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 * Does not apply to OBJT_MGTDEVICE (e.g. gpu / drm 1975 * subsystem). For OBJT_MGTDEVICE the pages are not 1976 * indexed in the VM object at all but instead directly 1977 * entered into the pmap. 1978 */ 1979 fs->mary[0] = NULL; 1980 if (fs->ba->object->type == OBJT_MGTDEVICE) 1981 goto readrest; 1982 1983 if (!vm_page_count_severe()) { 1984 fs->mary[0] = vm_page_alloc(fs->ba->object, 1985 pindex, 1986 ((fs->vp || fs->ba->backing_ba) ? 1987 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1988 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1989 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1990 } 1991 if (fs->mary[0] == NULL) { 1992 vm_object_pip_wakeup(fs->first_ba->object); 1993 unlock_things(fs); 1994 if (allow_nofault == 0 || 1995 (curthread->td_flags & TDF_NOFAULT) == 0) { 1996 thread_t td; 1997 1998 vm_wait_pfault(); 1999 td = curthread; 2000 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 2001 return (KERN_PROTECTION_FAILURE); 2002 } 2003 return (KERN_TRY_AGAIN); 2004 } 2005 2006 /* 2007 * Fall through to readrest. We have a new page which 2008 * will have to be paged (since m->valid will be 0). 2009 */ 2010 } 2011 2012 readrest: 2013 /* 2014 * We have found an invalid or partially valid page, a 2015 * page with a read-ahead mark which might be partially or 2016 * fully valid (and maybe dirty too), or we have allocated 2017 * a new page. 2018 * 2019 * Attempt to fault-in the page if there is a chance that the 2020 * pager has it, and potentially fault in additional pages 2021 * at the same time. 2022 * 2023 * If TRYPAGER is true then fs.mary[0] will be non-NULL and 2024 * busied for us. 2025 */ 2026 if (TRYPAGER(fs)) { 2027 u_char behavior = vm_map_entry_behavior(fs->entry); 2028 vm_object_t object; 2029 vm_page_t first_m; 2030 int seqaccess; 2031 int rv; 2032 2033 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 2034 seqaccess = 0; 2035 else 2036 seqaccess = -1; 2037 2038 /* 2039 * Doing I/O may synchronously insert additional 2040 * pages so we can't be shared at this point either. 2041 * 2042 * NOTE: We can't free fs->mary[0] here in the 2043 * allocated case (fs->ba != fs->first_ba) as 2044 * this would require an exclusively locked 2045 * VM object. 2046 */ 2047 if (fs->ba == fs->first_ba && fs->first_shared) { 2048 if (fs->mary[0]) { 2049 vm_page_deactivate(fs->mary[0]); 2050 vm_page_wakeup(fs->mary[0]); 2051 fs->mary[0]= NULL; 2052 } 2053 fs->first_shared = 0; 2054 vm_object_pip_wakeup(fs->first_ba->object); 2055 unlock_things(fs); 2056 return (KERN_TRY_AGAIN); 2057 } 2058 if (fs->ba != fs->first_ba && fs->shared) { 2059 if (fs->mary[0]) { 2060 vm_page_deactivate(fs->mary[0]); 2061 vm_page_wakeup(fs->mary[0]); 2062 fs->mary[0] = NULL; 2063 } 2064 fs->first_shared = 0; 2065 fs->shared = 0; 2066 vm_object_pip_wakeup(fs->first_ba->object); 2067 unlock_things(fs); 2068 return (KERN_TRY_AGAIN); 2069 } 2070 2071 object = fs->ba->object; 2072 first_m = NULL; 2073 2074 /* object is held, no more access to entry or ba's */ 2075 2076 /* 2077 * Acquire the page data. We still hold object 2078 * and the page has been BUSY's. 2079 * 2080 * We own the page, but we must re-issue the lookup 2081 * because the pager may have replaced it (for example, 2082 * in order to enter a fictitious page into the 2083 * object). In this situation the pager will have 2084 * cleaned up the old page and left the new one 2085 * busy for us. 2086 * 2087 * If we got here through a PG_RAM read-ahead 2088 * mark the page may be partially dirty and thus 2089 * not freeable. Don't bother checking to see 2090 * if the pager has the page because we can't free 2091 * it anyway. We have to depend on the get_page 2092 * operation filling in any gaps whether there is 2093 * backing store or not. 2094 * 2095 * We must dispose of the page (fs->mary[0]) and also 2096 * possibly first_m (the fronting layer). If 2097 * this is a write fault leave the page intact 2098 * because we will probably have to copy fs->mary[0] 2099 * to fs->first_m on the retry. If this is a 2100 * read fault we probably won't need the page. 2101 * 2102 * For OBJT_MGTDEVICE (and eventually all types), 2103 * fs->mary[0] is not pre-allocated and may be set 2104 * to a vm_page (busied for us) without being inserted 2105 * into the object. In this case we want to return 2106 * the vm_page directly so the caller can issue the 2107 * pmap_enter(). 2108 */ 2109 rv = vm_pager_get_page(object, pindex, 2110 &fs->mary[0], seqaccess); 2111 2112 if (rv == VM_PAGER_OK) { 2113 ++fs->hardfault; 2114 if (object->type == OBJT_MGTDEVICE) { 2115 break; 2116 } 2117 2118 fs->mary[0] = vm_page_lookup(object, pindex); 2119 if (fs->mary[0]) { 2120 vm_page_activate(fs->mary[0]); 2121 vm_page_wakeup(fs->mary[0]); 2122 fs->mary[0] = NULL; 2123 } 2124 2125 if (fs->mary[0]) { 2126 /* NOT REACHED */ 2127 /* have page */ 2128 break; 2129 } 2130 vm_object_pip_wakeup(fs->first_ba->object); 2131 unlock_things(fs); 2132 return (KERN_TRY_AGAIN); 2133 } 2134 2135 /* 2136 * If the pager doesn't have the page, continue on 2137 * to the next object. Retain the vm_page if this 2138 * is the first object, we may need to copy into 2139 * it later. 2140 */ 2141 if (rv == VM_PAGER_FAIL) { 2142 if (fs->ba != fs->first_ba) { 2143 if (fs->mary[0]) { 2144 vm_page_free(fs->mary[0]); 2145 fs->mary[0] = NULL; 2146 } 2147 } 2148 goto next; 2149 } 2150 2151 /* 2152 * Remove the bogus page (which does not exist at this 2153 * object/offset). 2154 * 2155 * Also wake up any other process that may want to bring 2156 * in this page. 2157 * 2158 * If this is the top-level object, we must leave the 2159 * busy page to prevent another process from rushing 2160 * past us, and inserting the page in that object at 2161 * the same time that we are. 2162 */ 2163 if (rv == VM_PAGER_ERROR) { 2164 if (curproc) { 2165 kprintf("vm_fault: pager read error, " 2166 "pid %d (%s)\n", 2167 curproc->p_pid, 2168 curproc->p_comm); 2169 } else { 2170 kprintf("vm_fault: pager read error, " 2171 "thread %p (%s)\n", 2172 curthread, 2173 curthread->td_comm); 2174 } 2175 } 2176 2177 /* 2178 * I/O error or data outside pager's range. 2179 */ 2180 if (fs->mary[0]) { 2181 vnode_pager_freepage(fs->mary[0]); 2182 fs->mary[0] = NULL; 2183 } 2184 if (first_m) { 2185 vm_page_free(first_m); 2186 first_m = NULL; /* safety */ 2187 } 2188 vm_object_pip_wakeup(object); 2189 unlock_things(fs); 2190 2191 switch(rv) { 2192 case VM_PAGER_ERROR: 2193 return (KERN_FAILURE); 2194 case VM_PAGER_BAD: 2195 return (KERN_PROTECTION_FAILURE); 2196 default: 2197 return (KERN_PROTECTION_FAILURE); 2198 } 2199 2200 #if 0 2201 /* 2202 * Data outside the range of the pager or an I/O error 2203 * 2204 * The page may have been wired during the pagein, 2205 * e.g. by the buffer cache, and cannot simply be 2206 * freed. Call vnode_pager_freepage() to deal with it. 2207 * 2208 * The object is not held shared so we can safely 2209 * free the page. 2210 */ 2211 if (fs->ba != fs->first_ba) { 2212 2213 /* 2214 * XXX - we cannot just fall out at this 2215 * point, m has been freed and is invalid! 2216 */ 2217 } 2218 2219 /* 2220 * XXX - the check for kernel_map is a kludge to work 2221 * around having the machine panic on a kernel space 2222 * fault w/ I/O error. 2223 */ 2224 if (((fs->map != &kernel_map) && 2225 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 2226 if (fs->m) { 2227 /* from just above */ 2228 KKASSERT(fs->first_shared == 0); 2229 vnode_pager_freepage(fs->m); 2230 fs->m = NULL; 2231 } 2232 /* NOT REACHED */ 2233 } 2234 #endif 2235 } 2236 2237 next: 2238 /* 2239 * We get here if the object has a default pager (or unwiring) 2240 * or the pager doesn't have the page. 2241 * 2242 * fs->first_m will be used for the COW unless we find a 2243 * deeper page to be mapped read-only, in which case the 2244 * unlock*(fs) will free first_m. 2245 */ 2246 if (fs->ba == fs->first_ba) 2247 fs->first_m = fs->mary[0]; 2248 2249 /* 2250 * Move on to the next object. The chain lock should prevent 2251 * the backing_object from getting ripped out from under us. 2252 * 2253 * The object lock for the next object is governed by 2254 * fs->shared. 2255 */ 2256 next_ba = fs->ba->backing_ba; 2257 if (next_ba == NULL) { 2258 /* 2259 * If there's no object left, fill the page in the top 2260 * object with zeros. 2261 */ 2262 if (fs->ba != fs->first_ba) { 2263 vm_object_pip_wakeup(fs->ba->object); 2264 vm_object_drop(fs->ba->object); 2265 fs->ba = fs->first_ba; 2266 pindex = first_pindex; 2267 fs->mary[0] = fs->first_m; 2268 } 2269 fs->first_m = NULL; 2270 2271 /* 2272 * Zero the page and mark it valid. 2273 */ 2274 vm_page_zero_fill(fs->mary[0]); 2275 mycpu->gd_cnt.v_zfod++; 2276 fs->mary[0]->valid = VM_PAGE_BITS_ALL; 2277 break; /* break to PAGE HAS BEEN FOUND */ 2278 } 2279 2280 if (fs->shared) 2281 vm_object_hold_shared(next_ba->object); 2282 else 2283 vm_object_hold(next_ba->object); 2284 KKASSERT(next_ba == fs->ba->backing_ba); 2285 pindex -= OFF_TO_IDX(fs->ba->offset); 2286 pindex += OFF_TO_IDX(next_ba->offset); 2287 2288 if (fs->ba != fs->first_ba) { 2289 vm_object_pip_wakeup(fs->ba->object); 2290 vm_object_lock_swap(); /* flip ba/next_ba */ 2291 vm_object_drop(fs->ba->object); 2292 } 2293 fs->ba = next_ba; 2294 vm_object_pip_add(next_ba->object, 1); 2295 } 2296 2297 /* 2298 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 2299 * is held.] 2300 * 2301 * object still held. 2302 * vm_map may not be locked (determined by fs->lookup_still_valid) 2303 * 2304 * local shared variable may be different from fs->shared. 2305 * 2306 * If the page is being written, but isn't already owned by the 2307 * top-level object, we have to copy it into a new page owned by the 2308 * top-level object. 2309 */ 2310 KASSERT((fs->mary[0]->busy_count & PBUSY_LOCKED) != 0, 2311 ("vm_fault: not busy after main loop")); 2312 2313 if (fs->ba != fs->first_ba) { 2314 /* 2315 * We only really need to copy if we want to write it. 2316 */ 2317 if (fault_type & VM_PROT_WRITE) { 2318 #if 0 2319 /* CODE REFACTOR IN PROGRESS, REMOVE OPTIMIZATION */ 2320 /* 2321 * This allows pages to be virtually copied from a 2322 * backing_object into the first_object, where the 2323 * backing object has no other refs to it, and cannot 2324 * gain any more refs. Instead of a bcopy, we just 2325 * move the page from the backing object to the 2326 * first object. Note that we must mark the page 2327 * dirty in the first object so that it will go out 2328 * to swap when needed. 2329 */ 2330 if (virtual_copy_ok(fs)) { 2331 /* 2332 * (first_m) and (m) are both busied. We have 2333 * move (m) into (first_m)'s object/pindex 2334 * in an atomic fashion, then free (first_m). 2335 * 2336 * first_object is held so second remove 2337 * followed by the rename should wind 2338 * up being atomic. vm_page_free() might 2339 * block so we don't do it until after the 2340 * rename. 2341 */ 2342 vm_page_protect(fs->first_m, VM_PROT_NONE); 2343 vm_page_remove(fs->first_m); 2344 vm_page_rename(fs->mary[0], 2345 fs->first_ba->object, 2346 first_pindex); 2347 vm_page_free(fs->first_m); 2348 fs->first_m = fs->mary[0]; 2349 fs->mary[0] = NULL; 2350 mycpu->gd_cnt.v_cow_optim++; 2351 } else 2352 #endif 2353 { 2354 /* 2355 * Oh, well, lets copy it. 2356 * 2357 * We used to unmap the original page here 2358 * because vm_fault_page() didn't and this 2359 * would cause havoc for the umtx*() code 2360 * and the procfs code. 2361 * 2362 * This is no longer necessary. The 2363 * vm_fault_page() routine will now unmap the 2364 * page after a COW, and the umtx code will 2365 * recover on its own. 2366 */ 2367 /* 2368 * NOTE: Since fs->mary[0] is a backing page, 2369 * it is read-only, so there isn't any 2370 * copy race vs writers. 2371 */ 2372 KKASSERT(fs->first_shared == 0); 2373 vm_page_copy(fs->mary[0], fs->first_m); 2374 /* pmap_remove_specific( 2375 &curthread->td_lwp->lwp_vmspace->vm_pmap, 2376 fs->mary[0]); */ 2377 } 2378 2379 /* 2380 * We no longer need the old page or object. 2381 */ 2382 if (fs->mary[0]) 2383 release_page(fs); 2384 2385 /* 2386 * fs->ba != fs->first_ba due to above conditional 2387 */ 2388 vm_object_pip_wakeup(fs->ba->object); 2389 vm_object_drop(fs->ba->object); 2390 fs->ba = fs->first_ba; 2391 2392 /* 2393 * Only use the new page below... 2394 */ 2395 mycpu->gd_cnt.v_cow_faults++; 2396 fs->mary[0] = fs->first_m; 2397 pindex = first_pindex; 2398 } else { 2399 /* 2400 * If it wasn't a write fault avoid having to copy 2401 * the page by mapping it read-only from backing 2402 * store. The process is not allowed to modify 2403 * backing pages. 2404 */ 2405 fs->prot &= ~VM_PROT_WRITE; 2406 } 2407 } 2408 2409 /* 2410 * Relock the map if necessary, then check the generation count. 2411 * relock_map() will update fs->timestamp to account for the 2412 * relocking if necessary. 2413 * 2414 * If the count has changed after relocking then all sorts of 2415 * crap may have happened and we have to retry. 2416 * 2417 * NOTE: The relock_map() can fail due to a deadlock against 2418 * the vm_page we are holding BUSY. 2419 */ 2420 KKASSERT(fs->lookup_still_valid != 0); 2421 #if 0 2422 if (fs->lookup_still_valid == 0 && fs->map) { 2423 if (relock_map(fs) || 2424 fs->map->timestamp != fs->map_generation) { 2425 release_page(fs); 2426 vm_object_pip_wakeup(fs->first_ba->object); 2427 unlock_things(fs); 2428 return (KERN_TRY_AGAIN); 2429 } 2430 } 2431 #endif 2432 2433 /* 2434 * If the fault is a write, we know that this page is being 2435 * written NOW so dirty it explicitly to save on pmap_is_modified() 2436 * calls later. 2437 * 2438 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 2439 * if the page is already dirty to prevent data written with 2440 * the expectation of being synced from not being synced. 2441 * Likewise if this entry does not request NOSYNC then make 2442 * sure the page isn't marked NOSYNC. Applications sharing 2443 * data should use the same flags to avoid ping ponging. 2444 * 2445 * Also tell the backing pager, if any, that it should remove 2446 * any swap backing since the page is now dirty. 2447 */ 2448 vm_page_activate(fs->mary[0]); 2449 if (fs->prot & VM_PROT_WRITE) { 2450 vm_object_set_writeable_dirty(fs->first_ba->object); 2451 vm_set_nosync(fs->mary[0], fs->entry); 2452 if (fs->fault_flags & VM_FAULT_DIRTY) { 2453 vm_page_dirty(fs->mary[0]); 2454 if (fs->mary[0]->flags & PG_SWAPPED) { 2455 /* 2456 * If the page is swapped out we have to call 2457 * swap_pager_unswapped() which requires an 2458 * exclusive object lock. If we are shared, 2459 * we must clear the shared flag and retry. 2460 */ 2461 if ((fs->ba == fs->first_ba && 2462 fs->first_shared) || 2463 (fs->ba != fs->first_ba && fs->shared)) { 2464 vm_page_wakeup(fs->mary[0]); 2465 fs->mary[0] = NULL; 2466 if (fs->ba == fs->first_ba) 2467 fs->first_shared = 0; 2468 else 2469 fs->shared = 0; 2470 vm_object_pip_wakeup( 2471 fs->first_ba->object); 2472 unlock_things(fs); 2473 return (KERN_TRY_AGAIN); 2474 } 2475 swap_pager_unswapped(fs->mary[0]); 2476 } 2477 } 2478 } 2479 2480 /* 2481 * We found our page at backing layer ba. Leave the layer state 2482 * intact. 2483 */ 2484 2485 vm_object_pip_wakeup(fs->first_ba->object); 2486 #if 0 2487 if (fs->ba != fs->first_ba) 2488 vm_object_drop(fs->ba->object); 2489 #endif 2490 2491 /* 2492 * Page had better still be busy. We are still locked up and 2493 * fs->ba->object will have another PIP reference for the case 2494 * where fs->ba != fs->first_ba. 2495 */ 2496 KASSERT(fs->mary[0]->busy_count & PBUSY_LOCKED, 2497 ("vm_fault: page %p not busy!", fs->mary[0])); 2498 2499 /* 2500 * Sanity check: page must be completely valid or it is not fit to 2501 * map into user space. vm_pager_get_pages() ensures this. 2502 */ 2503 if (fs->mary[0]->valid != VM_PAGE_BITS_ALL) { 2504 vm_page_zero_invalid(fs->mary[0], TRUE); 2505 kprintf("Warning: page %p partially invalid on fault\n", 2506 fs->mary[0]); 2507 } 2508 2509 return (KERN_SUCCESS); 2510 } 2511 2512 /* 2513 * Wire down a range of virtual addresses in a map. The entry in question 2514 * should be marked in-transition and the map must be locked. We must 2515 * release the map temporarily while faulting-in the page to avoid a 2516 * deadlock. Note that the entry may be clipped while we are blocked but 2517 * will never be freed. 2518 * 2519 * map must be locked on entry. 2520 */ 2521 int 2522 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, 2523 boolean_t user_wire, int kmflags) 2524 { 2525 boolean_t fictitious; 2526 vm_offset_t start; 2527 vm_offset_t end; 2528 vm_offset_t va; 2529 pmap_t pmap; 2530 int rv; 2531 int wire_prot; 2532 int fault_flags; 2533 vm_page_t m; 2534 2535 if (user_wire) { 2536 wire_prot = VM_PROT_READ; 2537 fault_flags = VM_FAULT_USER_WIRE; 2538 } else { 2539 wire_prot = VM_PROT_READ | VM_PROT_WRITE; 2540 fault_flags = VM_FAULT_CHANGE_WIRING; 2541 } 2542 if (kmflags & KM_NOTLBSYNC) 2543 wire_prot |= VM_PROT_NOSYNC; 2544 2545 pmap = vm_map_pmap(map); 2546 start = entry->ba.start; 2547 end = entry->ba.end; 2548 2549 switch(entry->maptype) { 2550 case VM_MAPTYPE_NORMAL: 2551 case VM_MAPTYPE_VPAGETABLE: 2552 fictitious = entry->ba.object && 2553 ((entry->ba.object->type == OBJT_DEVICE) || 2554 (entry->ba.object->type == OBJT_MGTDEVICE)); 2555 break; 2556 case VM_MAPTYPE_UKSMAP: 2557 fictitious = TRUE; 2558 break; 2559 default: 2560 fictitious = FALSE; 2561 break; 2562 } 2563 2564 if (entry->eflags & MAP_ENTRY_KSTACK) 2565 start += PAGE_SIZE; 2566 map->timestamp++; 2567 vm_map_unlock(map); 2568 2569 /* 2570 * We simulate a fault to get the page and enter it in the physical 2571 * map. 2572 */ 2573 for (va = start; va < end; va += PAGE_SIZE) { 2574 rv = vm_fault(map, va, wire_prot, fault_flags); 2575 if (rv) { 2576 while (va > start) { 2577 va -= PAGE_SIZE; 2578 m = pmap_unwire(pmap, va); 2579 if (m && !fictitious) { 2580 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2581 vm_page_unwire(m, 1); 2582 vm_page_wakeup(m); 2583 } 2584 } 2585 goto done; 2586 } 2587 } 2588 rv = KERN_SUCCESS; 2589 done: 2590 vm_map_lock(map); 2591 2592 return (rv); 2593 } 2594 2595 /* 2596 * Unwire a range of virtual addresses in a map. The map should be 2597 * locked. 2598 */ 2599 void 2600 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 2601 { 2602 boolean_t fictitious; 2603 vm_offset_t start; 2604 vm_offset_t end; 2605 vm_offset_t va; 2606 pmap_t pmap; 2607 vm_page_t m; 2608 2609 pmap = vm_map_pmap(map); 2610 start = entry->ba.start; 2611 end = entry->ba.end; 2612 fictitious = entry->ba.object && 2613 ((entry->ba.object->type == OBJT_DEVICE) || 2614 (entry->ba.object->type == OBJT_MGTDEVICE)); 2615 if (entry->eflags & MAP_ENTRY_KSTACK) 2616 start += PAGE_SIZE; 2617 2618 /* 2619 * Since the pages are wired down, we must be able to get their 2620 * mappings from the physical map system. 2621 */ 2622 for (va = start; va < end; va += PAGE_SIZE) { 2623 m = pmap_unwire(pmap, va); 2624 if (m && !fictitious) { 2625 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2626 vm_page_unwire(m, 1); 2627 vm_page_wakeup(m); 2628 } 2629 } 2630 } 2631 2632 /* 2633 * Simulate write faults to bring all data into the head object, return 2634 * KERN_SUCCESS on success (which should be always unless the system runs 2635 * out of memory). 2636 * 2637 * The caller will handle destroying the backing_ba's. 2638 */ 2639 int 2640 vm_fault_collapse(vm_map_t map, vm_map_entry_t entry) 2641 { 2642 struct faultstate fs; 2643 vm_ooffset_t scan; 2644 vm_pindex_t pindex; 2645 vm_object_t object; 2646 int rv; 2647 int all_shadowed; 2648 2649 bzero(&fs, sizeof(fs)); 2650 object = entry->ba.object; 2651 2652 fs.first_prot = entry->max_protection | /* optional VM_PROT_EXECUTE */ 2653 VM_PROT_READ | VM_PROT_WRITE | VM_PROT_OVERRIDE_WRITE; 2654 fs.fault_flags = VM_FAULT_NORMAL; 2655 fs.map = map; 2656 fs.entry = entry; 2657 fs.lookup_still_valid = -1; /* leave map atomically locked */ 2658 fs.first_ba = &entry->ba; 2659 fs.first_ba_held = -1; /* leave object held */ 2660 2661 /* fs.hardfault */ 2662 2663 vm_object_hold(object); 2664 rv = KERN_SUCCESS; 2665 2666 scan = entry->ba.start; 2667 all_shadowed = 1; 2668 2669 while (scan < entry->ba.end) { 2670 pindex = OFF_TO_IDX(entry->ba.offset + (scan - entry->ba.start)); 2671 2672 if (vm_page_lookup(object, pindex)) { 2673 scan += PAGE_SIZE; 2674 continue; 2675 } 2676 2677 all_shadowed = 0; 2678 fs.ba = fs.first_ba; 2679 fs.prot = fs.first_prot; 2680 2681 rv = vm_fault_object(&fs, pindex, fs.first_prot, 1); 2682 if (rv == KERN_TRY_AGAIN) 2683 continue; 2684 if (rv != KERN_SUCCESS) 2685 break; 2686 vm_page_flag_set(fs.mary[0], PG_REFERENCED); 2687 vm_page_activate(fs.mary[0]); 2688 vm_page_wakeup(fs.mary[0]); 2689 scan += PAGE_SIZE; 2690 } 2691 KKASSERT(entry->ba.object == object); 2692 vm_object_drop(object); 2693 2694 /* 2695 * If the fronting object did not have every page we have to clear 2696 * the pmap range due to the pages being changed so we can fault-in 2697 * the proper pages. 2698 */ 2699 if (all_shadowed == 0) 2700 pmap_remove(map->pmap, entry->ba.start, entry->ba.end); 2701 2702 return rv; 2703 } 2704 2705 /* 2706 * Copy all of the pages from one map entry to another. If the source 2707 * is wired down we just use vm_page_lookup(). If not we use 2708 * vm_fault_object(). 2709 * 2710 * The source and destination maps must be locked for write. 2711 * The source and destination maps token must be held 2712 * 2713 * No other requirements. 2714 * 2715 * XXX do segment optimization 2716 */ 2717 void 2718 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2719 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2720 { 2721 vm_object_t dst_object; 2722 vm_object_t src_object; 2723 vm_ooffset_t dst_offset; 2724 vm_ooffset_t src_offset; 2725 vm_prot_t prot; 2726 vm_offset_t vaddr; 2727 vm_page_t dst_m; 2728 vm_page_t src_m; 2729 2730 src_object = src_entry->ba.object; 2731 src_offset = src_entry->ba.offset; 2732 2733 /* 2734 * Create the top-level object for the destination entry. (Doesn't 2735 * actually shadow anything - we copy the pages directly.) 2736 */ 2737 vm_map_entry_allocate_object(dst_entry); 2738 dst_object = dst_entry->ba.object; 2739 2740 prot = dst_entry->max_protection; 2741 2742 /* 2743 * Loop through all of the pages in the entry's range, copying each 2744 * one from the source object (it should be there) to the destination 2745 * object. 2746 */ 2747 vm_object_hold(src_object); 2748 vm_object_hold(dst_object); 2749 2750 for (vaddr = dst_entry->ba.start, dst_offset = 0; 2751 vaddr < dst_entry->ba.end; 2752 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2753 2754 /* 2755 * Allocate a page in the destination object 2756 */ 2757 do { 2758 dst_m = vm_page_alloc(dst_object, 2759 OFF_TO_IDX(dst_offset), 2760 VM_ALLOC_NORMAL); 2761 if (dst_m == NULL) { 2762 vm_wait(0); 2763 } 2764 } while (dst_m == NULL); 2765 2766 /* 2767 * Find the page in the source object, and copy it in. 2768 * (Because the source is wired down, the page will be in 2769 * memory.) 2770 */ 2771 src_m = vm_page_lookup(src_object, 2772 OFF_TO_IDX(dst_offset + src_offset)); 2773 if (src_m == NULL) 2774 panic("vm_fault_copy_wired: page missing"); 2775 2776 vm_page_copy(src_m, dst_m); 2777 2778 /* 2779 * Enter it in the pmap... 2780 */ 2781 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2782 2783 /* 2784 * Mark it no longer busy, and put it on the active list. 2785 */ 2786 vm_page_activate(dst_m); 2787 vm_page_wakeup(dst_m); 2788 } 2789 vm_object_drop(dst_object); 2790 vm_object_drop(src_object); 2791 } 2792 2793 #if 0 2794 2795 /* 2796 * This routine checks around the requested page for other pages that 2797 * might be able to be faulted in. This routine brackets the viable 2798 * pages for the pages to be paged in. 2799 * 2800 * Inputs: 2801 * m, rbehind, rahead 2802 * 2803 * Outputs: 2804 * marray (array of vm_page_t), reqpage (index of requested page) 2805 * 2806 * Return value: 2807 * number of pages in marray 2808 */ 2809 static int 2810 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2811 vm_page_t *marray, int *reqpage) 2812 { 2813 int i,j; 2814 vm_object_t object; 2815 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2816 vm_page_t rtm; 2817 int cbehind, cahead; 2818 2819 object = m->object; 2820 pindex = m->pindex; 2821 2822 /* 2823 * we don't fault-ahead for device pager 2824 */ 2825 if ((object->type == OBJT_DEVICE) || 2826 (object->type == OBJT_MGTDEVICE)) { 2827 *reqpage = 0; 2828 marray[0] = m; 2829 return 1; 2830 } 2831 2832 /* 2833 * if the requested page is not available, then give up now 2834 */ 2835 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2836 *reqpage = 0; /* not used by caller, fix compiler warn */ 2837 return 0; 2838 } 2839 2840 if ((cbehind == 0) && (cahead == 0)) { 2841 *reqpage = 0; 2842 marray[0] = m; 2843 return 1; 2844 } 2845 2846 if (rahead > cahead) { 2847 rahead = cahead; 2848 } 2849 2850 if (rbehind > cbehind) { 2851 rbehind = cbehind; 2852 } 2853 2854 /* 2855 * Do not do any readahead if we have insufficient free memory. 2856 * 2857 * XXX code was broken disabled before and has instability 2858 * with this conditonal fixed, so shortcut for now. 2859 */ 2860 if (burst_fault == 0 || vm_page_count_severe()) { 2861 marray[0] = m; 2862 *reqpage = 0; 2863 return 1; 2864 } 2865 2866 /* 2867 * scan backward for the read behind pages -- in memory 2868 * 2869 * Assume that if the page is not found an interrupt will not 2870 * create it. Theoretically interrupts can only remove (busy) 2871 * pages, not create new associations. 2872 */ 2873 if (pindex > 0) { 2874 if (rbehind > pindex) { 2875 rbehind = pindex; 2876 startpindex = 0; 2877 } else { 2878 startpindex = pindex - rbehind; 2879 } 2880 2881 vm_object_hold(object); 2882 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2883 if (vm_page_lookup(object, tpindex - 1)) 2884 break; 2885 } 2886 2887 i = 0; 2888 while (tpindex < pindex) { 2889 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2890 VM_ALLOC_NULL_OK); 2891 if (rtm == NULL) { 2892 for (j = 0; j < i; j++) { 2893 vm_page_free(marray[j]); 2894 } 2895 vm_object_drop(object); 2896 marray[0] = m; 2897 *reqpage = 0; 2898 return 1; 2899 } 2900 marray[i] = rtm; 2901 ++i; 2902 ++tpindex; 2903 } 2904 vm_object_drop(object); 2905 } else { 2906 i = 0; 2907 } 2908 2909 /* 2910 * Assign requested page 2911 */ 2912 marray[i] = m; 2913 *reqpage = i; 2914 ++i; 2915 2916 /* 2917 * Scan forwards for read-ahead pages 2918 */ 2919 tpindex = pindex + 1; 2920 endpindex = tpindex + rahead; 2921 if (endpindex > object->size) 2922 endpindex = object->size; 2923 2924 vm_object_hold(object); 2925 while (tpindex < endpindex) { 2926 if (vm_page_lookup(object, tpindex)) 2927 break; 2928 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2929 VM_ALLOC_NULL_OK); 2930 if (rtm == NULL) 2931 break; 2932 marray[i] = rtm; 2933 ++i; 2934 ++tpindex; 2935 } 2936 vm_object_drop(object); 2937 2938 return (i); 2939 } 2940 2941 #endif 2942 2943 /* 2944 * vm_prefault() provides a quick way of clustering pagefaults into a 2945 * processes address space. It is a "cousin" of pmap_object_init_pt, 2946 * except it runs at page fault time instead of mmap time. 2947 * 2948 * vm.fast_fault Enables pre-faulting zero-fill pages 2949 * 2950 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2951 * prefault. Scan stops in either direction when 2952 * a page is found to already exist. 2953 * 2954 * This code used to be per-platform pmap_prefault(). It is now 2955 * machine-independent and enhanced to also pre-fault zero-fill pages 2956 * (see vm.fast_fault) as well as make them writable, which greatly 2957 * reduces the number of page faults programs incur. 2958 * 2959 * Application performance when pre-faulting zero-fill pages is heavily 2960 * dependent on the application. Very tiny applications like /bin/echo 2961 * lose a little performance while applications of any appreciable size 2962 * gain performance. Prefaulting multiple pages also reduces SMP 2963 * congestion and can improve SMP performance significantly. 2964 * 2965 * NOTE! prot may allow writing but this only applies to the top level 2966 * object. If we wind up mapping a page extracted from a backing 2967 * object we have to make sure it is read-only. 2968 * 2969 * NOTE! The caller has already handled any COW operations on the 2970 * vm_map_entry via the normal fault code. Do NOT call this 2971 * shortcut unless the normal fault code has run on this entry. 2972 * 2973 * The related map must be locked. 2974 * No other requirements. 2975 */ 2976 __read_mostly static int vm_prefault_pages = 8; 2977 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2978 "Maximum number of pages to pre-fault"); 2979 __read_mostly static int vm_fast_fault = 1; 2980 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2981 "Burst fault zero-fill regions"); 2982 2983 /* 2984 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2985 * is not already dirty by other means. This will prevent passive 2986 * filesystem syncing as well as 'sync' from writing out the page. 2987 */ 2988 static void 2989 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2990 { 2991 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2992 if (m->dirty == 0) 2993 vm_page_flag_set(m, PG_NOSYNC); 2994 } else { 2995 vm_page_flag_clear(m, PG_NOSYNC); 2996 } 2997 } 2998 2999 static void 3000 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 3001 int fault_flags) 3002 { 3003 vm_map_backing_t ba; /* first ba */ 3004 struct lwp *lp; 3005 vm_page_t m; 3006 vm_offset_t addr; 3007 vm_pindex_t index; 3008 vm_pindex_t pindex; 3009 vm_object_t object; 3010 int pprot; 3011 int i; 3012 int noneg; 3013 int nopos; 3014 int maxpages; 3015 3016 /* 3017 * Get stable max count value, disabled if set to 0 3018 */ 3019 maxpages = vm_prefault_pages; 3020 cpu_ccfence(); 3021 if (maxpages <= 0) 3022 return; 3023 3024 /* 3025 * We do not currently prefault mappings that use virtual page 3026 * tables. We do not prefault foreign pmaps. 3027 */ 3028 if (entry->maptype != VM_MAPTYPE_NORMAL) 3029 return; 3030 lp = curthread->td_lwp; 3031 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 3032 return; 3033 3034 /* 3035 * Limit pre-fault count to 1024 pages. 3036 */ 3037 if (maxpages > 1024) 3038 maxpages = 1024; 3039 3040 ba = &entry->ba; 3041 object = entry->ba.object; 3042 KKASSERT(object != NULL); 3043 3044 /* 3045 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively 3046 * now (or do something more complex XXX). 3047 */ 3048 vm_object_hold(object); 3049 3050 noneg = 0; 3051 nopos = 0; 3052 for (i = 0; i < maxpages; ++i) { 3053 vm_object_t lobject; 3054 vm_object_t nobject; 3055 vm_map_backing_t last_ba; /* last ba */ 3056 vm_map_backing_t next_ba; /* last ba */ 3057 int allocated = 0; 3058 int error; 3059 3060 /* 3061 * This can eat a lot of time on a heavily contended 3062 * machine so yield on the tick if needed. 3063 */ 3064 if ((i & 7) == 7) 3065 lwkt_yield(); 3066 3067 /* 3068 * Calculate the page to pre-fault, stopping the scan in 3069 * each direction separately if the limit is reached. 3070 */ 3071 if (i & 1) { 3072 if (noneg) 3073 continue; 3074 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 3075 } else { 3076 if (nopos) 3077 continue; 3078 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 3079 } 3080 if (addr < entry->ba.start) { 3081 noneg = 1; 3082 if (noneg && nopos) 3083 break; 3084 continue; 3085 } 3086 if (addr >= entry->ba.end) { 3087 nopos = 1; 3088 if (noneg && nopos) 3089 break; 3090 continue; 3091 } 3092 3093 /* 3094 * Skip pages already mapped, and stop scanning in that 3095 * direction. When the scan terminates in both directions 3096 * we are done. 3097 */ 3098 if (pmap_prefault_ok(pmap, addr) == 0) { 3099 if (i & 1) 3100 noneg = 1; 3101 else 3102 nopos = 1; 3103 if (noneg && nopos) 3104 break; 3105 continue; 3106 } 3107 3108 /* 3109 * Follow the backing layers to obtain the page to be mapped 3110 * into the pmap. 3111 * 3112 * If we reach the terminal object without finding a page 3113 * and we determine it would be advantageous, then allocate 3114 * a zero-fill page for the base object. The base object 3115 * is guaranteed to be OBJT_DEFAULT for this case. 3116 * 3117 * In order to not have to check the pager via *haspage*() 3118 * we stop if any non-default object is encountered. e.g. 3119 * a vnode or swap object would stop the loop. 3120 */ 3121 index = ((addr - entry->ba.start) + entry->ba.offset) >> 3122 PAGE_SHIFT; 3123 last_ba = ba; 3124 lobject = object; 3125 pindex = index; 3126 pprot = prot; 3127 3128 /*vm_object_hold(lobject); implied */ 3129 3130 while ((m = vm_page_lookup_busy_try(lobject, pindex, 3131 TRUE, &error)) == NULL) { 3132 if (lobject->type != OBJT_DEFAULT) 3133 break; 3134 if ((next_ba = last_ba->backing_ba) == NULL) { 3135 if (vm_fast_fault == 0) 3136 break; 3137 if ((prot & VM_PROT_WRITE) == 0 || 3138 vm_page_count_min(0)) { 3139 break; 3140 } 3141 3142 /* 3143 * NOTE: Allocated from base object 3144 */ 3145 m = vm_page_alloc(object, index, 3146 VM_ALLOC_NORMAL | 3147 VM_ALLOC_ZERO | 3148 VM_ALLOC_USE_GD | 3149 VM_ALLOC_NULL_OK); 3150 if (m == NULL) 3151 break; 3152 allocated = 1; 3153 pprot = prot; 3154 /* lobject = object .. not needed */ 3155 break; 3156 } 3157 if (next_ba->offset & PAGE_MASK) 3158 break; 3159 nobject = next_ba->object; 3160 vm_object_hold(nobject); 3161 pindex -= last_ba->offset >> PAGE_SHIFT; 3162 pindex += next_ba->offset >> PAGE_SHIFT; 3163 if (last_ba != ba) { 3164 vm_object_lock_swap(); 3165 vm_object_drop(lobject); 3166 } 3167 lobject = nobject; 3168 last_ba = next_ba; 3169 pprot &= ~VM_PROT_WRITE; 3170 } 3171 3172 /* 3173 * NOTE: A non-NULL (m) will be associated with lobject if 3174 * it was found there, otherwise it is probably a 3175 * zero-fill page associated with the base object. 3176 * 3177 * Give-up if no page is available. 3178 */ 3179 if (m == NULL) { 3180 if (last_ba != ba) 3181 vm_object_drop(lobject); 3182 break; 3183 } 3184 3185 /* 3186 * The object must be marked dirty if we are mapping a 3187 * writable page. Note that (m) does not have to be 3188 * entered into the object, so use lobject or object 3189 * as appropriate instead of m->object. 3190 * 3191 * Do this before we potentially drop the object. 3192 */ 3193 if (pprot & VM_PROT_WRITE) { 3194 vm_object_set_writeable_dirty( 3195 (allocated ? object : lobject)); 3196 } 3197 3198 /* 3199 * Do not conditionalize on PG_RAM. If pages are present in 3200 * the VM system we assume optimal caching. If caching is 3201 * not optimal the I/O gravy train will be restarted when we 3202 * hit an unavailable page. We do not want to try to restart 3203 * the gravy train now because we really don't know how much 3204 * of the object has been cached. The cost for restarting 3205 * the gravy train should be low (since accesses will likely 3206 * be I/O bound anyway). 3207 */ 3208 if (last_ba != ba) 3209 vm_object_drop(lobject); 3210 3211 /* 3212 * Enter the page into the pmap if appropriate. If we had 3213 * allocated the page we have to place it on a queue. If not 3214 * we just have to make sure it isn't on the cache queue 3215 * (pages on the cache queue are not allowed to be mapped). 3216 * 3217 * When allocated is TRUE, m corresponds to object, 3218 * not lobject. 3219 */ 3220 if (allocated) { 3221 /* 3222 * Page must be zerod. 3223 */ 3224 vm_page_zero_fill(m); 3225 mycpu->gd_cnt.v_zfod++; 3226 m->valid = VM_PAGE_BITS_ALL; 3227 3228 /* 3229 * Handle dirty page case 3230 */ 3231 if (pprot & VM_PROT_WRITE) 3232 vm_set_nosync(m, entry); 3233 pmap_enter(pmap, addr, m, pprot, 0, entry); 3234 #if 0 3235 /* REMOVE ME, a burst counts as one fault */ 3236 mycpu->gd_cnt.v_vm_faults++; 3237 if (curthread->td_lwp) 3238 ++curthread->td_lwp->lwp_ru.ru_minflt; 3239 #endif 3240 vm_page_deactivate(m); 3241 if (pprot & VM_PROT_WRITE) { 3242 /*vm_object_set_writeable_dirty(object);*/ 3243 vm_set_nosync(m, entry); 3244 if (fault_flags & VM_FAULT_DIRTY) { 3245 vm_page_dirty(m); 3246 /*XXX*/ 3247 swap_pager_unswapped(m); 3248 } 3249 } 3250 vm_page_wakeup(m); 3251 } else if (error) { 3252 /* couldn't busy page, no wakeup */ 3253 } else if ( 3254 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 3255 (m->flags & PG_FICTITIOUS) == 0) { 3256 /* 3257 * A fully valid page not undergoing soft I/O can 3258 * be immediately entered into the pmap. 3259 * 3260 * When allocated is false, m corresponds to lobject. 3261 */ 3262 if ((m->queue - m->pc) == PQ_CACHE) 3263 vm_page_deactivate(m); 3264 if (pprot & VM_PROT_WRITE) { 3265 /*vm_object_set_writeable_dirty(lobject);*/ 3266 vm_set_nosync(m, entry); 3267 if (fault_flags & VM_FAULT_DIRTY) { 3268 vm_page_dirty(m); 3269 /*XXX*/ 3270 swap_pager_unswapped(m); 3271 } 3272 } 3273 if (pprot & VM_PROT_WRITE) 3274 vm_set_nosync(m, entry); 3275 pmap_enter(pmap, addr, m, pprot, 0, entry); 3276 #if 0 3277 /* REMOVE ME, a burst counts as one fault */ 3278 mycpu->gd_cnt.v_vm_faults++; 3279 if (curthread->td_lwp) 3280 ++curthread->td_lwp->lwp_ru.ru_minflt; 3281 #endif 3282 vm_page_wakeup(m); 3283 } else { 3284 vm_page_wakeup(m); 3285 } 3286 } 3287 vm_object_drop(object); 3288 } 3289 3290 /* 3291 * Object can be held shared 3292 */ 3293 static void 3294 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 3295 vm_map_entry_t entry, int prot, int fault_flags) 3296 { 3297 struct lwp *lp; 3298 vm_page_t m; 3299 vm_offset_t addr; 3300 vm_pindex_t pindex; 3301 vm_object_t object; 3302 int i; 3303 int noneg; 3304 int nopos; 3305 int maxpages; 3306 3307 /* 3308 * Get stable max count value, disabled if set to 0 3309 */ 3310 maxpages = vm_prefault_pages; 3311 cpu_ccfence(); 3312 if (maxpages <= 0) 3313 return; 3314 3315 /* 3316 * We do not currently prefault mappings that use virtual page 3317 * tables. We do not prefault foreign pmaps. 3318 */ 3319 if (entry->maptype != VM_MAPTYPE_NORMAL) 3320 return; 3321 lp = curthread->td_lwp; 3322 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 3323 return; 3324 object = entry->ba.object; 3325 if (entry->ba.backing_ba != NULL) 3326 return; 3327 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 3328 3329 /* 3330 * Limit pre-fault count to 1024 pages. 3331 */ 3332 if (maxpages > 1024) 3333 maxpages = 1024; 3334 3335 noneg = 0; 3336 nopos = 0; 3337 for (i = 0; i < maxpages; ++i) { 3338 int error; 3339 3340 /* 3341 * Calculate the page to pre-fault, stopping the scan in 3342 * each direction separately if the limit is reached. 3343 */ 3344 if (i & 1) { 3345 if (noneg) 3346 continue; 3347 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 3348 } else { 3349 if (nopos) 3350 continue; 3351 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 3352 } 3353 if (addr < entry->ba.start) { 3354 noneg = 1; 3355 if (noneg && nopos) 3356 break; 3357 continue; 3358 } 3359 if (addr >= entry->ba.end) { 3360 nopos = 1; 3361 if (noneg && nopos) 3362 break; 3363 continue; 3364 } 3365 3366 /* 3367 * Follow the VM object chain to obtain the page to be mapped 3368 * into the pmap. This version of the prefault code only 3369 * works with terminal objects. 3370 * 3371 * The page must already exist. If we encounter a problem 3372 * we stop here. 3373 * 3374 * WARNING! We cannot call swap_pager_unswapped() or insert 3375 * a new vm_page with a shared token. 3376 */ 3377 pindex = ((addr - entry->ba.start) + entry->ba.offset) >> 3378 PAGE_SHIFT; 3379 3380 /* 3381 * Skip pages already mapped, and stop scanning in that 3382 * direction. When the scan terminates in both directions 3383 * we are done. 3384 */ 3385 if (pmap_prefault_ok(pmap, addr) == 0) { 3386 if (i & 1) 3387 noneg = 1; 3388 else 3389 nopos = 1; 3390 if (noneg && nopos) 3391 break; 3392 continue; 3393 } 3394 3395 /* 3396 * Shortcut the read-only mapping case using the far more 3397 * efficient vm_page_lookup_sbusy_try() function. This 3398 * allows us to acquire the page soft-busied only which 3399 * is especially nice for concurrent execs of the same 3400 * program. 3401 * 3402 * The lookup function also validates page suitability 3403 * (all valid bits set, and not fictitious). 3404 * 3405 * If the page is in PQ_CACHE we have to fall-through 3406 * and hard-busy it so we can move it out of PQ_CACHE. 3407 */ 3408 if ((prot & VM_PROT_WRITE) == 0) { 3409 m = vm_page_lookup_sbusy_try(object, pindex, 3410 0, PAGE_SIZE); 3411 if (m == NULL) 3412 break; 3413 if ((m->queue - m->pc) != PQ_CACHE) { 3414 pmap_enter(pmap, addr, m, prot, 0, entry); 3415 #if 0 3416 /* REMOVE ME, a burst counts as one fault */ 3417 mycpu->gd_cnt.v_vm_faults++; 3418 if (curthread->td_lwp) 3419 ++curthread->td_lwp->lwp_ru.ru_minflt; 3420 #endif 3421 vm_page_sbusy_drop(m); 3422 continue; 3423 } 3424 vm_page_sbusy_drop(m); 3425 } 3426 3427 /* 3428 * Fallback to normal vm_page lookup code. This code 3429 * hard-busies the page. Not only that, but the page 3430 * can remain in that state for a significant period 3431 * time due to pmap_enter()'s overhead. 3432 */ 3433 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3434 if (m == NULL || error) 3435 break; 3436 3437 /* 3438 * Stop if the page cannot be trivially entered into the 3439 * pmap. 3440 */ 3441 if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) || 3442 (m->flags & PG_FICTITIOUS) || 3443 ((m->flags & PG_SWAPPED) && 3444 (prot & VM_PROT_WRITE) && 3445 (fault_flags & VM_FAULT_DIRTY))) { 3446 vm_page_wakeup(m); 3447 break; 3448 } 3449 3450 /* 3451 * Enter the page into the pmap. The object might be held 3452 * shared so we can't do any (serious) modifying operation 3453 * on it. 3454 */ 3455 if ((m->queue - m->pc) == PQ_CACHE) 3456 vm_page_deactivate(m); 3457 if (prot & VM_PROT_WRITE) { 3458 vm_object_set_writeable_dirty(m->object); 3459 vm_set_nosync(m, entry); 3460 if (fault_flags & VM_FAULT_DIRTY) { 3461 vm_page_dirty(m); 3462 /* can't happeen due to conditional above */ 3463 /* swap_pager_unswapped(m); */ 3464 } 3465 } 3466 pmap_enter(pmap, addr, m, prot, 0, entry); 3467 #if 0 3468 /* REMOVE ME, a burst counts as one fault */ 3469 mycpu->gd_cnt.v_vm_faults++; 3470 if (curthread->td_lwp) 3471 ++curthread->td_lwp->lwp_ru.ru_minflt; 3472 #endif 3473 vm_page_wakeup(m); 3474 } 3475 } 3476