1 /* 2 * Copyright (c) 2003-2020 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/swap_pager.h> 127 #include <vm/vm_extern.h> 128 129 #include <vm/vm_page2.h> 130 131 #define VM_FAULT_MAX_QUICK 16 132 133 struct faultstate { 134 vm_page_t mary[VM_FAULT_MAX_QUICK]; 135 vm_map_backing_t ba; 136 vm_prot_t prot; 137 vm_page_t first_m; 138 vm_map_backing_t first_ba; 139 vm_prot_t first_prot; 140 vm_map_t map; 141 vm_map_entry_t entry; 142 int lookup_still_valid; /* 0=inv 1=valid/rel -1=valid/atomic */ 143 int hardfault; 144 int fault_flags; 145 int shared; 146 int msoftonly; 147 int first_shared; 148 int wflags; 149 int first_ba_held; /* 0=unlocked 1=locked/rel -1=lock/atomic */ 150 struct vnode *vp; 151 }; 152 153 __read_mostly static int debug_fault = 0; 154 SYSCTL_INT(_vm, OID_AUTO, debug_fault, CTLFLAG_RW, &debug_fault, 0, ""); 155 __read_mostly static int debug_cluster = 0; 156 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, ""); 157 #if 0 158 static int virtual_copy_enable = 1; 159 SYSCTL_INT(_vm, OID_AUTO, virtual_copy_enable, CTLFLAG_RW, 160 &virtual_copy_enable, 0, ""); 161 #endif 162 __read_mostly int vm_shared_fault = 1; 163 TUNABLE_INT("vm.shared_fault", &vm_shared_fault); 164 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, 165 &vm_shared_fault, 0, "Allow shared token on vm_object"); 166 __read_mostly static int vm_fault_bypass_count = 1; 167 TUNABLE_INT("vm.fault_bypass", &vm_fault_bypass_count); 168 SYSCTL_INT(_vm, OID_AUTO, fault_bypass, CTLFLAG_RW, 169 &vm_fault_bypass_count, 0, "Allow fast vm_fault shortcut"); 170 171 /* 172 * Define here for debugging ioctls. Note that these are globals, so 173 * they were cause a ton of cache line bouncing. Only use for debugging 174 * purposes. 175 */ 176 /*#define VM_FAULT_QUICK_DEBUG */ 177 #ifdef VM_FAULT_QUICK_DEBUG 178 static long vm_fault_bypass_success_count = 0; 179 SYSCTL_LONG(_vm, OID_AUTO, fault_bypass_success_count, CTLFLAG_RW, 180 &vm_fault_bypass_success_count, 0, ""); 181 static long vm_fault_bypass_failure_count1 = 0; 182 SYSCTL_LONG(_vm, OID_AUTO, fault_bypass_failure_count1, CTLFLAG_RW, 183 &vm_fault_bypass_failure_count1, 0, ""); 184 static long vm_fault_bypass_failure_count2 = 0; 185 SYSCTL_LONG(_vm, OID_AUTO, fault_bypass_failure_count2, CTLFLAG_RW, 186 &vm_fault_bypass_failure_count2, 0, ""); 187 static long vm_fault_bypass_failure_count3 = 0; 188 SYSCTL_LONG(_vm, OID_AUTO, fault_bypass_failure_count3, CTLFLAG_RW, 189 &vm_fault_bypass_failure_count3, 0, ""); 190 static long vm_fault_bypass_failure_count4 = 0; 191 SYSCTL_LONG(_vm, OID_AUTO, fault_bypass_failure_count4, CTLFLAG_RW, 192 &vm_fault_bypass_failure_count4, 0, ""); 193 #endif 194 195 static int vm_fault_bypass(struct faultstate *fs, vm_pindex_t first_pindex, 196 vm_pindex_t first_count, int *mextcountp, 197 vm_prot_t fault_type); 198 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int); 199 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, 200 vpte_t, int, int); 201 #if 0 202 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 203 #endif 204 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry); 205 static void vm_prefault(pmap_t pmap, vm_offset_t addra, 206 vm_map_entry_t entry, int prot, int fault_flags); 207 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 208 vm_map_entry_t entry, int prot, int fault_flags); 209 210 static __inline void 211 release_page(struct faultstate *fs) 212 { 213 vm_page_deactivate(fs->mary[0]); 214 vm_page_wakeup(fs->mary[0]); 215 fs->mary[0] = NULL; 216 } 217 218 static __inline void 219 unlock_map(struct faultstate *fs) 220 { 221 if (fs->ba != fs->first_ba) 222 vm_object_drop(fs->ba->object); 223 if (fs->first_ba && fs->first_ba_held == 1) { 224 vm_object_drop(fs->first_ba->object); 225 fs->first_ba_held = 0; 226 fs->first_ba = NULL; 227 } 228 fs->ba = NULL; 229 230 /* 231 * NOTE: If lookup_still_valid == -1 the map is assumed to be locked 232 * and caller expects it to remain locked atomically. 233 */ 234 if (fs->lookup_still_valid == 1 && fs->map) { 235 vm_map_lookup_done(fs->map, fs->entry, 0); 236 fs->lookup_still_valid = 0; 237 fs->entry = NULL; 238 } 239 } 240 241 /* 242 * Clean up after a successful call to vm_fault_object() so another call 243 * to vm_fault_object() can be made. 244 */ 245 static void 246 cleanup_fault(struct faultstate *fs) 247 { 248 /* 249 * We allocated a junk page for a COW operation that did 250 * not occur, the page must be freed. 251 */ 252 if (fs->ba != fs->first_ba) { 253 KKASSERT(fs->first_shared == 0); 254 255 /* 256 * first_m could be completely valid and we got here 257 * because of a PG_RAM, don't mistakenly free it! 258 */ 259 if ((fs->first_m->valid & VM_PAGE_BITS_ALL) == 260 VM_PAGE_BITS_ALL) { 261 vm_page_wakeup(fs->first_m); 262 } else { 263 vm_page_free(fs->first_m); 264 } 265 vm_object_pip_wakeup(fs->ba->object); 266 fs->first_m = NULL; 267 268 /* 269 * Reset fs->ba (used by vm_fault_vpagetahble() without 270 * calling unlock_map(), so we need a little duplication. 271 */ 272 vm_object_drop(fs->ba->object); 273 fs->ba = fs->first_ba; 274 } 275 } 276 277 static void 278 unlock_things(struct faultstate *fs) 279 { 280 cleanup_fault(fs); 281 unlock_map(fs); 282 if (fs->vp != NULL) { 283 vput(fs->vp); 284 fs->vp = NULL; 285 } 286 } 287 288 #if 0 289 /* 290 * Virtual copy tests. Used by the fault code to determine if a 291 * page can be moved from an orphan vm_object into its shadow 292 * instead of copying its contents. 293 */ 294 static __inline int 295 virtual_copy_test(struct faultstate *fs) 296 { 297 /* 298 * Must be holding exclusive locks 299 */ 300 if (fs->first_shared || fs->shared || virtual_copy_enable == 0) 301 return 0; 302 303 /* 304 * Map, if present, has not changed 305 */ 306 if (fs->map && fs->map_generation != fs->map->timestamp) 307 return 0; 308 309 /* 310 * No refs, except us 311 */ 312 if (fs->ba->object->ref_count != 1) 313 return 0; 314 315 /* 316 * No one else can look this object up 317 */ 318 if (fs->ba->object->handle != NULL) 319 return 0; 320 321 /* 322 * No other ways to look the object up 323 */ 324 if (fs->ba->object->type != OBJT_DEFAULT && 325 fs->ba->object->type != OBJT_SWAP) 326 return 0; 327 328 /* 329 * We don't chase down the shadow chain 330 */ 331 if (fs->ba != fs->first_ba->backing_ba) 332 return 0; 333 334 return 1; 335 } 336 337 static __inline int 338 virtual_copy_ok(struct faultstate *fs) 339 { 340 if (virtual_copy_test(fs)) { 341 /* 342 * Grab the lock and re-test changeable items. 343 */ 344 if (fs->lookup_still_valid == 0 && fs->map) { 345 if (lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT)) 346 return 0; 347 fs->lookup_still_valid = 1; 348 if (virtual_copy_test(fs)) { 349 fs->map_generation = ++fs->map->timestamp; 350 return 1; 351 } 352 fs->lookup_still_valid = 0; 353 lockmgr(&fs->map->lock, LK_RELEASE); 354 } 355 } 356 return 0; 357 } 358 #endif 359 360 /* 361 * TRYPAGER 362 * 363 * Determine if the pager for the current object *might* contain the page. 364 * 365 * We only need to try the pager if this is not a default object (default 366 * objects are zero-fill and have no real pager), and if we are not taking 367 * a wiring fault or if the FS entry is wired. 368 */ 369 #define TRYPAGER(fs) \ 370 (fs->ba->object->type != OBJT_DEFAULT && \ 371 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || \ 372 (fs->wflags & FW_WIRED))) 373 374 /* 375 * vm_fault: 376 * 377 * Handle a page fault occuring at the given address, requiring the given 378 * permissions, in the map specified. If successful, the page is inserted 379 * into the associated physical map. 380 * 381 * NOTE: The given address should be truncated to the proper page address. 382 * 383 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 384 * a standard error specifying why the fault is fatal is returned. 385 * 386 * The map in question must be referenced, and remains so. 387 * The caller may hold no locks. 388 * No other requirements. 389 */ 390 int 391 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 392 { 393 vm_pindex_t first_pindex; 394 vm_pindex_t first_count; 395 struct faultstate fs; 396 struct lwp *lp; 397 struct proc *p; 398 thread_t td; 399 struct vm_map_ilock ilock; 400 int mextcount; 401 int didilock; 402 int growstack; 403 int retry = 0; 404 int inherit_prot; 405 int result; 406 int n; 407 408 inherit_prot = fault_type & VM_PROT_NOSYNC; 409 fs.hardfault = 0; 410 fs.fault_flags = fault_flags; 411 fs.vp = NULL; 412 fs.shared = vm_shared_fault; 413 fs.first_shared = vm_shared_fault; 414 growstack = 1; 415 416 /* 417 * vm_map interactions 418 */ 419 td = curthread; 420 if ((lp = td->td_lwp) != NULL) 421 lp->lwp_flags |= LWP_PAGING; 422 423 RetryFault: 424 /* 425 * vm_fault_bypass() can shortcut us. 426 */ 427 fs.msoftonly = 0; 428 fs.first_ba_held = 0; 429 mextcount = 1; 430 431 /* 432 * Find the vm_map_entry representing the backing store and resolve 433 * the top level object and page index. This may have the side 434 * effect of executing a copy-on-write on the map entry, 435 * creating a shadow object, or splitting an anonymous entry for 436 * performance, but will not COW any actual VM pages. 437 * 438 * On success fs.map is left read-locked and various other fields 439 * are initialized but not otherwise referenced or locked. 440 * 441 * NOTE! vm_map_lookup will try to upgrade the fault_type to 442 * VM_FAULT_WRITE if the map entry is a virtual page table 443 * and also writable, so we can set the 'A'accessed bit in 444 * the virtual page table entry. 445 */ 446 fs.map = map; 447 result = vm_map_lookup(&fs.map, vaddr, fault_type, 448 &fs.entry, &fs.first_ba, 449 &first_pindex, &first_count, 450 &fs.first_prot, &fs.wflags); 451 452 /* 453 * If the lookup failed or the map protections are incompatible, 454 * the fault generally fails. 455 * 456 * The failure could be due to TDF_NOFAULT if vm_map_lookup() 457 * tried to do a COW fault. 458 * 459 * If the caller is trying to do a user wiring we have more work 460 * to do. 461 */ 462 if (result != KERN_SUCCESS) { 463 if (result == KERN_FAILURE_NOFAULT) { 464 result = KERN_FAILURE; 465 goto done; 466 } 467 if (result != KERN_PROTECTION_FAILURE || 468 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 469 { 470 if (result == KERN_INVALID_ADDRESS && growstack && 471 map != &kernel_map && curproc != NULL) { 472 result = vm_map_growstack(map, vaddr); 473 if (result == KERN_SUCCESS) { 474 growstack = 0; 475 ++retry; 476 goto RetryFault; 477 } 478 result = KERN_FAILURE; 479 } 480 goto done; 481 } 482 483 /* 484 * If we are user-wiring a r/w segment, and it is COW, then 485 * we need to do the COW operation. Note that we don't 486 * currently COW RO sections now, because it is NOT desirable 487 * to COW .text. We simply keep .text from ever being COW'ed 488 * and take the heat that one cannot debug wired .text sections. 489 * 490 * XXX Try to allow the above by specifying OVERRIDE_WRITE. 491 */ 492 result = vm_map_lookup(&fs.map, vaddr, 493 VM_PROT_READ|VM_PROT_WRITE| 494 VM_PROT_OVERRIDE_WRITE, 495 &fs.entry, &fs.first_ba, 496 &first_pindex, &first_count, 497 &fs.first_prot, &fs.wflags); 498 if (result != KERN_SUCCESS) { 499 /* could also be KERN_FAILURE_NOFAULT */ 500 result = KERN_FAILURE; 501 goto done; 502 } 503 504 /* 505 * If we don't COW now, on a user wire, the user will never 506 * be able to write to the mapping. If we don't make this 507 * restriction, the bookkeeping would be nearly impossible. 508 * 509 * XXX We have a shared lock, this will have a MP race but 510 * I don't see how it can hurt anything. 511 */ 512 if ((fs.entry->protection & VM_PROT_WRITE) == 0) { 513 atomic_clear_char(&fs.entry->max_protection, 514 VM_PROT_WRITE); 515 } 516 } 517 518 /* 519 * fs.map is read-locked 520 * 521 * Misc checks. Save the map generation number to detect races. 522 */ 523 fs.lookup_still_valid = 1; 524 fs.first_m = NULL; 525 fs.ba = fs.first_ba; /* so unlock_things() works */ 526 fs.prot = fs.first_prot; /* default (used by uksmap) */ 527 528 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) { 529 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 530 panic("vm_fault: fault on nofault entry, addr: %p", 531 (void *)vaddr); 532 } 533 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) && 534 vaddr >= fs.entry->ba.start && 535 vaddr < fs.entry->ba.start + PAGE_SIZE) { 536 panic("vm_fault: fault on stack guard, addr: %p", 537 (void *)vaddr); 538 } 539 } 540 541 /* 542 * A user-kernel shared map has no VM object and bypasses 543 * everything. We execute the uksmap function with a temporary 544 * fictitious vm_page. The address is directly mapped with no 545 * management. 546 */ 547 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 548 struct vm_page fakem; 549 550 bzero(&fakem, sizeof(fakem)); 551 fakem.pindex = first_pindex; 552 fakem.flags = PG_FICTITIOUS | PG_UNQUEUED; 553 fakem.busy_count = PBUSY_LOCKED; 554 fakem.valid = VM_PAGE_BITS_ALL; 555 fakem.pat_mode = VM_MEMATTR_DEFAULT; 556 if (fs.entry->ba.uksmap(&fs.entry->ba, UKSMAPOP_FAULT, 557 fs.entry->aux.dev, &fakem)) { 558 result = KERN_FAILURE; 559 unlock_things(&fs); 560 goto done2; 561 } 562 pmap_enter(fs.map->pmap, vaddr, &fakem, fs.prot | inherit_prot, 563 (fs.wflags & FW_WIRED), fs.entry); 564 goto done_success; 565 } 566 567 /* 568 * A system map entry may return a NULL object. No object means 569 * no pager means an unrecoverable kernel fault. 570 */ 571 if (fs.first_ba == NULL) { 572 panic("vm_fault: unrecoverable fault at %p in entry %p", 573 (void *)vaddr, fs.entry); 574 } 575 576 /* 577 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 578 * is set. 579 * 580 * Unfortunately a deadlock can occur if we are forced to page-in 581 * from swap, but diving all the way into the vm_pager_get_page() 582 * function to find out is too much. Just check the object type. 583 * 584 * The deadlock is a CAM deadlock on a busy VM page when trying 585 * to finish an I/O if another process gets stuck in 586 * vop_helper_read_shortcut() due to a swap fault. 587 */ 588 if ((td->td_flags & TDF_NOFAULT) && 589 (retry || 590 fs.first_ba->object->type == OBJT_VNODE || 591 fs.first_ba->object->type == OBJT_SWAP || 592 fs.first_ba->backing_ba)) { 593 result = KERN_FAILURE; 594 unlock_things(&fs); 595 goto done2; 596 } 597 598 /* 599 * If the entry is wired we cannot change the page protection. 600 */ 601 if (fs.wflags & FW_WIRED) 602 fault_type = fs.first_prot; 603 604 /* 605 * We generally want to avoid unnecessary exclusive modes on backing 606 * and terminal objects because this can seriously interfere with 607 * heavily fork()'d processes (particularly /bin/sh scripts). 608 * 609 * However, we also want to avoid unnecessary retries due to needed 610 * shared->exclusive promotion for common faults. Exclusive mode is 611 * always needed if any page insertion, rename, or free occurs in an 612 * object (and also indirectly if any I/O is done). 613 * 614 * The main issue here is going to be fs.first_shared. If the 615 * first_object has a backing object which isn't shadowed and the 616 * process is single-threaded we might as well use an exclusive 617 * lock/chain right off the bat. 618 */ 619 #if 0 620 /* WORK IN PROGRESS, CODE REMOVED */ 621 if (fs.first_shared && fs.first_object->backing_object && 622 LIST_EMPTY(&fs.first_object->shadow_head) && 623 td->td_proc && td->td_proc->p_nthreads == 1) { 624 fs.first_shared = 0; 625 } 626 #endif 627 628 /* 629 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 630 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 631 * we can try shared first. 632 */ 633 if (fault_flags & VM_FAULT_UNSWAP) 634 fs.first_shared = 0; 635 636 /* 637 * Try to shortcut the entire mess and run the fault lockless. 638 * This will burst in multiple pages via fs->mary[]. 639 */ 640 if (vm_fault_bypass_count && 641 vm_fault_bypass(&fs, first_pindex, first_count, 642 &mextcount, fault_type) == KERN_SUCCESS) { 643 didilock = 0; 644 fault_flags &= ~VM_FAULT_BURST; 645 goto success; 646 } 647 648 /* 649 * Exclusive heuristic (alloc page vs page exists) 650 */ 651 if (fs.first_ba->flags & VM_MAP_BACK_EXCL_HEUR) 652 fs.first_shared = 0; 653 654 /* 655 * Obtain a top-level object lock, shared or exclusive depending 656 * on fs.first_shared. If a shared lock winds up being insufficient 657 * we will retry with an exclusive lock. 658 * 659 * The vnode pager lock is always shared. 660 */ 661 if (fs.first_shared) 662 vm_object_hold_shared(fs.first_ba->object); 663 else 664 vm_object_hold(fs.first_ba->object); 665 if (fs.vp == NULL) 666 fs.vp = vnode_pager_lock(fs.first_ba); 667 fs.first_ba_held = 1; 668 669 /* 670 * The page we want is at (first_object, first_pindex), but if the 671 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 672 * page table to figure out the actual pindex. 673 * 674 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 675 * ONLY 676 */ 677 didilock = 0; 678 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 679 vm_map_interlock(fs.map, &ilock, vaddr, vaddr + PAGE_SIZE); 680 didilock = 1; 681 result = vm_fault_vpagetable(&fs, &first_pindex, 682 fs.entry->aux.master_pde, 683 fault_type, 1); 684 if (result == KERN_TRY_AGAIN) { 685 vm_map_deinterlock(fs.map, &ilock); 686 ++retry; 687 goto RetryFault; 688 } 689 if (result != KERN_SUCCESS) { 690 vm_map_deinterlock(fs.map, &ilock); 691 goto done; 692 } 693 } 694 695 /* 696 * Now we have the actual (object, pindex), fault in the page. If 697 * vm_fault_object() fails it will unlock and deallocate the FS 698 * data. If it succeeds everything remains locked and fs->ba->object 699 * will have an additional PIP count if fs->ba != fs->first_ba. 700 * 701 * vm_fault_object will set fs->prot for the pmap operation. It is 702 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 703 * page can be safely written. However, it will force a read-only 704 * mapping for a read fault if the memory is managed by a virtual 705 * page table. 706 * 707 * If the fault code uses the shared object lock shortcut 708 * we must not try to burst (we can't allocate VM pages). 709 */ 710 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 711 712 if (debug_fault > 0) { 713 --debug_fault; 714 kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x " 715 "fs.m=%p fs.prot=%02x fs.wflags=%02x fs.entry=%p\n", 716 result, (intmax_t)vaddr, fault_type, fault_flags, 717 fs.mary[0], fs.prot, fs.wflags, fs.entry); 718 } 719 720 if (result == KERN_TRY_AGAIN) { 721 if (didilock) 722 vm_map_deinterlock(fs.map, &ilock); 723 ++retry; 724 goto RetryFault; 725 } 726 if (result != KERN_SUCCESS) { 727 if (didilock) 728 vm_map_deinterlock(fs.map, &ilock); 729 goto done; 730 } 731 732 success: 733 /* 734 * On success vm_fault_object() does not unlock or deallocate, and fs.m 735 * will contain a busied page. It does drop fs->ba if appropriate. 736 * 737 * Enter the page into the pmap and do pmap-related adjustments. 738 * 739 * WARNING! Soft-busied fs.m's can only be manipulated in limited 740 * ways. 741 */ 742 KKASSERT(fs.lookup_still_valid != 0); 743 vm_page_flag_set(fs.mary[0], PG_REFERENCED); 744 745 for (n = 0; n < mextcount; ++n) { 746 pmap_enter(fs.map->pmap, vaddr + (n << PAGE_SHIFT), 747 fs.mary[n], fs.prot | inherit_prot, 748 fs.wflags & FW_WIRED, fs.entry); 749 } 750 751 if (didilock) 752 vm_map_deinterlock(fs.map, &ilock); 753 754 /* 755 * If the page is not wired down, then put it where the pageout daemon 756 * can find it. 757 * 758 * NOTE: We cannot safely wire, unwire, or adjust queues for a 759 * soft-busied page. 760 */ 761 for (n = 0; n < mextcount; ++n) { 762 if (fs.msoftonly) { 763 KKASSERT(fs.mary[n]->busy_count & PBUSY_MASK); 764 KKASSERT((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0); 765 vm_page_sbusy_drop(fs.mary[n]); 766 } else { 767 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 768 if (fs.wflags & FW_WIRED) 769 vm_page_wire(fs.mary[n]); 770 else 771 vm_page_unwire(fs.mary[n], 1); 772 } else { 773 vm_page_activate(fs.mary[n]); 774 } 775 KKASSERT(fs.mary[n]->busy_count & PBUSY_LOCKED); 776 vm_page_wakeup(fs.mary[n]); 777 } 778 } 779 780 /* 781 * Burst in a few more pages if possible. The fs.map should still 782 * be locked. To avoid interlocking against a vnode->getblk 783 * operation we had to be sure to unbusy our primary vm_page above 784 * first. 785 * 786 * A normal burst can continue down backing store, only execute 787 * if we are holding an exclusive lock, otherwise the exclusive 788 * locks the burst code gets might cause excessive SMP collisions. 789 * 790 * A quick burst can be utilized when there is no backing object 791 * (i.e. a shared file mmap). 792 */ 793 if ((fault_flags & VM_FAULT_BURST) && 794 (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 && 795 (fs.wflags & FW_WIRED) == 0) { 796 if (fs.first_shared == 0 && fs.shared == 0) { 797 vm_prefault(fs.map->pmap, vaddr, 798 fs.entry, fs.prot, fault_flags); 799 } else { 800 vm_prefault_quick(fs.map->pmap, vaddr, 801 fs.entry, fs.prot, fault_flags); 802 } 803 } 804 805 done_success: 806 /* 807 * Unlock everything, and return 808 */ 809 unlock_things(&fs); 810 811 mycpu->gd_cnt.v_vm_faults++; 812 if (td->td_lwp) { 813 if (fs.hardfault) { 814 ++td->td_lwp->lwp_ru.ru_majflt; 815 } else { 816 ++td->td_lwp->lwp_ru.ru_minflt; 817 } 818 } 819 820 /*vm_object_deallocate(fs.first_ba->object);*/ 821 /*fs.m = NULL; */ 822 823 result = KERN_SUCCESS; 824 done: 825 if (fs.first_ba && fs.first_ba->object && fs.first_ba_held == 1) { 826 vm_object_drop(fs.first_ba->object); 827 fs.first_ba_held = 0; 828 } 829 done2: 830 if (lp) 831 lp->lwp_flags &= ~LWP_PAGING; 832 833 #if !defined(NO_SWAPPING) 834 /* 835 * Check the process RSS limit and force deactivation and 836 * (asynchronous) paging if necessary. This is a complex operation, 837 * only do it for direct user-mode faults, for now. 838 * 839 * To reduce overhead implement approximately a ~16MB hysteresis. 840 */ 841 p = td->td_proc; 842 if ((fault_flags & VM_FAULT_USERMODE) && lp && 843 p->p_limit && map->pmap && vm_pageout_memuse_mode >= 1 && 844 map != &kernel_map) { 845 vm_pindex_t limit; 846 vm_pindex_t size; 847 848 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 849 p->p_rlimit[RLIMIT_RSS].rlim_max)); 850 size = pmap_resident_tlnw_count(map->pmap); 851 if (limit >= 0 && size > 4096 && size - 4096 >= limit) { 852 vm_pageout_map_deactivate_pages(map, limit); 853 } 854 } 855 #endif 856 857 if (result != KERN_SUCCESS && debug_fault < 0) { 858 kprintf("VM_FAULT %d:%d (%s) result %d " 859 "addr=%jx type=%02x flags=%02x " 860 "fs.m=%p fs.prot=%02x fs.wflags=%02x fs.entry=%p\n", 861 (curthread->td_proc ? curthread->td_proc->p_pid : -1), 862 (curthread->td_lwp ? curthread->td_lwp->lwp_tid : -1), 863 curthread->td_comm, 864 result, 865 (intmax_t)vaddr, fault_type, fault_flags, 866 fs.mary[0], fs.prot, fs.wflags, fs.entry); 867 while (debug_fault < 0 && (debug_fault & 1)) 868 tsleep(&debug_fault, 0, "DEBUG", hz); 869 } 870 871 return (result); 872 } 873 874 /* 875 * Attempt a lockless vm_fault() shortcut. The stars have to align for this 876 * to work. But if it does we can get our page only soft-busied and not 877 * have to touch the vm_object or vnode locks at all. 878 */ 879 static 880 int 881 vm_fault_bypass(struct faultstate *fs, vm_pindex_t first_pindex, 882 vm_pindex_t first_count, int *mextcountp, 883 vm_prot_t fault_type) 884 { 885 vm_page_t m; 886 vm_object_t obj; /* NOT LOCKED */ 887 int n; 888 int nlim; 889 890 /* 891 * Don't waste time if the object is only being used by one vm_map. 892 */ 893 obj = fs->first_ba->object; 894 #if 0 895 if (obj->flags & OBJ_ONEMAPPING) 896 return KERN_FAILURE; 897 #endif 898 899 /* 900 * This will try to wire/unwire a page, which can't be done with 901 * a soft-busied page. 902 */ 903 if (fs->fault_flags & VM_FAULT_WIRE_MASK) 904 return KERN_FAILURE; 905 906 /* 907 * Ick, can't handle this 908 */ 909 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 910 #ifdef VM_FAULT_QUICK_DEBUG 911 ++vm_fault_bypass_failure_count1; 912 #endif 913 return KERN_FAILURE; 914 } 915 916 /* 917 * Ok, try to get the vm_page quickly via the hash table. The 918 * page will be soft-busied on success (NOT hard-busied). 919 */ 920 m = vm_page_hash_get(obj, first_pindex); 921 if (m == NULL) { 922 #ifdef VM_FAULT_QUICK_DEBUG 923 ++vm_fault_bypass_failure_count2; 924 #endif 925 return KERN_FAILURE; 926 } 927 if ((obj->flags & OBJ_DEAD) || 928 m->valid != VM_PAGE_BITS_ALL || 929 m->queue - m->pc != PQ_ACTIVE || 930 (m->flags & PG_SWAPPED)) { 931 vm_page_sbusy_drop(m); 932 #ifdef VM_FAULT_QUICK_DEBUG 933 ++vm_fault_bypass_failure_count3; 934 #endif 935 return KERN_FAILURE; 936 } 937 938 /* 939 * The page is already fully valid, ACTIVE, and is not PG_SWAPPED. 940 * 941 * Don't map the page writable when emulating the dirty bit, a 942 * fault must be taken for proper emulation (vkernel). 943 */ 944 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 945 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 946 if ((fault_type & VM_PROT_WRITE) == 0) 947 fs->prot &= ~VM_PROT_WRITE; 948 } 949 950 /* 951 * If this is a write fault the object and the page must already 952 * be writable. Since we don't hold an object lock and only a 953 * soft-busy on the page, we cannot manipulate the object or 954 * the page state (other than the page queue). 955 */ 956 if (fs->prot & VM_PROT_WRITE) { 957 if ((obj->flags & (OBJ_WRITEABLE | OBJ_MIGHTBEDIRTY)) != 958 (OBJ_WRITEABLE | OBJ_MIGHTBEDIRTY) || 959 m->dirty != VM_PAGE_BITS_ALL) { 960 vm_page_sbusy_drop(m); 961 #ifdef VM_FAULT_QUICK_DEBUG 962 ++vm_fault_bypass_failure_count4; 963 #endif 964 return KERN_FAILURE; 965 } 966 vm_set_nosync(m, fs->entry); 967 } 968 969 /* 970 * Set page and potentially burst in more 971 * 972 * Even though we are only soft-busied we can still move pages 973 * around in the normal queue(s). The soft-busy prevents the 974 * page from being removed from the object, etc (normal operation). 975 * 976 * However, in this fast path it is excessively important to avoid 977 * any hard locks, so we use a special passive version of activate. 978 */ 979 fs->msoftonly = 1; 980 fs->mary[0] = m; 981 vm_page_soft_activate(m); 982 983 if (vm_fault_bypass_count > 1) { 984 nlim = vm_fault_bypass_count; 985 if (nlim > VM_FAULT_MAX_QUICK) /* array limit(+1) */ 986 nlim = VM_FAULT_MAX_QUICK; 987 if (nlim > first_count) /* user limit */ 988 nlim = first_count; 989 990 for (n = 1; n < nlim; ++n) { 991 m = vm_page_hash_get(obj, first_pindex + n); 992 if (m == NULL) 993 break; 994 if (m->valid != VM_PAGE_BITS_ALL || 995 m->queue - m->pc != PQ_ACTIVE || 996 (m->flags & PG_SWAPPED)) { 997 vm_page_sbusy_drop(m); 998 break; 999 } 1000 if (fs->prot & VM_PROT_WRITE) { 1001 if ((obj->flags & (OBJ_WRITEABLE | 1002 OBJ_MIGHTBEDIRTY)) != 1003 (OBJ_WRITEABLE | OBJ_MIGHTBEDIRTY) || 1004 m->dirty != VM_PAGE_BITS_ALL) { 1005 vm_page_sbusy_drop(m); 1006 break; 1007 } 1008 } 1009 vm_page_soft_activate(m); 1010 fs->mary[n] = m; 1011 } 1012 *mextcountp = n; 1013 } 1014 1015 #ifdef VM_FAULT_QUICK_DEBUG 1016 ++vm_fault_bypass_success_count; 1017 #endif 1018 1019 return KERN_SUCCESS; 1020 } 1021 1022 /* 1023 * Fault in the specified virtual address in the current process map, 1024 * returning a held VM page or NULL. See vm_fault_page() for more 1025 * information. 1026 * 1027 * No requirements. 1028 */ 1029 vm_page_t 1030 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, 1031 int *errorp, int *busyp) 1032 { 1033 struct lwp *lp = curthread->td_lwp; 1034 vm_page_t m; 1035 1036 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 1037 fault_type, VM_FAULT_NORMAL, 1038 errorp, busyp); 1039 return(m); 1040 } 1041 1042 /* 1043 * Fault in the specified virtual address in the specified map, doing all 1044 * necessary manipulation of the object store and all necessary I/O. Return 1045 * a held VM page or NULL, and set *errorp. The related pmap is not 1046 * updated. 1047 * 1048 * If busyp is not NULL then *busyp will be set to TRUE if this routine 1049 * decides to return a busied page (aka VM_PROT_WRITE), or FALSE if it 1050 * does not (VM_PROT_WRITE not specified or busyp is NULL). If busyp is 1051 * NULL the returned page is only held. 1052 * 1053 * If the caller has no intention of writing to the page's contents, busyp 1054 * can be passed as NULL along with VM_PROT_WRITE to force a COW operation 1055 * without busying the page. 1056 * 1057 * The returned page will also be marked PG_REFERENCED. 1058 * 1059 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 1060 * error will be returned. 1061 * 1062 * No requirements. 1063 */ 1064 vm_page_t 1065 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 1066 int fault_flags, int *errorp, int *busyp) 1067 { 1068 vm_pindex_t first_pindex; 1069 vm_pindex_t first_count; 1070 struct faultstate fs; 1071 int result; 1072 int retry; 1073 int growstack; 1074 int didcow; 1075 vm_prot_t orig_fault_type = fault_type; 1076 1077 retry = 0; 1078 didcow = 0; 1079 fs.hardfault = 0; 1080 fs.fault_flags = fault_flags; 1081 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 1082 1083 /* 1084 * Dive the pmap (concurrency possible). If we find the 1085 * appropriate page we can terminate early and quickly. 1086 * 1087 * This works great for normal programs but will always return 1088 * NULL for host lookups of vkernel maps in VMM mode. 1089 * 1090 * NOTE: pmap_fault_page_quick() might not busy the page. If 1091 * VM_PROT_WRITE is set in fault_type and pmap_fault_page_quick() 1092 * returns non-NULL, it will safely dirty the returned vm_page_t 1093 * for us. We cannot safely dirty it here (it might not be 1094 * busy). 1095 */ 1096 fs.mary[0] = pmap_fault_page_quick(map->pmap, vaddr, fault_type, busyp); 1097 if (fs.mary[0]) { 1098 *errorp = 0; 1099 return(fs.mary[0]); 1100 } 1101 1102 /* 1103 * Otherwise take a concurrency hit and do a formal page 1104 * fault. 1105 */ 1106 fs.vp = NULL; 1107 fs.shared = vm_shared_fault; 1108 fs.first_shared = vm_shared_fault; 1109 fs.msoftonly = 0; 1110 growstack = 1; 1111 1112 /* 1113 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 1114 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 1115 * we can try shared first. 1116 */ 1117 if (fault_flags & VM_FAULT_UNSWAP) { 1118 fs.first_shared = 0; 1119 } 1120 1121 RetryFault: 1122 /* 1123 * Find the vm_map_entry representing the backing store and resolve 1124 * the top level object and page index. This may have the side 1125 * effect of executing a copy-on-write on the map entry and/or 1126 * creating a shadow object, but will not COW any actual VM pages. 1127 * 1128 * On success fs.map is left read-locked and various other fields 1129 * are initialized but not otherwise referenced or locked. 1130 * 1131 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 1132 * if the map entry is a virtual page table and also writable, 1133 * so we can set the 'A'accessed bit in the virtual page table 1134 * entry. 1135 */ 1136 fs.map = map; 1137 fs.first_ba_held = 0; 1138 result = vm_map_lookup(&fs.map, vaddr, fault_type, 1139 &fs.entry, &fs.first_ba, 1140 &first_pindex, &first_count, 1141 &fs.first_prot, &fs.wflags); 1142 1143 if (result != KERN_SUCCESS) { 1144 if (result == KERN_FAILURE_NOFAULT) { 1145 *errorp = KERN_FAILURE; 1146 fs.mary[0] = NULL; 1147 goto done; 1148 } 1149 if (result != KERN_PROTECTION_FAILURE || 1150 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 1151 { 1152 if (result == KERN_INVALID_ADDRESS && growstack && 1153 map != &kernel_map && curproc != NULL) { 1154 result = vm_map_growstack(map, vaddr); 1155 if (result == KERN_SUCCESS) { 1156 growstack = 0; 1157 ++retry; 1158 goto RetryFault; 1159 } 1160 result = KERN_FAILURE; 1161 } 1162 fs.mary[0] = NULL; 1163 *errorp = result; 1164 goto done; 1165 } 1166 1167 /* 1168 * If we are user-wiring a r/w segment, and it is COW, then 1169 * we need to do the COW operation. Note that we don't 1170 * currently COW RO sections now, because it is NOT desirable 1171 * to COW .text. We simply keep .text from ever being COW'ed 1172 * and take the heat that one cannot debug wired .text sections. 1173 */ 1174 result = vm_map_lookup(&fs.map, vaddr, 1175 VM_PROT_READ|VM_PROT_WRITE| 1176 VM_PROT_OVERRIDE_WRITE, 1177 &fs.entry, &fs.first_ba, 1178 &first_pindex, &first_count, 1179 &fs.first_prot, &fs.wflags); 1180 if (result != KERN_SUCCESS) { 1181 /* could also be KERN_FAILURE_NOFAULT */ 1182 *errorp = KERN_FAILURE; 1183 fs.mary[0] = NULL; 1184 goto done; 1185 } 1186 1187 /* 1188 * If we don't COW now, on a user wire, the user will never 1189 * be able to write to the mapping. If we don't make this 1190 * restriction, the bookkeeping would be nearly impossible. 1191 * 1192 * XXX We have a shared lock, this will have a MP race but 1193 * I don't see how it can hurt anything. 1194 */ 1195 if ((fs.entry->protection & VM_PROT_WRITE) == 0) { 1196 atomic_clear_char(&fs.entry->max_protection, 1197 VM_PROT_WRITE); 1198 } 1199 } 1200 1201 /* 1202 * fs.map is read-locked 1203 * 1204 * Misc checks. Save the map generation number to detect races. 1205 */ 1206 fs.lookup_still_valid = 1; 1207 fs.first_m = NULL; 1208 fs.ba = fs.first_ba; 1209 1210 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 1211 panic("vm_fault: fault on nofault entry, addr: %lx", 1212 (u_long)vaddr); 1213 } 1214 1215 /* 1216 * A user-kernel shared map has no VM object and bypasses 1217 * everything. We execute the uksmap function with a temporary 1218 * fictitious vm_page. The address is directly mapped with no 1219 * management. 1220 */ 1221 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 1222 struct vm_page fakem; 1223 1224 bzero(&fakem, sizeof(fakem)); 1225 fakem.pindex = first_pindex; 1226 fakem.flags = PG_FICTITIOUS | PG_UNQUEUED; 1227 fakem.busy_count = PBUSY_LOCKED; 1228 fakem.valid = VM_PAGE_BITS_ALL; 1229 fakem.pat_mode = VM_MEMATTR_DEFAULT; 1230 if (fs.entry->ba.uksmap(&fs.entry->ba, UKSMAPOP_FAULT, 1231 fs.entry->aux.dev, &fakem)) { 1232 *errorp = KERN_FAILURE; 1233 fs.mary[0] = NULL; 1234 unlock_things(&fs); 1235 goto done2; 1236 } 1237 fs.mary[0] = PHYS_TO_VM_PAGE(fakem.phys_addr); 1238 vm_page_hold(fs.mary[0]); 1239 if (busyp) 1240 *busyp = 0; /* don't need to busy R or W */ 1241 unlock_things(&fs); 1242 *errorp = 0; 1243 goto done; 1244 } 1245 1246 1247 /* 1248 * A system map entry may return a NULL object. No object means 1249 * no pager means an unrecoverable kernel fault. 1250 */ 1251 if (fs.first_ba == NULL) { 1252 panic("vm_fault: unrecoverable fault at %p in entry %p", 1253 (void *)vaddr, fs.entry); 1254 } 1255 1256 /* 1257 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 1258 * is set. 1259 * 1260 * Unfortunately a deadlock can occur if we are forced to page-in 1261 * from swap, but diving all the way into the vm_pager_get_page() 1262 * function to find out is too much. Just check the object type. 1263 */ 1264 if ((curthread->td_flags & TDF_NOFAULT) && 1265 (retry || 1266 fs.first_ba->object->type == OBJT_VNODE || 1267 fs.first_ba->object->type == OBJT_SWAP || 1268 fs.first_ba->backing_ba)) { 1269 *errorp = KERN_FAILURE; 1270 unlock_things(&fs); 1271 fs.mary[0] = NULL; 1272 goto done2; 1273 } 1274 1275 /* 1276 * If the entry is wired we cannot change the page protection. 1277 */ 1278 if (fs.wflags & FW_WIRED) 1279 fault_type = fs.first_prot; 1280 1281 /* 1282 * Make a reference to this object to prevent its disposal while we 1283 * are messing with it. Once we have the reference, the map is free 1284 * to be diddled. Since objects reference their shadows (and copies), 1285 * they will stay around as well. 1286 * 1287 * The reference should also prevent an unexpected collapse of the 1288 * parent that might move pages from the current object into the 1289 * parent unexpectedly, resulting in corruption. 1290 * 1291 * Bump the paging-in-progress count to prevent size changes (e.g. 1292 * truncation operations) during I/O. This must be done after 1293 * obtaining the vnode lock in order to avoid possible deadlocks. 1294 */ 1295 if (fs.first_ba->flags & VM_MAP_BACK_EXCL_HEUR) 1296 fs.first_shared = 0; 1297 1298 if (fs.first_shared) 1299 vm_object_hold_shared(fs.first_ba->object); 1300 else 1301 vm_object_hold(fs.first_ba->object); 1302 fs.first_ba_held = 1; 1303 if (fs.vp == NULL) 1304 fs.vp = vnode_pager_lock(fs.first_ba); /* shared */ 1305 1306 /* 1307 * The page we want is at (first_object, first_pindex), but if the 1308 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 1309 * page table to figure out the actual pindex. 1310 * 1311 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 1312 * ONLY 1313 */ 1314 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1315 result = vm_fault_vpagetable(&fs, &first_pindex, 1316 fs.entry->aux.master_pde, 1317 fault_type, 1); 1318 first_count = 1; 1319 if (result == KERN_TRY_AGAIN) { 1320 ++retry; 1321 goto RetryFault; 1322 } 1323 if (result != KERN_SUCCESS) { 1324 *errorp = result; 1325 fs.mary[0] = NULL; 1326 goto done; 1327 } 1328 } 1329 1330 /* 1331 * Now we have the actual (object, pindex), fault in the page. If 1332 * vm_fault_object() fails it will unlock and deallocate the FS 1333 * data. If it succeeds everything remains locked and fs->ba->object 1334 * will have an additinal PIP count if fs->ba != fs->first_ba. 1335 */ 1336 fs.mary[0] = NULL; 1337 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 1338 1339 if (result == KERN_TRY_AGAIN) { 1340 KKASSERT(fs.first_ba_held == 0); 1341 ++retry; 1342 didcow |= fs.wflags & FW_DIDCOW; 1343 goto RetryFault; 1344 } 1345 if (result != KERN_SUCCESS) { 1346 *errorp = result; 1347 fs.mary[0] = NULL; 1348 goto done; 1349 } 1350 1351 if ((orig_fault_type & VM_PROT_WRITE) && 1352 (fs.prot & VM_PROT_WRITE) == 0) { 1353 *errorp = KERN_PROTECTION_FAILURE; 1354 unlock_things(&fs); 1355 fs.mary[0] = NULL; 1356 goto done; 1357 } 1358 1359 /* 1360 * Generally speaking we don't want to update the pmap because 1361 * this routine can be called many times for situations that do 1362 * not require updating the pmap, not to mention the page might 1363 * already be in the pmap. 1364 * 1365 * However, if our vm_map_lookup() results in a COW, we need to 1366 * at least remove the pte from the pmap to guarantee proper 1367 * visibility of modifications made to the process. For example, 1368 * modifications made by vkernel uiocopy/related routines and 1369 * modifications made by ptrace(). 1370 */ 1371 vm_page_flag_set(fs.mary[0], PG_REFERENCED); 1372 #if 0 1373 pmap_enter(fs.map->pmap, vaddr, fs.mary[0], fs.prot, 1374 fs.wflags & FW_WIRED, NULL); 1375 mycpu->gd_cnt.v_vm_faults++; 1376 if (curthread->td_lwp) 1377 ++curthread->td_lwp->lwp_ru.ru_minflt; 1378 #endif 1379 if ((fs.wflags | didcow) | FW_DIDCOW) { 1380 pmap_remove(fs.map->pmap, 1381 vaddr & ~PAGE_MASK, 1382 (vaddr & ~PAGE_MASK) + PAGE_SIZE); 1383 } 1384 1385 /* 1386 * On success vm_fault_object() does not unlock or deallocate, and 1387 * fs.mary[0] will contain a busied page. So we must unlock here 1388 * after having messed with the pmap. 1389 */ 1390 unlock_things(&fs); 1391 1392 /* 1393 * Return a held page. We are not doing any pmap manipulation so do 1394 * not set PG_MAPPED. However, adjust the page flags according to 1395 * the fault type because the caller may not use a managed pmapping 1396 * (so we don't want to lose the fact that the page will be dirtied 1397 * if a write fault was specified). 1398 */ 1399 if (fault_type & VM_PROT_WRITE) 1400 vm_page_dirty(fs.mary[0]); 1401 vm_page_activate(fs.mary[0]); 1402 1403 if (curthread->td_lwp) { 1404 if (fs.hardfault) { 1405 curthread->td_lwp->lwp_ru.ru_majflt++; 1406 } else { 1407 curthread->td_lwp->lwp_ru.ru_minflt++; 1408 } 1409 } 1410 1411 /* 1412 * Unlock everything, and return the held or busied page. 1413 */ 1414 if (busyp) { 1415 if (fault_type & VM_PROT_WRITE) { 1416 vm_page_dirty(fs.mary[0]); 1417 *busyp = 1; 1418 } else { 1419 *busyp = 0; 1420 vm_page_hold(fs.mary[0]); 1421 vm_page_wakeup(fs.mary[0]); 1422 } 1423 } else { 1424 vm_page_hold(fs.mary[0]); 1425 vm_page_wakeup(fs.mary[0]); 1426 } 1427 /*vm_object_deallocate(fs.first_ba->object);*/ 1428 *errorp = 0; 1429 1430 done: 1431 KKASSERT(fs.first_ba_held == 0); 1432 done2: 1433 return(fs.mary[0]); 1434 } 1435 1436 /* 1437 * Fault in the specified (object,offset), dirty the returned page as 1438 * needed. If the requested fault_type cannot be done NULL and an 1439 * error is returned. 1440 * 1441 * A held (but not busied) page is returned. 1442 * 1443 * The passed in object must be held as specified by the shared 1444 * argument. 1445 */ 1446 vm_page_t 1447 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 1448 vm_prot_t fault_type, int fault_flags, 1449 int *sharedp, int *errorp) 1450 { 1451 int result; 1452 vm_pindex_t first_pindex; 1453 vm_pindex_t first_count; 1454 struct faultstate fs; 1455 struct vm_map_entry entry; 1456 1457 /* 1458 * Since we aren't actually faulting the page into a 1459 * pmap we can just fake the entry.ba. 1460 */ 1461 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1462 bzero(&entry, sizeof(entry)); 1463 entry.maptype = VM_MAPTYPE_NORMAL; 1464 entry.protection = entry.max_protection = fault_type; 1465 entry.ba.backing_ba = NULL; 1466 entry.ba.object = object; 1467 entry.ba.offset = 0; 1468 1469 fs.hardfault = 0; 1470 fs.fault_flags = fault_flags; 1471 fs.map = NULL; 1472 fs.shared = vm_shared_fault; 1473 fs.first_shared = *sharedp; 1474 fs.msoftonly = 0; 1475 fs.vp = NULL; 1476 fs.first_ba_held = -1; /* object held across call, prevent drop */ 1477 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 1478 1479 /* 1480 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 1481 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 1482 * we can try shared first. 1483 */ 1484 if (fs.first_shared && (fault_flags & VM_FAULT_UNSWAP)) { 1485 fs.first_shared = 0; 1486 vm_object_upgrade(object); 1487 } 1488 1489 /* 1490 * Retry loop as needed (typically for shared->exclusive transitions) 1491 */ 1492 RetryFault: 1493 *sharedp = fs.first_shared; 1494 first_pindex = OFF_TO_IDX(offset); 1495 first_count = 1; 1496 fs.first_ba = &entry.ba; 1497 fs.ba = fs.first_ba; 1498 fs.entry = &entry; 1499 fs.first_prot = fault_type; 1500 fs.wflags = 0; 1501 1502 /* 1503 * Make a reference to this object to prevent its disposal while we 1504 * are messing with it. Once we have the reference, the map is free 1505 * to be diddled. Since objects reference their shadows (and copies), 1506 * they will stay around as well. 1507 * 1508 * The reference should also prevent an unexpected collapse of the 1509 * parent that might move pages from the current object into the 1510 * parent unexpectedly, resulting in corruption. 1511 * 1512 * Bump the paging-in-progress count to prevent size changes (e.g. 1513 * truncation operations) during I/O. This must be done after 1514 * obtaining the vnode lock in order to avoid possible deadlocks. 1515 */ 1516 if (fs.vp == NULL) 1517 fs.vp = vnode_pager_lock(fs.first_ba); 1518 1519 fs.lookup_still_valid = 1; 1520 fs.first_m = NULL; 1521 1522 #if 0 1523 /* XXX future - ability to operate on VM object using vpagetable */ 1524 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1525 result = vm_fault_vpagetable(&fs, &first_pindex, 1526 fs.entry->aux.master_pde, 1527 fault_type, 0); 1528 if (result == KERN_TRY_AGAIN) { 1529 if (fs.first_shared == 0 && *sharedp) 1530 vm_object_upgrade(object); 1531 goto RetryFault; 1532 } 1533 if (result != KERN_SUCCESS) { 1534 *errorp = result; 1535 return (NULL); 1536 } 1537 } 1538 #endif 1539 1540 /* 1541 * Now we have the actual (object, pindex), fault in the page. If 1542 * vm_fault_object() fails it will unlock and deallocate the FS 1543 * data. If it succeeds everything remains locked and fs->ba->object 1544 * will have an additinal PIP count if fs->ba != fs->first_ba. 1545 * 1546 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_ba intact. 1547 * We may have to upgrade its lock to handle the requested fault. 1548 */ 1549 result = vm_fault_object(&fs, first_pindex, fault_type, 0); 1550 1551 if (result == KERN_TRY_AGAIN) { 1552 if (fs.first_shared == 0 && *sharedp) 1553 vm_object_upgrade(object); 1554 goto RetryFault; 1555 } 1556 if (result != KERN_SUCCESS) { 1557 *errorp = result; 1558 return(NULL); 1559 } 1560 1561 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 1562 *errorp = KERN_PROTECTION_FAILURE; 1563 unlock_things(&fs); 1564 return(NULL); 1565 } 1566 1567 /* 1568 * On success vm_fault_object() does not unlock or deallocate, so we 1569 * do it here. Note that the returned fs.m will be busied. 1570 */ 1571 unlock_things(&fs); 1572 1573 /* 1574 * Return a held page. We are not doing any pmap manipulation so do 1575 * not set PG_MAPPED. However, adjust the page flags according to 1576 * the fault type because the caller may not use a managed pmapping 1577 * (so we don't want to lose the fact that the page will be dirtied 1578 * if a write fault was specified). 1579 */ 1580 vm_page_hold(fs.mary[0]); 1581 vm_page_activate(fs.mary[0]); 1582 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY)) 1583 vm_page_dirty(fs.mary[0]); 1584 if (fault_flags & VM_FAULT_UNSWAP) 1585 swap_pager_unswapped(fs.mary[0]); 1586 1587 /* 1588 * Indicate that the page was accessed. 1589 */ 1590 vm_page_flag_set(fs.mary[0], PG_REFERENCED); 1591 1592 if (curthread->td_lwp) { 1593 if (fs.hardfault) { 1594 curthread->td_lwp->lwp_ru.ru_majflt++; 1595 } else { 1596 curthread->td_lwp->lwp_ru.ru_minflt++; 1597 } 1598 } 1599 1600 /* 1601 * Unlock everything, and return the held page. 1602 */ 1603 vm_page_wakeup(fs.mary[0]); 1604 /*vm_object_deallocate(fs.first_ba->object);*/ 1605 1606 *errorp = 0; 1607 return(fs.mary[0]); 1608 } 1609 1610 /* 1611 * Translate the virtual page number (first_pindex) that is relative 1612 * to the address space into a logical page number that is relative to the 1613 * backing object. Use the virtual page table pointed to by (vpte). 1614 * 1615 * Possibly downgrade the protection based on the vpte bits. 1616 * 1617 * This implements an N-level page table. Any level can terminate the 1618 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 1619 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 1620 */ 1621 static 1622 int 1623 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 1624 vpte_t vpte, int fault_type, int allow_nofault) 1625 { 1626 struct lwbuf *lwb; 1627 struct lwbuf lwb_cache; 1628 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */ 1629 int result; 1630 vpte_t *ptep; 1631 1632 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_ba->object)); 1633 for (;;) { 1634 /* 1635 * We cannot proceed if the vpte is not valid, not readable 1636 * for a read fault, not writable for a write fault, or 1637 * not executable for an instruction execution fault. 1638 */ 1639 if ((vpte & VPTE_V) == 0) { 1640 unlock_things(fs); 1641 return (KERN_FAILURE); 1642 } 1643 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) { 1644 unlock_things(fs); 1645 return (KERN_FAILURE); 1646 } 1647 if ((fault_type & VM_PROT_EXECUTE) && (vpte & VPTE_NX)) { 1648 unlock_things(fs); 1649 return (KERN_FAILURE); 1650 } 1651 if ((vpte & VPTE_PS) || vshift == 0) 1652 break; 1653 1654 /* 1655 * Get the page table page. Nominally we only read the page 1656 * table, but since we are actively setting VPTE_M and VPTE_A, 1657 * tell vm_fault_object() that we are writing it. 1658 * 1659 * There is currently no real need to optimize this. 1660 */ 1661 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT, 1662 VM_PROT_READ|VM_PROT_WRITE, 1663 allow_nofault); 1664 if (result != KERN_SUCCESS) 1665 return (result); 1666 1667 /* 1668 * Process the returned fs.mary[0] and look up the page table 1669 * entry in the page table page. 1670 */ 1671 vshift -= VPTE_PAGE_BITS; 1672 lwb = lwbuf_alloc(fs->mary[0], &lwb_cache); 1673 ptep = ((vpte_t *)lwbuf_kva(lwb) + 1674 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 1675 vm_page_activate(fs->mary[0]); 1676 1677 /* 1678 * Page table write-back - entire operation including 1679 * validation of the pte must be atomic to avoid races 1680 * against the vkernel changing the pte. 1681 * 1682 * If the vpte is valid for the* requested operation, do 1683 * a write-back to the page table. 1684 * 1685 * XXX VPTE_M is not set properly for page directory pages. 1686 * It doesn't get set in the page directory if the page table 1687 * is modified during a read access. 1688 */ 1689 for (;;) { 1690 vpte_t nvpte; 1691 1692 /* 1693 * Reload for the cmpset, but make sure the pte is 1694 * still valid. 1695 */ 1696 vpte = *ptep; 1697 cpu_ccfence(); 1698 nvpte = vpte; 1699 1700 if ((vpte & VPTE_V) == 0) 1701 break; 1702 1703 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW)) 1704 nvpte |= VPTE_M | VPTE_A; 1705 if (fault_type & (VM_PROT_READ | VM_PROT_EXECUTE)) 1706 nvpte |= VPTE_A; 1707 if (vpte == nvpte) 1708 break; 1709 if (atomic_cmpset_long(ptep, vpte, nvpte)) { 1710 vm_page_dirty(fs->mary[0]); 1711 break; 1712 } 1713 } 1714 lwbuf_free(lwb); 1715 vm_page_flag_set(fs->mary[0], PG_REFERENCED); 1716 vm_page_wakeup(fs->mary[0]); 1717 fs->mary[0] = NULL; 1718 cleanup_fault(fs); 1719 } 1720 1721 /* 1722 * When the vkernel sets VPTE_RW it expects the real kernel to 1723 * reflect VPTE_M back when the page is modified via the mapping. 1724 * In order to accomplish this the real kernel must map the page 1725 * read-only for read faults and use write faults to reflect VPTE_M 1726 * back. 1727 * 1728 * Once VPTE_M has been set, the real kernel's pte allows writing. 1729 * If the vkernel clears VPTE_M the vkernel must be sure to 1730 * MADV_INVAL the real kernel's mappings to force the real kernel 1731 * to re-fault on the next write so oit can set VPTE_M again. 1732 */ 1733 if ((fault_type & VM_PROT_WRITE) == 0 && 1734 (vpte & (VPTE_RW | VPTE_M)) != (VPTE_RW | VPTE_M)) { 1735 fs->first_prot &= ~VM_PROT_WRITE; 1736 } 1737 1738 /* 1739 * Disable EXECUTE perms if NX bit is set. 1740 */ 1741 if (vpte & VPTE_NX) 1742 fs->first_prot &= ~VM_PROT_EXECUTE; 1743 1744 /* 1745 * Combine remaining address bits with the vpte. 1746 */ 1747 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) + 1748 (*pindex & ((1L << vshift) - 1)); 1749 return (KERN_SUCCESS); 1750 } 1751 1752 1753 /* 1754 * This is the core of the vm_fault code. 1755 * 1756 * Do all operations required to fault-in (fs.first_ba->object, pindex). 1757 * Run through the backing store as necessary and do required COW or virtual 1758 * copy operations. The caller has already fully resolved the vm_map_entry 1759 * and, if appropriate, has created a copy-on-write layer. All we need to 1760 * do is iterate the object chain. 1761 * 1762 * On failure (fs) is unlocked and deallocated and the caller may return or 1763 * retry depending on the failure code. On success (fs) is NOT unlocked or 1764 * deallocated, fs.mary[0] will contained a resolved, busied page, and fs.ba's 1765 * object will have an additional PIP count if it is not equal to 1766 * fs.first_ba. 1767 * 1768 * If locks based on fs->first_shared or fs->shared are insufficient, 1769 * clear the appropriate field(s) and return RETRY. COWs require that 1770 * first_shared be 0, while page allocations (or frees) require that 1771 * shared be 0. Renames require that both be 0. 1772 * 1773 * NOTE! fs->[first_]shared might be set with VM_FAULT_DIRTY also set. 1774 * we will have to retry with it exclusive if the vm_page is 1775 * PG_SWAPPED. 1776 * 1777 * fs->first_ba->object must be held on call. 1778 */ 1779 static 1780 int 1781 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex, 1782 vm_prot_t fault_type, int allow_nofault) 1783 { 1784 vm_map_backing_t next_ba; 1785 vm_pindex_t pindex; 1786 int error; 1787 1788 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_ba->object)); 1789 fs->prot = fs->first_prot; 1790 pindex = first_pindex; 1791 KKASSERT(fs->ba == fs->first_ba); 1792 1793 vm_object_pip_add(fs->first_ba->object, 1); 1794 1795 /* 1796 * If a read fault occurs we try to upgrade the page protection 1797 * and make it also writable if possible. There are three cases 1798 * where we cannot make the page mapping writable: 1799 * 1800 * (1) The mapping is read-only or the VM object is read-only, 1801 * fs->prot above will simply not have VM_PROT_WRITE set. 1802 * 1803 * (2) If the mapping is a virtual page table fs->first_prot will 1804 * have already been properly adjusted by vm_fault_vpagetable(). 1805 * to detect writes so we can set VPTE_M in the virtual page 1806 * table. Used by vkernels. 1807 * 1808 * (3) If the VM page is read-only or copy-on-write, upgrading would 1809 * just result in an unnecessary COW fault. 1810 * 1811 * (4) If the pmap specifically requests A/M bit emulation, downgrade 1812 * here. 1813 */ 1814 #if 0 1815 /* see vpagetable code */ 1816 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1817 if ((fault_type & VM_PROT_WRITE) == 0) 1818 fs->prot &= ~VM_PROT_WRITE; 1819 } 1820 #endif 1821 1822 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 1823 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 1824 if ((fault_type & VM_PROT_WRITE) == 0) 1825 fs->prot &= ~VM_PROT_WRITE; 1826 } 1827 1828 /* vm_object_hold(fs->ba->object); implied b/c ba == first_ba */ 1829 1830 for (;;) { 1831 /* 1832 * If the object is dead, we stop here 1833 */ 1834 if (fs->ba->object->flags & OBJ_DEAD) { 1835 vm_object_pip_wakeup(fs->first_ba->object); 1836 unlock_things(fs); 1837 return (KERN_PROTECTION_FAILURE); 1838 } 1839 1840 /* 1841 * See if the page is resident. Wait/Retry if the page is 1842 * busy (lots of stuff may have changed so we can't continue 1843 * in that case). 1844 * 1845 * We can theoretically allow the soft-busy case on a read 1846 * fault if the page is marked valid, but since such 1847 * pages are typically already pmap'd, putting that 1848 * special case in might be more effort then it is 1849 * worth. We cannot under any circumstances mess 1850 * around with a vm_page_t->busy page except, perhaps, 1851 * to pmap it. 1852 */ 1853 fs->mary[0] = vm_page_lookup_busy_try(fs->ba->object, pindex, 1854 TRUE, &error); 1855 if (error) { 1856 vm_object_pip_wakeup(fs->first_ba->object); 1857 unlock_things(fs); 1858 vm_page_sleep_busy(fs->mary[0], TRUE, "vmpfw"); 1859 mycpu->gd_cnt.v_intrans++; 1860 fs->mary[0] = NULL; 1861 return (KERN_TRY_AGAIN); 1862 } 1863 if (fs->mary[0]) { 1864 /* 1865 * The page is busied for us. 1866 * 1867 * If reactivating a page from PQ_CACHE we may have 1868 * to rate-limit. 1869 */ 1870 int queue = fs->mary[0]->queue; 1871 vm_page_unqueue_nowakeup(fs->mary[0]); 1872 1873 if ((queue - fs->mary[0]->pc) == PQ_CACHE && 1874 vm_page_count_severe()) { 1875 vm_page_activate(fs->mary[0]); 1876 vm_page_wakeup(fs->mary[0]); 1877 fs->mary[0] = NULL; 1878 vm_object_pip_wakeup(fs->first_ba->object); 1879 unlock_things(fs); 1880 if (allow_nofault == 0 || 1881 (curthread->td_flags & TDF_NOFAULT) == 0) { 1882 thread_t td; 1883 1884 vm_wait_pfault(); 1885 td = curthread; 1886 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1887 return (KERN_PROTECTION_FAILURE); 1888 } 1889 return (KERN_TRY_AGAIN); 1890 } 1891 1892 /* 1893 * If it still isn't completely valid (readable), 1894 * or if a read-ahead-mark is set on the VM page, 1895 * jump to readrest, else we found the page and 1896 * can return. 1897 * 1898 * We can release the spl once we have marked the 1899 * page busy. 1900 */ 1901 if (fs->mary[0]->object != &kernel_object) { 1902 if ((fs->mary[0]->valid & VM_PAGE_BITS_ALL) != 1903 VM_PAGE_BITS_ALL) { 1904 goto readrest; 1905 } 1906 if (fs->mary[0]->flags & PG_RAM) { 1907 if (debug_cluster) 1908 kprintf("R"); 1909 vm_page_flag_clear(fs->mary[0], PG_RAM); 1910 goto readrest; 1911 } 1912 } 1913 atomic_clear_int(&fs->first_ba->flags, 1914 VM_MAP_BACK_EXCL_HEUR); 1915 break; /* break to PAGE HAS BEEN FOUND */ 1916 } 1917 1918 /* 1919 * Page is not resident, If this is the search termination 1920 * or the pager might contain the page, allocate a new page. 1921 */ 1922 if (TRYPAGER(fs) || fs->ba == fs->first_ba) { 1923 /* 1924 * If this is a SWAP object we can use the shared 1925 * lock to check existence of a swap block. If 1926 * there isn't one we can skip to the next object. 1927 * 1928 * However, if this is the first object we allocate 1929 * a page now just in case we need to copy to it 1930 * later. 1931 */ 1932 if (fs->ba != fs->first_ba && 1933 fs->ba->object->type == OBJT_SWAP) { 1934 if (swap_pager_haspage_locked(fs->ba->object, 1935 pindex) == 0) { 1936 goto next; 1937 } 1938 } 1939 1940 /* 1941 * Allocating, must be exclusive. 1942 */ 1943 atomic_set_int(&fs->first_ba->flags, 1944 VM_MAP_BACK_EXCL_HEUR); 1945 if (fs->ba == fs->first_ba && fs->first_shared) { 1946 fs->first_shared = 0; 1947 vm_object_pip_wakeup(fs->first_ba->object); 1948 unlock_things(fs); 1949 return (KERN_TRY_AGAIN); 1950 } 1951 if (fs->ba != fs->first_ba && fs->shared) { 1952 fs->first_shared = 0; 1953 fs->shared = 0; 1954 vm_object_pip_wakeup(fs->first_ba->object); 1955 unlock_things(fs); 1956 return (KERN_TRY_AGAIN); 1957 } 1958 1959 /* 1960 * If the page is beyond the object size we fail 1961 */ 1962 if (pindex >= fs->ba->object->size) { 1963 vm_object_pip_wakeup(fs->first_ba->object); 1964 unlock_things(fs); 1965 return (KERN_PROTECTION_FAILURE); 1966 } 1967 1968 /* 1969 * Allocate a new page for this object/offset pair. 1970 * 1971 * It is possible for the allocation to race, so 1972 * handle the case. 1973 */ 1974 fs->mary[0] = NULL; 1975 if (!vm_page_count_severe()) { 1976 fs->mary[0] = vm_page_alloc(fs->ba->object, 1977 pindex, 1978 ((fs->vp || fs->ba->backing_ba) ? 1979 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1980 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1981 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1982 } 1983 if (fs->mary[0] == NULL) { 1984 vm_object_pip_wakeup(fs->first_ba->object); 1985 unlock_things(fs); 1986 if (allow_nofault == 0 || 1987 (curthread->td_flags & TDF_NOFAULT) == 0) { 1988 thread_t td; 1989 1990 vm_wait_pfault(); 1991 td = curthread; 1992 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1993 return (KERN_PROTECTION_FAILURE); 1994 } 1995 return (KERN_TRY_AGAIN); 1996 } 1997 1998 /* 1999 * Fall through to readrest. We have a new page which 2000 * will have to be paged (since m->valid will be 0). 2001 */ 2002 } 2003 2004 readrest: 2005 /* 2006 * We have found an invalid or partially valid page, a 2007 * page with a read-ahead mark which might be partially or 2008 * fully valid (and maybe dirty too), or we have allocated 2009 * a new page. 2010 * 2011 * Attempt to fault-in the page if there is a chance that the 2012 * pager has it, and potentially fault in additional pages 2013 * at the same time. 2014 * 2015 * If TRYPAGER is true then fs.mary[0] will be non-NULL and 2016 * busied for us. 2017 */ 2018 if (TRYPAGER(fs)) { 2019 u_char behavior = vm_map_entry_behavior(fs->entry); 2020 vm_object_t object; 2021 vm_page_t first_m; 2022 int seqaccess; 2023 int rv; 2024 2025 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 2026 seqaccess = 0; 2027 else 2028 seqaccess = -1; 2029 2030 /* 2031 * Doing I/O may synchronously insert additional 2032 * pages so we can't be shared at this point either. 2033 * 2034 * NOTE: We can't free fs->mary[0] here in the 2035 * allocated case (fs->ba != fs->first_ba) as 2036 * this would require an exclusively locked 2037 * VM object. 2038 */ 2039 if (fs->ba == fs->first_ba && fs->first_shared) { 2040 vm_page_deactivate(fs->mary[0]); 2041 vm_page_wakeup(fs->mary[0]); 2042 fs->mary[0]= NULL; 2043 fs->first_shared = 0; 2044 vm_object_pip_wakeup(fs->first_ba->object); 2045 unlock_things(fs); 2046 return (KERN_TRY_AGAIN); 2047 } 2048 if (fs->ba != fs->first_ba && fs->shared) { 2049 vm_page_deactivate(fs->mary[0]); 2050 vm_page_wakeup(fs->mary[0]); 2051 fs->mary[0] = NULL; 2052 fs->first_shared = 0; 2053 fs->shared = 0; 2054 vm_object_pip_wakeup(fs->first_ba->object); 2055 unlock_things(fs); 2056 return (KERN_TRY_AGAIN); 2057 } 2058 2059 object = fs->ba->object; 2060 first_m = NULL; 2061 2062 /* object is held, no more access to entry or ba's */ 2063 2064 /* 2065 * Acquire the page data. We still hold object 2066 * and the page has been BUSY's. 2067 * 2068 * We own the page, but we must re-issue the lookup 2069 * because the pager may have replaced it (for example, 2070 * in order to enter a fictitious page into the 2071 * object). In this situation the pager will have 2072 * cleaned up the old page and left the new one 2073 * busy for us. 2074 * 2075 * If we got here through a PG_RAM read-ahead 2076 * mark the page may be partially dirty and thus 2077 * not freeable. Don't bother checking to see 2078 * if the pager has the page because we can't free 2079 * it anyway. We have to depend on the get_page 2080 * operation filling in any gaps whether there is 2081 * backing store or not. 2082 * 2083 * We must dispose of the page (fs->mary[0]) and also 2084 * possibly first_m (the fronting layer). If 2085 * this is a write fault leave the page intact 2086 * because we will probably have to copy fs->mary[0] 2087 * to fs->first_m on the retry. If this is a 2088 * read fault we probably won't need the page. 2089 */ 2090 rv = vm_pager_get_page(object, &fs->mary[0], seqaccess); 2091 2092 if (rv == VM_PAGER_OK) { 2093 ++fs->hardfault; 2094 fs->mary[0] = vm_page_lookup(object, pindex); 2095 if (fs->mary[0]) { 2096 vm_page_activate(fs->mary[0]); 2097 vm_page_wakeup(fs->mary[0]); 2098 fs->mary[0] = NULL; 2099 } 2100 2101 if (fs->mary[0]) { 2102 /* have page */ 2103 break; 2104 } 2105 vm_object_pip_wakeup(fs->first_ba->object); 2106 unlock_things(fs); 2107 return (KERN_TRY_AGAIN); 2108 } 2109 2110 /* 2111 * If the pager doesn't have the page, continue on 2112 * to the next object. Retain the vm_page if this 2113 * is the first object, we may need to copy into 2114 * it later. 2115 */ 2116 if (rv == VM_PAGER_FAIL) { 2117 if (fs->ba != fs->first_ba) { 2118 vm_page_free(fs->mary[0]); 2119 fs->mary[0] = NULL; 2120 } 2121 goto next; 2122 } 2123 2124 /* 2125 * Remove the bogus page (which does not exist at this 2126 * object/offset). 2127 * 2128 * Also wake up any other process that may want to bring 2129 * in this page. 2130 * 2131 * If this is the top-level object, we must leave the 2132 * busy page to prevent another process from rushing 2133 * past us, and inserting the page in that object at 2134 * the same time that we are. 2135 */ 2136 if (rv == VM_PAGER_ERROR) { 2137 if (curproc) { 2138 kprintf("vm_fault: pager read error, " 2139 "pid %d (%s)\n", 2140 curproc->p_pid, 2141 curproc->p_comm); 2142 } else { 2143 kprintf("vm_fault: pager read error, " 2144 "thread %p (%s)\n", 2145 curthread, 2146 curthread->td_comm); 2147 } 2148 } 2149 2150 /* 2151 * I/O error or data outside pager's range. 2152 */ 2153 if (fs->mary[0]) { 2154 vnode_pager_freepage(fs->mary[0]); 2155 fs->mary[0] = NULL; 2156 } 2157 if (first_m) { 2158 vm_page_free(first_m); 2159 first_m = NULL; /* safety */ 2160 } 2161 vm_object_pip_wakeup(object); 2162 unlock_things(fs); 2163 2164 switch(rv) { 2165 case VM_PAGER_ERROR: 2166 return (KERN_FAILURE); 2167 case VM_PAGER_BAD: 2168 return (KERN_PROTECTION_FAILURE); 2169 default: 2170 return (KERN_PROTECTION_FAILURE); 2171 } 2172 2173 #if 0 2174 /* 2175 * Data outside the range of the pager or an I/O error 2176 * 2177 * The page may have been wired during the pagein, 2178 * e.g. by the buffer cache, and cannot simply be 2179 * freed. Call vnode_pager_freepage() to deal with it. 2180 * 2181 * The object is not held shared so we can safely 2182 * free the page. 2183 */ 2184 if (fs->ba != fs->first_ba) { 2185 2186 /* 2187 * XXX - we cannot just fall out at this 2188 * point, m has been freed and is invalid! 2189 */ 2190 } 2191 2192 /* 2193 * XXX - the check for kernel_map is a kludge to work 2194 * around having the machine panic on a kernel space 2195 * fault w/ I/O error. 2196 */ 2197 if (((fs->map != &kernel_map) && 2198 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 2199 if (fs->m) { 2200 /* from just above */ 2201 KKASSERT(fs->first_shared == 0); 2202 vnode_pager_freepage(fs->m); 2203 fs->m = NULL; 2204 } 2205 /* NOT REACHED */ 2206 } 2207 #endif 2208 } 2209 2210 next: 2211 /* 2212 * We get here if the object has a default pager (or unwiring) 2213 * or the pager doesn't have the page. 2214 * 2215 * fs->first_m will be used for the COW unless we find a 2216 * deeper page to be mapped read-only, in which case the 2217 * unlock*(fs) will free first_m. 2218 */ 2219 if (fs->ba == fs->first_ba) 2220 fs->first_m = fs->mary[0]; 2221 2222 /* 2223 * Move on to the next object. The chain lock should prevent 2224 * the backing_object from getting ripped out from under us. 2225 * 2226 * The object lock for the next object is governed by 2227 * fs->shared. 2228 */ 2229 next_ba = fs->ba->backing_ba; 2230 if (next_ba == NULL) { 2231 /* 2232 * If there's no object left, fill the page in the top 2233 * object with zeros. 2234 */ 2235 if (fs->ba != fs->first_ba) { 2236 vm_object_pip_wakeup(fs->ba->object); 2237 vm_object_drop(fs->ba->object); 2238 fs->ba = fs->first_ba; 2239 pindex = first_pindex; 2240 fs->mary[0] = fs->first_m; 2241 } 2242 fs->first_m = NULL; 2243 2244 /* 2245 * Zero the page and mark it valid. 2246 */ 2247 vm_page_zero_fill(fs->mary[0]); 2248 mycpu->gd_cnt.v_zfod++; 2249 fs->mary[0]->valid = VM_PAGE_BITS_ALL; 2250 break; /* break to PAGE HAS BEEN FOUND */ 2251 } 2252 2253 if (fs->shared) 2254 vm_object_hold_shared(next_ba->object); 2255 else 2256 vm_object_hold(next_ba->object); 2257 KKASSERT(next_ba == fs->ba->backing_ba); 2258 pindex -= OFF_TO_IDX(fs->ba->offset); 2259 pindex += OFF_TO_IDX(next_ba->offset); 2260 2261 if (fs->ba != fs->first_ba) { 2262 vm_object_pip_wakeup(fs->ba->object); 2263 vm_object_lock_swap(); /* flip ba/next_ba */ 2264 vm_object_drop(fs->ba->object); 2265 } 2266 fs->ba = next_ba; 2267 vm_object_pip_add(next_ba->object, 1); 2268 } 2269 2270 /* 2271 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 2272 * is held.] 2273 * 2274 * object still held. 2275 * vm_map may not be locked (determined by fs->lookup_still_valid) 2276 * 2277 * local shared variable may be different from fs->shared. 2278 * 2279 * If the page is being written, but isn't already owned by the 2280 * top-level object, we have to copy it into a new page owned by the 2281 * top-level object. 2282 */ 2283 KASSERT((fs->mary[0]->busy_count & PBUSY_LOCKED) != 0, 2284 ("vm_fault: not busy after main loop")); 2285 2286 if (fs->ba != fs->first_ba) { 2287 /* 2288 * We only really need to copy if we want to write it. 2289 */ 2290 if (fault_type & VM_PROT_WRITE) { 2291 #if 0 2292 /* CODE REFACTOR IN PROGRESS, REMOVE OPTIMIZATION */ 2293 /* 2294 * This allows pages to be virtually copied from a 2295 * backing_object into the first_object, where the 2296 * backing object has no other refs to it, and cannot 2297 * gain any more refs. Instead of a bcopy, we just 2298 * move the page from the backing object to the 2299 * first object. Note that we must mark the page 2300 * dirty in the first object so that it will go out 2301 * to swap when needed. 2302 */ 2303 if (virtual_copy_ok(fs)) { 2304 /* 2305 * (first_m) and (m) are both busied. We have 2306 * move (m) into (first_m)'s object/pindex 2307 * in an atomic fashion, then free (first_m). 2308 * 2309 * first_object is held so second remove 2310 * followed by the rename should wind 2311 * up being atomic. vm_page_free() might 2312 * block so we don't do it until after the 2313 * rename. 2314 */ 2315 vm_page_protect(fs->first_m, VM_PROT_NONE); 2316 vm_page_remove(fs->first_m); 2317 vm_page_rename(fs->mary[0], 2318 fs->first_ba->object, 2319 first_pindex); 2320 vm_page_free(fs->first_m); 2321 fs->first_m = fs->mary[0]; 2322 fs->mary[0] = NULL; 2323 mycpu->gd_cnt.v_cow_optim++; 2324 } else 2325 #endif 2326 { 2327 /* 2328 * Oh, well, lets copy it. 2329 * 2330 * We used to unmap the original page here 2331 * because vm_fault_page() didn't and this 2332 * would cause havoc for the umtx*() code 2333 * and the procfs code. 2334 * 2335 * This is no longer necessary. The 2336 * vm_fault_page() routine will now unmap the 2337 * page after a COW, and the umtx code will 2338 * recover on its own. 2339 */ 2340 /* 2341 * NOTE: Since fs->mary[0] is a backing page, 2342 * it is read-only, so there isn't any 2343 * copy race vs writers. 2344 */ 2345 KKASSERT(fs->first_shared == 0); 2346 vm_page_copy(fs->mary[0], fs->first_m); 2347 /* pmap_remove_specific( 2348 &curthread->td_lwp->lwp_vmspace->vm_pmap, 2349 fs->mary[0]); */ 2350 } 2351 2352 /* 2353 * We no longer need the old page or object. 2354 */ 2355 if (fs->mary[0]) 2356 release_page(fs); 2357 2358 /* 2359 * fs->ba != fs->first_ba due to above conditional 2360 */ 2361 vm_object_pip_wakeup(fs->ba->object); 2362 vm_object_drop(fs->ba->object); 2363 fs->ba = fs->first_ba; 2364 2365 /* 2366 * Only use the new page below... 2367 */ 2368 mycpu->gd_cnt.v_cow_faults++; 2369 fs->mary[0] = fs->first_m; 2370 pindex = first_pindex; 2371 } else { 2372 /* 2373 * If it wasn't a write fault avoid having to copy 2374 * the page by mapping it read-only from backing 2375 * store. The process is not allowed to modify 2376 * backing pages. 2377 */ 2378 fs->prot &= ~VM_PROT_WRITE; 2379 } 2380 } 2381 2382 /* 2383 * Relock the map if necessary, then check the generation count. 2384 * relock_map() will update fs->timestamp to account for the 2385 * relocking if necessary. 2386 * 2387 * If the count has changed after relocking then all sorts of 2388 * crap may have happened and we have to retry. 2389 * 2390 * NOTE: The relock_map() can fail due to a deadlock against 2391 * the vm_page we are holding BUSY. 2392 */ 2393 KKASSERT(fs->lookup_still_valid != 0); 2394 #if 0 2395 if (fs->lookup_still_valid == 0 && fs->map) { 2396 if (relock_map(fs) || 2397 fs->map->timestamp != fs->map_generation) { 2398 release_page(fs); 2399 vm_object_pip_wakeup(fs->first_ba->object); 2400 unlock_things(fs); 2401 return (KERN_TRY_AGAIN); 2402 } 2403 } 2404 #endif 2405 2406 /* 2407 * If the fault is a write, we know that this page is being 2408 * written NOW so dirty it explicitly to save on pmap_is_modified() 2409 * calls later. 2410 * 2411 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 2412 * if the page is already dirty to prevent data written with 2413 * the expectation of being synced from not being synced. 2414 * Likewise if this entry does not request NOSYNC then make 2415 * sure the page isn't marked NOSYNC. Applications sharing 2416 * data should use the same flags to avoid ping ponging. 2417 * 2418 * Also tell the backing pager, if any, that it should remove 2419 * any swap backing since the page is now dirty. 2420 */ 2421 vm_page_activate(fs->mary[0]); 2422 if (fs->prot & VM_PROT_WRITE) { 2423 vm_object_set_writeable_dirty(fs->mary[0]->object); 2424 vm_set_nosync(fs->mary[0], fs->entry); 2425 if (fs->fault_flags & VM_FAULT_DIRTY) { 2426 vm_page_dirty(fs->mary[0]); 2427 if (fs->mary[0]->flags & PG_SWAPPED) { 2428 /* 2429 * If the page is swapped out we have to call 2430 * swap_pager_unswapped() which requires an 2431 * exclusive object lock. If we are shared, 2432 * we must clear the shared flag and retry. 2433 */ 2434 if ((fs->ba == fs->first_ba && 2435 fs->first_shared) || 2436 (fs->ba != fs->first_ba && fs->shared)) { 2437 vm_page_wakeup(fs->mary[0]); 2438 fs->mary[0] = NULL; 2439 if (fs->ba == fs->first_ba) 2440 fs->first_shared = 0; 2441 else 2442 fs->shared = 0; 2443 vm_object_pip_wakeup( 2444 fs->first_ba->object); 2445 unlock_things(fs); 2446 return (KERN_TRY_AGAIN); 2447 } 2448 swap_pager_unswapped(fs->mary[0]); 2449 } 2450 } 2451 } 2452 2453 /* 2454 * We found our page at backing layer ba. Leave the layer state 2455 * intact. 2456 */ 2457 2458 vm_object_pip_wakeup(fs->first_ba->object); 2459 #if 0 2460 if (fs->ba != fs->first_ba) 2461 vm_object_drop(fs->ba->object); 2462 #endif 2463 2464 /* 2465 * Page had better still be busy. We are still locked up and 2466 * fs->ba->object will have another PIP reference for the case 2467 * where fs->ba != fs->first_ba. 2468 */ 2469 KASSERT(fs->mary[0]->busy_count & PBUSY_LOCKED, 2470 ("vm_fault: page %p not busy!", fs->mary[0])); 2471 2472 /* 2473 * Sanity check: page must be completely valid or it is not fit to 2474 * map into user space. vm_pager_get_pages() ensures this. 2475 */ 2476 if (fs->mary[0]->valid != VM_PAGE_BITS_ALL) { 2477 vm_page_zero_invalid(fs->mary[0], TRUE); 2478 kprintf("Warning: page %p partially invalid on fault\n", 2479 fs->mary[0]); 2480 } 2481 2482 return (KERN_SUCCESS); 2483 } 2484 2485 /* 2486 * Wire down a range of virtual addresses in a map. The entry in question 2487 * should be marked in-transition and the map must be locked. We must 2488 * release the map temporarily while faulting-in the page to avoid a 2489 * deadlock. Note that the entry may be clipped while we are blocked but 2490 * will never be freed. 2491 * 2492 * map must be locked on entry. 2493 */ 2494 int 2495 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, 2496 boolean_t user_wire, int kmflags) 2497 { 2498 boolean_t fictitious; 2499 vm_offset_t start; 2500 vm_offset_t end; 2501 vm_offset_t va; 2502 pmap_t pmap; 2503 int rv; 2504 int wire_prot; 2505 int fault_flags; 2506 vm_page_t m; 2507 2508 if (user_wire) { 2509 wire_prot = VM_PROT_READ; 2510 fault_flags = VM_FAULT_USER_WIRE; 2511 } else { 2512 wire_prot = VM_PROT_READ | VM_PROT_WRITE; 2513 fault_flags = VM_FAULT_CHANGE_WIRING; 2514 } 2515 if (kmflags & KM_NOTLBSYNC) 2516 wire_prot |= VM_PROT_NOSYNC; 2517 2518 pmap = vm_map_pmap(map); 2519 start = entry->ba.start; 2520 end = entry->ba.end; 2521 2522 switch(entry->maptype) { 2523 case VM_MAPTYPE_NORMAL: 2524 case VM_MAPTYPE_VPAGETABLE: 2525 fictitious = entry->ba.object && 2526 ((entry->ba.object->type == OBJT_DEVICE) || 2527 (entry->ba.object->type == OBJT_MGTDEVICE)); 2528 break; 2529 case VM_MAPTYPE_UKSMAP: 2530 fictitious = TRUE; 2531 break; 2532 default: 2533 fictitious = FALSE; 2534 break; 2535 } 2536 2537 if (entry->eflags & MAP_ENTRY_KSTACK) 2538 start += PAGE_SIZE; 2539 map->timestamp++; 2540 vm_map_unlock(map); 2541 2542 /* 2543 * We simulate a fault to get the page and enter it in the physical 2544 * map. 2545 */ 2546 for (va = start; va < end; va += PAGE_SIZE) { 2547 rv = vm_fault(map, va, wire_prot, fault_flags); 2548 if (rv) { 2549 while (va > start) { 2550 va -= PAGE_SIZE; 2551 m = pmap_unwire(pmap, va); 2552 if (m && !fictitious) { 2553 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2554 vm_page_unwire(m, 1); 2555 vm_page_wakeup(m); 2556 } 2557 } 2558 goto done; 2559 } 2560 } 2561 rv = KERN_SUCCESS; 2562 done: 2563 vm_map_lock(map); 2564 2565 return (rv); 2566 } 2567 2568 /* 2569 * Unwire a range of virtual addresses in a map. The map should be 2570 * locked. 2571 */ 2572 void 2573 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 2574 { 2575 boolean_t fictitious; 2576 vm_offset_t start; 2577 vm_offset_t end; 2578 vm_offset_t va; 2579 pmap_t pmap; 2580 vm_page_t m; 2581 2582 pmap = vm_map_pmap(map); 2583 start = entry->ba.start; 2584 end = entry->ba.end; 2585 fictitious = entry->ba.object && 2586 ((entry->ba.object->type == OBJT_DEVICE) || 2587 (entry->ba.object->type == OBJT_MGTDEVICE)); 2588 if (entry->eflags & MAP_ENTRY_KSTACK) 2589 start += PAGE_SIZE; 2590 2591 /* 2592 * Since the pages are wired down, we must be able to get their 2593 * mappings from the physical map system. 2594 */ 2595 for (va = start; va < end; va += PAGE_SIZE) { 2596 m = pmap_unwire(pmap, va); 2597 if (m && !fictitious) { 2598 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2599 vm_page_unwire(m, 1); 2600 vm_page_wakeup(m); 2601 } 2602 } 2603 } 2604 2605 /* 2606 * Simulate write faults to bring all data into the head object, return 2607 * KERN_SUCCESS on success (which should be always unless the system runs 2608 * out of memory). 2609 * 2610 * The caller will handle destroying the backing_ba's. 2611 */ 2612 int 2613 vm_fault_collapse(vm_map_t map, vm_map_entry_t entry) 2614 { 2615 struct faultstate fs; 2616 vm_ooffset_t scan; 2617 vm_pindex_t pindex; 2618 vm_object_t object; 2619 int rv; 2620 int all_shadowed; 2621 2622 bzero(&fs, sizeof(fs)); 2623 object = entry->ba.object; 2624 2625 fs.first_prot = entry->max_protection | /* optional VM_PROT_EXECUTE */ 2626 VM_PROT_READ | VM_PROT_WRITE | VM_PROT_OVERRIDE_WRITE; 2627 fs.fault_flags = VM_FAULT_NORMAL; 2628 fs.map = map; 2629 fs.entry = entry; 2630 fs.lookup_still_valid = -1; /* leave map atomically locked */ 2631 fs.first_ba = &entry->ba; 2632 fs.first_ba_held = -1; /* leave object held */ 2633 2634 /* fs.hardfault */ 2635 2636 vm_object_hold(object); 2637 rv = KERN_SUCCESS; 2638 2639 scan = entry->ba.start; 2640 all_shadowed = 1; 2641 2642 while (scan < entry->ba.end) { 2643 pindex = OFF_TO_IDX(entry->ba.offset + (scan - entry->ba.start)); 2644 2645 if (vm_page_lookup(object, pindex)) { 2646 scan += PAGE_SIZE; 2647 continue; 2648 } 2649 2650 all_shadowed = 0; 2651 fs.ba = fs.first_ba; 2652 fs.prot = fs.first_prot; 2653 2654 rv = vm_fault_object(&fs, pindex, fs.first_prot, 1); 2655 if (rv == KERN_TRY_AGAIN) 2656 continue; 2657 if (rv != KERN_SUCCESS) 2658 break; 2659 vm_page_flag_set(fs.mary[0], PG_REFERENCED); 2660 vm_page_activate(fs.mary[0]); 2661 vm_page_wakeup(fs.mary[0]); 2662 scan += PAGE_SIZE; 2663 } 2664 KKASSERT(entry->ba.object == object); 2665 vm_object_drop(object); 2666 2667 /* 2668 * If the fronting object did not have every page we have to clear 2669 * the pmap range due to the pages being changed so we can fault-in 2670 * the proper pages. 2671 */ 2672 if (all_shadowed == 0) 2673 pmap_remove(map->pmap, entry->ba.start, entry->ba.end); 2674 2675 return rv; 2676 } 2677 2678 /* 2679 * Copy all of the pages from one map entry to another. If the source 2680 * is wired down we just use vm_page_lookup(). If not we use 2681 * vm_fault_object(). 2682 * 2683 * The source and destination maps must be locked for write. 2684 * The source and destination maps token must be held 2685 * 2686 * No other requirements. 2687 * 2688 * XXX do segment optimization 2689 */ 2690 void 2691 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2692 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2693 { 2694 vm_object_t dst_object; 2695 vm_object_t src_object; 2696 vm_ooffset_t dst_offset; 2697 vm_ooffset_t src_offset; 2698 vm_prot_t prot; 2699 vm_offset_t vaddr; 2700 vm_page_t dst_m; 2701 vm_page_t src_m; 2702 2703 src_object = src_entry->ba.object; 2704 src_offset = src_entry->ba.offset; 2705 2706 /* 2707 * Create the top-level object for the destination entry. (Doesn't 2708 * actually shadow anything - we copy the pages directly.) 2709 */ 2710 vm_map_entry_allocate_object(dst_entry); 2711 dst_object = dst_entry->ba.object; 2712 2713 prot = dst_entry->max_protection; 2714 2715 /* 2716 * Loop through all of the pages in the entry's range, copying each 2717 * one from the source object (it should be there) to the destination 2718 * object. 2719 */ 2720 vm_object_hold(src_object); 2721 vm_object_hold(dst_object); 2722 2723 for (vaddr = dst_entry->ba.start, dst_offset = 0; 2724 vaddr < dst_entry->ba.end; 2725 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2726 2727 /* 2728 * Allocate a page in the destination object 2729 */ 2730 do { 2731 dst_m = vm_page_alloc(dst_object, 2732 OFF_TO_IDX(dst_offset), 2733 VM_ALLOC_NORMAL); 2734 if (dst_m == NULL) { 2735 vm_wait(0); 2736 } 2737 } while (dst_m == NULL); 2738 2739 /* 2740 * Find the page in the source object, and copy it in. 2741 * (Because the source is wired down, the page will be in 2742 * memory.) 2743 */ 2744 src_m = vm_page_lookup(src_object, 2745 OFF_TO_IDX(dst_offset + src_offset)); 2746 if (src_m == NULL) 2747 panic("vm_fault_copy_wired: page missing"); 2748 2749 vm_page_copy(src_m, dst_m); 2750 2751 /* 2752 * Enter it in the pmap... 2753 */ 2754 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2755 2756 /* 2757 * Mark it no longer busy, and put it on the active list. 2758 */ 2759 vm_page_activate(dst_m); 2760 vm_page_wakeup(dst_m); 2761 } 2762 vm_object_drop(dst_object); 2763 vm_object_drop(src_object); 2764 } 2765 2766 #if 0 2767 2768 /* 2769 * This routine checks around the requested page for other pages that 2770 * might be able to be faulted in. This routine brackets the viable 2771 * pages for the pages to be paged in. 2772 * 2773 * Inputs: 2774 * m, rbehind, rahead 2775 * 2776 * Outputs: 2777 * marray (array of vm_page_t), reqpage (index of requested page) 2778 * 2779 * Return value: 2780 * number of pages in marray 2781 */ 2782 static int 2783 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2784 vm_page_t *marray, int *reqpage) 2785 { 2786 int i,j; 2787 vm_object_t object; 2788 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2789 vm_page_t rtm; 2790 int cbehind, cahead; 2791 2792 object = m->object; 2793 pindex = m->pindex; 2794 2795 /* 2796 * we don't fault-ahead for device pager 2797 */ 2798 if ((object->type == OBJT_DEVICE) || 2799 (object->type == OBJT_MGTDEVICE)) { 2800 *reqpage = 0; 2801 marray[0] = m; 2802 return 1; 2803 } 2804 2805 /* 2806 * if the requested page is not available, then give up now 2807 */ 2808 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2809 *reqpage = 0; /* not used by caller, fix compiler warn */ 2810 return 0; 2811 } 2812 2813 if ((cbehind == 0) && (cahead == 0)) { 2814 *reqpage = 0; 2815 marray[0] = m; 2816 return 1; 2817 } 2818 2819 if (rahead > cahead) { 2820 rahead = cahead; 2821 } 2822 2823 if (rbehind > cbehind) { 2824 rbehind = cbehind; 2825 } 2826 2827 /* 2828 * Do not do any readahead if we have insufficient free memory. 2829 * 2830 * XXX code was broken disabled before and has instability 2831 * with this conditonal fixed, so shortcut for now. 2832 */ 2833 if (burst_fault == 0 || vm_page_count_severe()) { 2834 marray[0] = m; 2835 *reqpage = 0; 2836 return 1; 2837 } 2838 2839 /* 2840 * scan backward for the read behind pages -- in memory 2841 * 2842 * Assume that if the page is not found an interrupt will not 2843 * create it. Theoretically interrupts can only remove (busy) 2844 * pages, not create new associations. 2845 */ 2846 if (pindex > 0) { 2847 if (rbehind > pindex) { 2848 rbehind = pindex; 2849 startpindex = 0; 2850 } else { 2851 startpindex = pindex - rbehind; 2852 } 2853 2854 vm_object_hold(object); 2855 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2856 if (vm_page_lookup(object, tpindex - 1)) 2857 break; 2858 } 2859 2860 i = 0; 2861 while (tpindex < pindex) { 2862 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2863 VM_ALLOC_NULL_OK); 2864 if (rtm == NULL) { 2865 for (j = 0; j < i; j++) { 2866 vm_page_free(marray[j]); 2867 } 2868 vm_object_drop(object); 2869 marray[0] = m; 2870 *reqpage = 0; 2871 return 1; 2872 } 2873 marray[i] = rtm; 2874 ++i; 2875 ++tpindex; 2876 } 2877 vm_object_drop(object); 2878 } else { 2879 i = 0; 2880 } 2881 2882 /* 2883 * Assign requested page 2884 */ 2885 marray[i] = m; 2886 *reqpage = i; 2887 ++i; 2888 2889 /* 2890 * Scan forwards for read-ahead pages 2891 */ 2892 tpindex = pindex + 1; 2893 endpindex = tpindex + rahead; 2894 if (endpindex > object->size) 2895 endpindex = object->size; 2896 2897 vm_object_hold(object); 2898 while (tpindex < endpindex) { 2899 if (vm_page_lookup(object, tpindex)) 2900 break; 2901 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2902 VM_ALLOC_NULL_OK); 2903 if (rtm == NULL) 2904 break; 2905 marray[i] = rtm; 2906 ++i; 2907 ++tpindex; 2908 } 2909 vm_object_drop(object); 2910 2911 return (i); 2912 } 2913 2914 #endif 2915 2916 /* 2917 * vm_prefault() provides a quick way of clustering pagefaults into a 2918 * processes address space. It is a "cousin" of pmap_object_init_pt, 2919 * except it runs at page fault time instead of mmap time. 2920 * 2921 * vm.fast_fault Enables pre-faulting zero-fill pages 2922 * 2923 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2924 * prefault. Scan stops in either direction when 2925 * a page is found to already exist. 2926 * 2927 * This code used to be per-platform pmap_prefault(). It is now 2928 * machine-independent and enhanced to also pre-fault zero-fill pages 2929 * (see vm.fast_fault) as well as make them writable, which greatly 2930 * reduces the number of page faults programs incur. 2931 * 2932 * Application performance when pre-faulting zero-fill pages is heavily 2933 * dependent on the application. Very tiny applications like /bin/echo 2934 * lose a little performance while applications of any appreciable size 2935 * gain performance. Prefaulting multiple pages also reduces SMP 2936 * congestion and can improve SMP performance significantly. 2937 * 2938 * NOTE! prot may allow writing but this only applies to the top level 2939 * object. If we wind up mapping a page extracted from a backing 2940 * object we have to make sure it is read-only. 2941 * 2942 * NOTE! The caller has already handled any COW operations on the 2943 * vm_map_entry via the normal fault code. Do NOT call this 2944 * shortcut unless the normal fault code has run on this entry. 2945 * 2946 * The related map must be locked. 2947 * No other requirements. 2948 */ 2949 __read_mostly static int vm_prefault_pages = 8; 2950 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2951 "Maximum number of pages to pre-fault"); 2952 __read_mostly static int vm_fast_fault = 1; 2953 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2954 "Burst fault zero-fill regions"); 2955 2956 /* 2957 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2958 * is not already dirty by other means. This will prevent passive 2959 * filesystem syncing as well as 'sync' from writing out the page. 2960 */ 2961 static void 2962 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2963 { 2964 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2965 if (m->dirty == 0) 2966 vm_page_flag_set(m, PG_NOSYNC); 2967 } else { 2968 vm_page_flag_clear(m, PG_NOSYNC); 2969 } 2970 } 2971 2972 static void 2973 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 2974 int fault_flags) 2975 { 2976 vm_map_backing_t ba; /* first ba */ 2977 struct lwp *lp; 2978 vm_page_t m; 2979 vm_offset_t addr; 2980 vm_pindex_t index; 2981 vm_pindex_t pindex; 2982 vm_object_t object; 2983 int pprot; 2984 int i; 2985 int noneg; 2986 int nopos; 2987 int maxpages; 2988 2989 /* 2990 * Get stable max count value, disabled if set to 0 2991 */ 2992 maxpages = vm_prefault_pages; 2993 cpu_ccfence(); 2994 if (maxpages <= 0) 2995 return; 2996 2997 /* 2998 * We do not currently prefault mappings that use virtual page 2999 * tables. We do not prefault foreign pmaps. 3000 */ 3001 if (entry->maptype != VM_MAPTYPE_NORMAL) 3002 return; 3003 lp = curthread->td_lwp; 3004 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 3005 return; 3006 3007 /* 3008 * Limit pre-fault count to 1024 pages. 3009 */ 3010 if (maxpages > 1024) 3011 maxpages = 1024; 3012 3013 ba = &entry->ba; 3014 object = entry->ba.object; 3015 KKASSERT(object != NULL); 3016 3017 /* 3018 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively 3019 * now (or do something more complex XXX). 3020 */ 3021 vm_object_hold(object); 3022 3023 noneg = 0; 3024 nopos = 0; 3025 for (i = 0; i < maxpages; ++i) { 3026 vm_object_t lobject; 3027 vm_object_t nobject; 3028 vm_map_backing_t last_ba; /* last ba */ 3029 vm_map_backing_t next_ba; /* last ba */ 3030 int allocated = 0; 3031 int error; 3032 3033 /* 3034 * This can eat a lot of time on a heavily contended 3035 * machine so yield on the tick if needed. 3036 */ 3037 if ((i & 7) == 7) 3038 lwkt_yield(); 3039 3040 /* 3041 * Calculate the page to pre-fault, stopping the scan in 3042 * each direction separately if the limit is reached. 3043 */ 3044 if (i & 1) { 3045 if (noneg) 3046 continue; 3047 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 3048 } else { 3049 if (nopos) 3050 continue; 3051 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 3052 } 3053 if (addr < entry->ba.start) { 3054 noneg = 1; 3055 if (noneg && nopos) 3056 break; 3057 continue; 3058 } 3059 if (addr >= entry->ba.end) { 3060 nopos = 1; 3061 if (noneg && nopos) 3062 break; 3063 continue; 3064 } 3065 3066 /* 3067 * Skip pages already mapped, and stop scanning in that 3068 * direction. When the scan terminates in both directions 3069 * we are done. 3070 */ 3071 if (pmap_prefault_ok(pmap, addr) == 0) { 3072 if (i & 1) 3073 noneg = 1; 3074 else 3075 nopos = 1; 3076 if (noneg && nopos) 3077 break; 3078 continue; 3079 } 3080 3081 /* 3082 * Follow the backing layers to obtain the page to be mapped 3083 * into the pmap. 3084 * 3085 * If we reach the terminal object without finding a page 3086 * and we determine it would be advantageous, then allocate 3087 * a zero-fill page for the base object. The base object 3088 * is guaranteed to be OBJT_DEFAULT for this case. 3089 * 3090 * In order to not have to check the pager via *haspage*() 3091 * we stop if any non-default object is encountered. e.g. 3092 * a vnode or swap object would stop the loop. 3093 */ 3094 index = ((addr - entry->ba.start) + entry->ba.offset) >> 3095 PAGE_SHIFT; 3096 last_ba = ba; 3097 lobject = object; 3098 pindex = index; 3099 pprot = prot; 3100 3101 /*vm_object_hold(lobject); implied */ 3102 3103 while ((m = vm_page_lookup_busy_try(lobject, pindex, 3104 TRUE, &error)) == NULL) { 3105 if (lobject->type != OBJT_DEFAULT) 3106 break; 3107 if ((next_ba = last_ba->backing_ba) == NULL) { 3108 if (vm_fast_fault == 0) 3109 break; 3110 if ((prot & VM_PROT_WRITE) == 0 || 3111 vm_page_count_min(0)) { 3112 break; 3113 } 3114 3115 /* 3116 * NOTE: Allocated from base object 3117 */ 3118 m = vm_page_alloc(object, index, 3119 VM_ALLOC_NORMAL | 3120 VM_ALLOC_ZERO | 3121 VM_ALLOC_USE_GD | 3122 VM_ALLOC_NULL_OK); 3123 if (m == NULL) 3124 break; 3125 allocated = 1; 3126 pprot = prot; 3127 /* lobject = object .. not needed */ 3128 break; 3129 } 3130 if (next_ba->offset & PAGE_MASK) 3131 break; 3132 nobject = next_ba->object; 3133 vm_object_hold(nobject); 3134 pindex -= last_ba->offset >> PAGE_SHIFT; 3135 pindex += next_ba->offset >> PAGE_SHIFT; 3136 if (last_ba != ba) { 3137 vm_object_lock_swap(); 3138 vm_object_drop(lobject); 3139 } 3140 lobject = nobject; 3141 last_ba = next_ba; 3142 pprot &= ~VM_PROT_WRITE; 3143 } 3144 3145 /* 3146 * NOTE: A non-NULL (m) will be associated with lobject if 3147 * it was found there, otherwise it is probably a 3148 * zero-fill page associated with the base object. 3149 * 3150 * Give-up if no page is available. 3151 */ 3152 if (m == NULL) { 3153 if (last_ba != ba) 3154 vm_object_drop(lobject); 3155 break; 3156 } 3157 3158 /* 3159 * The object must be marked dirty if we are mapping a 3160 * writable page. m->object is either lobject or object, 3161 * both of which are still held. Do this before we 3162 * potentially drop the object. 3163 */ 3164 if (pprot & VM_PROT_WRITE) 3165 vm_object_set_writeable_dirty(m->object); 3166 3167 /* 3168 * Do not conditionalize on PG_RAM. If pages are present in 3169 * the VM system we assume optimal caching. If caching is 3170 * not optimal the I/O gravy train will be restarted when we 3171 * hit an unavailable page. We do not want to try to restart 3172 * the gravy train now because we really don't know how much 3173 * of the object has been cached. The cost for restarting 3174 * the gravy train should be low (since accesses will likely 3175 * be I/O bound anyway). 3176 */ 3177 if (last_ba != ba) 3178 vm_object_drop(lobject); 3179 3180 /* 3181 * Enter the page into the pmap if appropriate. If we had 3182 * allocated the page we have to place it on a queue. If not 3183 * we just have to make sure it isn't on the cache queue 3184 * (pages on the cache queue are not allowed to be mapped). 3185 */ 3186 if (allocated) { 3187 /* 3188 * Page must be zerod. 3189 */ 3190 vm_page_zero_fill(m); 3191 mycpu->gd_cnt.v_zfod++; 3192 m->valid = VM_PAGE_BITS_ALL; 3193 3194 /* 3195 * Handle dirty page case 3196 */ 3197 if (pprot & VM_PROT_WRITE) 3198 vm_set_nosync(m, entry); 3199 pmap_enter(pmap, addr, m, pprot, 0, entry); 3200 #if 0 3201 /* REMOVE ME, a burst counts as one fault */ 3202 mycpu->gd_cnt.v_vm_faults++; 3203 if (curthread->td_lwp) 3204 ++curthread->td_lwp->lwp_ru.ru_minflt; 3205 #endif 3206 vm_page_deactivate(m); 3207 if (pprot & VM_PROT_WRITE) { 3208 /*vm_object_set_writeable_dirty(m->object);*/ 3209 vm_set_nosync(m, entry); 3210 if (fault_flags & VM_FAULT_DIRTY) { 3211 vm_page_dirty(m); 3212 /*XXX*/ 3213 swap_pager_unswapped(m); 3214 } 3215 } 3216 vm_page_wakeup(m); 3217 } else if (error) { 3218 /* couldn't busy page, no wakeup */ 3219 } else if ( 3220 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 3221 (m->flags & PG_FICTITIOUS) == 0) { 3222 /* 3223 * A fully valid page not undergoing soft I/O can 3224 * be immediately entered into the pmap. 3225 */ 3226 if ((m->queue - m->pc) == PQ_CACHE) 3227 vm_page_deactivate(m); 3228 if (pprot & VM_PROT_WRITE) { 3229 /*vm_object_set_writeable_dirty(m->object);*/ 3230 vm_set_nosync(m, entry); 3231 if (fault_flags & VM_FAULT_DIRTY) { 3232 vm_page_dirty(m); 3233 /*XXX*/ 3234 swap_pager_unswapped(m); 3235 } 3236 } 3237 if (pprot & VM_PROT_WRITE) 3238 vm_set_nosync(m, entry); 3239 pmap_enter(pmap, addr, m, pprot, 0, entry); 3240 #if 0 3241 /* REMOVE ME, a burst counts as one fault */ 3242 mycpu->gd_cnt.v_vm_faults++; 3243 if (curthread->td_lwp) 3244 ++curthread->td_lwp->lwp_ru.ru_minflt; 3245 #endif 3246 vm_page_wakeup(m); 3247 } else { 3248 vm_page_wakeup(m); 3249 } 3250 } 3251 vm_object_drop(object); 3252 } 3253 3254 /* 3255 * Object can be held shared 3256 */ 3257 static void 3258 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 3259 vm_map_entry_t entry, int prot, int fault_flags) 3260 { 3261 struct lwp *lp; 3262 vm_page_t m; 3263 vm_offset_t addr; 3264 vm_pindex_t pindex; 3265 vm_object_t object; 3266 int i; 3267 int noneg; 3268 int nopos; 3269 int maxpages; 3270 3271 /* 3272 * Get stable max count value, disabled if set to 0 3273 */ 3274 maxpages = vm_prefault_pages; 3275 cpu_ccfence(); 3276 if (maxpages <= 0) 3277 return; 3278 3279 /* 3280 * We do not currently prefault mappings that use virtual page 3281 * tables. We do not prefault foreign pmaps. 3282 */ 3283 if (entry->maptype != VM_MAPTYPE_NORMAL) 3284 return; 3285 lp = curthread->td_lwp; 3286 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 3287 return; 3288 object = entry->ba.object; 3289 if (entry->ba.backing_ba != NULL) 3290 return; 3291 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 3292 3293 /* 3294 * Limit pre-fault count to 1024 pages. 3295 */ 3296 if (maxpages > 1024) 3297 maxpages = 1024; 3298 3299 noneg = 0; 3300 nopos = 0; 3301 for (i = 0; i < maxpages; ++i) { 3302 int error; 3303 3304 /* 3305 * Calculate the page to pre-fault, stopping the scan in 3306 * each direction separately if the limit is reached. 3307 */ 3308 if (i & 1) { 3309 if (noneg) 3310 continue; 3311 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 3312 } else { 3313 if (nopos) 3314 continue; 3315 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 3316 } 3317 if (addr < entry->ba.start) { 3318 noneg = 1; 3319 if (noneg && nopos) 3320 break; 3321 continue; 3322 } 3323 if (addr >= entry->ba.end) { 3324 nopos = 1; 3325 if (noneg && nopos) 3326 break; 3327 continue; 3328 } 3329 3330 /* 3331 * Follow the VM object chain to obtain the page to be mapped 3332 * into the pmap. This version of the prefault code only 3333 * works with terminal objects. 3334 * 3335 * The page must already exist. If we encounter a problem 3336 * we stop here. 3337 * 3338 * WARNING! We cannot call swap_pager_unswapped() or insert 3339 * a new vm_page with a shared token. 3340 */ 3341 pindex = ((addr - entry->ba.start) + entry->ba.offset) >> 3342 PAGE_SHIFT; 3343 3344 /* 3345 * Skip pages already mapped, and stop scanning in that 3346 * direction. When the scan terminates in both directions 3347 * we are done. 3348 */ 3349 if (pmap_prefault_ok(pmap, addr) == 0) { 3350 if (i & 1) 3351 noneg = 1; 3352 else 3353 nopos = 1; 3354 if (noneg && nopos) 3355 break; 3356 continue; 3357 } 3358 3359 /* 3360 * Shortcut the read-only mapping case using the far more 3361 * efficient vm_page_lookup_sbusy_try() function. This 3362 * allows us to acquire the page soft-busied only which 3363 * is especially nice for concurrent execs of the same 3364 * program. 3365 * 3366 * The lookup function also validates page suitability 3367 * (all valid bits set, and not fictitious). 3368 * 3369 * If the page is in PQ_CACHE we have to fall-through 3370 * and hard-busy it so we can move it out of PQ_CACHE. 3371 */ 3372 if ((prot & VM_PROT_WRITE) == 0) { 3373 m = vm_page_lookup_sbusy_try(object, pindex, 3374 0, PAGE_SIZE); 3375 if (m == NULL) 3376 break; 3377 if ((m->queue - m->pc) != PQ_CACHE) { 3378 pmap_enter(pmap, addr, m, prot, 0, entry); 3379 #if 0 3380 /* REMOVE ME, a burst counts as one fault */ 3381 mycpu->gd_cnt.v_vm_faults++; 3382 if (curthread->td_lwp) 3383 ++curthread->td_lwp->lwp_ru.ru_minflt; 3384 #endif 3385 vm_page_sbusy_drop(m); 3386 continue; 3387 } 3388 vm_page_sbusy_drop(m); 3389 } 3390 3391 /* 3392 * Fallback to normal vm_page lookup code. This code 3393 * hard-busies the page. Not only that, but the page 3394 * can remain in that state for a significant period 3395 * time due to pmap_enter()'s overhead. 3396 */ 3397 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3398 if (m == NULL || error) 3399 break; 3400 3401 /* 3402 * Stop if the page cannot be trivially entered into the 3403 * pmap. 3404 */ 3405 if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) || 3406 (m->flags & PG_FICTITIOUS) || 3407 ((m->flags & PG_SWAPPED) && 3408 (prot & VM_PROT_WRITE) && 3409 (fault_flags & VM_FAULT_DIRTY))) { 3410 vm_page_wakeup(m); 3411 break; 3412 } 3413 3414 /* 3415 * Enter the page into the pmap. The object might be held 3416 * shared so we can't do any (serious) modifying operation 3417 * on it. 3418 */ 3419 if ((m->queue - m->pc) == PQ_CACHE) 3420 vm_page_deactivate(m); 3421 if (prot & VM_PROT_WRITE) { 3422 vm_object_set_writeable_dirty(m->object); 3423 vm_set_nosync(m, entry); 3424 if (fault_flags & VM_FAULT_DIRTY) { 3425 vm_page_dirty(m); 3426 /* can't happeen due to conditional above */ 3427 /* swap_pager_unswapped(m); */ 3428 } 3429 } 3430 pmap_enter(pmap, addr, m, prot, 0, entry); 3431 #if 0 3432 /* REMOVE ME, a burst counts as one fault */ 3433 mycpu->gd_cnt.v_vm_faults++; 3434 if (curthread->td_lwp) 3435 ++curthread->td_lwp->lwp_ru.ru_minflt; 3436 #endif 3437 vm_page_wakeup(m); 3438 } 3439 } 3440