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