also writable if possible. There are three cases 1553 * where we cannot make the page mapping writable: 1554 * 1555 * (1) The mapping is read-only or the VM object is read-only, 1556 * fs->prot above will simply not have VM_PROT_WRITE set. 1557 * 1558 * (2) If the mapping is a virtual page table fs->first_prot will 1559 * have already been properly adjusted by vm_fault_vpagetable(). 1560 * to detect writes so we can set VPTE_M in the virtual page 1561 * table. Used by vkernels. 1562 * 1563 * (3) If the VM page is read-only or copy-on-write, upgrading would 1564 * just result in an unnecessary COW fault. 1565 * 1566 * (4) If the pmap specifically requests A/M bit emulation, downgrade 1567 * here. 1568 */ 1569 #if 0 1570 /* see vpagetable code */ 1571 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1572 if ((fault_type & VM_PROT_WRITE) == 0) 1573 fs->prot &= ~VM_PROT_WRITE; 1574 } 1575 #endif 1576 1577 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 1578 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 1579 if ((fault_type & VM_PROT_WRITE) == 0) 1580 fs->prot &= ~VM_PROT_WRITE; 1581 } 1582 1583 /* vm_object_hold(fs->object); implied b/c object == first_object */ 1584 1585 for (;;) { 1586 /* 1587 * The entire backing chain from first_object to object 1588 * inclusive is chainlocked. 1589 * 1590 * If the object is dead, we stop here 1591 */ 1592 if (fs->object->flags & OBJ_DEAD) { 1593 vm_object_pip_wakeup(fs->first_object); 1594 vm_object_chain_release_all(fs->first_object, 1595 fs->object); 1596 if (fs->object != fs->first_object) 1597 vm_object_drop(fs->object); 1598 unlock_and_deallocate(fs); 1599 return (KERN_PROTECTION_FAILURE); 1600 } 1601 1602 /* 1603 * See if the page is resident. Wait/Retry if the page is 1604 * busy (lots of stuff may have changed so we can't continue 1605 * in that case). 1606 * 1607 * We can theoretically allow the soft-busy case on a read 1608 * fault if the page is marked valid, but since such 1609 * pages are typically already pmap'd, putting that 1610 * special case in might be more effort then it is 1611 * worth. We cannot under any circumstances mess 1612 * around with a vm_page_t->busy page except, perhaps, 1613 * to pmap it. 1614 */ 1615 fs->m = vm_page_lookup_busy_try(fs->object, pindex, 1616 TRUE, &error); 1617 if (error) { 1618 vm_object_pip_wakeup(fs->first_object); 1619 vm_object_chain_release_all(fs->first_object, 1620 fs->object); 1621 if (fs->object != fs->first_object) 1622 vm_object_drop(fs->object); 1623 unlock_things(fs); 1624 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 1625 mycpu->gd_cnt.v_intrans++; 1626 /*vm_object_deallocate(fs->first_object);*/ 1627 /*fs->first_object = NULL;*/ 1628 fs->m = NULL; 1629 return (KERN_TRY_AGAIN); 1630 } 1631 if (fs->m) { 1632 /* 1633 * The page is busied for us. 1634 * 1635 * If reactivating a page from PQ_CACHE we may have 1636 * to rate-limit. 1637 */ 1638 int queue = fs->m->queue; 1639 vm_page_unqueue_nowakeup(fs->m); 1640 1641 if ((queue - fs->m->pc) == PQ_CACHE && 1642 vm_page_count_severe()) { 1643 vm_page_activate(fs->m); 1644 vm_page_wakeup(fs->m); 1645 fs->m = NULL; 1646 vm_object_pip_wakeup(fs->first_object); 1647 vm_object_chain_release_all(fs->first_object, 1648 fs->object); 1649 if (fs->object != fs->first_object) 1650 vm_object_drop(fs->object); 1651 unlock_and_deallocate(fs); 1652 if (allow_nofault == 0 || 1653 (curthread->td_flags & TDF_NOFAULT) == 0) { 1654 thread_t td; 1655 1656 vm_wait_pfault(); 1657 td = curthread; 1658 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1659 return (KERN_PROTECTION_FAILURE); 1660 } 1661 return (KERN_TRY_AGAIN); 1662 } 1663 1664 /* 1665 * If it still isn't completely valid (readable), 1666 * or if a read-ahead-mark is set on the VM page, 1667 * jump to readrest, else we found the page and 1668 * can return. 1669 * 1670 * We can release the spl once we have marked the 1671 * page busy. 1672 */ 1673 if (fs->m->object != &kernel_object) { 1674 if ((fs->m->valid & VM_PAGE_BITS_ALL) != 1675 VM_PAGE_BITS_ALL) { 1676 goto readrest; 1677 } 1678 if (fs->m->flags & PG_RAM) { 1679 if (debug_cluster) 1680 kprintf("R"); 1681 vm_page_flag_clear(fs->m, PG_RAM); 1682 goto readrest; 1683 } 1684 } 1685 break; /* break to PAGE HAS BEEN FOUND */ 1686 } 1687 1688 /* 1689 * Page is not resident, If this is the search termination 1690 * or the pager might contain the page, allocate a new page. 1691 */ 1692 if (TRYPAGER(fs) || fs->object == fs->first_object) { 1693 /* 1694 * Allocating, must be exclusive. 1695 */ 1696 if (fs->object == fs->first_object && 1697 fs->first_shared) { 1698 fs->first_shared = 0; 1699 vm_object_pip_wakeup(fs->first_object); 1700 vm_object_chain_release_all(fs->first_object, 1701 fs->object); 1702 if (fs->object != fs->first_object) 1703 vm_object_drop(fs->object); 1704 unlock_and_deallocate(fs); 1705 return (KERN_TRY_AGAIN); 1706 } 1707 if (fs->object != fs->first_object && 1708 fs->shared) { 1709 fs->first_shared = 0; 1710 fs->shared = 0; 1711 vm_object_pip_wakeup(fs->first_object); 1712 vm_object_chain_release_all(fs->first_object, 1713 fs->object); 1714 if (fs->object != fs->first_object) 1715 vm_object_drop(fs->object); 1716 unlock_and_deallocate(fs); 1717 return (KERN_TRY_AGAIN); 1718 } 1719 1720 /* 1721 * If the page is beyond the object size we fail 1722 */ 1723 if (pindex >= fs->object->size) { 1724 vm_object_pip_wakeup(fs->first_object); 1725 vm_object_chain_release_all(fs->first_object, 1726 fs->object); 1727 if (fs->object != fs->first_object) 1728 vm_object_drop(fs->object); 1729 unlock_and_deallocate(fs); 1730 return (KERN_PROTECTION_FAILURE); 1731 } 1732 1733 /* 1734 * Allocate a new page for this object/offset pair. 1735 * 1736 * It is possible for the allocation to race, so 1737 * handle the case. 1738 */ 1739 fs->m = NULL; 1740 if (!vm_page_count_severe()) { 1741 fs->m = vm_page_alloc(fs->object, pindex, 1742 ((fs->vp || fs->object->backing_object) ? 1743 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1744 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1745 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1746 } 1747 if (fs->m == NULL) { 1748 vm_object_pip_wakeup(fs->first_object); 1749 vm_object_chain_release_all(fs->first_object, 1750 fs->object); 1751 if (fs->object != fs->first_object) 1752 vm_object_drop(fs->object); 1753 unlock_and_deallocate(fs); 1754 if (allow_nofault == 0 || 1755 (curthread->td_flags & TDF_NOFAULT) == 0) { 1756 thread_t td; 1757 1758 vm_wait_pfault(); 1759 td = curthread; 1760 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1761 return (KERN_PROTECTION_FAILURE); 1762 } 1763 return (KERN_TRY_AGAIN); 1764 } 1765 1766 /* 1767 * Fall through to readrest. We have a new page which 1768 * will have to be paged (since m->valid will be 0). 1769 */ 1770 } 1771 1772 readrest: 1773 /* 1774 * We have found an invalid or partially valid page, a 1775 * page with a read-ahead mark which might be partially or 1776 * fully valid (and maybe dirty too), or we have allocated 1777 * a new page. 1778 * 1779 * Attempt to fault-in the page if there is a chance that the 1780 * pager has it, and potentially fault in additional pages 1781 * at the same time. 1782 * 1783 * If TRYPAGER is true then fs.m will be non-NULL and busied 1784 * for us. 1785 */ 1786 if (TRYPAGER(fs)) { 1787 int rv; 1788 int seqaccess; 1789 u_char behavior = vm_map_entry_behavior(fs->entry); 1790 1791 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 1792 seqaccess = 0; 1793 else 1794 seqaccess = -1; 1795 1796 /* 1797 * Doing I/O may synchronously insert additional 1798 * pages so we can't be shared at this point either. 1799 * 1800 * NOTE: We can't free fs->m here in the allocated 1801 * case (fs->object != fs->first_object) as 1802 * this would require an exclusively locked 1803 * VM object. 1804 */ 1805 if (fs->object == fs->first_object && 1806 fs->first_shared) { 1807 vm_page_deactivate(fs->m); 1808 vm_page_wakeup(fs->m); 1809 fs->m = NULL; 1810 fs->first_shared = 0; 1811 vm_object_pip_wakeup(fs->first_object); 1812 vm_object_chain_release_all(fs->first_object, 1813 fs->object); 1814 if (fs->object != fs->first_object) 1815 vm_object_drop(fs->object); 1816 unlock_and_deallocate(fs); 1817 return (KERN_TRY_AGAIN); 1818 } 1819 if (fs->object != fs->first_object && 1820 fs->shared) { 1821 vm_page_deactivate(fs->m); 1822 vm_page_wakeup(fs->m); 1823 fs->m = NULL; 1824 fs->first_shared = 0; 1825 fs->shared = 0; 1826 vm_object_pip_wakeup(fs->first_object); 1827 vm_object_chain_release_all(fs->first_object, 1828 fs->object); 1829 if (fs->object != fs->first_object) 1830 vm_object_drop(fs->object); 1831 unlock_and_deallocate(fs); 1832 return (KERN_TRY_AGAIN); 1833 } 1834 1835 /* 1836 * Avoid deadlocking against the map when doing I/O. 1837 * fs.object and the page is BUSY'd. 1838 * 1839 * NOTE: Once unlocked, fs->entry can become stale 1840 * so this will NULL it out. 1841 * 1842 * NOTE: fs->entry is invalid until we relock the 1843 * map and verify that the timestamp has not 1844 * changed. 1845 */ 1846 unlock_map(fs); 1847 1848 /* 1849 * Acquire the page data. We still hold a ref on 1850 * fs.object and the page has been BUSY's. 1851 * 1852 * The pager may replace the page (for example, in 1853 * order to enter a fictitious page into the 1854 * object). If it does so it is responsible for 1855 * cleaning up the passed page and properly setting 1856 * the new page BUSY. 1857 * 1858 * If we got here through a PG_RAM read-ahead 1859 * mark the page may be partially dirty and thus 1860 * not freeable. Don't bother checking to see 1861 * if the pager has the page because we can't free 1862 * it anyway. We have to depend on the get_page 1863 * operation filling in any gaps whether there is 1864 * backing store or not. 1865 */ 1866 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess); 1867 1868 if (rv == VM_PAGER_OK) { 1869 /* 1870 * Relookup in case pager changed page. Pager 1871 * is responsible for disposition of old page 1872 * if moved. 1873 * 1874 * XXX other code segments do relookups too. 1875 * It's a bad abstraction that needs to be 1876 * fixed/removed. 1877 */ 1878 fs->m = vm_page_lookup(fs->object, pindex); 1879 if (fs->m == NULL) { 1880 vm_object_pip_wakeup(fs->first_object); 1881 vm_object_chain_release_all( 1882 fs->first_object, fs->object); 1883 if (fs->object != fs->first_object) 1884 vm_object_drop(fs->object); 1885 unlock_and_deallocate(fs); 1886 return (KERN_TRY_AGAIN); 1887 } 1888 ++fs->hardfault; 1889 break; /* break to PAGE HAS BEEN FOUND */ 1890 } 1891 1892 /* 1893 * Remove the bogus page (which does not exist at this 1894 * object/offset); before doing so, we must get back 1895 * our object lock to preserve our invariant. 1896 * 1897 * Also wake up any other process that may want to bring 1898 * in this page. 1899 * 1900 * If this is the top-level object, we must leave the 1901 * busy page to prevent another process from rushing 1902 * past us, and inserting the page in that object at 1903 * the same time that we are. 1904 */ 1905 if (rv == VM_PAGER_ERROR) { 1906 if (curproc) { 1907 kprintf("vm_fault: pager read error, " 1908 "pid %d (%s)\n", 1909 curproc->p_pid, 1910 curproc->p_comm); 1911 } else { 1912 kprintf("vm_fault: pager read error, " 1913 "thread %p (%s)\n", 1914 curthread, 1915 curproc->p_comm); 1916 } 1917 } 1918 1919 /* 1920 * Data outside the range of the pager or an I/O error 1921 * 1922 * The page may have been wired during the pagein, 1923 * e.g. by the buffer cache, and cannot simply be 1924 * freed. Call vnode_pager_freepage() to deal with it. 1925 * 1926 * Also note that we cannot free the page if we are 1927 * holding the related object shared. XXX not sure 1928 * what to do in that case. 1929 */ 1930 if (fs->object != fs->first_object) { 1931 /* 1932 * Scrap the page. Check to see if the 1933 * vm_pager_get_page() call has already 1934 * dealt with it. 1935 */ 1936 if (fs->m) { 1937 vnode_pager_freepage(fs->m); 1938 fs->m = NULL; 1939 } 1940 1941 /* 1942 * XXX - we cannot just fall out at this 1943 * point, m has been freed and is invalid! 1944 */ 1945 } 1946 /* 1947 * XXX - the check for kernel_map is a kludge to work 1948 * around having the machine panic on a kernel space 1949 * fault w/ I/O error. 1950 */ 1951 if (((fs->map != &kernel_map) && 1952 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 1953 if (fs->m) { 1954 if (fs->first_shared) { 1955 vm_page_deactivate(fs->m); 1956 vm_page_wakeup(fs->m); 1957 } else { 1958 vnode_pager_freepage(fs->m); 1959 } 1960 fs->m = NULL; 1961 } 1962 vm_object_pip_wakeup(fs->first_object); 1963 vm_object_chain_release_all(fs->first_object, 1964 fs->object); 1965 if (fs->object != fs->first_object) 1966 vm_object_drop(fs->object); 1967 unlock_and_deallocate(fs); 1968 if (rv == VM_PAGER_ERROR) 1969 return (KERN_FAILURE); 1970 else 1971 return (KERN_PROTECTION_FAILURE); 1972 /* NOT REACHED */ 1973 } 1974 } 1975 1976 /* 1977 * We get here if the object has a default pager (or unwiring) 1978 * or the pager doesn't have the page. 1979 * 1980 * fs->first_m will be used for the COW unless we find a 1981 * deeper page to be mapped read-only, in which case the 1982 * unlock*(fs) will free first_m. 1983 */ 1984 if (fs->object == fs->first_object) 1985 fs->first_m = fs->m; 1986 1987 /* 1988 * Move on to the next object. The chain lock should prevent 1989 * the backing_object from getting ripped out from under us. 1990 * 1991 * The object lock for the next object is governed by 1992 * fs->shared. 1993 */ 1994 if ((next_object = fs->object->backing_object) != NULL) { 1995 if (fs->shared) 1996 vm_object_hold_shared(next_object); 1997 else 1998 vm_object_hold(next_object); 1999 vm_object_chain_acquire(next_object, fs->shared); 2000 KKASSERT(next_object == fs->object->backing_object); 2001 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 2002 } 2003 2004 if (next_object == NULL) { 2005 /* 2006 * If there's no object left, fill the page in the top 2007 * object with zeros. 2008 */ 2009 if (fs->object != fs->first_object) { 2010 #if 0 2011 if (fs->first_object->backing_object != 2012 fs->object) { 2013 vm_object_hold(fs->first_object->backing_object); 2014 } 2015 #endif 2016 vm_object_chain_release_all( 2017 fs->first_object->backing_object, 2018 fs->object); 2019 #if 0 2020 if (fs->first_object->backing_object != 2021 fs->object) { 2022 vm_object_drop(fs->first_object->backing_object); 2023 } 2024 #endif 2025 vm_object_pip_wakeup(fs->object); 2026 vm_object_drop(fs->object); 2027 fs->object = fs->first_object; 2028 pindex = first_pindex; 2029 fs->m = fs->first_m; 2030 } 2031 fs->first_m = NULL; 2032 2033 /* 2034 * Zero the page and mark it valid. 2035 */ 2036 vm_page_zero_fill(fs->m); 2037 mycpu->gd_cnt.v_zfod++; 2038 fs->m->valid = VM_PAGE_BITS_ALL; 2039 break; /* break to PAGE HAS BEEN FOUND */ 2040 } 2041 if (fs->object != fs->first_object) { 2042 vm_object_pip_wakeup(fs->object); 2043 vm_object_lock_swap(); 2044 vm_object_drop(fs->object); 2045 } 2046 KASSERT(fs->object != next_object, 2047 ("object loop %p", next_object)); 2048 fs->object = next_object; 2049 vm_object_pip_add(fs->object, 1); 2050 } 2051 2052 /* 2053 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 2054 * is held.] 2055 * 2056 * object still held. 2057 * vm_map may not be locked (determined by fs->lookup_still_valid) 2058 * 2059 * local shared variable may be different from fs->shared. 2060 * 2061 * If the page is being written, but isn't already owned by the 2062 * top-level object, we have to copy it into a new page owned by the 2063 * top-level object. 2064 */ 2065 KASSERT((fs->m->busy_count & PBUSY_LOCKED) != 0, 2066 ("vm_fault: not busy after main loop")); 2067 2068 if (fs->object != fs->first_object) { 2069 /* 2070 * We only really need to copy if we want to write it. 2071 */ 2072 if (fault_type & VM_PROT_WRITE) { 2073 /* 2074 * This allows pages to be virtually copied from a 2075 * backing_object into the first_object, where the 2076 * backing object has no other refs to it, and cannot 2077 * gain any more refs. Instead of a bcopy, we just 2078 * move the page from the backing object to the 2079 * first object. Note that we must mark the page 2080 * dirty in the first object so that it will go out 2081 * to swap when needed. 2082 */ 2083 if (virtual_copy_ok(fs)) { 2084 /* 2085 * (first_m) and (m) are both busied. We have 2086 * move (m) into (first_m)'s object/pindex 2087 * in an atomic fashion, then free (first_m). 2088 * 2089 * first_object is held so second remove 2090 * followed by the rename should wind 2091 * up being atomic. vm_page_free() might 2092 * block so we don't do it until after the 2093 * rename. 2094 */ 2095 vm_page_protect(fs->first_m, VM_PROT_NONE); 2096 vm_page_remove(fs->first_m); 2097 vm_page_rename(fs->m, fs->first_object, 2098 first_pindex); 2099 vm_page_free(fs->first_m); 2100 fs->first_m = fs->m; 2101 fs->m = NULL; 2102 mycpu->gd_cnt.v_cow_optim++; 2103 } else { 2104 /* 2105 * Oh, well, lets copy it. 2106 * 2107 * Why are we unmapping the original page 2108 * here? Well, in short, not all accessors 2109 * of user memory go through the pmap. The 2110 * procfs code doesn't have access user memory 2111 * via a local pmap, so vm_fault_page*() 2112 * can't call pmap_enter(). And the umtx*() 2113 * code may modify the COW'd page via a DMAP 2114 * or kernel mapping and not via the pmap, 2115 * leaving the original page still mapped 2116 * read-only into the pmap. 2117 * 2118 * So we have to remove the page from at 2119 * least the current pmap if it is in it. 2120 * 2121 * We used to just remove it from all pmaps 2122 * but that creates inefficiencies on SMP, 2123 * particularly for COW program & library 2124 * mappings that are concurrently exec'd. 2125 * Only remove the page from the current 2126 * pmap. 2127 */ 2128 KKASSERT(fs->first_shared == 0); 2129 vm_page_copy(fs->m, fs->first_m); 2130 /*vm_page_protect(fs->m, VM_PROT_NONE);*/ 2131 pmap_remove_specific( 2132 &curthread->td_lwp->lwp_vmspace->vm_pmap, 2133 fs->m); 2134 } 2135 2136 /* 2137 * We no longer need the old page or object. 2138 */ 2139 if (fs->m) 2140 release_page(fs); 2141 2142 /* 2143 * We intend to revert to first_object, undo the 2144 * chain lock through to that. 2145 */ 2146 #if 0 2147 if (fs->first_object->backing_object != fs->object) 2148 vm_object_hold(fs->first_object->backing_object); 2149 #endif 2150 vm_object_chain_release_all( 2151 fs->first_object->backing_object, 2152 fs->object); 2153 #if 0 2154 if (fs->first_object->backing_object != fs->object) 2155 vm_object_drop(fs->first_object->backing_object); 2156 #endif 2157 2158 /* 2159 * fs->object != fs->first_object due to above 2160 * conditional 2161 */ 2162 vm_object_pip_wakeup(fs->object); 2163 vm_object_drop(fs->object); 2164 2165 /* 2166 * Only use the new page below... 2167 */ 2168 mycpu->gd_cnt.v_cow_faults++; 2169 fs->m = fs->first_m; 2170 fs->object = fs->first_object; 2171 pindex = first_pindex; 2172 } else { 2173 /* 2174 * If it wasn't a write fault avoid having to copy 2175 * the page by mapping it read-only. 2176 */ 2177 fs->prot &= ~VM_PROT_WRITE; 2178 } 2179 } 2180 2181 /* 2182 * Relock the map if necessary, then check the generation count. 2183 * relock_map() will update fs->timestamp to account for the 2184 * relocking if necessary. 2185 * 2186 * If the count has changed after relocking then all sorts of 2187 * crap may have happened and we have to retry. 2188 * 2189 * NOTE: The relock_map() can fail due to a deadlock against 2190 * the vm_page we are holding BUSY. 2191 */ 2192 if (fs->lookup_still_valid == FALSE && fs->map) { 2193 if (relock_map(fs) || 2194 fs->map->timestamp != fs->map_generation) { 2195 release_page(fs); 2196 vm_object_pip_wakeup(fs->first_object); 2197 vm_object_chain_release_all(fs->first_object, 2198 fs->object); 2199 if (fs->object != fs->first_object) 2200 vm_object_drop(fs->object); 2201 unlock_and_deallocate(fs); 2202 return (KERN_TRY_AGAIN); 2203 } 2204 } 2205 2206 /* 2207 * If the fault is a write, we know that this page is being 2208 * written NOW so dirty it explicitly to save on pmap_is_modified() 2209 * calls later. 2210 * 2211 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 2212 * if the page is already dirty to prevent data written with 2213 * the expectation of being synced from not being synced. 2214 * Likewise if this entry does not request NOSYNC then make 2215 * sure the page isn't marked NOSYNC. Applications sharing 2216 * data should use the same flags to avoid ping ponging. 2217 * 2218 * Also tell the backing pager, if any, that it should remove 2219 * any swap backing since the page is now dirty. 2220 */ 2221 vm_page_activate(fs->m); 2222 if (fs->prot & VM_PROT_WRITE) { 2223 vm_object_set_writeable_dirty(fs->m->object); 2224 vm_set_nosync(fs->m, fs->entry); 2225 if (fs->fault_flags & VM_FAULT_DIRTY) { 2226 vm_page_dirty(fs->m); 2227 if (fs->m->flags & PG_SWAPPED) { 2228 /* 2229 * If the page is swapped out we have to call 2230 * swap_pager_unswapped() which requires an 2231 * exclusive object lock. If we are shared, 2232 * we must clear the shared flag and retry. 2233 */ 2234 if ((fs->object == fs->first_object && 2235 fs->first_shared) || 2236 (fs->object != fs->first_object && 2237 fs->shared)) { 2238 vm_page_wakeup(fs->m); 2239 fs->m = NULL; 2240 if (fs->object == fs->first_object) 2241 fs->first_shared = 0; 2242 else 2243 fs->shared = 0; 2244 vm_object_pip_wakeup(fs->first_object); 2245 vm_object_chain_release_all( 2246 fs->first_object, fs->object); 2247 if (fs->object != fs->first_object) 2248 vm_object_drop(fs->object); 2249 unlock_and_deallocate(fs); 2250 return (KERN_TRY_AGAIN); 2251 } 2252 swap_pager_unswapped(fs->m); 2253 } 2254 } 2255 } 2256 2257 vm_object_pip_wakeup(fs->first_object); 2258 vm_object_chain_release_all(fs->first_object, fs->object); 2259 if (fs->object != fs->first_object) 2260 vm_object_drop(fs->object); 2261 2262 /* 2263 * Page had better still be busy. We are still locked up and 2264 * fs->object will have another PIP reference if it is not equal 2265 * to fs->first_object. 2266 */ 2267 KASSERT(fs->m->busy_count & PBUSY_LOCKED, 2268 ("vm_fault: page %p not busy!", fs->m)); 2269 2270 /* 2271 * Sanity check: page must be completely valid or it is not fit to 2272 * map into user space. vm_pager_get_pages() ensures this. 2273 */ 2274 if (fs->m->valid != VM_PAGE_BITS_ALL) { 2275 vm_page_zero_invalid(fs->m, TRUE); 2276 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 2277 } 2278 2279 return (KERN_SUCCESS); 2280 } 2281 2282 /* 2283 * Wire down a range of virtual addresses in a map. The entry in question 2284 * should be marked in-transition and the map must be locked. We must 2285 * release the map temporarily while faulting-in the page to avoid a 2286 * deadlock. Note that the entry may be clipped while we are blocked but 2287 * will never be freed. 2288 * 2289 * No requirements. 2290 */ 2291 int 2292 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, 2293 boolean_t user_wire, int kmflags) 2294 { 2295 boolean_t fictitious; 2296 vm_offset_t start; 2297 vm_offset_t end; 2298 vm_offset_t va; 2299 pmap_t pmap; 2300 int rv; 2301 int wire_prot; 2302 int fault_flags; 2303 vm_page_t m; 2304 2305 if (user_wire) { 2306 wire_prot = VM_PROT_READ; 2307 fault_flags = VM_FAULT_USER_WIRE; 2308 } else { 2309 wire_prot = VM_PROT_READ | VM_PROT_WRITE; 2310 fault_flags = VM_FAULT_CHANGE_WIRING; 2311 } 2312 if (kmflags & KM_NOTLBSYNC) 2313 wire_prot |= VM_PROT_NOSYNC; 2314 2315 pmap = vm_map_pmap(map); 2316 start = entry->start; 2317 end = entry->end; 2318 2319 switch(entry->maptype) { 2320 case VM_MAPTYPE_NORMAL: 2321 case VM_MAPTYPE_VPAGETABLE: 2322 fictitious = entry->object.vm_object && 2323 ((entry->object.vm_object->type == OBJT_DEVICE) || 2324 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2325 break; 2326 case VM_MAPTYPE_UKSMAP: 2327 fictitious = TRUE; 2328 break; 2329 default: 2330 fictitious = FALSE; 2331 break; 2332 } 2333 2334 if (entry->eflags & MAP_ENTRY_KSTACK) 2335 start += PAGE_SIZE; 2336 map->timestamp++; 2337 vm_map_unlock(map); 2338 2339 /* 2340 * We simulate a fault to get the page and enter it in the physical 2341 * map. 2342 */ 2343 for (va = start; va < end; va += PAGE_SIZE) { 2344 rv = vm_fault(map, va, wire_prot, fault_flags); 2345 if (rv) { 2346 while (va > start) { 2347 va -= PAGE_SIZE; 2348 m = pmap_unwire(pmap, va); 2349 if (m && !fictitious) { 2350 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2351 vm_page_unwire(m, 1); 2352 vm_page_wakeup(m); 2353 } 2354 } 2355 goto done; 2356 } 2357 } 2358 rv = KERN_SUCCESS; 2359 done: 2360 vm_map_lock(map); 2361 2362 return (rv); 2363 } 2364 2365 /* 2366 * Unwire a range of virtual addresses in a map. The map should be 2367 * locked. 2368 */ 2369 void 2370 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 2371 { 2372 boolean_t fictitious; 2373 vm_offset_t start; 2374 vm_offset_t end; 2375 vm_offset_t va; 2376 pmap_t pmap; 2377 vm_page_t m; 2378 2379 pmap = vm_map_pmap(map); 2380 start = entry->start; 2381 end = entry->end; 2382 fictitious = entry->object.vm_object && 2383 ((entry->object.vm_object->type == OBJT_DEVICE) || 2384 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2385 if (entry->eflags & MAP_ENTRY_KSTACK) 2386 start += PAGE_SIZE; 2387 2388 /* 2389 * Since the pages are wired down, we must be able to get their 2390 * mappings from the physical map system. 2391 */ 2392 for (va = start; va < end; va += PAGE_SIZE) { 2393 m = pmap_unwire(pmap, va); 2394 if (m && !fictitious) { 2395 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2396 vm_page_unwire(m, 1); 2397 vm_page_wakeup(m); 2398 } 2399 } 2400 } 2401 2402 /* 2403 * Copy all of the pages from a wired-down map entry to another. 2404 * 2405 * The source and destination maps must be locked for write. 2406 * The source and destination maps token must be held 2407 * The source map entry must be wired down (or be a sharing map 2408 * entry corresponding to a main map entry that is wired down). 2409 * 2410 * No other requirements. 2411 * 2412 * XXX do segment optimization 2413 */ 2414 void 2415 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2416 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2417 { 2418 vm_object_t dst_object; 2419 vm_object_t src_object; 2420 vm_ooffset_t dst_offset; 2421 vm_ooffset_t src_offset; 2422 vm_prot_t prot; 2423 vm_offset_t vaddr; 2424 vm_page_t dst_m; 2425 vm_page_t src_m; 2426 2427 src_object = src_entry->object.vm_object; 2428 src_offset = src_entry->offset; 2429 2430 /* 2431 * Create the top-level object for the destination entry. (Doesn't 2432 * actually shadow anything - we copy the pages directly.) 2433 */ 2434 vm_map_entry_allocate_object(dst_entry); 2435 dst_object = dst_entry->object.vm_object; 2436 2437 prot = dst_entry->max_protection; 2438 2439 /* 2440 * Loop through all of the pages in the entry's range, copying each 2441 * one from the source object (it should be there) to the destination 2442 * object. 2443 */ 2444 vm_object_hold(src_object); 2445 vm_object_hold(dst_object); 2446 2447 for (vaddr = dst_entry->start, dst_offset = 0; 2448 vaddr < dst_entry->end; 2449 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2450 2451 /* 2452 * Allocate a page in the destination object 2453 */ 2454 do { 2455 dst_m = vm_page_alloc(dst_object, 2456 OFF_TO_IDX(dst_offset), 2457 VM_ALLOC_NORMAL); 2458 if (dst_m == NULL) { 2459 vm_wait(0); 2460 } 2461 } while (dst_m == NULL); 2462 2463 /* 2464 * Find the page in the source object, and copy it in. 2465 * (Because the source is wired down, the page will be in 2466 * memory.) 2467 */ 2468 src_m = vm_page_lookup(src_object, 2469 OFF_TO_IDX(dst_offset + src_offset)); 2470 if (src_m == NULL) 2471 panic("vm_fault_copy_wired: page missing"); 2472 2473 vm_page_copy(src_m, dst_m); 2474 2475 /* 2476 * Enter it in the pmap... 2477 */ 2478 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2479 2480 /* 2481 * Mark it no longer busy, and put it on the active list. 2482 */ 2483 vm_page_activate(dst_m); 2484 vm_page_wakeup(dst_m); 2485 } 2486 vm_object_drop(dst_object); 2487 vm_object_drop(src_object); 2488 } 2489 2490 #if 0 2491 2492 /* 2493 * This routine checks around the requested page for other pages that 2494 * might be able to be faulted in. This routine brackets the viable 2495 * pages for the pages to be paged in. 2496 * 2497 * Inputs: 2498 * m, rbehind, rahead 2499 * 2500 * Outputs: 2501 * marray (array of vm_page_t), reqpage (index of requested page) 2502 * 2503 * Return value: 2504 * number of pages in marray 2505 */ 2506 static int 2507 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2508 vm_page_t *marray, int *reqpage) 2509 { 2510 int i,j; 2511 vm_object_t object; 2512 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2513 vm_page_t rtm; 2514 int cbehind, cahead; 2515 2516 object = m->object; 2517 pindex = m->pindex; 2518 2519 /* 2520 * we don't fault-ahead for device pager 2521 */ 2522 if ((object->type == OBJT_DEVICE) || 2523 (object->type == OBJT_MGTDEVICE)) { 2524 *reqpage = 0; 2525 marray[0] = m; 2526 return 1; 2527 } 2528 2529 /* 2530 * if the requested page is not available, then give up now 2531 */ 2532 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2533 *reqpage = 0; /* not used by caller, fix compiler warn */ 2534 return 0; 2535 } 2536 2537 if ((cbehind == 0) && (cahead == 0)) { 2538 *reqpage = 0; 2539 marray[0] = m; 2540 return 1; 2541 } 2542 2543 if (rahead > cahead) { 2544 rahead = cahead; 2545 } 2546 2547 if (rbehind > cbehind) { 2548 rbehind = cbehind; 2549 } 2550 2551 /* 2552 * Do not do any readahead if we have insufficient free memory. 2553 * 2554 * XXX code was broken disabled before and has instability 2555 * with this conditonal fixed, so shortcut for now. 2556 */ 2557 if (burst_fault == 0 || vm_page_count_severe()) { 2558 marray[0] = m; 2559 *reqpage = 0; 2560 return 1; 2561 } 2562 2563 /* 2564 * scan backward for the read behind pages -- in memory 2565 * 2566 * Assume that if the page is not found an interrupt will not 2567 * create it. Theoretically interrupts can only remove (busy) 2568 * pages, not create new associations. 2569 */ 2570 if (pindex > 0) { 2571 if (rbehind > pindex) { 2572 rbehind = pindex; 2573 startpindex = 0; 2574 } else { 2575 startpindex = pindex - rbehind; 2576 } 2577 2578 vm_object_hold(object); 2579 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2580 if (vm_page_lookup(object, tpindex - 1)) 2581 break; 2582 } 2583 2584 i = 0; 2585 while (tpindex < pindex) { 2586 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2587 VM_ALLOC_NULL_OK); 2588 if (rtm == NULL) { 2589 for (j = 0; j < i; j++) { 2590 vm_page_free(marray[j]); 2591 } 2592 vm_object_drop(object); 2593 marray[0] = m; 2594 *reqpage = 0; 2595 return 1; 2596 } 2597 marray[i] = rtm; 2598 ++i; 2599 ++tpindex; 2600 } 2601 vm_object_drop(object); 2602 } else { 2603 i = 0; 2604 } 2605 2606 /* 2607 * Assign requested page 2608 */ 2609 marray[i] = m; 2610 *reqpage = i; 2611 ++i; 2612 2613 /* 2614 * Scan forwards for read-ahead pages 2615 */ 2616 tpindex = pindex + 1; 2617 endpindex = tpindex + rahead; 2618 if (endpindex > object->size) 2619 endpindex = object->size; 2620 2621 vm_object_hold(object); 2622 while (tpindex < endpindex) { 2623 if (vm_page_lookup(object, tpindex)) 2624 break; 2625 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2626 VM_ALLOC_NULL_OK); 2627 if (rtm == NULL) 2628 break; 2629 marray[i] = rtm; 2630 ++i; 2631 ++tpindex; 2632 } 2633 vm_object_drop(object); 2634 2635 return (i); 2636 } 2637 2638 #endif 2639 2640 /* 2641 * vm_prefault() provides a quick way of clustering pagefaults into a 2642 * processes address space. It is a "cousin" of pmap_object_init_pt, 2643 * except it runs at page fault time instead of mmap time. 2644 * 2645 * vm.fast_fault Enables pre-faulting zero-fill pages 2646 * 2647 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2648 * prefault. Scan stops in either direction when 2649 * a page is found to already exist. 2650 * 2651 * This code used to be per-platform pmap_prefault(). It is now 2652 * machine-independent and enhanced to also pre-fault zero-fill pages 2653 * (see vm.fast_fault) as well as make them writable, which greatly 2654 * reduces the number of page faults programs incur. 2655 * 2656 * Application performance when pre-faulting zero-fill pages is heavily 2657 * dependent on the application. Very tiny applications like /bin/echo 2658 * lose a little performance while applications of any appreciable size 2659 * gain performance. Prefaulting multiple pages also reduces SMP 2660 * congestion and can improve SMP performance significantly. 2661 * 2662 * NOTE! prot may allow writing but this only applies to the top level 2663 * object. If we wind up mapping a page extracted from a backing 2664 * object we have to make sure it is read-only. 2665 * 2666 * NOTE! The caller has already handled any COW operations on the 2667 * vm_map_entry via the normal fault code. Do NOT call this 2668 * shortcut unless the normal fault code has run on this entry. 2669 * 2670 * The related map must be locked. 2671 * No other requirements. 2672 */ 2673 static int vm_prefault_pages = 8; 2674 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2675 "Maximum number of pages to pre-fault"); 2676 static int vm_fast_fault = 1; 2677 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2678 "Burst fault zero-fill regions"); 2679 2680 /* 2681 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2682 * is not already dirty by other means. This will prevent passive 2683 * filesystem syncing as well as 'sync' from writing out the page. 2684 */ 2685 static void 2686 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2687 { 2688 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2689 if (m->dirty == 0) 2690 vm_page_flag_set(m, PG_NOSYNC); 2691 } else { 2692 vm_page_flag_clear(m, PG_NOSYNC); 2693 } 2694 } 2695 2696 static void 2697 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 2698 int fault_flags) 2699 { 2700 struct lwp *lp; 2701 vm_page_t m; 2702 vm_offset_t addr; 2703 vm_pindex_t index; 2704 vm_pindex_t pindex; 2705 vm_object_t object; 2706 int pprot; 2707 int i; 2708 int noneg; 2709 int nopos; 2710 int maxpages; 2711 2712 /* 2713 * Get stable max count value, disabled if set to 0 2714 */ 2715 maxpages = vm_prefault_pages; 2716 cpu_ccfence(); 2717 if (maxpages <= 0) 2718 return; 2719 2720 /* 2721 * We do not currently prefault mappings that use virtual page 2722 * tables. We do not prefault foreign pmaps. 2723 */ 2724 if (entry->maptype != VM_MAPTYPE_NORMAL) 2725 return; 2726 lp = curthread->td_lwp; 2727 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2728 return; 2729 2730 /* 2731 * Limit pre-fault count to 1024 pages. 2732 */ 2733 if (maxpages > 1024) 2734 maxpages = 1024; 2735 2736 object = entry->object.vm_object; 2737 KKASSERT(object != NULL); 2738 KKASSERT(object == entry->object.vm_object); 2739 2740 /* 2741 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively 2742 * now (or do something more complex XXX). 2743 */ 2744 vm_object_hold(object); 2745 vm_object_chain_acquire(object, 0); 2746 2747 noneg = 0; 2748 nopos = 0; 2749 for (i = 0; i < maxpages; ++i) { 2750 vm_object_t lobject; 2751 vm_object_t nobject; 2752 int allocated = 0; 2753 int error; 2754 2755 /* 2756 * This can eat a lot of time on a heavily contended 2757 * machine so yield on the tick if needed. 2758 */ 2759 if ((i & 7) == 7) 2760 lwkt_yield(); 2761 2762 /* 2763 * Calculate the page to pre-fault, stopping the scan in 2764 * each direction separately if the limit is reached. 2765 */ 2766 if (i & 1) { 2767 if (noneg) 2768 continue; 2769 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2770 } else { 2771 if (nopos) 2772 continue; 2773 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2774 } 2775 if (addr < entry->start) { 2776 noneg = 1; 2777 if (noneg && nopos) 2778 break; 2779 continue; 2780 } 2781 if (addr >= entry->end) { 2782 nopos = 1; 2783 if (noneg && nopos) 2784 break; 2785 continue; 2786 } 2787 2788 /* 2789 * Skip pages already mapped, and stop scanning in that 2790 * direction. When the scan terminates in both directions 2791 * we are done. 2792 */ 2793 if (pmap_prefault_ok(pmap, addr) == 0) { 2794 if (i & 1) 2795 noneg = 1; 2796 else 2797 nopos = 1; 2798 if (noneg && nopos) 2799 break; 2800 continue; 2801 } 2802 2803 /* 2804 * Follow the VM object chain to obtain the page to be mapped 2805 * into the pmap. 2806 * 2807 * If we reach the terminal object without finding a page 2808 * and we determine it would be advantageous, then allocate 2809 * a zero-fill page for the base object. The base object 2810 * is guaranteed to be OBJT_DEFAULT for this case. 2811 * 2812 * In order to not have to check the pager via *haspage*() 2813 * we stop if any non-default object is encountered. e.g. 2814 * a vnode or swap object would stop the loop. 2815 */ 2816 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2817 lobject = object; 2818 pindex = index; 2819 pprot = prot; 2820 2821 KKASSERT(lobject == entry->object.vm_object); 2822 /*vm_object_hold(lobject); implied */ 2823 2824 while ((m = vm_page_lookup_busy_try(lobject, pindex, 2825 TRUE, &error)) == NULL) { 2826 if (lobject->type != OBJT_DEFAULT) 2827 break; 2828 if (lobject->backing_object == NULL) { 2829 if (vm_fast_fault == 0) 2830 break; 2831 if ((prot & VM_PROT_WRITE) == 0 || 2832 vm_page_count_min(0)) { 2833 break; 2834 } 2835 2836 /* 2837 * NOTE: Allocated from base object 2838 */ 2839 m = vm_page_alloc(object, index, 2840 VM_ALLOC_NORMAL | 2841 VM_ALLOC_ZERO | 2842 VM_ALLOC_USE_GD | 2843 VM_ALLOC_NULL_OK); 2844 if (m == NULL) 2845 break; 2846 allocated = 1; 2847 pprot = prot; 2848 /* lobject = object .. not needed */ 2849 break; 2850 } 2851 if (lobject->backing_object_offset & PAGE_MASK) 2852 break; 2853 nobject = lobject->backing_object; 2854 vm_object_hold(nobject); 2855 KKASSERT(nobject == lobject->backing_object); 2856 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 2857 if (lobject != object) { 2858 vm_object_lock_swap(); 2859 vm_object_drop(lobject); 2860 } 2861 lobject = nobject; 2862 pprot &= ~VM_PROT_WRITE; 2863 vm_object_chain_acquire(lobject, 0); 2864 } 2865 2866 /* 2867 * NOTE: A non-NULL (m) will be associated with lobject if 2868 * it was found there, otherwise it is probably a 2869 * zero-fill page associated with the base object. 2870 * 2871 * Give-up if no page is available. 2872 */ 2873 if (m == NULL) { 2874 if (lobject != object) { 2875 #if 0 2876 if (object->backing_object != lobject) 2877 vm_object_hold(object->backing_object); 2878 #endif 2879 vm_object_chain_release_all( 2880 object->backing_object, lobject); 2881 #if 0 2882 if (object->backing_object != lobject) 2883 vm_object_drop(object->backing_object); 2884 #endif 2885 vm_object_drop(lobject); 2886 } 2887 break; 2888 } 2889 2890 /* 2891 * The object must be marked dirty if we are mapping a 2892 * writable page. m->object is either lobject or object, 2893 * both of which are still held. Do this before we 2894 * potentially drop the object. 2895 */ 2896 if (pprot & VM_PROT_WRITE) 2897 vm_object_set_writeable_dirty(m->object); 2898 2899 /* 2900 * Do not conditionalize on PG_RAM. If pages are present in 2901 * the VM system we assume optimal caching. If caching is 2902 * not optimal the I/O gravy train will be restarted when we 2903 * hit an unavailable page. We do not want to try to restart 2904 * the gravy train now because we really don't know how much 2905 * of the object has been cached. The cost for restarting 2906 * the gravy train should be low (since accesses will likely 2907 * be I/O bound anyway). 2908 */ 2909 if (lobject != object) { 2910 #if 0 2911 if (object->backing_object != lobject) 2912 vm_object_hold(object->backing_object); 2913 #endif 2914 vm_object_chain_release_all(object->backing_object, 2915 lobject); 2916 #if 0 2917 if (object->backing_object != lobject) 2918 vm_object_drop(object->backing_object); 2919 #endif 2920 vm_object_drop(lobject); 2921 } 2922 2923 /* 2924 * Enter the page into the pmap if appropriate. If we had 2925 * allocated the page we have to place it on a queue. If not 2926 * we just have to make sure it isn't on the cache queue 2927 * (pages on the cache queue are not allowed to be mapped). 2928 */ 2929 if (allocated) { 2930 /* 2931 * Page must be zerod. 2932 */ 2933 vm_page_zero_fill(m); 2934 mycpu->gd_cnt.v_zfod++; 2935 m->valid = VM_PAGE_BITS_ALL; 2936 2937 /* 2938 * Handle dirty page case 2939 */ 2940 if (pprot & VM_PROT_WRITE) 2941 vm_set_nosync(m, entry); 2942 pmap_enter(pmap, addr, m, pprot, 0, entry); 2943 mycpu->gd_cnt.v_vm_faults++; 2944 if (curthread->td_lwp) 2945 ++curthread->td_lwp->lwp_ru.ru_minflt; 2946 vm_page_deactivate(m); 2947 if (pprot & VM_PROT_WRITE) { 2948 /*vm_object_set_writeable_dirty(m->object);*/ 2949 vm_set_nosync(m, entry); 2950 if (fault_flags & VM_FAULT_DIRTY) { 2951 vm_page_dirty(m); 2952 /*XXX*/ 2953 swap_pager_unswapped(m); 2954 } 2955 } 2956 vm_page_wakeup(m); 2957 } else if (error) { 2958 /* couldn't busy page, no wakeup */ 2959 } else if ( 2960 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2961 (m->flags & PG_FICTITIOUS) == 0) { 2962 /* 2963 * A fully valid page not undergoing soft I/O can 2964 * be immediately entered into the pmap. 2965 */ 2966 if ((m->queue - m->pc) == PQ_CACHE) 2967 vm_page_deactivate(m); 2968 if (pprot & VM_PROT_WRITE) { 2969 /*vm_object_set_writeable_dirty(m->object);*/ 2970 vm_set_nosync(m, entry); 2971 if (fault_flags & VM_FAULT_DIRTY) { 2972 vm_page_dirty(m); 2973 /*XXX*/ 2974 swap_pager_unswapped(m); 2975 } 2976 } 2977 if (pprot & VM_PROT_WRITE) 2978 vm_set_nosync(m, entry); 2979 pmap_enter(pmap, addr, m, pprot, 0, entry); 2980 mycpu->gd_cnt.v_vm_faults++; 2981 if (curthread->td_lwp) 2982 ++curthread->td_lwp->lwp_ru.ru_minflt; 2983 vm_page_wakeup(m); 2984 } else { 2985 vm_page_wakeup(m); 2986 } 2987 } 2988 vm_object_chain_release(object); 2989 vm_object_drop(object); 2990 } 2991 2992 /* 2993 * Object can be held shared 2994 */ 2995 static void 2996 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 2997 vm_map_entry_t entry, int prot, int fault_flags) 2998 { 2999 struct lwp *lp; 3000 vm_page_t m; 3001 vm_offset_t addr; 3002 vm_pindex_t pindex; 3003 vm_object_t object; 3004 int i; 3005 int noneg; 3006 int nopos; 3007 int maxpages; 3008 3009 /* 3010 * Get stable max count value, disabled if set to 0 3011 */ 3012 maxpages = vm_prefault_pages; 3013 cpu_ccfence(); 3014 if (maxpages <= 0) 3015 return; 3016 3017 /* 3018 * We do not currently prefault mappings that use virtual page 3019 * tables. We do not prefault foreign pmaps. 3020 */ 3021 if (entry->maptype != VM_MAPTYPE_NORMAL) 3022 return; 3023 lp = curthread->td_lwp; 3024 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 3025 return; 3026 object = entry->object.vm_object; 3027 if (object->backing_object != NULL) 3028 return; 3029 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 3030 3031 /* 3032 * Limit pre-fault count to 1024 pages. 3033 */ 3034 if (maxpages > 1024) 3035 maxpages = 1024; 3036 3037 noneg = 0; 3038 nopos = 0; 3039 for (i = 0; i < maxpages; ++i) { 3040 int error; 3041 3042 /* 3043 * Calculate the page to pre-fault, stopping the scan in 3044 * each direction separately if the limit is reached. 3045 */ 3046 if (i & 1) { 3047 if (noneg) 3048 continue; 3049 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 3050 } else { 3051 if (nopos) 3052 continue; 3053 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 3054 } 3055 if (addr < entry->start) { 3056 noneg = 1; 3057 if (noneg && nopos) 3058 break; 3059 continue; 3060 } 3061 if (addr >= entry->end) { 3062 nopos = 1; 3063 if (noneg && nopos) 3064 break; 3065 continue; 3066 } 3067 3068 /* 3069 * Follow the VM object chain to obtain the page to be mapped 3070 * into the pmap. This version of the prefault code only 3071 * works with terminal objects. 3072 * 3073 * The page must already exist. If we encounter a problem 3074 * we stop here. 3075 * 3076 * WARNING! We cannot call swap_pager_unswapped() or insert 3077 * a new vm_page with a shared token. 3078 */ 3079 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 3080 3081 /* 3082 * Skip pages already mapped, and stop scanning in that 3083 * direction. When the scan terminates in both directions 3084 * we are done. 3085 */ 3086 if (pmap_prefault_ok(pmap, addr) == 0) { 3087 if (i & 1) 3088 noneg = 1; 3089 else 3090 nopos = 1; 3091 if (noneg && nopos) 3092 break; 3093 continue; 3094 } 3095 3096 /* 3097 * Shortcut the read-only mapping case using the far more 3098 * efficient vm_page_lookup_sbusy_try() function. This 3099 * allows us to acquire the page soft-busied only which 3100 * is especially nice for concurrent execs of the same 3101 * program. 3102 * 3103 * The lookup function also validates page suitability 3104 * (all valid bits set, and not fictitious). 3105 * 3106 * If the page is in PQ_CACHE we have to fall-through 3107 * and hard-busy it so we can move it out of PQ_CACHE. 3108 */ 3109 if ((prot & (VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE)) == 0) { 3110 m = vm_page_lookup_sbusy_try(object, pindex, 3111 0, PAGE_SIZE); 3112 if (m == NULL) 3113 break; 3114 if ((m->queue - m->pc) != PQ_CACHE) { 3115 pmap_enter(pmap, addr, m, prot, 0, entry); 3116 mycpu->gd_cnt.v_vm_faults++; 3117 if (curthread->td_lwp) 3118 ++curthread->td_lwp->lwp_ru.ru_minflt; 3119 vm_page_sbusy_drop(m); 3120 continue; 3121 } 3122 vm_page_sbusy_drop(m); 3123 } 3124 3125 /* 3126 * Fallback to normal vm_page lookup code. This code 3127 * hard-busies the page. Not only that, but the page 3128 * can remain in that state for a significant period 3129 * time due to pmap_enter()'s overhead. 3130 */ 3131 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3132 if (m == NULL || error) 3133 break; 3134 3135 /* 3136 * Stop if the page cannot be trivially entered into the 3137 * pmap. 3138 */ 3139 if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) || 3140 (m->flags & PG_FICTITIOUS) || 3141 ((m->flags & PG_SWAPPED) && 3142 (prot & VM_PROT_WRITE) && 3143 (fault_flags & VM_FAULT_DIRTY))) { 3144 vm_page_wakeup(m); 3145 break; 3146 } 3147 3148 /* 3149 * Enter the page into the pmap. The object might be held 3150 * shared so we can't do any (serious) modifying operation 3151 * on it. 3152 */ 3153 if ((m->queue - m->pc) == PQ_CACHE) 3154 vm_page_deactivate(m); 3155 if (prot & VM_PROT_WRITE) { 3156 vm_object_set_writeable_dirty(m->object); 3157 vm_set_nosync(m, entry); 3158 if (fault_flags & VM_FAULT_DIRTY) { 3159 vm_page_dirty(m); 3160 /* can't happeen due to conditional above */ 3161 /* swap_pager_unswapped(m); */ 3162 } 3163 } 3164 pmap_enter(pmap, addr, m, prot, 0, entry); 3165 mycpu->gd_cnt.v_vm_faults++; 3166 if (curthread->td_lwp) 3167 ++curthread->td_lwp->lwp_ru.ru_minflt; 3168 vm_page_wakeup(m); 3169 } 3170 } 3171