1 /* 2 * Copyright (c) 2003-2014 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * --- 35 * 36 * Copyright (c) 1991, 1993 37 * The Regents of the University of California. All rights reserved. 38 * Copyright (c) 1994 John S. Dyson 39 * All rights reserved. 40 * Copyright (c) 1994 David Greenman 41 * All rights reserved. 42 * 43 * 44 * This code is derived from software contributed to Berkeley by 45 * The Mach Operating System project at Carnegie-Mellon University. 46 * 47 * Redistribution and use in source and binary forms, with or without 48 * modification, are permitted provided that the following conditions 49 * are met: 50 * 1. Redistributions of source code must retain the above copyright 51 * notice, this list of conditions and the following disclaimer. 52 * 2. Redistributions in binary form must reproduce the above copyright 53 * notice, this list of conditions and the following disclaimer in the 54 * documentation and/or other materials provided with the distribution. 55 * 3. Neither the name of the University nor the names of its contributors 56 * may be used to endorse or promote products derived from this software 57 * without specific prior written permission. 58 * 59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 69 * SUCH DAMAGE. 70 * 71 * --- 72 * 73 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 74 * All rights reserved. 75 * 76 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 77 * 78 * Permission to use, copy, modify and distribute this software and 79 * its documentation is hereby granted, provided that both the copyright 80 * notice and this permission notice appear in all copies of the 81 * software, derivative works or modified versions, and any portions 82 * thereof, and that both notices appear in supporting documentation. 83 * 84 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 85 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 86 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 87 * 88 * Carnegie Mellon requests users of this software to return to 89 * 90 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 91 * School of Computer Science 92 * Carnegie Mellon University 93 * Pittsburgh PA 15213-3890 94 * 95 * any improvements or extensions that they make and grant Carnegie the 96 * rights to redistribute these changes. 97 */ 98 99 /* 100 * Page fault handling module. 101 */ 102 103 #include <sys/param.h> 104 #include <sys/systm.h> 105 #include <sys/kernel.h> 106 #include <sys/proc.h> 107 #include <sys/vnode.h> 108 #include <sys/resourcevar.h> 109 #include <sys/vmmeter.h> 110 #include <sys/vkernel.h> 111 #include <sys/lock.h> 112 #include <sys/sysctl.h> 113 114 #include <cpu/lwbuf.h> 115 116 #include <vm/vm.h> 117 #include <vm/vm_param.h> 118 #include <vm/pmap.h> 119 #include <vm/vm_map.h> 120 #include <vm/vm_object.h> 121 #include <vm/vm_page.h> 122 #include <vm/vm_pageout.h> 123 #include <vm/vm_kern.h> 124 #include <vm/vm_pager.h> 125 #include <vm/vnode_pager.h> 126 #include <vm/vm_extern.h> 127 128 #include <sys/thread2.h> 129 #include <vm/vm_page2.h> 130 131 struct faultstate { 132 vm_page_t m; 133 vm_object_t object; 134 vm_pindex_t pindex; 135 vm_prot_t prot; 136 vm_page_t first_m; 137 vm_object_t first_object; 138 vm_prot_t first_prot; 139 vm_map_t map; 140 vm_map_entry_t entry; 141 int lookup_still_valid; 142 int hardfault; 143 int fault_flags; 144 int map_generation; 145 int shared; 146 int first_shared; 147 boolean_t wired; 148 struct vnode *vp; 149 }; 150 151 static int debug_fault = 0; 152 SYSCTL_INT(_vm, OID_AUTO, debug_fault, CTLFLAG_RW, &debug_fault, 0, ""); 153 static int debug_cluster = 0; 154 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, ""); 155 int vm_shared_fault = 1; 156 TUNABLE_INT("vm.shared_fault", &vm_shared_fault); 157 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, &vm_shared_fault, 0, 158 "Allow shared token on vm_object"); 159 160 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int); 161 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, 162 vpte_t, int, int); 163 #if 0 164 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 165 #endif 166 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry); 167 static void vm_prefault(pmap_t pmap, vm_offset_t addra, 168 vm_map_entry_t entry, int prot, int fault_flags); 169 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 170 vm_map_entry_t entry, int prot, int fault_flags); 171 172 static __inline void 173 release_page(struct faultstate *fs) 174 { 175 vm_page_deactivate(fs->m); 176 vm_page_wakeup(fs->m); 177 fs->m = NULL; 178 } 179 180 /* 181 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse 182 * requires relocking and then checking the timestamp. 183 * 184 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do 185 * not have to update fs->map_generation here. 186 * 187 * NOTE: This function can fail due to a deadlock against the caller's 188 * holding of a vm_page BUSY. 189 */ 190 static __inline int 191 relock_map(struct faultstate *fs) 192 { 193 int error; 194 195 if (fs->lookup_still_valid == FALSE && fs->map) { 196 error = vm_map_lock_read_to(fs->map); 197 if (error == 0) 198 fs->lookup_still_valid = TRUE; 199 } else { 200 error = 0; 201 } 202 return error; 203 } 204 205 static __inline void 206 unlock_map(struct faultstate *fs) 207 { 208 if (fs->lookup_still_valid && fs->map) { 209 vm_map_lookup_done(fs->map, fs->entry, 0); 210 fs->lookup_still_valid = FALSE; 211 } 212 } 213 214 /* 215 * Clean up after a successful call to vm_fault_object() so another call 216 * to vm_fault_object() can be made. 217 */ 218 static void 219 _cleanup_successful_fault(struct faultstate *fs, int relock) 220 { 221 /* 222 * We allocated a junk page for a COW operation that did 223 * not occur, the page must be freed. 224 */ 225 if (fs->object != fs->first_object) { 226 KKASSERT(fs->first_shared == 0); 227 vm_page_free(fs->first_m); 228 vm_object_pip_wakeup(fs->object); 229 fs->first_m = NULL; 230 } 231 232 /* 233 * Reset fs->object. 234 */ 235 fs->object = fs->first_object; 236 if (relock && fs->lookup_still_valid == FALSE) { 237 if (fs->map) 238 vm_map_lock_read(fs->map); 239 fs->lookup_still_valid = TRUE; 240 } 241 } 242 243 static void 244 _unlock_things(struct faultstate *fs, int dealloc) 245 { 246 _cleanup_successful_fault(fs, 0); 247 if (dealloc) { 248 /*vm_object_deallocate(fs->first_object);*/ 249 /*fs->first_object = NULL; drop used later on */ 250 } 251 unlock_map(fs); 252 if (fs->vp != NULL) { 253 vput(fs->vp); 254 fs->vp = NULL; 255 } 256 } 257 258 #define unlock_things(fs) _unlock_things(fs, 0) 259 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 260 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 261 262 /* 263 * TRYPAGER 264 * 265 * Determine if the pager for the current object *might* contain the page. 266 * 267 * We only need to try the pager if this is not a default object (default 268 * objects are zero-fill and have no real pager), and if we are not taking 269 * a wiring fault or if the FS entry is wired. 270 */ 271 #define TRYPAGER(fs) \ 272 (fs->object->type != OBJT_DEFAULT && \ 273 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired)) 274 275 /* 276 * vm_fault: 277 * 278 * Handle a page fault occuring at the given address, requiring the given 279 * permissions, in the map specified. If successful, the page is inserted 280 * into the associated physical map. 281 * 282 * NOTE: The given address should be truncated to the proper page address. 283 * 284 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 285 * a standard error specifying why the fault is fatal is returned. 286 * 287 * The map in question must be referenced, and remains so. 288 * The caller may hold no locks. 289 * No other requirements. 290 */ 291 int 292 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 293 { 294 int result; 295 vm_pindex_t first_pindex; 296 struct faultstate fs; 297 struct lwp *lp; 298 struct proc *p; 299 thread_t td; 300 int growstack; 301 int retry = 0; 302 int inherit_prot; 303 304 inherit_prot = fault_type & VM_PROT_NOSYNC; 305 fs.hardfault = 0; 306 fs.fault_flags = fault_flags; 307 fs.vp = NULL; 308 fs.shared = vm_shared_fault; 309 fs.first_shared = vm_shared_fault; 310 growstack = 1; 311 312 /* 313 * vm_map interactions 314 */ 315 td = curthread; 316 if ((lp = td->td_lwp) != NULL) 317 lp->lwp_flags |= LWP_PAGING; 318 lwkt_gettoken(&map->token); 319 320 RetryFault: 321 /* 322 * Find the vm_map_entry representing the backing store and resolve 323 * the top level object and page index. This may have the side 324 * effect of executing a copy-on-write on the map entry and/or 325 * creating a shadow object, but will not COW any actual VM pages. 326 * 327 * On success fs.map is left read-locked and various other fields 328 * are initialized but not otherwise referenced or locked. 329 * 330 * NOTE! vm_map_lookup will try to upgrade the fault_type to 331 * VM_FAULT_WRITE if the map entry is a virtual page table and also 332 * writable, so we can set the 'A'accessed bit in the virtual page 333 * table entry. 334 */ 335 fs.map = map; 336 result = vm_map_lookup(&fs.map, vaddr, fault_type, 337 &fs.entry, &fs.first_object, 338 &first_pindex, &fs.first_prot, &fs.wired); 339 340 /* 341 * If the lookup failed or the map protections are incompatible, 342 * the fault generally fails. 343 * 344 * The failure could be due to TDF_NOFAULT if vm_map_lookup() 345 * tried to do a COW fault. 346 * 347 * If the caller is trying to do a user wiring we have more work 348 * to do. 349 */ 350 if (result != KERN_SUCCESS) { 351 if (result == KERN_FAILURE_NOFAULT) { 352 result = KERN_FAILURE; 353 goto done; 354 } 355 if (result != KERN_PROTECTION_FAILURE || 356 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 357 { 358 if (result == KERN_INVALID_ADDRESS && growstack && 359 map != &kernel_map && curproc != NULL) { 360 result = vm_map_growstack(curproc, vaddr); 361 if (result == KERN_SUCCESS) { 362 growstack = 0; 363 ++retry; 364 goto RetryFault; 365 } 366 result = KERN_FAILURE; 367 } 368 goto done; 369 } 370 371 /* 372 * If we are user-wiring a r/w segment, and it is COW, then 373 * we need to do the COW operation. Note that we don't 374 * currently COW RO sections now, because it is NOT desirable 375 * to COW .text. We simply keep .text from ever being COW'ed 376 * and take the heat that one cannot debug wired .text sections. 377 */ 378 result = vm_map_lookup(&fs.map, vaddr, 379 VM_PROT_READ|VM_PROT_WRITE| 380 VM_PROT_OVERRIDE_WRITE, 381 &fs.entry, &fs.first_object, 382 &first_pindex, &fs.first_prot, 383 &fs.wired); 384 if (result != KERN_SUCCESS) { 385 /* could also be KERN_FAILURE_NOFAULT */ 386 result = KERN_FAILURE; 387 goto done; 388 } 389 390 /* 391 * If we don't COW now, on a user wire, the user will never 392 * be able to write to the mapping. If we don't make this 393 * restriction, the bookkeeping would be nearly impossible. 394 * 395 * XXX We have a shared lock, this will have a MP race but 396 * I don't see how it can hurt anything. 397 */ 398 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 399 fs.entry->max_protection &= ~VM_PROT_WRITE; 400 } 401 402 /* 403 * fs.map is read-locked 404 * 405 * Misc checks. Save the map generation number to detect races. 406 */ 407 fs.map_generation = fs.map->timestamp; 408 fs.lookup_still_valid = TRUE; 409 fs.first_m = NULL; 410 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 411 fs.prot = fs.first_prot; /* default (used by uksmap) */ 412 413 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) { 414 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 415 panic("vm_fault: fault on nofault entry, addr: %p", 416 (void *)vaddr); 417 } 418 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) && 419 vaddr >= fs.entry->start && 420 vaddr < fs.entry->start + PAGE_SIZE) { 421 panic("vm_fault: fault on stack guard, addr: %p", 422 (void *)vaddr); 423 } 424 } 425 426 /* 427 * A user-kernel shared map has no VM object and bypasses 428 * everything. We execute the uksmap function with a temporary 429 * fictitious vm_page. The address is directly mapped with no 430 * management. 431 */ 432 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 433 struct vm_page fakem; 434 435 bzero(&fakem, sizeof(fakem)); 436 fakem.pindex = first_pindex; 437 fakem.flags = PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED; 438 fakem.valid = VM_PAGE_BITS_ALL; 439 fakem.pat_mode = VM_MEMATTR_DEFAULT; 440 if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) { 441 result = KERN_FAILURE; 442 unlock_things(&fs); 443 goto done2; 444 } 445 pmap_enter(fs.map->pmap, vaddr, &fakem, fs.prot | inherit_prot, 446 fs.wired, fs.entry); 447 goto done_success; 448 } 449 450 /* 451 * A system map entry may return a NULL object. No object means 452 * no pager means an unrecoverable kernel fault. 453 */ 454 if (fs.first_object == NULL) { 455 panic("vm_fault: unrecoverable fault at %p in entry %p", 456 (void *)vaddr, fs.entry); 457 } 458 459 /* 460 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 461 * is set. 462 * 463 * Unfortunately a deadlock can occur if we are forced to page-in 464 * from swap, but diving all the way into the vm_pager_get_page() 465 * function to find out is too much. Just check the object type. 466 * 467 * The deadlock is a CAM deadlock on a busy VM page when trying 468 * to finish an I/O if another process gets stuck in 469 * vop_helper_read_shortcut() due to a swap fault. 470 */ 471 if ((td->td_flags & TDF_NOFAULT) && 472 (retry || 473 fs.first_object->type == OBJT_VNODE || 474 fs.first_object->type == OBJT_SWAP || 475 fs.first_object->backing_object)) { 476 result = KERN_FAILURE; 477 unlock_things(&fs); 478 goto done2; 479 } 480 481 /* 482 * If the entry is wired we cannot change the page protection. 483 */ 484 if (fs.wired) 485 fault_type = fs.first_prot; 486 487 /* 488 * We generally want to avoid unnecessary exclusive modes on backing 489 * and terminal objects because this can seriously interfere with 490 * heavily fork()'d processes (particularly /bin/sh scripts). 491 * 492 * However, we also want to avoid unnecessary retries due to needed 493 * shared->exclusive promotion for common faults. Exclusive mode is 494 * always needed if any page insertion, rename, or free occurs in an 495 * object (and also indirectly if any I/O is done). 496 * 497 * The main issue here is going to be fs.first_shared. If the 498 * first_object has a backing object which isn't shadowed and the 499 * process is single-threaded we might as well use an exclusive 500 * lock/chain right off the bat. 501 */ 502 if (fs.first_shared && fs.first_object->backing_object && 503 LIST_EMPTY(&fs.first_object->shadow_head) && 504 td->td_proc && td->td_proc->p_nthreads == 1) { 505 fs.first_shared = 0; 506 } 507 508 /* 509 * swap_pager_unswapped() needs an exclusive object 510 */ 511 if (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY)) { 512 fs.first_shared = 0; 513 } 514 515 /* 516 * Obtain a top-level object lock, shared or exclusive depending 517 * on fs.first_shared. If a shared lock winds up being insufficient 518 * we will retry with an exclusive lock. 519 * 520 * The vnode pager lock is always shared. 521 */ 522 if (fs.first_shared) 523 vm_object_hold_shared(fs.first_object); 524 else 525 vm_object_hold(fs.first_object); 526 if (fs.vp == NULL) 527 fs.vp = vnode_pager_lock(fs.first_object); 528 529 /* 530 * The page we want is at (first_object, first_pindex), but if the 531 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 532 * page table to figure out the actual pindex. 533 * 534 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 535 * ONLY 536 */ 537 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 538 result = vm_fault_vpagetable(&fs, &first_pindex, 539 fs.entry->aux.master_pde, 540 fault_type, 1); 541 if (result == KERN_TRY_AGAIN) { 542 vm_object_drop(fs.first_object); 543 ++retry; 544 goto RetryFault; 545 } 546 if (result != KERN_SUCCESS) 547 goto done; 548 } 549 550 /* 551 * Now we have the actual (object, pindex), fault in the page. If 552 * vm_fault_object() fails it will unlock and deallocate the FS 553 * data. If it succeeds everything remains locked and fs->object 554 * will have an additional PIP count if it is not equal to 555 * fs->first_object 556 * 557 * vm_fault_object will set fs->prot for the pmap operation. It is 558 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 559 * page can be safely written. However, it will force a read-only 560 * mapping for a read fault if the memory is managed by a virtual 561 * page table. 562 * 563 * If the fault code uses the shared object lock shortcut 564 * we must not try to burst (we can't allocate VM pages). 565 */ 566 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 567 568 if (debug_fault > 0) { 569 --debug_fault; 570 kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x " 571 "fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n", 572 result, (intmax_t)vaddr, fault_type, fault_flags, 573 fs.m, fs.prot, fs.wired, fs.entry); 574 } 575 576 if (result == KERN_TRY_AGAIN) { 577 vm_object_drop(fs.first_object); 578 ++retry; 579 goto RetryFault; 580 } 581 if (result != KERN_SUCCESS) 582 goto done; 583 584 /* 585 * On success vm_fault_object() does not unlock or deallocate, and fs.m 586 * will contain a busied page. 587 * 588 * Enter the page into the pmap and do pmap-related adjustments. 589 */ 590 KKASSERT(fs.lookup_still_valid == TRUE); 591 vm_page_flag_set(fs.m, PG_REFERENCED); 592 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot | inherit_prot, 593 fs.wired, fs.entry); 594 595 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */ 596 KKASSERT(fs.m->flags & PG_BUSY); 597 598 /* 599 * If the page is not wired down, then put it where the pageout daemon 600 * can find it. 601 */ 602 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 603 if (fs.wired) 604 vm_page_wire(fs.m); 605 else 606 vm_page_unwire(fs.m, 1); 607 } else { 608 vm_page_activate(fs.m); 609 } 610 vm_page_wakeup(fs.m); 611 612 /* 613 * Burst in a few more pages if possible. The fs.map should still 614 * be locked. To avoid interlocking against a vnode->getblk 615 * operation we had to be sure to unbusy our primary vm_page above 616 * first. 617 * 618 * A normal burst can continue down backing store, only execute 619 * if we are holding an exclusive lock, otherwise the exclusive 620 * locks the burst code gets might cause excessive SMP collisions. 621 * 622 * A quick burst can be utilized when there is no backing object 623 * (i.e. a shared file mmap). 624 */ 625 if ((fault_flags & VM_FAULT_BURST) && 626 (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 && 627 fs.wired == 0) { 628 if (fs.first_shared == 0 && fs.shared == 0) { 629 vm_prefault(fs.map->pmap, vaddr, 630 fs.entry, fs.prot, fault_flags); 631 } else { 632 vm_prefault_quick(fs.map->pmap, vaddr, 633 fs.entry, fs.prot, fault_flags); 634 } 635 } 636 637 done_success: 638 mycpu->gd_cnt.v_vm_faults++; 639 if (td->td_lwp) 640 ++td->td_lwp->lwp_ru.ru_minflt; 641 642 /* 643 * Unlock everything, and return 644 */ 645 unlock_things(&fs); 646 647 if (td->td_lwp) { 648 if (fs.hardfault) { 649 td->td_lwp->lwp_ru.ru_majflt++; 650 } else { 651 td->td_lwp->lwp_ru.ru_minflt++; 652 } 653 } 654 655 /*vm_object_deallocate(fs.first_object);*/ 656 /*fs.m = NULL; */ 657 /*fs.first_object = NULL; must still drop later */ 658 659 result = KERN_SUCCESS; 660 done: 661 if (fs.first_object) 662 vm_object_drop(fs.first_object); 663 done2: 664 lwkt_reltoken(&map->token); 665 if (lp) 666 lp->lwp_flags &= ~LWP_PAGING; 667 668 #if !defined(NO_SWAPPING) 669 /* 670 * Check the process RSS limit and force deactivation and 671 * (asynchronous) paging if necessary. This is a complex operation, 672 * only do it for direct user-mode faults, for now. 673 * 674 * To reduce overhead implement approximately a ~16MB hysteresis. 675 */ 676 p = td->td_proc; 677 if ((fault_flags & VM_FAULT_USERMODE) && lp && 678 p->p_limit && map->pmap && vm_pageout_memuse_mode >= 1 && 679 map != &kernel_map) { 680 vm_pindex_t limit; 681 vm_pindex_t size; 682 683 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 684 p->p_rlimit[RLIMIT_RSS].rlim_max)); 685 size = pmap_resident_tlnw_count(map->pmap); 686 if (limit >= 0 && size > 4096 && size - 4096 >= limit) { 687 vm_pageout_map_deactivate_pages(map, limit); 688 } 689 } 690 #endif 691 692 return (result); 693 } 694 695 /* 696 * Fault in the specified virtual address in the current process map, 697 * returning a held VM page or NULL. See vm_fault_page() for more 698 * information. 699 * 700 * No requirements. 701 */ 702 vm_page_t 703 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp) 704 { 705 struct lwp *lp = curthread->td_lwp; 706 vm_page_t m; 707 708 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 709 fault_type, VM_FAULT_NORMAL, errorp); 710 return(m); 711 } 712 713 /* 714 * Fault in the specified virtual address in the specified map, doing all 715 * necessary manipulation of the object store and all necessary I/O. Return 716 * a held VM page or NULL, and set *errorp. The related pmap is not 717 * updated. 718 * 719 * The returned page will be properly dirtied if VM_PROT_WRITE was specified, 720 * and marked PG_REFERENCED as well. 721 * 722 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 723 * error will be returned. 724 * 725 * No requirements. 726 */ 727 vm_page_t 728 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 729 int fault_flags, int *errorp) 730 { 731 vm_pindex_t first_pindex; 732 struct faultstate fs; 733 int result; 734 int retry = 0; 735 vm_prot_t orig_fault_type = fault_type; 736 737 fs.hardfault = 0; 738 fs.fault_flags = fault_flags; 739 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 740 741 /* 742 * Dive the pmap (concurrency possible). If we find the 743 * appropriate page we can terminate early and quickly. 744 */ 745 fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type); 746 if (fs.m) { 747 *errorp = 0; 748 return(fs.m); 749 } 750 751 /* 752 * Otherwise take a concurrency hit and do a formal page 753 * fault. 754 */ 755 fs.shared = vm_shared_fault; 756 fs.first_shared = vm_shared_fault; 757 fs.vp = NULL; 758 lwkt_gettoken(&map->token); 759 760 /* 761 * swap_pager_unswapped() needs an exclusive object 762 */ 763 if (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY)) { 764 fs.first_shared = 0; 765 } 766 767 RetryFault: 768 /* 769 * Find the vm_map_entry representing the backing store and resolve 770 * the top level object and page index. This may have the side 771 * effect of executing a copy-on-write on the map entry and/or 772 * creating a shadow object, but will not COW any actual VM pages. 773 * 774 * On success fs.map is left read-locked and various other fields 775 * are initialized but not otherwise referenced or locked. 776 * 777 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 778 * if the map entry is a virtual page table and also writable, 779 * so we can set the 'A'accessed bit in the virtual page table entry. 780 */ 781 fs.map = map; 782 result = vm_map_lookup(&fs.map, vaddr, fault_type, 783 &fs.entry, &fs.first_object, 784 &first_pindex, &fs.first_prot, &fs.wired); 785 786 if (result != KERN_SUCCESS) { 787 *errorp = result; 788 fs.m = NULL; 789 goto done; 790 } 791 792 /* 793 * fs.map is read-locked 794 * 795 * Misc checks. Save the map generation number to detect races. 796 */ 797 fs.map_generation = fs.map->timestamp; 798 fs.lookup_still_valid = TRUE; 799 fs.first_m = NULL; 800 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 801 802 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 803 panic("vm_fault: fault on nofault entry, addr: %lx", 804 (u_long)vaddr); 805 } 806 807 /* 808 * A user-kernel shared map has no VM object and bypasses 809 * everything. We execute the uksmap function with a temporary 810 * fictitious vm_page. The address is directly mapped with no 811 * management. 812 */ 813 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 814 struct vm_page fakem; 815 816 bzero(&fakem, sizeof(fakem)); 817 fakem.pindex = first_pindex; 818 fakem.flags = PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED; 819 fakem.valid = VM_PAGE_BITS_ALL; 820 fakem.pat_mode = VM_MEMATTR_DEFAULT; 821 if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) { 822 *errorp = KERN_FAILURE; 823 fs.m = NULL; 824 unlock_things(&fs); 825 goto done2; 826 } 827 fs.m = PHYS_TO_VM_PAGE(fakem.phys_addr); 828 vm_page_hold(fs.m); 829 830 unlock_things(&fs); 831 *errorp = 0; 832 goto done; 833 } 834 835 836 /* 837 * A system map entry may return a NULL object. No object means 838 * no pager means an unrecoverable kernel fault. 839 */ 840 if (fs.first_object == NULL) { 841 panic("vm_fault: unrecoverable fault at %p in entry %p", 842 (void *)vaddr, fs.entry); 843 } 844 845 /* 846 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 847 * is set. 848 * 849 * Unfortunately a deadlock can occur if we are forced to page-in 850 * from swap, but diving all the way into the vm_pager_get_page() 851 * function to find out is too much. Just check the object type. 852 */ 853 if ((curthread->td_flags & TDF_NOFAULT) && 854 (retry || 855 fs.first_object->type == OBJT_VNODE || 856 fs.first_object->type == OBJT_SWAP || 857 fs.first_object->backing_object)) { 858 *errorp = KERN_FAILURE; 859 unlock_things(&fs); 860 goto done2; 861 } 862 863 /* 864 * If the entry is wired we cannot change the page protection. 865 */ 866 if (fs.wired) 867 fault_type = fs.first_prot; 868 869 /* 870 * Make a reference to this object to prevent its disposal while we 871 * are messing with it. Once we have the reference, the map is free 872 * to be diddled. Since objects reference their shadows (and copies), 873 * they will stay around as well. 874 * 875 * The reference should also prevent an unexpected collapse of the 876 * parent that might move pages from the current object into the 877 * parent unexpectedly, resulting in corruption. 878 * 879 * Bump the paging-in-progress count to prevent size changes (e.g. 880 * truncation operations) during I/O. This must be done after 881 * obtaining the vnode lock in order to avoid possible deadlocks. 882 */ 883 if (fs.first_shared) 884 vm_object_hold_shared(fs.first_object); 885 else 886 vm_object_hold(fs.first_object); 887 if (fs.vp == NULL) 888 fs.vp = vnode_pager_lock(fs.first_object); /* shared */ 889 890 /* 891 * The page we want is at (first_object, first_pindex), but if the 892 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 893 * page table to figure out the actual pindex. 894 * 895 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 896 * ONLY 897 */ 898 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 899 result = vm_fault_vpagetable(&fs, &first_pindex, 900 fs.entry->aux.master_pde, 901 fault_type, 1); 902 if (result == KERN_TRY_AGAIN) { 903 vm_object_drop(fs.first_object); 904 ++retry; 905 goto RetryFault; 906 } 907 if (result != KERN_SUCCESS) { 908 *errorp = result; 909 fs.m = NULL; 910 goto done; 911 } 912 } 913 914 /* 915 * Now we have the actual (object, pindex), fault in the page. If 916 * vm_fault_object() fails it will unlock and deallocate the FS 917 * data. If it succeeds everything remains locked and fs->object 918 * will have an additinal PIP count if it is not equal to 919 * fs->first_object 920 */ 921 fs.m = NULL; 922 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 923 924 if (result == KERN_TRY_AGAIN) { 925 vm_object_drop(fs.first_object); 926 ++retry; 927 goto RetryFault; 928 } 929 if (result != KERN_SUCCESS) { 930 *errorp = result; 931 fs.m = NULL; 932 goto done; 933 } 934 935 if ((orig_fault_type & VM_PROT_WRITE) && 936 (fs.prot & VM_PROT_WRITE) == 0) { 937 *errorp = KERN_PROTECTION_FAILURE; 938 unlock_and_deallocate(&fs); 939 fs.m = NULL; 940 goto done; 941 } 942 943 /* 944 * DO NOT UPDATE THE PMAP!!! This function may be called for 945 * a pmap unrelated to the current process pmap, in which case 946 * the current cpu core will not be listed in the pmap's pm_active 947 * mask. Thus invalidation interlocks will fail to work properly. 948 * 949 * (for example, 'ps' uses procfs to read program arguments from 950 * each process's stack). 951 * 952 * In addition to the above this function will be called to acquire 953 * a page that might already be faulted in, re-faulting it 954 * continuously is a waste of time. 955 * 956 * XXX could this have been the cause of our random seg-fault 957 * issues? procfs accesses user stacks. 958 */ 959 vm_page_flag_set(fs.m, PG_REFERENCED); 960 #if 0 961 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL); 962 mycpu->gd_cnt.v_vm_faults++; 963 if (curthread->td_lwp) 964 ++curthread->td_lwp->lwp_ru.ru_minflt; 965 #endif 966 967 /* 968 * On success vm_fault_object() does not unlock or deallocate, and fs.m 969 * will contain a busied page. So we must unlock here after having 970 * messed with the pmap. 971 */ 972 unlock_things(&fs); 973 974 /* 975 * Return a held page. We are not doing any pmap manipulation so do 976 * not set PG_MAPPED. However, adjust the page flags according to 977 * the fault type because the caller may not use a managed pmapping 978 * (so we don't want to lose the fact that the page will be dirtied 979 * if a write fault was specified). 980 */ 981 vm_page_hold(fs.m); 982 vm_page_activate(fs.m); 983 if (fault_type & VM_PROT_WRITE) 984 vm_page_dirty(fs.m); 985 986 if (curthread->td_lwp) { 987 if (fs.hardfault) { 988 curthread->td_lwp->lwp_ru.ru_majflt++; 989 } else { 990 curthread->td_lwp->lwp_ru.ru_minflt++; 991 } 992 } 993 994 /* 995 * Unlock everything, and return the held page. 996 */ 997 vm_page_wakeup(fs.m); 998 /*vm_object_deallocate(fs.first_object);*/ 999 /*fs.first_object = NULL; */ 1000 *errorp = 0; 1001 1002 done: 1003 if (fs.first_object) 1004 vm_object_drop(fs.first_object); 1005 done2: 1006 lwkt_reltoken(&map->token); 1007 return(fs.m); 1008 } 1009 1010 /* 1011 * Fault in the specified (object,offset), dirty the returned page as 1012 * needed. If the requested fault_type cannot be done NULL and an 1013 * error is returned. 1014 * 1015 * A held (but not busied) page is returned. 1016 * 1017 * The passed in object must be held as specified by the shared 1018 * argument. 1019 */ 1020 vm_page_t 1021 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 1022 vm_prot_t fault_type, int fault_flags, 1023 int *sharedp, int *errorp) 1024 { 1025 int result; 1026 vm_pindex_t first_pindex; 1027 struct faultstate fs; 1028 struct vm_map_entry entry; 1029 1030 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1031 bzero(&entry, sizeof(entry)); 1032 entry.object.vm_object = object; 1033 entry.maptype = VM_MAPTYPE_NORMAL; 1034 entry.protection = entry.max_protection = fault_type; 1035 1036 fs.hardfault = 0; 1037 fs.fault_flags = fault_flags; 1038 fs.map = NULL; 1039 fs.shared = vm_shared_fault; 1040 fs.first_shared = *sharedp; 1041 fs.vp = NULL; 1042 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 1043 1044 /* 1045 * Might require swap block adjustments 1046 */ 1047 if (fs.first_shared && (fault_flags & (VM_FAULT_UNSWAP | VM_FAULT_DIRTY))) { 1048 fs.first_shared = 0; 1049 vm_object_upgrade(object); 1050 } 1051 1052 /* 1053 * Retry loop as needed (typically for shared->exclusive transitions) 1054 */ 1055 RetryFault: 1056 *sharedp = fs.first_shared; 1057 first_pindex = OFF_TO_IDX(offset); 1058 fs.first_object = object; 1059 fs.entry = &entry; 1060 fs.first_prot = fault_type; 1061 fs.wired = 0; 1062 /*fs.map_generation = 0; unused */ 1063 1064 /* 1065 * Make a reference to this object to prevent its disposal while we 1066 * are messing with it. Once we have the reference, the map is free 1067 * to be diddled. Since objects reference their shadows (and copies), 1068 * they will stay around as well. 1069 * 1070 * The reference should also prevent an unexpected collapse of the 1071 * parent that might move pages from the current object into the 1072 * parent unexpectedly, resulting in corruption. 1073 * 1074 * Bump the paging-in-progress count to prevent size changes (e.g. 1075 * truncation operations) during I/O. This must be done after 1076 * obtaining the vnode lock in order to avoid possible deadlocks. 1077 */ 1078 if (fs.vp == NULL) 1079 fs.vp = vnode_pager_lock(fs.first_object); 1080 1081 fs.lookup_still_valid = TRUE; 1082 fs.first_m = NULL; 1083 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 1084 1085 #if 0 1086 /* XXX future - ability to operate on VM object using vpagetable */ 1087 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1088 result = vm_fault_vpagetable(&fs, &first_pindex, 1089 fs.entry->aux.master_pde, 1090 fault_type, 0); 1091 if (result == KERN_TRY_AGAIN) { 1092 if (fs.first_shared == 0 && *sharedp) 1093 vm_object_upgrade(object); 1094 goto RetryFault; 1095 } 1096 if (result != KERN_SUCCESS) { 1097 *errorp = result; 1098 return (NULL); 1099 } 1100 } 1101 #endif 1102 1103 /* 1104 * Now we have the actual (object, pindex), fault in the page. If 1105 * vm_fault_object() fails it will unlock and deallocate the FS 1106 * data. If it succeeds everything remains locked and fs->object 1107 * will have an additinal PIP count if it is not equal to 1108 * fs->first_object 1109 * 1110 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact. 1111 * We may have to upgrade its lock to handle the requested fault. 1112 */ 1113 result = vm_fault_object(&fs, first_pindex, fault_type, 0); 1114 1115 if (result == KERN_TRY_AGAIN) { 1116 if (fs.first_shared == 0 && *sharedp) 1117 vm_object_upgrade(object); 1118 goto RetryFault; 1119 } 1120 if (result != KERN_SUCCESS) { 1121 *errorp = result; 1122 return(NULL); 1123 } 1124 1125 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 1126 *errorp = KERN_PROTECTION_FAILURE; 1127 unlock_and_deallocate(&fs); 1128 return(NULL); 1129 } 1130 1131 /* 1132 * On success vm_fault_object() does not unlock or deallocate, so we 1133 * do it here. Note that the returned fs.m will be busied. 1134 */ 1135 unlock_things(&fs); 1136 1137 /* 1138 * Return a held page. We are not doing any pmap manipulation so do 1139 * not set PG_MAPPED. However, adjust the page flags according to 1140 * the fault type because the caller may not use a managed pmapping 1141 * (so we don't want to lose the fact that the page will be dirtied 1142 * if a write fault was specified). 1143 */ 1144 vm_page_hold(fs.m); 1145 vm_page_activate(fs.m); 1146 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY)) 1147 vm_page_dirty(fs.m); 1148 if (fault_flags & VM_FAULT_UNSWAP) 1149 swap_pager_unswapped(fs.m); 1150 1151 /* 1152 * Indicate that the page was accessed. 1153 */ 1154 vm_page_flag_set(fs.m, PG_REFERENCED); 1155 1156 if (curthread->td_lwp) { 1157 if (fs.hardfault) { 1158 curthread->td_lwp->lwp_ru.ru_majflt++; 1159 } else { 1160 curthread->td_lwp->lwp_ru.ru_minflt++; 1161 } 1162 } 1163 1164 /* 1165 * Unlock everything, and return the held page. 1166 */ 1167 vm_page_wakeup(fs.m); 1168 /*vm_object_deallocate(fs.first_object);*/ 1169 /*fs.first_object = NULL; */ 1170 1171 *errorp = 0; 1172 return(fs.m); 1173 } 1174 1175 /* 1176 * Translate the virtual page number (first_pindex) that is relative 1177 * to the address space into a logical page number that is relative to the 1178 * backing object. Use the virtual page table pointed to by (vpte). 1179 * 1180 * This implements an N-level page table. Any level can terminate the 1181 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 1182 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 1183 */ 1184 static 1185 int 1186 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 1187 vpte_t vpte, int fault_type, int allow_nofault) 1188 { 1189 struct lwbuf *lwb; 1190 struct lwbuf lwb_cache; 1191 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */ 1192 int result = KERN_SUCCESS; 1193 vpte_t *ptep; 1194 1195 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1196 for (;;) { 1197 /* 1198 * We cannot proceed if the vpte is not valid, not readable 1199 * for a read fault, or not writable for a write fault. 1200 */ 1201 if ((vpte & VPTE_V) == 0) { 1202 unlock_and_deallocate(fs); 1203 return (KERN_FAILURE); 1204 } 1205 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) { 1206 unlock_and_deallocate(fs); 1207 return (KERN_FAILURE); 1208 } 1209 if ((vpte & VPTE_PS) || vshift == 0) 1210 break; 1211 KKASSERT(vshift >= VPTE_PAGE_BITS); 1212 1213 /* 1214 * Get the page table page. Nominally we only read the page 1215 * table, but since we are actively setting VPTE_M and VPTE_A, 1216 * tell vm_fault_object() that we are writing it. 1217 * 1218 * There is currently no real need to optimize this. 1219 */ 1220 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT, 1221 VM_PROT_READ|VM_PROT_WRITE, 1222 allow_nofault); 1223 if (result != KERN_SUCCESS) 1224 return (result); 1225 1226 /* 1227 * Process the returned fs.m and look up the page table 1228 * entry in the page table page. 1229 */ 1230 vshift -= VPTE_PAGE_BITS; 1231 lwb = lwbuf_alloc(fs->m, &lwb_cache); 1232 ptep = ((vpte_t *)lwbuf_kva(lwb) + 1233 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 1234 vpte = *ptep; 1235 1236 /* 1237 * Page table write-back. If the vpte is valid for the 1238 * requested operation, do a write-back to the page table. 1239 * 1240 * XXX VPTE_M is not set properly for page directory pages. 1241 * It doesn't get set in the page directory if the page table 1242 * is modified during a read access. 1243 */ 1244 vm_page_activate(fs->m); 1245 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) && 1246 (vpte & VPTE_RW)) { 1247 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) { 1248 atomic_set_long(ptep, VPTE_M | VPTE_A); 1249 vm_page_dirty(fs->m); 1250 } 1251 } 1252 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V)) { 1253 if ((vpte & VPTE_A) == 0) { 1254 atomic_set_long(ptep, VPTE_A); 1255 vm_page_dirty(fs->m); 1256 } 1257 } 1258 lwbuf_free(lwb); 1259 vm_page_flag_set(fs->m, PG_REFERENCED); 1260 vm_page_wakeup(fs->m); 1261 fs->m = NULL; 1262 cleanup_successful_fault(fs); 1263 } 1264 /* 1265 * Combine remaining address bits with the vpte. 1266 */ 1267 /* JG how many bits from each? */ 1268 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) + 1269 (*pindex & ((1L << vshift) - 1)); 1270 return (KERN_SUCCESS); 1271 } 1272 1273 1274 /* 1275 * This is the core of the vm_fault code. 1276 * 1277 * Do all operations required to fault-in (fs.first_object, pindex). Run 1278 * through the shadow chain as necessary and do required COW or virtual 1279 * copy operations. The caller has already fully resolved the vm_map_entry 1280 * and, if appropriate, has created a copy-on-write layer. All we need to 1281 * do is iterate the object chain. 1282 * 1283 * On failure (fs) is unlocked and deallocated and the caller may return or 1284 * retry depending on the failure code. On success (fs) is NOT unlocked or 1285 * deallocated, fs.m will contained a resolved, busied page, and fs.object 1286 * will have an additional PIP count if it is not equal to fs.first_object. 1287 * 1288 * If locks based on fs->first_shared or fs->shared are insufficient, 1289 * clear the appropriate field(s) and return RETRY. COWs require that 1290 * first_shared be 0, while page allocations (or frees) require that 1291 * shared be 0. Renames require that both be 0. 1292 * 1293 * fs->first_object must be held on call. 1294 */ 1295 static 1296 int 1297 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex, 1298 vm_prot_t fault_type, int allow_nofault) 1299 { 1300 vm_object_t next_object; 1301 vm_pindex_t pindex; 1302 int error; 1303 1304 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1305 fs->prot = fs->first_prot; 1306 fs->object = fs->first_object; 1307 pindex = first_pindex; 1308 1309 vm_object_chain_acquire(fs->first_object, fs->shared); 1310 vm_object_pip_add(fs->first_object, 1); 1311 1312 /* 1313 * If a read fault occurs we try to make the page writable if 1314 * possible. There are three cases where we cannot make the 1315 * page mapping writable: 1316 * 1317 * (1) The mapping is read-only or the VM object is read-only, 1318 * fs->prot above will simply not have VM_PROT_WRITE set. 1319 * 1320 * (2) If the mapping is a virtual page table we need to be able 1321 * to detect writes so we can set VPTE_M in the virtual page 1322 * table. 1323 * 1324 * (3) If the VM page is read-only or copy-on-write, upgrading would 1325 * just result in an unnecessary COW fault. 1326 * 1327 * VM_PROT_VPAGED is set if faulting via a virtual page table and 1328 * causes adjustments to the 'M'odify bit to also turn off write 1329 * access to force a re-fault. 1330 */ 1331 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1332 if ((fault_type & VM_PROT_WRITE) == 0) 1333 fs->prot &= ~VM_PROT_WRITE; 1334 } 1335 1336 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 1337 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 1338 if ((fault_type & VM_PROT_WRITE) == 0) 1339 fs->prot &= ~VM_PROT_WRITE; 1340 } 1341 1342 /* vm_object_hold(fs->object); implied b/c object == first_object */ 1343 1344 for (;;) { 1345 /* 1346 * The entire backing chain from first_object to object 1347 * inclusive is chainlocked. 1348 * 1349 * If the object is dead, we stop here 1350 */ 1351 if (fs->object->flags & OBJ_DEAD) { 1352 vm_object_pip_wakeup(fs->first_object); 1353 vm_object_chain_release_all(fs->first_object, 1354 fs->object); 1355 if (fs->object != fs->first_object) 1356 vm_object_drop(fs->object); 1357 unlock_and_deallocate(fs); 1358 return (KERN_PROTECTION_FAILURE); 1359 } 1360 1361 /* 1362 * See if the page is resident. Wait/Retry if the page is 1363 * busy (lots of stuff may have changed so we can't continue 1364 * in that case). 1365 * 1366 * We can theoretically allow the soft-busy case on a read 1367 * fault if the page is marked valid, but since such 1368 * pages are typically already pmap'd, putting that 1369 * special case in might be more effort then it is 1370 * worth. We cannot under any circumstances mess 1371 * around with a vm_page_t->busy page except, perhaps, 1372 * to pmap it. 1373 */ 1374 fs->m = vm_page_lookup_busy_try(fs->object, pindex, 1375 TRUE, &error); 1376 if (error) { 1377 vm_object_pip_wakeup(fs->first_object); 1378 vm_object_chain_release_all(fs->first_object, 1379 fs->object); 1380 if (fs->object != fs->first_object) 1381 vm_object_drop(fs->object); 1382 unlock_things(fs); 1383 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 1384 mycpu->gd_cnt.v_intrans++; 1385 /*vm_object_deallocate(fs->first_object);*/ 1386 /*fs->first_object = NULL;*/ 1387 fs->m = NULL; 1388 return (KERN_TRY_AGAIN); 1389 } 1390 if (fs->m) { 1391 /* 1392 * The page is busied for us. 1393 * 1394 * If reactivating a page from PQ_CACHE we may have 1395 * to rate-limit. 1396 */ 1397 int queue = fs->m->queue; 1398 vm_page_unqueue_nowakeup(fs->m); 1399 1400 if ((queue - fs->m->pc) == PQ_CACHE && 1401 vm_page_count_severe()) { 1402 vm_page_activate(fs->m); 1403 vm_page_wakeup(fs->m); 1404 fs->m = NULL; 1405 vm_object_pip_wakeup(fs->first_object); 1406 vm_object_chain_release_all(fs->first_object, 1407 fs->object); 1408 if (fs->object != fs->first_object) 1409 vm_object_drop(fs->object); 1410 unlock_and_deallocate(fs); 1411 if (allow_nofault == 0 || 1412 (curthread->td_flags & TDF_NOFAULT) == 0) { 1413 thread_t td; 1414 1415 vm_wait_pfault(); 1416 td = curthread; 1417 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1418 return (KERN_PROTECTION_FAILURE); 1419 } 1420 return (KERN_TRY_AGAIN); 1421 } 1422 1423 /* 1424 * If it still isn't completely valid (readable), 1425 * or if a read-ahead-mark is set on the VM page, 1426 * jump to readrest, else we found the page and 1427 * can return. 1428 * 1429 * We can release the spl once we have marked the 1430 * page busy. 1431 */ 1432 if (fs->m->object != &kernel_object) { 1433 if ((fs->m->valid & VM_PAGE_BITS_ALL) != 1434 VM_PAGE_BITS_ALL) { 1435 goto readrest; 1436 } 1437 if (fs->m->flags & PG_RAM) { 1438 if (debug_cluster) 1439 kprintf("R"); 1440 vm_page_flag_clear(fs->m, PG_RAM); 1441 goto readrest; 1442 } 1443 } 1444 break; /* break to PAGE HAS BEEN FOUND */ 1445 } 1446 1447 /* 1448 * Page is not resident, If this is the search termination 1449 * or the pager might contain the page, allocate a new page. 1450 */ 1451 if (TRYPAGER(fs) || fs->object == fs->first_object) { 1452 /* 1453 * Allocating, must be exclusive. 1454 */ 1455 if (fs->object == fs->first_object && 1456 fs->first_shared) { 1457 fs->first_shared = 0; 1458 vm_object_pip_wakeup(fs->first_object); 1459 vm_object_chain_release_all(fs->first_object, 1460 fs->object); 1461 if (fs->object != fs->first_object) 1462 vm_object_drop(fs->object); 1463 unlock_and_deallocate(fs); 1464 return (KERN_TRY_AGAIN); 1465 } 1466 if (fs->object != fs->first_object && 1467 fs->shared) { 1468 fs->first_shared = 0; 1469 fs->shared = 0; 1470 vm_object_pip_wakeup(fs->first_object); 1471 vm_object_chain_release_all(fs->first_object, 1472 fs->object); 1473 if (fs->object != fs->first_object) 1474 vm_object_drop(fs->object); 1475 unlock_and_deallocate(fs); 1476 return (KERN_TRY_AGAIN); 1477 } 1478 1479 /* 1480 * If the page is beyond the object size we fail 1481 */ 1482 if (pindex >= fs->object->size) { 1483 vm_object_pip_wakeup(fs->first_object); 1484 vm_object_chain_release_all(fs->first_object, 1485 fs->object); 1486 if (fs->object != fs->first_object) 1487 vm_object_drop(fs->object); 1488 unlock_and_deallocate(fs); 1489 return (KERN_PROTECTION_FAILURE); 1490 } 1491 1492 /* 1493 * Allocate a new page for this object/offset pair. 1494 * 1495 * It is possible for the allocation to race, so 1496 * handle the case. 1497 */ 1498 fs->m = NULL; 1499 if (!vm_page_count_severe()) { 1500 fs->m = vm_page_alloc(fs->object, pindex, 1501 ((fs->vp || fs->object->backing_object) ? 1502 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1503 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1504 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1505 } 1506 if (fs->m == NULL) { 1507 vm_object_pip_wakeup(fs->first_object); 1508 vm_object_chain_release_all(fs->first_object, 1509 fs->object); 1510 if (fs->object != fs->first_object) 1511 vm_object_drop(fs->object); 1512 unlock_and_deallocate(fs); 1513 if (allow_nofault == 0 || 1514 (curthread->td_flags & TDF_NOFAULT) == 0) { 1515 thread_t td; 1516 1517 vm_wait_pfault(); 1518 td = curthread; 1519 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1520 return (KERN_PROTECTION_FAILURE); 1521 } 1522 return (KERN_TRY_AGAIN); 1523 } 1524 1525 /* 1526 * Fall through to readrest. We have a new page which 1527 * will have to be paged (since m->valid will be 0). 1528 */ 1529 } 1530 1531 readrest: 1532 /* 1533 * We have found an invalid or partially valid page, a 1534 * page with a read-ahead mark which might be partially or 1535 * fully valid (and maybe dirty too), or we have allocated 1536 * a new page. 1537 * 1538 * Attempt to fault-in the page if there is a chance that the 1539 * pager has it, and potentially fault in additional pages 1540 * at the same time. 1541 * 1542 * If TRYPAGER is true then fs.m will be non-NULL and busied 1543 * for us. 1544 */ 1545 if (TRYPAGER(fs)) { 1546 int rv; 1547 int seqaccess; 1548 u_char behavior = vm_map_entry_behavior(fs->entry); 1549 1550 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 1551 seqaccess = 0; 1552 else 1553 seqaccess = -1; 1554 1555 /* 1556 * Doing I/O may synchronously insert additional 1557 * pages so we can't be shared at this point either. 1558 * 1559 * NOTE: We can't free fs->m here in the allocated 1560 * case (fs->object != fs->first_object) as 1561 * this would require an exclusively locked 1562 * VM object. 1563 */ 1564 if (fs->object == fs->first_object && 1565 fs->first_shared) { 1566 vm_page_deactivate(fs->m); 1567 vm_page_wakeup(fs->m); 1568 fs->m = NULL; 1569 fs->first_shared = 0; 1570 vm_object_pip_wakeup(fs->first_object); 1571 vm_object_chain_release_all(fs->first_object, 1572 fs->object); 1573 if (fs->object != fs->first_object) 1574 vm_object_drop(fs->object); 1575 unlock_and_deallocate(fs); 1576 return (KERN_TRY_AGAIN); 1577 } 1578 if (fs->object != fs->first_object && 1579 fs->shared) { 1580 vm_page_deactivate(fs->m); 1581 vm_page_wakeup(fs->m); 1582 fs->m = NULL; 1583 fs->first_shared = 0; 1584 fs->shared = 0; 1585 vm_object_pip_wakeup(fs->first_object); 1586 vm_object_chain_release_all(fs->first_object, 1587 fs->object); 1588 if (fs->object != fs->first_object) 1589 vm_object_drop(fs->object); 1590 unlock_and_deallocate(fs); 1591 return (KERN_TRY_AGAIN); 1592 } 1593 1594 /* 1595 * Avoid deadlocking against the map when doing I/O. 1596 * fs.object and the page is PG_BUSY'd. 1597 * 1598 * NOTE: Once unlocked, fs->entry can become stale 1599 * so this will NULL it out. 1600 * 1601 * NOTE: fs->entry is invalid until we relock the 1602 * map and verify that the timestamp has not 1603 * changed. 1604 */ 1605 unlock_map(fs); 1606 1607 /* 1608 * Acquire the page data. We still hold a ref on 1609 * fs.object and the page has been PG_BUSY's. 1610 * 1611 * The pager may replace the page (for example, in 1612 * order to enter a fictitious page into the 1613 * object). If it does so it is responsible for 1614 * cleaning up the passed page and properly setting 1615 * the new page PG_BUSY. 1616 * 1617 * If we got here through a PG_RAM read-ahead 1618 * mark the page may be partially dirty and thus 1619 * not freeable. Don't bother checking to see 1620 * if the pager has the page because we can't free 1621 * it anyway. We have to depend on the get_page 1622 * operation filling in any gaps whether there is 1623 * backing store or not. 1624 */ 1625 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess); 1626 1627 if (rv == VM_PAGER_OK) { 1628 /* 1629 * Relookup in case pager changed page. Pager 1630 * is responsible for disposition of old page 1631 * if moved. 1632 * 1633 * XXX other code segments do relookups too. 1634 * It's a bad abstraction that needs to be 1635 * fixed/removed. 1636 */ 1637 fs->m = vm_page_lookup(fs->object, pindex); 1638 if (fs->m == NULL) { 1639 vm_object_pip_wakeup(fs->first_object); 1640 vm_object_chain_release_all( 1641 fs->first_object, fs->object); 1642 if (fs->object != fs->first_object) 1643 vm_object_drop(fs->object); 1644 unlock_and_deallocate(fs); 1645 return (KERN_TRY_AGAIN); 1646 } 1647 ++fs->hardfault; 1648 break; /* break to PAGE HAS BEEN FOUND */ 1649 } 1650 1651 /* 1652 * Remove the bogus page (which does not exist at this 1653 * object/offset); before doing so, we must get back 1654 * our object lock to preserve our invariant. 1655 * 1656 * Also wake up any other process that may want to bring 1657 * in this page. 1658 * 1659 * If this is the top-level object, we must leave the 1660 * busy page to prevent another process from rushing 1661 * past us, and inserting the page in that object at 1662 * the same time that we are. 1663 */ 1664 if (rv == VM_PAGER_ERROR) { 1665 if (curproc) { 1666 kprintf("vm_fault: pager read error, " 1667 "pid %d (%s)\n", 1668 curproc->p_pid, 1669 curproc->p_comm); 1670 } else { 1671 kprintf("vm_fault: pager read error, " 1672 "thread %p (%s)\n", 1673 curthread, 1674 curproc->p_comm); 1675 } 1676 } 1677 1678 /* 1679 * Data outside the range of the pager or an I/O error 1680 * 1681 * The page may have been wired during the pagein, 1682 * e.g. by the buffer cache, and cannot simply be 1683 * freed. Call vnode_pager_freepage() to deal with it. 1684 * 1685 * Also note that we cannot free the page if we are 1686 * holding the related object shared. XXX not sure 1687 * what to do in that case. 1688 */ 1689 if (fs->object != fs->first_object) { 1690 vnode_pager_freepage(fs->m); 1691 fs->m = NULL; 1692 /* 1693 * XXX - we cannot just fall out at this 1694 * point, m has been freed and is invalid! 1695 */ 1696 } 1697 /* 1698 * XXX - the check for kernel_map is a kludge to work 1699 * around having the machine panic on a kernel space 1700 * fault w/ I/O error. 1701 */ 1702 if (((fs->map != &kernel_map) && 1703 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 1704 if (fs->m) { 1705 if (fs->first_shared) { 1706 vm_page_deactivate(fs->m); 1707 vm_page_wakeup(fs->m); 1708 } else { 1709 vnode_pager_freepage(fs->m); 1710 } 1711 fs->m = NULL; 1712 } 1713 vm_object_pip_wakeup(fs->first_object); 1714 vm_object_chain_release_all(fs->first_object, 1715 fs->object); 1716 if (fs->object != fs->first_object) 1717 vm_object_drop(fs->object); 1718 unlock_and_deallocate(fs); 1719 if (rv == VM_PAGER_ERROR) 1720 return (KERN_FAILURE); 1721 else 1722 return (KERN_PROTECTION_FAILURE); 1723 /* NOT REACHED */ 1724 } 1725 } 1726 1727 /* 1728 * We get here if the object has a default pager (or unwiring) 1729 * or the pager doesn't have the page. 1730 * 1731 * fs->first_m will be used for the COW unless we find a 1732 * deeper page to be mapped read-only, in which case the 1733 * unlock*(fs) will free first_m. 1734 */ 1735 if (fs->object == fs->first_object) 1736 fs->first_m = fs->m; 1737 1738 /* 1739 * Move on to the next object. The chain lock should prevent 1740 * the backing_object from getting ripped out from under us. 1741 * 1742 * The object lock for the next object is governed by 1743 * fs->shared. 1744 */ 1745 if ((next_object = fs->object->backing_object) != NULL) { 1746 if (fs->shared) 1747 vm_object_hold_shared(next_object); 1748 else 1749 vm_object_hold(next_object); 1750 vm_object_chain_acquire(next_object, fs->shared); 1751 KKASSERT(next_object == fs->object->backing_object); 1752 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1753 } 1754 1755 if (next_object == NULL) { 1756 /* 1757 * If there's no object left, fill the page in the top 1758 * object with zeros. 1759 */ 1760 if (fs->object != fs->first_object) { 1761 #if 0 1762 if (fs->first_object->backing_object != 1763 fs->object) { 1764 vm_object_hold(fs->first_object->backing_object); 1765 } 1766 #endif 1767 vm_object_chain_release_all( 1768 fs->first_object->backing_object, 1769 fs->object); 1770 #if 0 1771 if (fs->first_object->backing_object != 1772 fs->object) { 1773 vm_object_drop(fs->first_object->backing_object); 1774 } 1775 #endif 1776 vm_object_pip_wakeup(fs->object); 1777 vm_object_drop(fs->object); 1778 fs->object = fs->first_object; 1779 pindex = first_pindex; 1780 fs->m = fs->first_m; 1781 } 1782 fs->first_m = NULL; 1783 1784 /* 1785 * Zero the page and mark it valid. 1786 */ 1787 vm_page_zero_fill(fs->m); 1788 mycpu->gd_cnt.v_zfod++; 1789 fs->m->valid = VM_PAGE_BITS_ALL; 1790 break; /* break to PAGE HAS BEEN FOUND */ 1791 } 1792 if (fs->object != fs->first_object) { 1793 vm_object_pip_wakeup(fs->object); 1794 vm_object_lock_swap(); 1795 vm_object_drop(fs->object); 1796 } 1797 KASSERT(fs->object != next_object, 1798 ("object loop %p", next_object)); 1799 fs->object = next_object; 1800 vm_object_pip_add(fs->object, 1); 1801 } 1802 1803 /* 1804 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1805 * is held.] 1806 * 1807 * object still held. 1808 * 1809 * local shared variable may be different from fs->shared. 1810 * 1811 * If the page is being written, but isn't already owned by the 1812 * top-level object, we have to copy it into a new page owned by the 1813 * top-level object. 1814 */ 1815 KASSERT((fs->m->flags & PG_BUSY) != 0, 1816 ("vm_fault: not busy after main loop")); 1817 1818 if (fs->object != fs->first_object) { 1819 /* 1820 * We only really need to copy if we want to write it. 1821 */ 1822 if (fault_type & VM_PROT_WRITE) { 1823 /* 1824 * This allows pages to be virtually copied from a 1825 * backing_object into the first_object, where the 1826 * backing object has no other refs to it, and cannot 1827 * gain any more refs. Instead of a bcopy, we just 1828 * move the page from the backing object to the 1829 * first object. Note that we must mark the page 1830 * dirty in the first object so that it will go out 1831 * to swap when needed. 1832 */ 1833 if ( 1834 /* 1835 * Must be holding exclusive locks 1836 */ 1837 fs->first_shared == 0 && 1838 fs->shared == 0 && 1839 /* 1840 * Map, if present, has not changed 1841 */ 1842 (fs->map == NULL || 1843 fs->map_generation == fs->map->timestamp) && 1844 /* 1845 * Only one shadow object 1846 */ 1847 (fs->object->shadow_count == 1) && 1848 /* 1849 * No COW refs, except us 1850 */ 1851 (fs->object->ref_count == 1) && 1852 /* 1853 * No one else can look this object up 1854 */ 1855 (fs->object->handle == NULL) && 1856 /* 1857 * No other ways to look the object up 1858 */ 1859 ((fs->object->type == OBJT_DEFAULT) || 1860 (fs->object->type == OBJT_SWAP)) && 1861 /* 1862 * We don't chase down the shadow chain 1863 */ 1864 (fs->object == fs->first_object->backing_object) && 1865 1866 /* 1867 * grab the lock if we need to 1868 */ 1869 (fs->lookup_still_valid || 1870 fs->map == NULL || 1871 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 1872 ) { 1873 /* 1874 * (first_m) and (m) are both busied. We have 1875 * move (m) into (first_m)'s object/pindex 1876 * in an atomic fashion, then free (first_m). 1877 * 1878 * first_object is held so second remove 1879 * followed by the rename should wind 1880 * up being atomic. vm_page_free() might 1881 * block so we don't do it until after the 1882 * rename. 1883 */ 1884 fs->lookup_still_valid = 1; 1885 vm_page_protect(fs->first_m, VM_PROT_NONE); 1886 vm_page_remove(fs->first_m); 1887 vm_page_rename(fs->m, fs->first_object, 1888 first_pindex); 1889 vm_page_free(fs->first_m); 1890 fs->first_m = fs->m; 1891 fs->m = NULL; 1892 mycpu->gd_cnt.v_cow_optim++; 1893 } else { 1894 /* 1895 * Oh, well, lets copy it. 1896 * 1897 * Why are we unmapping the original page 1898 * here? Well, in short, not all accessors 1899 * of user memory go through the pmap. The 1900 * procfs code doesn't have access user memory 1901 * via a local pmap, so vm_fault_page*() 1902 * can't call pmap_enter(). And the umtx*() 1903 * code may modify the COW'd page via a DMAP 1904 * or kernel mapping and not via the pmap, 1905 * leaving the original page still mapped 1906 * read-only into the pmap. 1907 * 1908 * So we have to remove the page from at 1909 * least the current pmap if it is in it. 1910 * Just remove it from all pmaps. 1911 */ 1912 KKASSERT(fs->first_shared == 0); 1913 vm_page_copy(fs->m, fs->first_m); 1914 vm_page_protect(fs->m, VM_PROT_NONE); 1915 vm_page_event(fs->m, VMEVENT_COW); 1916 } 1917 1918 /* 1919 * We no longer need the old page or object. 1920 */ 1921 if (fs->m) 1922 release_page(fs); 1923 1924 /* 1925 * We intend to revert to first_object, undo the 1926 * chain lock through to that. 1927 */ 1928 #if 0 1929 if (fs->first_object->backing_object != fs->object) 1930 vm_object_hold(fs->first_object->backing_object); 1931 #endif 1932 vm_object_chain_release_all( 1933 fs->first_object->backing_object, 1934 fs->object); 1935 #if 0 1936 if (fs->first_object->backing_object != fs->object) 1937 vm_object_drop(fs->first_object->backing_object); 1938 #endif 1939 1940 /* 1941 * fs->object != fs->first_object due to above 1942 * conditional 1943 */ 1944 vm_object_pip_wakeup(fs->object); 1945 vm_object_drop(fs->object); 1946 1947 /* 1948 * Only use the new page below... 1949 */ 1950 mycpu->gd_cnt.v_cow_faults++; 1951 fs->m = fs->first_m; 1952 fs->object = fs->first_object; 1953 pindex = first_pindex; 1954 } else { 1955 /* 1956 * If it wasn't a write fault avoid having to copy 1957 * the page by mapping it read-only. 1958 */ 1959 fs->prot &= ~VM_PROT_WRITE; 1960 } 1961 } 1962 1963 /* 1964 * Relock the map if necessary, then check the generation count. 1965 * relock_map() will update fs->timestamp to account for the 1966 * relocking if necessary. 1967 * 1968 * If the count has changed after relocking then all sorts of 1969 * crap may have happened and we have to retry. 1970 * 1971 * NOTE: The relock_map() can fail due to a deadlock against 1972 * the vm_page we are holding BUSY. 1973 */ 1974 if (fs->lookup_still_valid == FALSE && fs->map) { 1975 if (relock_map(fs) || 1976 fs->map->timestamp != fs->map_generation) { 1977 release_page(fs); 1978 vm_object_pip_wakeup(fs->first_object); 1979 vm_object_chain_release_all(fs->first_object, 1980 fs->object); 1981 if (fs->object != fs->first_object) 1982 vm_object_drop(fs->object); 1983 unlock_and_deallocate(fs); 1984 return (KERN_TRY_AGAIN); 1985 } 1986 } 1987 1988 /* 1989 * If the fault is a write, we know that this page is being 1990 * written NOW so dirty it explicitly to save on pmap_is_modified() 1991 * calls later. 1992 * 1993 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 1994 * if the page is already dirty to prevent data written with 1995 * the expectation of being synced from not being synced. 1996 * Likewise if this entry does not request NOSYNC then make 1997 * sure the page isn't marked NOSYNC. Applications sharing 1998 * data should use the same flags to avoid ping ponging. 1999 * 2000 * Also tell the backing pager, if any, that it should remove 2001 * any swap backing since the page is now dirty. 2002 */ 2003 vm_page_activate(fs->m); 2004 if (fs->prot & VM_PROT_WRITE) { 2005 vm_object_set_writeable_dirty(fs->m->object); 2006 vm_set_nosync(fs->m, fs->entry); 2007 if (fs->fault_flags & VM_FAULT_DIRTY) { 2008 vm_page_dirty(fs->m); 2009 swap_pager_unswapped(fs->m); 2010 } 2011 } 2012 2013 vm_object_pip_wakeup(fs->first_object); 2014 vm_object_chain_release_all(fs->first_object, fs->object); 2015 if (fs->object != fs->first_object) 2016 vm_object_drop(fs->object); 2017 2018 /* 2019 * Page had better still be busy. We are still locked up and 2020 * fs->object will have another PIP reference if it is not equal 2021 * to fs->first_object. 2022 */ 2023 KASSERT(fs->m->flags & PG_BUSY, 2024 ("vm_fault: page %p not busy!", fs->m)); 2025 2026 /* 2027 * Sanity check: page must be completely valid or it is not fit to 2028 * map into user space. vm_pager_get_pages() ensures this. 2029 */ 2030 if (fs->m->valid != VM_PAGE_BITS_ALL) { 2031 vm_page_zero_invalid(fs->m, TRUE); 2032 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 2033 } 2034 2035 return (KERN_SUCCESS); 2036 } 2037 2038 /* 2039 * Hold each of the physical pages that are mapped by the specified range of 2040 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 2041 * and allow the specified types of access, "prot". If all of the implied 2042 * pages are successfully held, then the number of held pages is returned 2043 * together with pointers to those pages in the array "ma". However, if any 2044 * of the pages cannot be held, -1 is returned. 2045 */ 2046 int 2047 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 2048 vm_prot_t prot, vm_page_t *ma, int max_count) 2049 { 2050 vm_offset_t start, end; 2051 int i, npages, error; 2052 2053 start = trunc_page(addr); 2054 end = round_page(addr + len); 2055 2056 npages = howmany(end - start, PAGE_SIZE); 2057 2058 if (npages > max_count) 2059 return -1; 2060 2061 for (i = 0; i < npages; i++) { 2062 // XXX error handling 2063 ma[i] = vm_fault_page_quick(start + (i * PAGE_SIZE), 2064 prot, 2065 &error); 2066 } 2067 2068 return npages; 2069 } 2070 2071 /* 2072 * Wire down a range of virtual addresses in a map. The entry in question 2073 * should be marked in-transition and the map must be locked. We must 2074 * release the map temporarily while faulting-in the page to avoid a 2075 * deadlock. Note that the entry may be clipped while we are blocked but 2076 * will never be freed. 2077 * 2078 * No requirements. 2079 */ 2080 int 2081 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, 2082 boolean_t user_wire, int kmflags) 2083 { 2084 boolean_t fictitious; 2085 vm_offset_t start; 2086 vm_offset_t end; 2087 vm_offset_t va; 2088 vm_paddr_t pa; 2089 vm_page_t m; 2090 pmap_t pmap; 2091 int rv; 2092 int wire_prot; 2093 int fault_flags; 2094 2095 lwkt_gettoken(&map->token); 2096 2097 if (user_wire) { 2098 wire_prot = VM_PROT_READ; 2099 fault_flags = VM_FAULT_USER_WIRE; 2100 } else { 2101 wire_prot = VM_PROT_READ | VM_PROT_WRITE; 2102 fault_flags = VM_FAULT_CHANGE_WIRING; 2103 } 2104 if (kmflags & KM_NOTLBSYNC) 2105 wire_prot |= VM_PROT_NOSYNC; 2106 2107 pmap = vm_map_pmap(map); 2108 start = entry->start; 2109 end = entry->end; 2110 switch(entry->maptype) { 2111 case VM_MAPTYPE_NORMAL: 2112 case VM_MAPTYPE_VPAGETABLE: 2113 fictitious = entry->object.vm_object && 2114 ((entry->object.vm_object->type == OBJT_DEVICE) || 2115 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2116 break; 2117 case VM_MAPTYPE_UKSMAP: 2118 fictitious = TRUE; 2119 break; 2120 default: 2121 fictitious = FALSE; 2122 break; 2123 } 2124 2125 if (entry->eflags & MAP_ENTRY_KSTACK) 2126 start += PAGE_SIZE; 2127 map->timestamp++; 2128 vm_map_unlock(map); 2129 2130 /* 2131 * We simulate a fault to get the page and enter it in the physical 2132 * map. 2133 */ 2134 for (va = start; va < end; va += PAGE_SIZE) { 2135 rv = vm_fault(map, va, wire_prot, fault_flags); 2136 if (rv) { 2137 while (va > start) { 2138 va -= PAGE_SIZE; 2139 if ((pa = pmap_extract(pmap, va)) == 0) 2140 continue; 2141 pmap_change_wiring(pmap, va, FALSE, entry); 2142 if (!fictitious) { 2143 m = PHYS_TO_VM_PAGE(pa); 2144 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2145 vm_page_unwire(m, 1); 2146 vm_page_wakeup(m); 2147 } 2148 } 2149 goto done; 2150 } 2151 } 2152 rv = KERN_SUCCESS; 2153 done: 2154 vm_map_lock(map); 2155 lwkt_reltoken(&map->token); 2156 return (rv); 2157 } 2158 2159 /* 2160 * Unwire a range of virtual addresses in a map. The map should be 2161 * locked. 2162 */ 2163 void 2164 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 2165 { 2166 boolean_t fictitious; 2167 vm_offset_t start; 2168 vm_offset_t end; 2169 vm_offset_t va; 2170 vm_paddr_t pa; 2171 vm_page_t m; 2172 pmap_t pmap; 2173 2174 lwkt_gettoken(&map->token); 2175 2176 pmap = vm_map_pmap(map); 2177 start = entry->start; 2178 end = entry->end; 2179 fictitious = entry->object.vm_object && 2180 ((entry->object.vm_object->type == OBJT_DEVICE) || 2181 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2182 if (entry->eflags & MAP_ENTRY_KSTACK) 2183 start += PAGE_SIZE; 2184 2185 /* 2186 * Since the pages are wired down, we must be able to get their 2187 * mappings from the physical map system. 2188 */ 2189 for (va = start; va < end; va += PAGE_SIZE) { 2190 pa = pmap_extract(pmap, va); 2191 if (pa != 0) { 2192 pmap_change_wiring(pmap, va, FALSE, entry); 2193 if (!fictitious) { 2194 m = PHYS_TO_VM_PAGE(pa); 2195 vm_page_busy_wait(m, FALSE, "vmwupg"); 2196 vm_page_unwire(m, 1); 2197 vm_page_wakeup(m); 2198 } 2199 } 2200 } 2201 lwkt_reltoken(&map->token); 2202 } 2203 2204 /* 2205 * Copy all of the pages from a wired-down map entry to another. 2206 * 2207 * The source and destination maps must be locked for write. 2208 * The source and destination maps token must be held 2209 * The source map entry must be wired down (or be a sharing map 2210 * entry corresponding to a main map entry that is wired down). 2211 * 2212 * No other requirements. 2213 * 2214 * XXX do segment optimization 2215 */ 2216 void 2217 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2218 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2219 { 2220 vm_object_t dst_object; 2221 vm_object_t src_object; 2222 vm_ooffset_t dst_offset; 2223 vm_ooffset_t src_offset; 2224 vm_prot_t prot; 2225 vm_offset_t vaddr; 2226 vm_page_t dst_m; 2227 vm_page_t src_m; 2228 2229 src_object = src_entry->object.vm_object; 2230 src_offset = src_entry->offset; 2231 2232 /* 2233 * Create the top-level object for the destination entry. (Doesn't 2234 * actually shadow anything - we copy the pages directly.) 2235 */ 2236 vm_map_entry_allocate_object(dst_entry); 2237 dst_object = dst_entry->object.vm_object; 2238 2239 prot = dst_entry->max_protection; 2240 2241 /* 2242 * Loop through all of the pages in the entry's range, copying each 2243 * one from the source object (it should be there) to the destination 2244 * object. 2245 */ 2246 vm_object_hold(src_object); 2247 vm_object_hold(dst_object); 2248 for (vaddr = dst_entry->start, dst_offset = 0; 2249 vaddr < dst_entry->end; 2250 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2251 2252 /* 2253 * Allocate a page in the destination object 2254 */ 2255 do { 2256 dst_m = vm_page_alloc(dst_object, 2257 OFF_TO_IDX(dst_offset), 2258 VM_ALLOC_NORMAL); 2259 if (dst_m == NULL) { 2260 vm_wait(0); 2261 } 2262 } while (dst_m == NULL); 2263 2264 /* 2265 * Find the page in the source object, and copy it in. 2266 * (Because the source is wired down, the page will be in 2267 * memory.) 2268 */ 2269 src_m = vm_page_lookup(src_object, 2270 OFF_TO_IDX(dst_offset + src_offset)); 2271 if (src_m == NULL) 2272 panic("vm_fault_copy_wired: page missing"); 2273 2274 vm_page_copy(src_m, dst_m); 2275 vm_page_event(src_m, VMEVENT_COW); 2276 2277 /* 2278 * Enter it in the pmap... 2279 */ 2280 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2281 2282 /* 2283 * Mark it no longer busy, and put it on the active list. 2284 */ 2285 vm_page_activate(dst_m); 2286 vm_page_wakeup(dst_m); 2287 } 2288 vm_object_drop(dst_object); 2289 vm_object_drop(src_object); 2290 } 2291 2292 #if 0 2293 2294 /* 2295 * This routine checks around the requested page for other pages that 2296 * might be able to be faulted in. This routine brackets the viable 2297 * pages for the pages to be paged in. 2298 * 2299 * Inputs: 2300 * m, rbehind, rahead 2301 * 2302 * Outputs: 2303 * marray (array of vm_page_t), reqpage (index of requested page) 2304 * 2305 * Return value: 2306 * number of pages in marray 2307 */ 2308 static int 2309 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2310 vm_page_t *marray, int *reqpage) 2311 { 2312 int i,j; 2313 vm_object_t object; 2314 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2315 vm_page_t rtm; 2316 int cbehind, cahead; 2317 2318 object = m->object; 2319 pindex = m->pindex; 2320 2321 /* 2322 * we don't fault-ahead for device pager 2323 */ 2324 if ((object->type == OBJT_DEVICE) || 2325 (object->type == OBJT_MGTDEVICE)) { 2326 *reqpage = 0; 2327 marray[0] = m; 2328 return 1; 2329 } 2330 2331 /* 2332 * if the requested page is not available, then give up now 2333 */ 2334 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2335 *reqpage = 0; /* not used by caller, fix compiler warn */ 2336 return 0; 2337 } 2338 2339 if ((cbehind == 0) && (cahead == 0)) { 2340 *reqpage = 0; 2341 marray[0] = m; 2342 return 1; 2343 } 2344 2345 if (rahead > cahead) { 2346 rahead = cahead; 2347 } 2348 2349 if (rbehind > cbehind) { 2350 rbehind = cbehind; 2351 } 2352 2353 /* 2354 * Do not do any readahead if we have insufficient free memory. 2355 * 2356 * XXX code was broken disabled before and has instability 2357 * with this conditonal fixed, so shortcut for now. 2358 */ 2359 if (burst_fault == 0 || vm_page_count_severe()) { 2360 marray[0] = m; 2361 *reqpage = 0; 2362 return 1; 2363 } 2364 2365 /* 2366 * scan backward for the read behind pages -- in memory 2367 * 2368 * Assume that if the page is not found an interrupt will not 2369 * create it. Theoretically interrupts can only remove (busy) 2370 * pages, not create new associations. 2371 */ 2372 if (pindex > 0) { 2373 if (rbehind > pindex) { 2374 rbehind = pindex; 2375 startpindex = 0; 2376 } else { 2377 startpindex = pindex - rbehind; 2378 } 2379 2380 vm_object_hold(object); 2381 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2382 if (vm_page_lookup(object, tpindex - 1)) 2383 break; 2384 } 2385 2386 i = 0; 2387 while (tpindex < pindex) { 2388 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2389 VM_ALLOC_NULL_OK); 2390 if (rtm == NULL) { 2391 for (j = 0; j < i; j++) { 2392 vm_page_free(marray[j]); 2393 } 2394 vm_object_drop(object); 2395 marray[0] = m; 2396 *reqpage = 0; 2397 return 1; 2398 } 2399 marray[i] = rtm; 2400 ++i; 2401 ++tpindex; 2402 } 2403 vm_object_drop(object); 2404 } else { 2405 i = 0; 2406 } 2407 2408 /* 2409 * Assign requested page 2410 */ 2411 marray[i] = m; 2412 *reqpage = i; 2413 ++i; 2414 2415 /* 2416 * Scan forwards for read-ahead pages 2417 */ 2418 tpindex = pindex + 1; 2419 endpindex = tpindex + rahead; 2420 if (endpindex > object->size) 2421 endpindex = object->size; 2422 2423 vm_object_hold(object); 2424 while (tpindex < endpindex) { 2425 if (vm_page_lookup(object, tpindex)) 2426 break; 2427 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2428 VM_ALLOC_NULL_OK); 2429 if (rtm == NULL) 2430 break; 2431 marray[i] = rtm; 2432 ++i; 2433 ++tpindex; 2434 } 2435 vm_object_drop(object); 2436 2437 return (i); 2438 } 2439 2440 #endif 2441 2442 /* 2443 * vm_prefault() provides a quick way of clustering pagefaults into a 2444 * processes address space. It is a "cousin" of pmap_object_init_pt, 2445 * except it runs at page fault time instead of mmap time. 2446 * 2447 * vm.fast_fault Enables pre-faulting zero-fill pages 2448 * 2449 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2450 * prefault. Scan stops in either direction when 2451 * a page is found to already exist. 2452 * 2453 * This code used to be per-platform pmap_prefault(). It is now 2454 * machine-independent and enhanced to also pre-fault zero-fill pages 2455 * (see vm.fast_fault) as well as make them writable, which greatly 2456 * reduces the number of page faults programs incur. 2457 * 2458 * Application performance when pre-faulting zero-fill pages is heavily 2459 * dependent on the application. Very tiny applications like /bin/echo 2460 * lose a little performance while applications of any appreciable size 2461 * gain performance. Prefaulting multiple pages also reduces SMP 2462 * congestion and can improve SMP performance significantly. 2463 * 2464 * NOTE! prot may allow writing but this only applies to the top level 2465 * object. If we wind up mapping a page extracted from a backing 2466 * object we have to make sure it is read-only. 2467 * 2468 * NOTE! The caller has already handled any COW operations on the 2469 * vm_map_entry via the normal fault code. Do NOT call this 2470 * shortcut unless the normal fault code has run on this entry. 2471 * 2472 * The related map must be locked. 2473 * No other requirements. 2474 */ 2475 static int vm_prefault_pages = 8; 2476 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2477 "Maximum number of pages to pre-fault"); 2478 static int vm_fast_fault = 1; 2479 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2480 "Burst fault zero-fill regions"); 2481 2482 /* 2483 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2484 * is not already dirty by other means. This will prevent passive 2485 * filesystem syncing as well as 'sync' from writing out the page. 2486 */ 2487 static void 2488 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2489 { 2490 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2491 if (m->dirty == 0) 2492 vm_page_flag_set(m, PG_NOSYNC); 2493 } else { 2494 vm_page_flag_clear(m, PG_NOSYNC); 2495 } 2496 } 2497 2498 static void 2499 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 2500 int fault_flags) 2501 { 2502 struct lwp *lp; 2503 vm_page_t m; 2504 vm_offset_t addr; 2505 vm_pindex_t index; 2506 vm_pindex_t pindex; 2507 vm_object_t object; 2508 int pprot; 2509 int i; 2510 int noneg; 2511 int nopos; 2512 int maxpages; 2513 2514 /* 2515 * Get stable max count value, disabled if set to 0 2516 */ 2517 maxpages = vm_prefault_pages; 2518 cpu_ccfence(); 2519 if (maxpages <= 0) 2520 return; 2521 2522 /* 2523 * We do not currently prefault mappings that use virtual page 2524 * tables. We do not prefault foreign pmaps. 2525 */ 2526 if (entry->maptype != VM_MAPTYPE_NORMAL) 2527 return; 2528 lp = curthread->td_lwp; 2529 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2530 return; 2531 2532 /* 2533 * Limit pre-fault count to 1024 pages. 2534 */ 2535 if (maxpages > 1024) 2536 maxpages = 1024; 2537 2538 object = entry->object.vm_object; 2539 KKASSERT(object != NULL); 2540 KKASSERT(object == entry->object.vm_object); 2541 vm_object_hold(object); 2542 vm_object_chain_acquire(object, 0); 2543 2544 noneg = 0; 2545 nopos = 0; 2546 for (i = 0; i < maxpages; ++i) { 2547 vm_object_t lobject; 2548 vm_object_t nobject; 2549 int allocated = 0; 2550 int error; 2551 2552 /* 2553 * This can eat a lot of time on a heavily contended 2554 * machine so yield on the tick if needed. 2555 */ 2556 if ((i & 7) == 7) 2557 lwkt_yield(); 2558 2559 /* 2560 * Calculate the page to pre-fault, stopping the scan in 2561 * each direction separately if the limit is reached. 2562 */ 2563 if (i & 1) { 2564 if (noneg) 2565 continue; 2566 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2567 } else { 2568 if (nopos) 2569 continue; 2570 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2571 } 2572 if (addr < entry->start) { 2573 noneg = 1; 2574 if (noneg && nopos) 2575 break; 2576 continue; 2577 } 2578 if (addr >= entry->end) { 2579 nopos = 1; 2580 if (noneg && nopos) 2581 break; 2582 continue; 2583 } 2584 2585 /* 2586 * Skip pages already mapped, and stop scanning in that 2587 * direction. When the scan terminates in both directions 2588 * we are done. 2589 */ 2590 if (pmap_prefault_ok(pmap, addr) == 0) { 2591 if (i & 1) 2592 noneg = 1; 2593 else 2594 nopos = 1; 2595 if (noneg && nopos) 2596 break; 2597 continue; 2598 } 2599 2600 /* 2601 * Follow the VM object chain to obtain the page to be mapped 2602 * into the pmap. 2603 * 2604 * If we reach the terminal object without finding a page 2605 * and we determine it would be advantageous, then allocate 2606 * a zero-fill page for the base object. The base object 2607 * is guaranteed to be OBJT_DEFAULT for this case. 2608 * 2609 * In order to not have to check the pager via *haspage*() 2610 * we stop if any non-default object is encountered. e.g. 2611 * a vnode or swap object would stop the loop. 2612 */ 2613 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2614 lobject = object; 2615 pindex = index; 2616 pprot = prot; 2617 2618 KKASSERT(lobject == entry->object.vm_object); 2619 /*vm_object_hold(lobject); implied */ 2620 2621 while ((m = vm_page_lookup_busy_try(lobject, pindex, 2622 TRUE, &error)) == NULL) { 2623 if (lobject->type != OBJT_DEFAULT) 2624 break; 2625 if (lobject->backing_object == NULL) { 2626 if (vm_fast_fault == 0) 2627 break; 2628 if ((prot & VM_PROT_WRITE) == 0 || 2629 vm_page_count_min(0)) { 2630 break; 2631 } 2632 2633 /* 2634 * NOTE: Allocated from base object 2635 */ 2636 m = vm_page_alloc(object, index, 2637 VM_ALLOC_NORMAL | 2638 VM_ALLOC_ZERO | 2639 VM_ALLOC_USE_GD | 2640 VM_ALLOC_NULL_OK); 2641 if (m == NULL) 2642 break; 2643 allocated = 1; 2644 pprot = prot; 2645 /* lobject = object .. not needed */ 2646 break; 2647 } 2648 if (lobject->backing_object_offset & PAGE_MASK) 2649 break; 2650 nobject = lobject->backing_object; 2651 vm_object_hold(nobject); 2652 KKASSERT(nobject == lobject->backing_object); 2653 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 2654 if (lobject != object) { 2655 vm_object_lock_swap(); 2656 vm_object_drop(lobject); 2657 } 2658 lobject = nobject; 2659 pprot &= ~VM_PROT_WRITE; 2660 vm_object_chain_acquire(lobject, 0); 2661 } 2662 2663 /* 2664 * NOTE: A non-NULL (m) will be associated with lobject if 2665 * it was found there, otherwise it is probably a 2666 * zero-fill page associated with the base object. 2667 * 2668 * Give-up if no page is available. 2669 */ 2670 if (m == NULL) { 2671 if (lobject != object) { 2672 #if 0 2673 if (object->backing_object != lobject) 2674 vm_object_hold(object->backing_object); 2675 #endif 2676 vm_object_chain_release_all( 2677 object->backing_object, lobject); 2678 #if 0 2679 if (object->backing_object != lobject) 2680 vm_object_drop(object->backing_object); 2681 #endif 2682 vm_object_drop(lobject); 2683 } 2684 break; 2685 } 2686 2687 /* 2688 * The object must be marked dirty if we are mapping a 2689 * writable page. m->object is either lobject or object, 2690 * both of which are still held. Do this before we 2691 * potentially drop the object. 2692 */ 2693 if (pprot & VM_PROT_WRITE) 2694 vm_object_set_writeable_dirty(m->object); 2695 2696 /* 2697 * Do not conditionalize on PG_RAM. If pages are present in 2698 * the VM system we assume optimal caching. If caching is 2699 * not optimal the I/O gravy train will be restarted when we 2700 * hit an unavailable page. We do not want to try to restart 2701 * the gravy train now because we really don't know how much 2702 * of the object has been cached. The cost for restarting 2703 * the gravy train should be low (since accesses will likely 2704 * be I/O bound anyway). 2705 */ 2706 if (lobject != object) { 2707 #if 0 2708 if (object->backing_object != lobject) 2709 vm_object_hold(object->backing_object); 2710 #endif 2711 vm_object_chain_release_all(object->backing_object, 2712 lobject); 2713 #if 0 2714 if (object->backing_object != lobject) 2715 vm_object_drop(object->backing_object); 2716 #endif 2717 vm_object_drop(lobject); 2718 } 2719 2720 /* 2721 * Enter the page into the pmap if appropriate. If we had 2722 * allocated the page we have to place it on a queue. If not 2723 * we just have to make sure it isn't on the cache queue 2724 * (pages on the cache queue are not allowed to be mapped). 2725 */ 2726 if (allocated) { 2727 /* 2728 * Page must be zerod. 2729 */ 2730 vm_page_zero_fill(m); 2731 mycpu->gd_cnt.v_zfod++; 2732 m->valid = VM_PAGE_BITS_ALL; 2733 2734 /* 2735 * Handle dirty page case 2736 */ 2737 if (pprot & VM_PROT_WRITE) 2738 vm_set_nosync(m, entry); 2739 pmap_enter(pmap, addr, m, pprot, 0, entry); 2740 mycpu->gd_cnt.v_vm_faults++; 2741 if (curthread->td_lwp) 2742 ++curthread->td_lwp->lwp_ru.ru_minflt; 2743 vm_page_deactivate(m); 2744 if (pprot & VM_PROT_WRITE) { 2745 /*vm_object_set_writeable_dirty(m->object);*/ 2746 vm_set_nosync(m, entry); 2747 if (fault_flags & VM_FAULT_DIRTY) { 2748 vm_page_dirty(m); 2749 /*XXX*/ 2750 swap_pager_unswapped(m); 2751 } 2752 } 2753 vm_page_wakeup(m); 2754 } else if (error) { 2755 /* couldn't busy page, no wakeup */ 2756 } else if ( 2757 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2758 (m->flags & PG_FICTITIOUS) == 0) { 2759 /* 2760 * A fully valid page not undergoing soft I/O can 2761 * be immediately entered into the pmap. 2762 */ 2763 if ((m->queue - m->pc) == PQ_CACHE) 2764 vm_page_deactivate(m); 2765 if (pprot & VM_PROT_WRITE) { 2766 /*vm_object_set_writeable_dirty(m->object);*/ 2767 vm_set_nosync(m, entry); 2768 if (fault_flags & VM_FAULT_DIRTY) { 2769 vm_page_dirty(m); 2770 /*XXX*/ 2771 swap_pager_unswapped(m); 2772 } 2773 } 2774 if (pprot & VM_PROT_WRITE) 2775 vm_set_nosync(m, entry); 2776 pmap_enter(pmap, addr, m, pprot, 0, entry); 2777 mycpu->gd_cnt.v_vm_faults++; 2778 if (curthread->td_lwp) 2779 ++curthread->td_lwp->lwp_ru.ru_minflt; 2780 vm_page_wakeup(m); 2781 } else { 2782 vm_page_wakeup(m); 2783 } 2784 } 2785 vm_object_chain_release(object); 2786 vm_object_drop(object); 2787 } 2788 2789 /* 2790 * Object can be held shared 2791 */ 2792 static void 2793 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 2794 vm_map_entry_t entry, int prot, int fault_flags) 2795 { 2796 struct lwp *lp; 2797 vm_page_t m; 2798 vm_offset_t addr; 2799 vm_pindex_t pindex; 2800 vm_object_t object; 2801 int i; 2802 int noneg; 2803 int nopos; 2804 int maxpages; 2805 2806 /* 2807 * Get stable max count value, disabled if set to 0 2808 */ 2809 maxpages = vm_prefault_pages; 2810 cpu_ccfence(); 2811 if (maxpages <= 0) 2812 return; 2813 2814 /* 2815 * We do not currently prefault mappings that use virtual page 2816 * tables. We do not prefault foreign pmaps. 2817 */ 2818 if (entry->maptype != VM_MAPTYPE_NORMAL) 2819 return; 2820 lp = curthread->td_lwp; 2821 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2822 return; 2823 object = entry->object.vm_object; 2824 if (object->backing_object != NULL) 2825 return; 2826 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2827 2828 /* 2829 * Limit pre-fault count to 1024 pages. 2830 */ 2831 if (maxpages > 1024) 2832 maxpages = 1024; 2833 2834 noneg = 0; 2835 nopos = 0; 2836 for (i = 0; i < maxpages; ++i) { 2837 int error; 2838 2839 /* 2840 * Calculate the page to pre-fault, stopping the scan in 2841 * each direction separately if the limit is reached. 2842 */ 2843 if (i & 1) { 2844 if (noneg) 2845 continue; 2846 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2847 } else { 2848 if (nopos) 2849 continue; 2850 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2851 } 2852 if (addr < entry->start) { 2853 noneg = 1; 2854 if (noneg && nopos) 2855 break; 2856 continue; 2857 } 2858 if (addr >= entry->end) { 2859 nopos = 1; 2860 if (noneg && nopos) 2861 break; 2862 continue; 2863 } 2864 2865 /* 2866 * Follow the VM object chain to obtain the page to be mapped 2867 * into the pmap. This version of the prefault code only 2868 * works with terminal objects. 2869 * 2870 * The page must already exist. If we encounter a problem 2871 * we stop here. 2872 * 2873 * WARNING! We cannot call swap_pager_unswapped() or insert 2874 * a new vm_page with a shared token. 2875 */ 2876 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2877 2878 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2879 if (m == NULL || error) 2880 break; 2881 2882 /* 2883 * Skip pages already mapped, and stop scanning in that 2884 * direction. When the scan terminates in both directions 2885 * we are done. 2886 */ 2887 if (pmap_prefault_ok(pmap, addr) == 0) { 2888 vm_page_wakeup(m); 2889 if (i & 1) 2890 noneg = 1; 2891 else 2892 nopos = 1; 2893 if (noneg && nopos) 2894 break; 2895 continue; 2896 } 2897 2898 /* 2899 * Stop if the page cannot be trivially entered into the 2900 * pmap. 2901 */ 2902 if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) || 2903 (m->flags & PG_FICTITIOUS) || 2904 ((m->flags & PG_SWAPPED) && 2905 (prot & VM_PROT_WRITE) && 2906 (fault_flags & VM_FAULT_DIRTY))) { 2907 vm_page_wakeup(m); 2908 break; 2909 } 2910 2911 /* 2912 * Enter the page into the pmap. The object might be held 2913 * shared so we can't do any (serious) modifying operation 2914 * on it. 2915 */ 2916 if ((m->queue - m->pc) == PQ_CACHE) 2917 vm_page_deactivate(m); 2918 if (prot & VM_PROT_WRITE) { 2919 vm_object_set_writeable_dirty(m->object); 2920 vm_set_nosync(m, entry); 2921 if (fault_flags & VM_FAULT_DIRTY) { 2922 vm_page_dirty(m); 2923 /* can't happeen due to conditional above */ 2924 /* swap_pager_unswapped(m); */ 2925 } 2926 } 2927 pmap_enter(pmap, addr, m, prot, 0, entry); 2928 mycpu->gd_cnt.v_vm_faults++; 2929 if (curthread->td_lwp) 2930 ++curthread->td_lwp->lwp_ru.ru_minflt; 2931 vm_page_wakeup(m); 2932 } 2933 } 2934