1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * Copyright (c) 1994 John S. Dyson 7 * All rights reserved. 8 * Copyright (c) 1994 David Greenman 9 * All rights reserved. 10 * 11 * 12 * This code is derived from software contributed to Berkeley by 13 * The Mach Operating System project at Carnegie-Mellon University. 14 * 15 * Redistribution and use in source and binary forms, with or without 16 * modification, are permitted provided that the following conditions 17 * are met: 18 * 1. Redistributions of source code must retain the above copyright 19 * notice, this list of conditions and the following disclaimer. 20 * 2. Redistributions in binary form must reproduce the above copyright 21 * notice, this list of conditions and the following disclaimer in the 22 * documentation and/or other materials provided with the distribution. 23 * 3. Neither the name of the University nor the names of its contributors 24 * may be used to endorse or promote products derived from this software 25 * without specific prior written permission. 26 * 27 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 28 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 29 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 30 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 31 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 32 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 33 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 34 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 35 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 36 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 37 * SUCH DAMAGE. 38 * 39 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94 40 * 41 * 42 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 43 * All rights reserved. 44 * 45 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 46 * 47 * Permission to use, copy, modify and distribute this software and 48 * its documentation is hereby granted, provided that both the copyright 49 * notice and this permission notice appear in all copies of the 50 * software, derivative works or modified versions, and any portions 51 * thereof, and that both notices appear in supporting documentation. 52 * 53 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 54 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 55 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 56 * 57 * Carnegie Mellon requests users of this software to return to 58 * 59 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 60 * School of Computer Science 61 * Carnegie Mellon University 62 * Pittsburgh PA 15213-3890 63 * 64 * any improvements or extensions that they make and grant Carnegie the 65 * rights to redistribute these changes. 66 * 67 * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $ 68 * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $ 69 */ 70 71 /* 72 * Page fault handling module. 73 */ 74 75 #include <sys/param.h> 76 #include <sys/systm.h> 77 #include <sys/kernel.h> 78 #include <sys/proc.h> 79 #include <sys/vnode.h> 80 #include <sys/resourcevar.h> 81 #include <sys/vmmeter.h> 82 #include <sys/vkernel.h> 83 #include <sys/lock.h> 84 #include <sys/sysctl.h> 85 86 #include <cpu/lwbuf.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vnode_pager.h> 98 #include <vm/vm_extern.h> 99 100 #include <sys/thread2.h> 101 #include <vm/vm_page2.h> 102 103 struct faultstate { 104 vm_page_t m; 105 vm_object_t object; 106 vm_pindex_t pindex; 107 vm_prot_t prot; 108 vm_page_t first_m; 109 vm_object_t first_object; 110 vm_prot_t first_prot; 111 vm_map_t map; 112 vm_map_entry_t entry; 113 int lookup_still_valid; 114 int hardfault; 115 int fault_flags; 116 int map_generation; 117 int shared; 118 boolean_t wired; 119 struct vnode *vp; 120 }; 121 122 static int debug_cluster = 0; 123 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, ""); 124 int vm_shared_fault = 1; 125 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, &vm_shared_fault, 0, 126 "Allow shared token on vm_object"); 127 static long vm_shared_hit = 0; 128 SYSCTL_LONG(_vm, OID_AUTO, shared_hit, CTLFLAG_RW, &vm_shared_hit, 0, 129 "Successful shared faults"); 130 static long vm_shared_miss = 0; 131 SYSCTL_LONG(_vm, OID_AUTO, shared_miss, CTLFLAG_RW, &vm_shared_miss, 0, 132 "Unsuccessful shared faults"); 133 134 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int); 135 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, 136 vpte_t, int, int); 137 #if 0 138 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 139 #endif 140 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry); 141 static void vm_prefault(pmap_t pmap, vm_offset_t addra, 142 vm_map_entry_t entry, int prot, int fault_flags); 143 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 144 vm_map_entry_t entry, int prot, int fault_flags); 145 146 static __inline void 147 release_page(struct faultstate *fs) 148 { 149 vm_page_deactivate(fs->m); 150 vm_page_wakeup(fs->m); 151 fs->m = NULL; 152 } 153 154 /* 155 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse 156 * requires relocking and then checking the timestamp. 157 * 158 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do 159 * not have to update fs->map_generation here. 160 * 161 * NOTE: This function can fail due to a deadlock against the caller's 162 * holding of a vm_page BUSY. 163 */ 164 static __inline int 165 relock_map(struct faultstate *fs) 166 { 167 int error; 168 169 if (fs->lookup_still_valid == FALSE && fs->map) { 170 error = vm_map_lock_read_to(fs->map); 171 if (error == 0) 172 fs->lookup_still_valid = TRUE; 173 } else { 174 error = 0; 175 } 176 return error; 177 } 178 179 static __inline void 180 unlock_map(struct faultstate *fs) 181 { 182 if (fs->lookup_still_valid && fs->map) { 183 vm_map_lookup_done(fs->map, fs->entry, 0); 184 fs->lookup_still_valid = FALSE; 185 } 186 } 187 188 /* 189 * Clean up after a successful call to vm_fault_object() so another call 190 * to vm_fault_object() can be made. 191 */ 192 static void 193 _cleanup_successful_fault(struct faultstate *fs, int relock) 194 { 195 if (fs->object != fs->first_object) { 196 vm_page_free(fs->first_m); 197 vm_object_pip_wakeup(fs->object); 198 fs->first_m = NULL; 199 } 200 fs->object = fs->first_object; 201 if (relock && fs->lookup_still_valid == FALSE) { 202 if (fs->map) 203 vm_map_lock_read(fs->map); 204 fs->lookup_still_valid = TRUE; 205 } 206 } 207 208 static void 209 _unlock_things(struct faultstate *fs, int dealloc) 210 { 211 _cleanup_successful_fault(fs, 0); 212 if (dealloc) { 213 /*vm_object_deallocate(fs->first_object);*/ 214 /*fs->first_object = NULL; drop used later on */ 215 } 216 unlock_map(fs); 217 if (fs->vp != NULL) { 218 vput(fs->vp); 219 fs->vp = NULL; 220 } 221 } 222 223 #define unlock_things(fs) _unlock_things(fs, 0) 224 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 225 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 226 227 /* 228 * TRYPAGER 229 * 230 * Determine if the pager for the current object *might* contain the page. 231 * 232 * We only need to try the pager if this is not a default object (default 233 * objects are zero-fill and have no real pager), and if we are not taking 234 * a wiring fault or if the FS entry is wired. 235 */ 236 #define TRYPAGER(fs) \ 237 (fs->object->type != OBJT_DEFAULT && \ 238 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired)) 239 240 /* 241 * vm_fault: 242 * 243 * Handle a page fault occuring at the given address, requiring the given 244 * permissions, in the map specified. If successful, the page is inserted 245 * into the associated physical map. 246 * 247 * NOTE: The given address should be truncated to the proper page address. 248 * 249 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 250 * a standard error specifying why the fault is fatal is returned. 251 * 252 * The map in question must be referenced, and remains so. 253 * The caller may hold no locks. 254 * No other requirements. 255 */ 256 int 257 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 258 { 259 int result; 260 vm_pindex_t first_pindex; 261 struct faultstate fs; 262 struct lwp *lp; 263 int growstack; 264 int retry = 0; 265 266 vm_page_pcpu_cache(); 267 fs.hardfault = 0; 268 fs.fault_flags = fault_flags; 269 fs.vp = NULL; 270 growstack = 1; 271 272 if ((lp = curthread->td_lwp) != NULL) 273 lp->lwp_flags |= LWP_PAGING; 274 275 lwkt_gettoken(&map->token); 276 277 RetryFault: 278 /* 279 * Find the vm_map_entry representing the backing store and resolve 280 * the top level object and page index. This may have the side 281 * effect of executing a copy-on-write on the map entry and/or 282 * creating a shadow object, but will not COW any actual VM pages. 283 * 284 * On success fs.map is left read-locked and various other fields 285 * are initialized but not otherwise referenced or locked. 286 * 287 * NOTE! vm_map_lookup will try to upgrade the fault_type to 288 * VM_FAULT_WRITE if the map entry is a virtual page table and also 289 * writable, so we can set the 'A'accessed bit in the virtual page 290 * table entry. 291 */ 292 fs.map = map; 293 result = vm_map_lookup(&fs.map, vaddr, fault_type, 294 &fs.entry, &fs.first_object, 295 &first_pindex, &fs.first_prot, &fs.wired); 296 297 /* 298 * If the lookup failed or the map protections are incompatible, 299 * the fault generally fails. However, if the caller is trying 300 * to do a user wiring we have more work to do. 301 */ 302 if (result != KERN_SUCCESS) { 303 if (result != KERN_PROTECTION_FAILURE || 304 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 305 { 306 if (result == KERN_INVALID_ADDRESS && growstack && 307 map != &kernel_map && curproc != NULL) { 308 result = vm_map_growstack(curproc, vaddr); 309 if (result == KERN_SUCCESS) { 310 growstack = 0; 311 ++retry; 312 goto RetryFault; 313 } 314 result = KERN_FAILURE; 315 } 316 goto done; 317 } 318 319 /* 320 * If we are user-wiring a r/w segment, and it is COW, then 321 * we need to do the COW operation. Note that we don't 322 * currently COW RO sections now, because it is NOT desirable 323 * to COW .text. We simply keep .text from ever being COW'ed 324 * and take the heat that one cannot debug wired .text sections. 325 */ 326 result = vm_map_lookup(&fs.map, vaddr, 327 VM_PROT_READ|VM_PROT_WRITE| 328 VM_PROT_OVERRIDE_WRITE, 329 &fs.entry, &fs.first_object, 330 &first_pindex, &fs.first_prot, 331 &fs.wired); 332 if (result != KERN_SUCCESS) { 333 result = KERN_FAILURE; 334 goto done; 335 } 336 337 /* 338 * If we don't COW now, on a user wire, the user will never 339 * be able to write to the mapping. If we don't make this 340 * restriction, the bookkeeping would be nearly impossible. 341 * 342 * XXX We have a shared lock, this will have a MP race but 343 * I don't see how it can hurt anything. 344 */ 345 if ((fs.entry->protection & VM_PROT_WRITE) == 0) 346 fs.entry->max_protection &= ~VM_PROT_WRITE; 347 } 348 349 /* 350 * fs.map is read-locked 351 * 352 * Misc checks. Save the map generation number to detect races. 353 */ 354 fs.map_generation = fs.map->timestamp; 355 fs.lookup_still_valid = TRUE; 356 fs.first_m = NULL; 357 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 358 fs.shared = 0; 359 fs.vp = NULL; 360 361 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) { 362 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 363 panic("vm_fault: fault on nofault entry, addr: %p", 364 (void *)vaddr); 365 } 366 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) && 367 vaddr >= fs.entry->start && 368 vaddr < fs.entry->start + PAGE_SIZE) { 369 panic("vm_fault: fault on stack guard, addr: %p", 370 (void *)vaddr); 371 } 372 } 373 374 /* 375 * A system map entry may return a NULL object. No object means 376 * no pager means an unrecoverable kernel fault. 377 */ 378 if (fs.first_object == NULL) { 379 panic("vm_fault: unrecoverable fault at %p in entry %p", 380 (void *)vaddr, fs.entry); 381 } 382 383 /* 384 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 385 * is set. 386 */ 387 if ((curthread->td_flags & TDF_NOFAULT) && 388 (retry || 389 fs.first_object->type == OBJT_VNODE || 390 fs.first_object->backing_object)) { 391 result = KERN_FAILURE; 392 unlock_things(&fs); 393 goto done2; 394 } 395 396 /* 397 * Attempt to shortcut the fault if the lookup returns a 398 * terminal object and the page is present. This allows us 399 * to obtain a shared token on the object instead of an exclusive 400 * token, which theoretically should allow concurrent faults. 401 * 402 * We cannot acquire a shared token on kernel_map, at least not 403 * on i386, because the i386 pmap code uses the kernel_object for 404 * its page table page management, resulting in a shared->exclusive 405 * sequence which will deadlock. This will not happen normally 406 * anyway, except on well cached pageable kmem (like pipe buffers), 407 * so it should not impact performance. 408 */ 409 if (vm_shared_fault && 410 fs.first_object->backing_object == NULL && 411 fs.entry->maptype == VM_MAPTYPE_NORMAL && 412 fs.map != &kernel_map) { 413 int error; 414 vm_object_hold_shared(fs.first_object); 415 /*fs.vp = vnode_pager_lock(fs.first_object);*/ 416 fs.m = vm_page_lookup_busy_try(fs.first_object, 417 first_pindex, 418 TRUE, &error); 419 if (error == 0 && fs.m) { 420 /* 421 * Activate the page and figure out if we can 422 * short-cut a quick mapping. 423 * 424 * WARNING! We cannot call swap_pager_unswapped() 425 * with a shared token! Note that we 426 * have to test fs.first_prot here. 427 */ 428 vm_page_activate(fs.m); 429 if (fs.m->valid == VM_PAGE_BITS_ALL && 430 ((fs.m->flags & PG_SWAPPED) == 0 || 431 (fs.first_prot & VM_PROT_WRITE) == 0 || 432 (fs.fault_flags & VM_FAULT_DIRTY) == 0)) { 433 fs.lookup_still_valid = TRUE; 434 fs.first_m = NULL; 435 fs.object = fs.first_object; 436 fs.prot = fs.first_prot; 437 if (fs.wired) 438 fault_type = fs.first_prot; 439 if (fs.prot & VM_PROT_WRITE) { 440 vm_object_set_writeable_dirty( 441 fs.m->object); 442 vm_set_nosync(fs.m, fs.entry); 443 if (fs.fault_flags & VM_FAULT_DIRTY) { 444 vm_page_dirty(fs.m); 445 /*XXX*/ 446 swap_pager_unswapped(fs.m); 447 } 448 } 449 result = KERN_SUCCESS; 450 fault_flags |= VM_FAULT_BURST_QUICK; 451 fault_flags &= ~VM_FAULT_BURST; 452 ++vm_shared_hit; 453 goto quick; 454 } 455 vm_page_wakeup(fs.m); 456 fs.m = NULL; 457 } 458 vm_object_drop(fs.first_object); /* XXX drop on shared tok?*/ 459 } 460 ++vm_shared_miss; 461 462 /* 463 * Bump the paging-in-progress count to prevent size changes (e.g. 464 * truncation operations) during I/O. This must be done after 465 * obtaining the vnode lock in order to avoid possible deadlocks. 466 */ 467 vm_object_hold(fs.first_object); 468 if (fs.vp == NULL) 469 fs.vp = vnode_pager_lock(fs.first_object); 470 471 #if 0 472 fs.lookup_still_valid = TRUE; 473 fs.first_m = NULL; 474 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 475 fs.shared = 0; 476 #endif 477 478 /* 479 * If the entry is wired we cannot change the page protection. 480 */ 481 if (fs.wired) 482 fault_type = fs.first_prot; 483 484 /* 485 * The page we want is at (first_object, first_pindex), but if the 486 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 487 * page table to figure out the actual pindex. 488 * 489 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 490 * ONLY 491 */ 492 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 493 result = vm_fault_vpagetable(&fs, &first_pindex, 494 fs.entry->aux.master_pde, 495 fault_type, 1); 496 if (result == KERN_TRY_AGAIN) { 497 vm_object_drop(fs.first_object); 498 ++retry; 499 goto RetryFault; 500 } 501 if (result != KERN_SUCCESS) 502 goto done; 503 } 504 505 /* 506 * Now we have the actual (object, pindex), fault in the page. If 507 * vm_fault_object() fails it will unlock and deallocate the FS 508 * data. If it succeeds everything remains locked and fs->object 509 * will have an additional PIP count if it is not equal to 510 * fs->first_object 511 * 512 * vm_fault_object will set fs->prot for the pmap operation. It is 513 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 514 * page can be safely written. However, it will force a read-only 515 * mapping for a read fault if the memory is managed by a virtual 516 * page table. 517 * 518 * If the fault code uses the shared object lock shortcut 519 * we must not try to burst (we can't allocate VM pages). 520 */ 521 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 522 if (fs.shared) 523 fault_flags &= ~VM_FAULT_BURST; 524 525 if (result == KERN_TRY_AGAIN) { 526 vm_object_drop(fs.first_object); 527 ++retry; 528 goto RetryFault; 529 } 530 if (result != KERN_SUCCESS) 531 goto done; 532 533 quick: 534 /* 535 * On success vm_fault_object() does not unlock or deallocate, and fs.m 536 * will contain a busied page. 537 * 538 * Enter the page into the pmap and do pmap-related adjustments. 539 */ 540 KKASSERT(fs.lookup_still_valid == TRUE); 541 vm_page_flag_set(fs.m, PG_REFERENCED); 542 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, fs.entry); 543 mycpu->gd_cnt.v_vm_faults++; 544 if (curthread->td_lwp) 545 ++curthread->td_lwp->lwp_ru.ru_minflt; 546 547 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */ 548 KKASSERT(fs.m->flags & PG_BUSY); 549 550 /* 551 * If the page is not wired down, then put it where the pageout daemon 552 * can find it. 553 */ 554 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 555 if (fs.wired) 556 vm_page_wire(fs.m); 557 else 558 vm_page_unwire(fs.m, 1); 559 } else { 560 vm_page_activate(fs.m); 561 } 562 vm_page_wakeup(fs.m); 563 564 /* 565 * Burst in a few more pages if possible. The fs.map should still 566 * be locked. To avoid interlocking against a vnode->getblk 567 * operation we had to be sure to unbusy our primary vm_page above 568 * first. 569 */ 570 if (fault_flags & VM_FAULT_BURST) { 571 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 572 && fs.wired == 0) { 573 vm_prefault(fs.map->pmap, vaddr, 574 fs.entry, fs.prot, fault_flags); 575 } 576 } 577 if (fault_flags & VM_FAULT_BURST_QUICK) { 578 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 579 && fs.wired == 0) { 580 vm_prefault_quick(fs.map->pmap, vaddr, 581 fs.entry, fs.prot, fault_flags); 582 } 583 } 584 585 /* 586 * Unlock everything, and return 587 */ 588 unlock_things(&fs); 589 590 if (curthread->td_lwp) { 591 if (fs.hardfault) { 592 curthread->td_lwp->lwp_ru.ru_majflt++; 593 } else { 594 curthread->td_lwp->lwp_ru.ru_minflt++; 595 } 596 } 597 598 /*vm_object_deallocate(fs.first_object);*/ 599 /*fs.m = NULL; */ 600 /*fs.first_object = NULL; must still drop later */ 601 602 result = KERN_SUCCESS; 603 done: 604 if (fs.first_object) 605 vm_object_drop(fs.first_object); 606 done2: 607 lwkt_reltoken(&map->token); 608 if (lp) 609 lp->lwp_flags &= ~LWP_PAGING; 610 return (result); 611 } 612 613 /* 614 * Fault in the specified virtual address in the current process map, 615 * returning a held VM page or NULL. See vm_fault_page() for more 616 * information. 617 * 618 * No requirements. 619 */ 620 vm_page_t 621 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp) 622 { 623 struct lwp *lp = curthread->td_lwp; 624 vm_page_t m; 625 626 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 627 fault_type, VM_FAULT_NORMAL, errorp); 628 return(m); 629 } 630 631 /* 632 * Fault in the specified virtual address in the specified map, doing all 633 * necessary manipulation of the object store and all necessary I/O. Return 634 * a held VM page or NULL, and set *errorp. The related pmap is not 635 * updated. 636 * 637 * The returned page will be properly dirtied if VM_PROT_WRITE was specified, 638 * and marked PG_REFERENCED as well. 639 * 640 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 641 * error will be returned. 642 * 643 * No requirements. 644 */ 645 vm_page_t 646 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 647 int fault_flags, int *errorp) 648 { 649 vm_pindex_t first_pindex; 650 struct faultstate fs; 651 int result; 652 int retry = 0; 653 vm_prot_t orig_fault_type = fault_type; 654 655 fs.hardfault = 0; 656 fs.fault_flags = fault_flags; 657 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 658 659 /* 660 * Dive the pmap (concurrency possible). If we find the 661 * appropriate page we can terminate early and quickly. 662 */ 663 fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type); 664 if (fs.m) { 665 *errorp = 0; 666 return(fs.m); 667 } 668 669 /* 670 * Otherwise take a concurrency hit and do a formal page 671 * fault. 672 */ 673 lwkt_gettoken(&map->token); 674 675 RetryFault: 676 /* 677 * Find the vm_map_entry representing the backing store and resolve 678 * the top level object and page index. This may have the side 679 * effect of executing a copy-on-write on the map entry and/or 680 * creating a shadow object, but will not COW any actual VM pages. 681 * 682 * On success fs.map is left read-locked and various other fields 683 * are initialized but not otherwise referenced or locked. 684 * 685 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 686 * if the map entry is a virtual page table and also writable, 687 * so we can set the 'A'accessed bit in the virtual page table entry. 688 */ 689 fs.map = map; 690 result = vm_map_lookup(&fs.map, vaddr, fault_type, 691 &fs.entry, &fs.first_object, 692 &first_pindex, &fs.first_prot, &fs.wired); 693 694 if (result != KERN_SUCCESS) { 695 *errorp = result; 696 fs.m = NULL; 697 goto done; 698 } 699 700 /* 701 * fs.map is read-locked 702 * 703 * Misc checks. Save the map generation number to detect races. 704 */ 705 fs.map_generation = fs.map->timestamp; 706 fs.lookup_still_valid = TRUE; 707 fs.first_m = NULL; 708 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 709 fs.shared = 0; 710 fs.vp = NULL; 711 712 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 713 panic("vm_fault: fault on nofault entry, addr: %lx", 714 (u_long)vaddr); 715 } 716 717 /* 718 * A system map entry may return a NULL object. No object means 719 * no pager means an unrecoverable kernel fault. 720 */ 721 if (fs.first_object == NULL) { 722 panic("vm_fault: unrecoverable fault at %p in entry %p", 723 (void *)vaddr, fs.entry); 724 } 725 726 /* 727 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 728 * is set. 729 */ 730 if ((curthread->td_flags & TDF_NOFAULT) && 731 (retry || 732 fs.first_object->type == OBJT_VNODE || 733 fs.first_object->backing_object)) { 734 *errorp = KERN_FAILURE; 735 unlock_things(&fs); 736 goto done2; 737 } 738 739 /* 740 * Make a reference to this object to prevent its disposal while we 741 * are messing with it. Once we have the reference, the map is free 742 * to be diddled. Since objects reference their shadows (and copies), 743 * they will stay around as well. 744 * 745 * The reference should also prevent an unexpected collapse of the 746 * parent that might move pages from the current object into the 747 * parent unexpectedly, resulting in corruption. 748 * 749 * Bump the paging-in-progress count to prevent size changes (e.g. 750 * truncation operations) during I/O. This must be done after 751 * obtaining the vnode lock in order to avoid possible deadlocks. 752 */ 753 vm_object_hold(fs.first_object); 754 fs.vp = vnode_pager_lock(fs.first_object); 755 756 #if 0 757 fs.lookup_still_valid = TRUE; 758 fs.first_m = NULL; 759 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 760 fs.shared = 0; 761 #endif 762 763 /* 764 * If the entry is wired we cannot change the page protection. 765 */ 766 if (fs.wired) 767 fault_type = fs.first_prot; 768 769 /* 770 * The page we want is at (first_object, first_pindex), but if the 771 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 772 * page table to figure out the actual pindex. 773 * 774 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 775 * ONLY 776 */ 777 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 778 result = vm_fault_vpagetable(&fs, &first_pindex, 779 fs.entry->aux.master_pde, 780 fault_type, 1); 781 if (result == KERN_TRY_AGAIN) { 782 vm_object_drop(fs.first_object); 783 ++retry; 784 goto RetryFault; 785 } 786 if (result != KERN_SUCCESS) { 787 *errorp = result; 788 fs.m = NULL; 789 goto done; 790 } 791 } 792 793 /* 794 * Now we have the actual (object, pindex), fault in the page. If 795 * vm_fault_object() fails it will unlock and deallocate the FS 796 * data. If it succeeds everything remains locked and fs->object 797 * will have an additinal PIP count if it is not equal to 798 * fs->first_object 799 */ 800 fs.m = NULL; 801 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 802 803 if (result == KERN_TRY_AGAIN) { 804 vm_object_drop(fs.first_object); 805 ++retry; 806 goto RetryFault; 807 } 808 if (result != KERN_SUCCESS) { 809 *errorp = result; 810 fs.m = NULL; 811 goto done; 812 } 813 814 if ((orig_fault_type & VM_PROT_WRITE) && 815 (fs.prot & VM_PROT_WRITE) == 0) { 816 *errorp = KERN_PROTECTION_FAILURE; 817 unlock_and_deallocate(&fs); 818 fs.m = NULL; 819 goto done; 820 } 821 822 /* 823 * DO NOT UPDATE THE PMAP!!! This function may be called for 824 * a pmap unrelated to the current process pmap, in which case 825 * the current cpu core will not be listed in the pmap's pm_active 826 * mask. Thus invalidation interlocks will fail to work properly. 827 * 828 * (for example, 'ps' uses procfs to read program arguments from 829 * each process's stack). 830 * 831 * In addition to the above this function will be called to acquire 832 * a page that might already be faulted in, re-faulting it 833 * continuously is a waste of time. 834 * 835 * XXX could this have been the cause of our random seg-fault 836 * issues? procfs accesses user stacks. 837 */ 838 vm_page_flag_set(fs.m, PG_REFERENCED); 839 #if 0 840 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL); 841 mycpu->gd_cnt.v_vm_faults++; 842 if (curthread->td_lwp) 843 ++curthread->td_lwp->lwp_ru.ru_minflt; 844 #endif 845 846 /* 847 * On success vm_fault_object() does not unlock or deallocate, and fs.m 848 * will contain a busied page. So we must unlock here after having 849 * messed with the pmap. 850 */ 851 unlock_things(&fs); 852 853 /* 854 * Return a held page. We are not doing any pmap manipulation so do 855 * not set PG_MAPPED. However, adjust the page flags according to 856 * the fault type because the caller may not use a managed pmapping 857 * (so we don't want to lose the fact that the page will be dirtied 858 * if a write fault was specified). 859 */ 860 vm_page_hold(fs.m); 861 vm_page_activate(fs.m); 862 if (fault_type & VM_PROT_WRITE) 863 vm_page_dirty(fs.m); 864 865 if (curthread->td_lwp) { 866 if (fs.hardfault) { 867 curthread->td_lwp->lwp_ru.ru_majflt++; 868 } else { 869 curthread->td_lwp->lwp_ru.ru_minflt++; 870 } 871 } 872 873 /* 874 * Unlock everything, and return the held page. 875 */ 876 vm_page_wakeup(fs.m); 877 /*vm_object_deallocate(fs.first_object);*/ 878 /*fs.first_object = NULL; */ 879 *errorp = 0; 880 881 done: 882 if (fs.first_object) 883 vm_object_drop(fs.first_object); 884 done2: 885 lwkt_reltoken(&map->token); 886 return(fs.m); 887 } 888 889 /* 890 * Fault in the specified (object,offset), dirty the returned page as 891 * needed. If the requested fault_type cannot be done NULL and an 892 * error is returned. 893 * 894 * A held (but not busied) page is returned. 895 * 896 * No requirements. 897 */ 898 vm_page_t 899 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 900 vm_prot_t fault_type, int fault_flags, 901 int shared, int *errorp) 902 { 903 int result; 904 vm_pindex_t first_pindex; 905 struct faultstate fs; 906 struct vm_map_entry entry; 907 908 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 909 bzero(&entry, sizeof(entry)); 910 entry.object.vm_object = object; 911 entry.maptype = VM_MAPTYPE_NORMAL; 912 entry.protection = entry.max_protection = fault_type; 913 914 fs.hardfault = 0; 915 fs.fault_flags = fault_flags; 916 fs.map = NULL; 917 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 918 919 RetryFault: 920 921 fs.first_object = object; 922 first_pindex = OFF_TO_IDX(offset); 923 fs.entry = &entry; 924 fs.first_prot = fault_type; 925 fs.wired = 0; 926 fs.shared = shared; 927 /*fs.map_generation = 0; unused */ 928 929 /* 930 * Make a reference to this object to prevent its disposal while we 931 * are messing with it. Once we have the reference, the map is free 932 * to be diddled. Since objects reference their shadows (and copies), 933 * they will stay around as well. 934 * 935 * The reference should also prevent an unexpected collapse of the 936 * parent that might move pages from the current object into the 937 * parent unexpectedly, resulting in corruption. 938 * 939 * Bump the paging-in-progress count to prevent size changes (e.g. 940 * truncation operations) during I/O. This must be done after 941 * obtaining the vnode lock in order to avoid possible deadlocks. 942 */ 943 fs.vp = vnode_pager_lock(fs.first_object); 944 945 fs.lookup_still_valid = TRUE; 946 fs.first_m = NULL; 947 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 948 949 #if 0 950 /* XXX future - ability to operate on VM object using vpagetable */ 951 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 952 result = vm_fault_vpagetable(&fs, &first_pindex, 953 fs.entry->aux.master_pde, 954 fault_type, 0); 955 if (result == KERN_TRY_AGAIN) 956 goto RetryFault; 957 if (result != KERN_SUCCESS) { 958 *errorp = result; 959 return (NULL); 960 } 961 } 962 #endif 963 964 /* 965 * Now we have the actual (object, pindex), fault in the page. If 966 * vm_fault_object() fails it will unlock and deallocate the FS 967 * data. If it succeeds everything remains locked and fs->object 968 * will have an additinal PIP count if it is not equal to 969 * fs->first_object 970 */ 971 result = vm_fault_object(&fs, first_pindex, fault_type, 0); 972 973 if (result == KERN_TRY_AGAIN) 974 goto RetryFault; 975 if (result != KERN_SUCCESS) { 976 *errorp = result; 977 return(NULL); 978 } 979 980 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 981 *errorp = KERN_PROTECTION_FAILURE; 982 unlock_and_deallocate(&fs); 983 return(NULL); 984 } 985 986 /* 987 * On success vm_fault_object() does not unlock or deallocate, so we 988 * do it here. Note that the returned fs.m will be busied. 989 */ 990 unlock_things(&fs); 991 992 /* 993 * Return a held page. We are not doing any pmap manipulation so do 994 * not set PG_MAPPED. However, adjust the page flags according to 995 * the fault type because the caller may not use a managed pmapping 996 * (so we don't want to lose the fact that the page will be dirtied 997 * if a write fault was specified). 998 */ 999 vm_page_hold(fs.m); 1000 vm_page_activate(fs.m); 1001 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY)) 1002 vm_page_dirty(fs.m); 1003 if (fault_flags & VM_FAULT_UNSWAP) 1004 swap_pager_unswapped(fs.m); 1005 1006 /* 1007 * Indicate that the page was accessed. 1008 */ 1009 vm_page_flag_set(fs.m, PG_REFERENCED); 1010 1011 if (curthread->td_lwp) { 1012 if (fs.hardfault) { 1013 curthread->td_lwp->lwp_ru.ru_majflt++; 1014 } else { 1015 curthread->td_lwp->lwp_ru.ru_minflt++; 1016 } 1017 } 1018 1019 /* 1020 * Unlock everything, and return the held page. 1021 */ 1022 vm_page_wakeup(fs.m); 1023 /*vm_object_deallocate(fs.first_object);*/ 1024 /*fs.first_object = NULL; */ 1025 1026 *errorp = 0; 1027 return(fs.m); 1028 } 1029 1030 /* 1031 * Translate the virtual page number (first_pindex) that is relative 1032 * to the address space into a logical page number that is relative to the 1033 * backing object. Use the virtual page table pointed to by (vpte). 1034 * 1035 * This implements an N-level page table. Any level can terminate the 1036 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 1037 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 1038 */ 1039 static 1040 int 1041 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 1042 vpte_t vpte, int fault_type, int allow_nofault) 1043 { 1044 struct lwbuf *lwb; 1045 struct lwbuf lwb_cache; 1046 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */ 1047 int result = KERN_SUCCESS; 1048 vpte_t *ptep; 1049 1050 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1051 for (;;) { 1052 /* 1053 * We cannot proceed if the vpte is not valid, not readable 1054 * for a read fault, or not writable for a write fault. 1055 */ 1056 if ((vpte & VPTE_V) == 0) { 1057 unlock_and_deallocate(fs); 1058 return (KERN_FAILURE); 1059 } 1060 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) { 1061 unlock_and_deallocate(fs); 1062 return (KERN_FAILURE); 1063 } 1064 if ((vpte & VPTE_PS) || vshift == 0) 1065 break; 1066 KKASSERT(vshift >= VPTE_PAGE_BITS); 1067 1068 /* 1069 * Get the page table page. Nominally we only read the page 1070 * table, but since we are actively setting VPTE_M and VPTE_A, 1071 * tell vm_fault_object() that we are writing it. 1072 * 1073 * There is currently no real need to optimize this. 1074 */ 1075 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT, 1076 VM_PROT_READ|VM_PROT_WRITE, 1077 allow_nofault); 1078 if (result != KERN_SUCCESS) 1079 return (result); 1080 1081 /* 1082 * Process the returned fs.m and look up the page table 1083 * entry in the page table page. 1084 */ 1085 vshift -= VPTE_PAGE_BITS; 1086 lwb = lwbuf_alloc(fs->m, &lwb_cache); 1087 ptep = ((vpte_t *)lwbuf_kva(lwb) + 1088 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 1089 vpte = *ptep; 1090 1091 /* 1092 * Page table write-back. If the vpte is valid for the 1093 * requested operation, do a write-back to the page table. 1094 * 1095 * XXX VPTE_M is not set properly for page directory pages. 1096 * It doesn't get set in the page directory if the page table 1097 * is modified during a read access. 1098 */ 1099 vm_page_activate(fs->m); 1100 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) && 1101 (vpte & VPTE_RW)) { 1102 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) { 1103 atomic_set_long(ptep, VPTE_M | VPTE_A); 1104 vm_page_dirty(fs->m); 1105 } 1106 } 1107 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V)) { 1108 if ((vpte & VPTE_A) == 0) { 1109 atomic_set_long(ptep, VPTE_A); 1110 vm_page_dirty(fs->m); 1111 } 1112 } 1113 lwbuf_free(lwb); 1114 vm_page_flag_set(fs->m, PG_REFERENCED); 1115 vm_page_wakeup(fs->m); 1116 fs->m = NULL; 1117 cleanup_successful_fault(fs); 1118 } 1119 /* 1120 * Combine remaining address bits with the vpte. 1121 */ 1122 /* JG how many bits from each? */ 1123 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) + 1124 (*pindex & ((1L << vshift) - 1)); 1125 return (KERN_SUCCESS); 1126 } 1127 1128 1129 /* 1130 * This is the core of the vm_fault code. 1131 * 1132 * Do all operations required to fault-in (fs.first_object, pindex). Run 1133 * through the shadow chain as necessary and do required COW or virtual 1134 * copy operations. The caller has already fully resolved the vm_map_entry 1135 * and, if appropriate, has created a copy-on-write layer. All we need to 1136 * do is iterate the object chain. 1137 * 1138 * On failure (fs) is unlocked and deallocated and the caller may return or 1139 * retry depending on the failure code. On success (fs) is NOT unlocked or 1140 * deallocated, fs.m will contained a resolved, busied page, and fs.object 1141 * will have an additional PIP count if it is not equal to fs.first_object. 1142 * 1143 * fs->first_object must be held on call. 1144 */ 1145 static 1146 int 1147 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex, 1148 vm_prot_t fault_type, int allow_nofault) 1149 { 1150 vm_object_t next_object; 1151 vm_pindex_t pindex; 1152 int error; 1153 1154 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1155 fs->prot = fs->first_prot; 1156 fs->object = fs->first_object; 1157 pindex = first_pindex; 1158 1159 vm_object_chain_acquire(fs->first_object); 1160 vm_object_pip_add(fs->first_object, 1); 1161 1162 /* 1163 * If a read fault occurs we try to make the page writable if 1164 * possible. There are three cases where we cannot make the 1165 * page mapping writable: 1166 * 1167 * (1) The mapping is read-only or the VM object is read-only, 1168 * fs->prot above will simply not have VM_PROT_WRITE set. 1169 * 1170 * (2) If the mapping is a virtual page table we need to be able 1171 * to detect writes so we can set VPTE_M in the virtual page 1172 * table. 1173 * 1174 * (3) If the VM page is read-only or copy-on-write, upgrading would 1175 * just result in an unnecessary COW fault. 1176 * 1177 * VM_PROT_VPAGED is set if faulting via a virtual page table and 1178 * causes adjustments to the 'M'odify bit to also turn off write 1179 * access to force a re-fault. 1180 */ 1181 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1182 if ((fault_type & VM_PROT_WRITE) == 0) 1183 fs->prot &= ~VM_PROT_WRITE; 1184 } 1185 1186 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 1187 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 1188 if ((fault_type & VM_PROT_WRITE) == 0) 1189 fs->prot &= ~VM_PROT_WRITE; 1190 } 1191 1192 /* vm_object_hold(fs->object); implied b/c object == first_object */ 1193 1194 for (;;) { 1195 /* 1196 * The entire backing chain from first_object to object 1197 * inclusive is chainlocked. 1198 * 1199 * If the object is dead, we stop here 1200 * 1201 * vm_shared_fault (fs->shared != 0) case: nothing special. 1202 */ 1203 if (fs->object->flags & OBJ_DEAD) { 1204 vm_object_pip_wakeup(fs->first_object); 1205 vm_object_chain_release_all(fs->first_object, 1206 fs->object); 1207 if (fs->object != fs->first_object) 1208 vm_object_drop(fs->object); 1209 unlock_and_deallocate(fs); 1210 return (KERN_PROTECTION_FAILURE); 1211 } 1212 1213 /* 1214 * See if the page is resident. Wait/Retry if the page is 1215 * busy (lots of stuff may have changed so we can't continue 1216 * in that case). 1217 * 1218 * We can theoretically allow the soft-busy case on a read 1219 * fault if the page is marked valid, but since such 1220 * pages are typically already pmap'd, putting that 1221 * special case in might be more effort then it is 1222 * worth. We cannot under any circumstances mess 1223 * around with a vm_page_t->busy page except, perhaps, 1224 * to pmap it. 1225 * 1226 * vm_shared_fault (fs->shared != 0) case: 1227 * error nothing special 1228 * fs->m relock excl if I/O needed 1229 * NULL relock excl 1230 */ 1231 fs->m = vm_page_lookup_busy_try(fs->object, pindex, 1232 TRUE, &error); 1233 if (error) { 1234 vm_object_pip_wakeup(fs->first_object); 1235 vm_object_chain_release_all(fs->first_object, 1236 fs->object); 1237 if (fs->object != fs->first_object) 1238 vm_object_drop(fs->object); 1239 unlock_things(fs); 1240 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 1241 mycpu->gd_cnt.v_intrans++; 1242 /*vm_object_deallocate(fs->first_object);*/ 1243 /*fs->first_object = NULL;*/ 1244 fs->m = NULL; 1245 return (KERN_TRY_AGAIN); 1246 } 1247 if (fs->m) { 1248 /* 1249 * The page is busied for us. 1250 * 1251 * If reactivating a page from PQ_CACHE we may have 1252 * to rate-limit. 1253 */ 1254 int queue = fs->m->queue; 1255 vm_page_unqueue_nowakeup(fs->m); 1256 1257 if ((queue - fs->m->pc) == PQ_CACHE && 1258 vm_page_count_severe()) { 1259 vm_page_activate(fs->m); 1260 vm_page_wakeup(fs->m); 1261 fs->m = NULL; 1262 vm_object_pip_wakeup(fs->first_object); 1263 vm_object_chain_release_all(fs->first_object, 1264 fs->object); 1265 if (fs->object != fs->first_object) 1266 vm_object_drop(fs->object); 1267 unlock_and_deallocate(fs); 1268 if (allow_nofault == 0 || 1269 (curthread->td_flags & TDF_NOFAULT) == 0) { 1270 vm_wait_pfault(); 1271 } 1272 return (KERN_TRY_AGAIN); 1273 } 1274 1275 /* 1276 * If it still isn't completely valid (readable), 1277 * or if a read-ahead-mark is set on the VM page, 1278 * jump to readrest, else we found the page and 1279 * can return. 1280 * 1281 * We can release the spl once we have marked the 1282 * page busy. 1283 */ 1284 if (fs->m->object != &kernel_object) { 1285 if ((fs->m->valid & VM_PAGE_BITS_ALL) != 1286 VM_PAGE_BITS_ALL) { 1287 if (fs->shared) { 1288 vm_object_drop(fs->object); 1289 vm_object_hold(fs->object); 1290 fs->shared = 0; 1291 } 1292 goto readrest; 1293 } 1294 if (fs->m->flags & PG_RAM) { 1295 if (debug_cluster) 1296 kprintf("R"); 1297 vm_page_flag_clear(fs->m, PG_RAM); 1298 if (fs->shared) { 1299 vm_object_drop(fs->object); 1300 vm_object_hold(fs->object); 1301 fs->shared = 0; 1302 } 1303 goto readrest; 1304 } 1305 } 1306 break; /* break to PAGE HAS BEEN FOUND */ 1307 } 1308 1309 if (fs->shared) { 1310 vm_object_drop(fs->object); 1311 vm_object_hold(fs->object); 1312 fs->shared = 0; 1313 } 1314 1315 /* 1316 * Page is not resident, If this is the search termination 1317 * or the pager might contain the page, allocate a new page. 1318 */ 1319 if (TRYPAGER(fs) || fs->object == fs->first_object) { 1320 /* 1321 * If the page is beyond the object size we fail 1322 */ 1323 if (pindex >= fs->object->size) { 1324 vm_object_pip_wakeup(fs->first_object); 1325 vm_object_chain_release_all(fs->first_object, 1326 fs->object); 1327 if (fs->object != fs->first_object) 1328 vm_object_drop(fs->object); 1329 unlock_and_deallocate(fs); 1330 return (KERN_PROTECTION_FAILURE); 1331 } 1332 1333 /* 1334 * Allocate a new page for this object/offset pair. 1335 * 1336 * It is possible for the allocation to race, so 1337 * handle the case. 1338 */ 1339 fs->m = NULL; 1340 if (!vm_page_count_severe()) { 1341 fs->m = vm_page_alloc(fs->object, pindex, 1342 ((fs->vp || fs->object->backing_object) ? 1343 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1344 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1345 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1346 } 1347 if (fs->m == NULL) { 1348 vm_object_pip_wakeup(fs->first_object); 1349 vm_object_chain_release_all(fs->first_object, 1350 fs->object); 1351 if (fs->object != fs->first_object) 1352 vm_object_drop(fs->object); 1353 unlock_and_deallocate(fs); 1354 if (allow_nofault == 0 || 1355 (curthread->td_flags & TDF_NOFAULT) == 0) { 1356 vm_wait_pfault(); 1357 } 1358 return (KERN_TRY_AGAIN); 1359 } 1360 1361 /* 1362 * Fall through to readrest. We have a new page which 1363 * will have to be paged (since m->valid will be 0). 1364 */ 1365 } 1366 1367 readrest: 1368 /* 1369 * We have found an invalid or partially valid page, a 1370 * page with a read-ahead mark which might be partially or 1371 * fully valid (and maybe dirty too), or we have allocated 1372 * a new page. 1373 * 1374 * Attempt to fault-in the page if there is a chance that the 1375 * pager has it, and potentially fault in additional pages 1376 * at the same time. 1377 * 1378 * If TRYPAGER is true then fs.m will be non-NULL and busied 1379 * for us. 1380 */ 1381 if (TRYPAGER(fs)) { 1382 int rv; 1383 int seqaccess; 1384 u_char behavior = vm_map_entry_behavior(fs->entry); 1385 1386 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 1387 seqaccess = 0; 1388 else 1389 seqaccess = -1; 1390 1391 #if 0 1392 /* 1393 * If sequential access is detected then attempt 1394 * to deactivate/cache pages behind the scan to 1395 * prevent resource hogging. 1396 * 1397 * Use of PG_RAM to detect sequential access 1398 * also simulates multi-zone sequential access 1399 * detection for free. 1400 * 1401 * NOTE: Partially valid dirty pages cannot be 1402 * deactivated without causing NFS picemeal 1403 * writes to barf. 1404 */ 1405 if ((fs->first_object->type != OBJT_DEVICE) && 1406 (fs->first_object->type != OBJT_MGTDEVICE) && 1407 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 1408 (behavior != MAP_ENTRY_BEHAV_RANDOM && 1409 (fs->m->flags & PG_RAM))) 1410 ) { 1411 vm_pindex_t scan_pindex; 1412 int scan_count = 16; 1413 1414 if (first_pindex < 16) { 1415 scan_pindex = 0; 1416 scan_count = 0; 1417 } else { 1418 scan_pindex = first_pindex - 16; 1419 if (scan_pindex < 16) 1420 scan_count = scan_pindex; 1421 else 1422 scan_count = 16; 1423 } 1424 1425 while (scan_count) { 1426 vm_page_t mt; 1427 1428 mt = vm_page_lookup(fs->first_object, 1429 scan_pindex); 1430 if (mt == NULL) 1431 break; 1432 if (vm_page_busy_try(mt, TRUE)) 1433 goto skip; 1434 1435 if (mt->valid != VM_PAGE_BITS_ALL) { 1436 vm_page_wakeup(mt); 1437 break; 1438 } 1439 if ((mt->flags & 1440 (PG_FICTITIOUS | PG_UNMANAGED | 1441 PG_NEED_COMMIT)) || 1442 mt->hold_count || 1443 mt->wire_count) { 1444 vm_page_wakeup(mt); 1445 goto skip; 1446 } 1447 if (mt->dirty == 0) 1448 vm_page_test_dirty(mt); 1449 if (mt->dirty) { 1450 vm_page_protect(mt, 1451 VM_PROT_NONE); 1452 vm_page_deactivate(mt); 1453 vm_page_wakeup(mt); 1454 } else { 1455 vm_page_cache(mt); 1456 } 1457 skip: 1458 --scan_count; 1459 --scan_pindex; 1460 } 1461 1462 seqaccess = 1; 1463 } 1464 #endif 1465 1466 /* 1467 * Avoid deadlocking against the map when doing I/O. 1468 * fs.object and the page is PG_BUSY'd. 1469 * 1470 * NOTE: Once unlocked, fs->entry can become stale 1471 * so this will NULL it out. 1472 * 1473 * NOTE: fs->entry is invalid until we relock the 1474 * map and verify that the timestamp has not 1475 * changed. 1476 */ 1477 unlock_map(fs); 1478 1479 /* 1480 * Acquire the page data. We still hold a ref on 1481 * fs.object and the page has been PG_BUSY's. 1482 * 1483 * The pager may replace the page (for example, in 1484 * order to enter a fictitious page into the 1485 * object). If it does so it is responsible for 1486 * cleaning up the passed page and properly setting 1487 * the new page PG_BUSY. 1488 * 1489 * If we got here through a PG_RAM read-ahead 1490 * mark the page may be partially dirty and thus 1491 * not freeable. Don't bother checking to see 1492 * if the pager has the page because we can't free 1493 * it anyway. We have to depend on the get_page 1494 * operation filling in any gaps whether there is 1495 * backing store or not. 1496 */ 1497 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess); 1498 1499 if (rv == VM_PAGER_OK) { 1500 /* 1501 * Relookup in case pager changed page. Pager 1502 * is responsible for disposition of old page 1503 * if moved. 1504 * 1505 * XXX other code segments do relookups too. 1506 * It's a bad abstraction that needs to be 1507 * fixed/removed. 1508 */ 1509 fs->m = vm_page_lookup(fs->object, pindex); 1510 if (fs->m == NULL) { 1511 vm_object_pip_wakeup(fs->first_object); 1512 vm_object_chain_release_all( 1513 fs->first_object, fs->object); 1514 if (fs->object != fs->first_object) 1515 vm_object_drop(fs->object); 1516 unlock_and_deallocate(fs); 1517 return (KERN_TRY_AGAIN); 1518 } 1519 1520 ++fs->hardfault; 1521 break; /* break to PAGE HAS BEEN FOUND */ 1522 } 1523 1524 /* 1525 * Remove the bogus page (which does not exist at this 1526 * object/offset); before doing so, we must get back 1527 * our object lock to preserve our invariant. 1528 * 1529 * Also wake up any other process that may want to bring 1530 * in this page. 1531 * 1532 * If this is the top-level object, we must leave the 1533 * busy page to prevent another process from rushing 1534 * past us, and inserting the page in that object at 1535 * the same time that we are. 1536 */ 1537 if (rv == VM_PAGER_ERROR) { 1538 if (curproc) { 1539 kprintf("vm_fault: pager read error, " 1540 "pid %d (%s)\n", 1541 curproc->p_pid, 1542 curproc->p_comm); 1543 } else { 1544 kprintf("vm_fault: pager read error, " 1545 "thread %p (%s)\n", 1546 curthread, 1547 curproc->p_comm); 1548 } 1549 } 1550 1551 /* 1552 * Data outside the range of the pager or an I/O error 1553 * 1554 * The page may have been wired during the pagein, 1555 * e.g. by the buffer cache, and cannot simply be 1556 * freed. Call vnode_pager_freepage() to deal with it. 1557 */ 1558 /* 1559 * XXX - the check for kernel_map is a kludge to work 1560 * around having the machine panic on a kernel space 1561 * fault w/ I/O error. 1562 */ 1563 if (((fs->map != &kernel_map) && 1564 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 1565 vnode_pager_freepage(fs->m); 1566 fs->m = NULL; 1567 vm_object_pip_wakeup(fs->first_object); 1568 vm_object_chain_release_all(fs->first_object, 1569 fs->object); 1570 if (fs->object != fs->first_object) 1571 vm_object_drop(fs->object); 1572 unlock_and_deallocate(fs); 1573 if (rv == VM_PAGER_ERROR) 1574 return (KERN_FAILURE); 1575 else 1576 return (KERN_PROTECTION_FAILURE); 1577 /* NOT REACHED */ 1578 } 1579 if (fs->object != fs->first_object) { 1580 vnode_pager_freepage(fs->m); 1581 fs->m = NULL; 1582 /* 1583 * XXX - we cannot just fall out at this 1584 * point, m has been freed and is invalid! 1585 */ 1586 } 1587 } 1588 1589 /* 1590 * We get here if the object has a default pager (or unwiring) 1591 * or the pager doesn't have the page. 1592 */ 1593 if (fs->object == fs->first_object) 1594 fs->first_m = fs->m; 1595 1596 /* 1597 * Move on to the next object. The chain lock should prevent 1598 * the backing_object from getting ripped out from under us. 1599 * 1600 * vm_shared_fault case: 1601 * 1602 * If the next object is the last object and 1603 * vnode-backed (thus possibly shared), we can try a 1604 * shared object lock. There is no 'chain' for this 1605 * last object if vnode-backed (otherwise we would 1606 * need an exclusive lock). 1607 * 1608 * fs->shared mode is very fragile and only works 1609 * under certain specific conditions, and is only 1610 * handled for those conditions in our loop. Essentially 1611 * it is designed only to be able to 'dip into' the 1612 * vnode's object and extract an already-cached page. 1613 */ 1614 fs->shared = 0; 1615 if ((next_object = fs->object->backing_object) != NULL) { 1616 fs->shared = vm_object_hold_maybe_shared(next_object); 1617 vm_object_chain_acquire(next_object); 1618 KKASSERT(next_object == fs->object->backing_object); 1619 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1620 } 1621 1622 if (next_object == NULL) { 1623 /* 1624 * If there's no object left, fill the page in the top 1625 * object with zeros. 1626 */ 1627 if (fs->object != fs->first_object) { 1628 if (fs->first_object->backing_object != 1629 fs->object) { 1630 vm_object_hold(fs->first_object->backing_object); 1631 } 1632 vm_object_chain_release_all( 1633 fs->first_object->backing_object, 1634 fs->object); 1635 if (fs->first_object->backing_object != 1636 fs->object) { 1637 vm_object_drop(fs->first_object->backing_object); 1638 } 1639 vm_object_pip_wakeup(fs->object); 1640 vm_object_drop(fs->object); 1641 fs->object = fs->first_object; 1642 pindex = first_pindex; 1643 fs->m = fs->first_m; 1644 } 1645 fs->first_m = NULL; 1646 1647 /* 1648 * Zero the page if necessary and mark it valid. 1649 */ 1650 if ((fs->m->flags & PG_ZERO) == 0) { 1651 vm_page_zero_fill(fs->m); 1652 } else { 1653 #ifdef PMAP_DEBUG 1654 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m)); 1655 #endif 1656 vm_page_flag_clear(fs->m, PG_ZERO); 1657 mycpu->gd_cnt.v_ozfod++; 1658 } 1659 mycpu->gd_cnt.v_zfod++; 1660 fs->m->valid = VM_PAGE_BITS_ALL; 1661 break; /* break to PAGE HAS BEEN FOUND */ 1662 } 1663 if (fs->object != fs->first_object) { 1664 vm_object_pip_wakeup(fs->object); 1665 vm_object_lock_swap(); 1666 vm_object_drop(fs->object); 1667 } 1668 KASSERT(fs->object != next_object, 1669 ("object loop %p", next_object)); 1670 fs->object = next_object; 1671 vm_object_pip_add(fs->object, 1); 1672 } 1673 1674 /* 1675 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1676 * is held.] 1677 * 1678 * object still held. 1679 * 1680 * If the page is being written, but isn't already owned by the 1681 * top-level object, we have to copy it into a new page owned by the 1682 * top-level object. 1683 */ 1684 KASSERT((fs->m->flags & PG_BUSY) != 0, 1685 ("vm_fault: not busy after main loop")); 1686 1687 if (fs->object != fs->first_object) { 1688 /* 1689 * We only really need to copy if we want to write it. 1690 */ 1691 if (fault_type & VM_PROT_WRITE) { 1692 /* 1693 * This allows pages to be virtually copied from a 1694 * backing_object into the first_object, where the 1695 * backing object has no other refs to it, and cannot 1696 * gain any more refs. Instead of a bcopy, we just 1697 * move the page from the backing object to the 1698 * first object. Note that we must mark the page 1699 * dirty in the first object so that it will go out 1700 * to swap when needed. 1701 */ 1702 if ( 1703 /* 1704 * Map, if present, has not changed 1705 */ 1706 (fs->map == NULL || 1707 fs->map_generation == fs->map->timestamp) && 1708 /* 1709 * Only one shadow object 1710 */ 1711 (fs->object->shadow_count == 1) && 1712 /* 1713 * No COW refs, except us 1714 */ 1715 (fs->object->ref_count == 1) && 1716 /* 1717 * No one else can look this object up 1718 */ 1719 (fs->object->handle == NULL) && 1720 /* 1721 * No other ways to look the object up 1722 */ 1723 ((fs->object->type == OBJT_DEFAULT) || 1724 (fs->object->type == OBJT_SWAP)) && 1725 /* 1726 * We don't chase down the shadow chain 1727 */ 1728 (fs->object == fs->first_object->backing_object) && 1729 1730 /* 1731 * grab the lock if we need to 1732 */ 1733 (fs->lookup_still_valid || 1734 fs->map == NULL || 1735 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 1736 ) { 1737 /* 1738 * (first_m) and (m) are both busied. We have 1739 * move (m) into (first_m)'s object/pindex 1740 * in an atomic fashion, then free (first_m). 1741 * 1742 * first_object is held so second remove 1743 * followed by the rename should wind 1744 * up being atomic. vm_page_free() might 1745 * block so we don't do it until after the 1746 * rename. 1747 */ 1748 fs->lookup_still_valid = 1; 1749 vm_page_protect(fs->first_m, VM_PROT_NONE); 1750 vm_page_remove(fs->first_m); 1751 vm_page_rename(fs->m, fs->first_object, 1752 first_pindex); 1753 vm_page_free(fs->first_m); 1754 fs->first_m = fs->m; 1755 fs->m = NULL; 1756 mycpu->gd_cnt.v_cow_optim++; 1757 } else { 1758 /* 1759 * Oh, well, lets copy it. 1760 * 1761 * Why are we unmapping the original page 1762 * here? Well, in short, not all accessors 1763 * of user memory go through the pmap. The 1764 * procfs code doesn't have access user memory 1765 * via a local pmap, so vm_fault_page*() 1766 * can't call pmap_enter(). And the umtx*() 1767 * code may modify the COW'd page via a DMAP 1768 * or kernel mapping and not via the pmap, 1769 * leaving the original page still mapped 1770 * read-only into the pmap. 1771 * 1772 * So we have to remove the page from at 1773 * least the current pmap if it is in it. 1774 * Just remove it from all pmaps. 1775 */ 1776 vm_page_copy(fs->m, fs->first_m); 1777 vm_page_protect(fs->m, VM_PROT_NONE); 1778 vm_page_event(fs->m, VMEVENT_COW); 1779 } 1780 1781 if (fs->m) { 1782 /* 1783 * We no longer need the old page or object. 1784 */ 1785 release_page(fs); 1786 } 1787 1788 /* 1789 * We intend to revert to first_object, undo the 1790 * chain lock through to that. 1791 */ 1792 if (fs->first_object->backing_object != fs->object) 1793 vm_object_hold(fs->first_object->backing_object); 1794 vm_object_chain_release_all( 1795 fs->first_object->backing_object, 1796 fs->object); 1797 if (fs->first_object->backing_object != fs->object) 1798 vm_object_drop(fs->first_object->backing_object); 1799 1800 /* 1801 * fs->object != fs->first_object due to above 1802 * conditional 1803 */ 1804 vm_object_pip_wakeup(fs->object); 1805 vm_object_drop(fs->object); 1806 1807 /* 1808 * Only use the new page below... 1809 */ 1810 1811 mycpu->gd_cnt.v_cow_faults++; 1812 fs->m = fs->first_m; 1813 fs->object = fs->first_object; 1814 pindex = first_pindex; 1815 } else { 1816 /* 1817 * If it wasn't a write fault avoid having to copy 1818 * the page by mapping it read-only. 1819 */ 1820 fs->prot &= ~VM_PROT_WRITE; 1821 } 1822 } 1823 1824 /* 1825 * Relock the map if necessary, then check the generation count. 1826 * relock_map() will update fs->timestamp to account for the 1827 * relocking if necessary. 1828 * 1829 * If the count has changed after relocking then all sorts of 1830 * crap may have happened and we have to retry. 1831 * 1832 * NOTE: The relock_map() can fail due to a deadlock against 1833 * the vm_page we are holding BUSY. 1834 */ 1835 if (fs->lookup_still_valid == FALSE && fs->map) { 1836 if (relock_map(fs) || 1837 fs->map->timestamp != fs->map_generation) { 1838 release_page(fs); 1839 vm_object_pip_wakeup(fs->first_object); 1840 vm_object_chain_release_all(fs->first_object, 1841 fs->object); 1842 if (fs->object != fs->first_object) 1843 vm_object_drop(fs->object); 1844 unlock_and_deallocate(fs); 1845 return (KERN_TRY_AGAIN); 1846 } 1847 } 1848 1849 /* 1850 * If the fault is a write, we know that this page is being 1851 * written NOW so dirty it explicitly to save on pmap_is_modified() 1852 * calls later. 1853 * 1854 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 1855 * if the page is already dirty to prevent data written with 1856 * the expectation of being synced from not being synced. 1857 * Likewise if this entry does not request NOSYNC then make 1858 * sure the page isn't marked NOSYNC. Applications sharing 1859 * data should use the same flags to avoid ping ponging. 1860 * 1861 * Also tell the backing pager, if any, that it should remove 1862 * any swap backing since the page is now dirty. 1863 */ 1864 vm_page_activate(fs->m); 1865 if (fs->prot & VM_PROT_WRITE) { 1866 vm_object_set_writeable_dirty(fs->m->object); 1867 vm_set_nosync(fs->m, fs->entry); 1868 if (fs->fault_flags & VM_FAULT_DIRTY) { 1869 vm_page_dirty(fs->m); 1870 swap_pager_unswapped(fs->m); 1871 } 1872 } 1873 1874 vm_object_pip_wakeup(fs->first_object); 1875 vm_object_chain_release_all(fs->first_object, fs->object); 1876 if (fs->object != fs->first_object) 1877 vm_object_drop(fs->object); 1878 1879 /* 1880 * Page had better still be busy. We are still locked up and 1881 * fs->object will have another PIP reference if it is not equal 1882 * to fs->first_object. 1883 */ 1884 KASSERT(fs->m->flags & PG_BUSY, 1885 ("vm_fault: page %p not busy!", fs->m)); 1886 1887 /* 1888 * Sanity check: page must be completely valid or it is not fit to 1889 * map into user space. vm_pager_get_pages() ensures this. 1890 */ 1891 if (fs->m->valid != VM_PAGE_BITS_ALL) { 1892 vm_page_zero_invalid(fs->m, TRUE); 1893 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 1894 } 1895 vm_page_flag_clear(fs->m, PG_ZERO); 1896 1897 return (KERN_SUCCESS); 1898 } 1899 1900 /* 1901 * Hold each of the physical pages that are mapped by the specified range of 1902 * virtual addresses, ["addr", "addr" + "len"), if those mappings are valid 1903 * and allow the specified types of access, "prot". If all of the implied 1904 * pages are successfully held, then the number of held pages is returned 1905 * together with pointers to those pages in the array "ma". However, if any 1906 * of the pages cannot be held, -1 is returned. 1907 */ 1908 int 1909 vm_fault_quick_hold_pages(vm_map_t map, vm_offset_t addr, vm_size_t len, 1910 vm_prot_t prot, vm_page_t *ma, int max_count) 1911 { 1912 vm_offset_t start, end; 1913 int i, npages, error; 1914 1915 start = trunc_page(addr); 1916 end = round_page(addr + len); 1917 1918 npages = howmany(end - start, PAGE_SIZE); 1919 1920 if (npages > max_count) 1921 return -1; 1922 1923 for (i = 0; i < npages; i++) { 1924 // XXX error handling 1925 ma[i] = vm_fault_page_quick(start + (i * PAGE_SIZE), 1926 prot, 1927 &error); 1928 } 1929 1930 return npages; 1931 } 1932 1933 /* 1934 * Wire down a range of virtual addresses in a map. The entry in question 1935 * should be marked in-transition and the map must be locked. We must 1936 * release the map temporarily while faulting-in the page to avoid a 1937 * deadlock. Note that the entry may be clipped while we are blocked but 1938 * will never be freed. 1939 * 1940 * No requirements. 1941 */ 1942 int 1943 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 1944 { 1945 boolean_t fictitious; 1946 vm_offset_t start; 1947 vm_offset_t end; 1948 vm_offset_t va; 1949 vm_paddr_t pa; 1950 vm_page_t m; 1951 pmap_t pmap; 1952 int rv; 1953 1954 lwkt_gettoken(&map->token); 1955 1956 pmap = vm_map_pmap(map); 1957 start = entry->start; 1958 end = entry->end; 1959 fictitious = entry->object.vm_object && 1960 ((entry->object.vm_object->type == OBJT_DEVICE) || 1961 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 1962 if (entry->eflags & MAP_ENTRY_KSTACK) 1963 start += PAGE_SIZE; 1964 map->timestamp++; 1965 vm_map_unlock(map); 1966 1967 /* 1968 * We simulate a fault to get the page and enter it in the physical 1969 * map. 1970 */ 1971 for (va = start; va < end; va += PAGE_SIZE) { 1972 if (user_wire) { 1973 rv = vm_fault(map, va, VM_PROT_READ, 1974 VM_FAULT_USER_WIRE); 1975 } else { 1976 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 1977 VM_FAULT_CHANGE_WIRING); 1978 } 1979 if (rv) { 1980 while (va > start) { 1981 va -= PAGE_SIZE; 1982 if ((pa = pmap_extract(pmap, va)) == 0) 1983 continue; 1984 pmap_change_wiring(pmap, va, FALSE, entry); 1985 if (!fictitious) { 1986 m = PHYS_TO_VM_PAGE(pa); 1987 vm_page_busy_wait(m, FALSE, "vmwrpg"); 1988 vm_page_unwire(m, 1); 1989 vm_page_wakeup(m); 1990 } 1991 } 1992 goto done; 1993 } 1994 } 1995 rv = KERN_SUCCESS; 1996 done: 1997 vm_map_lock(map); 1998 lwkt_reltoken(&map->token); 1999 return (rv); 2000 } 2001 2002 /* 2003 * Unwire a range of virtual addresses in a map. The map should be 2004 * locked. 2005 */ 2006 void 2007 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 2008 { 2009 boolean_t fictitious; 2010 vm_offset_t start; 2011 vm_offset_t end; 2012 vm_offset_t va; 2013 vm_paddr_t pa; 2014 vm_page_t m; 2015 pmap_t pmap; 2016 2017 lwkt_gettoken(&map->token); 2018 2019 pmap = vm_map_pmap(map); 2020 start = entry->start; 2021 end = entry->end; 2022 fictitious = entry->object.vm_object && 2023 ((entry->object.vm_object->type == OBJT_DEVICE) || 2024 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2025 if (entry->eflags & MAP_ENTRY_KSTACK) 2026 start += PAGE_SIZE; 2027 2028 /* 2029 * Since the pages are wired down, we must be able to get their 2030 * mappings from the physical map system. 2031 */ 2032 for (va = start; va < end; va += PAGE_SIZE) { 2033 pa = pmap_extract(pmap, va); 2034 if (pa != 0) { 2035 pmap_change_wiring(pmap, va, FALSE, entry); 2036 if (!fictitious) { 2037 m = PHYS_TO_VM_PAGE(pa); 2038 vm_page_busy_wait(m, FALSE, "vmwupg"); 2039 vm_page_unwire(m, 1); 2040 vm_page_wakeup(m); 2041 } 2042 } 2043 } 2044 lwkt_reltoken(&map->token); 2045 } 2046 2047 /* 2048 * Copy all of the pages from a wired-down map entry to another. 2049 * 2050 * The source and destination maps must be locked for write. 2051 * The source and destination maps token must be held 2052 * The source map entry must be wired down (or be a sharing map 2053 * entry corresponding to a main map entry that is wired down). 2054 * 2055 * No other requirements. 2056 * 2057 * XXX do segment optimization 2058 */ 2059 void 2060 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2061 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2062 { 2063 vm_object_t dst_object; 2064 vm_object_t src_object; 2065 vm_ooffset_t dst_offset; 2066 vm_ooffset_t src_offset; 2067 vm_prot_t prot; 2068 vm_offset_t vaddr; 2069 vm_page_t dst_m; 2070 vm_page_t src_m; 2071 2072 src_object = src_entry->object.vm_object; 2073 src_offset = src_entry->offset; 2074 2075 /* 2076 * Create the top-level object for the destination entry. (Doesn't 2077 * actually shadow anything - we copy the pages directly.) 2078 */ 2079 vm_map_entry_allocate_object(dst_entry); 2080 dst_object = dst_entry->object.vm_object; 2081 2082 prot = dst_entry->max_protection; 2083 2084 /* 2085 * Loop through all of the pages in the entry's range, copying each 2086 * one from the source object (it should be there) to the destination 2087 * object. 2088 */ 2089 vm_object_hold(src_object); 2090 vm_object_hold(dst_object); 2091 for (vaddr = dst_entry->start, dst_offset = 0; 2092 vaddr < dst_entry->end; 2093 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2094 2095 /* 2096 * Allocate a page in the destination object 2097 */ 2098 do { 2099 dst_m = vm_page_alloc(dst_object, 2100 OFF_TO_IDX(dst_offset), 2101 VM_ALLOC_NORMAL); 2102 if (dst_m == NULL) { 2103 vm_wait(0); 2104 } 2105 } while (dst_m == NULL); 2106 2107 /* 2108 * Find the page in the source object, and copy it in. 2109 * (Because the source is wired down, the page will be in 2110 * memory.) 2111 */ 2112 src_m = vm_page_lookup(src_object, 2113 OFF_TO_IDX(dst_offset + src_offset)); 2114 if (src_m == NULL) 2115 panic("vm_fault_copy_wired: page missing"); 2116 2117 vm_page_copy(src_m, dst_m); 2118 vm_page_event(src_m, VMEVENT_COW); 2119 2120 /* 2121 * Enter it in the pmap... 2122 */ 2123 2124 vm_page_flag_clear(dst_m, PG_ZERO); 2125 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2126 2127 /* 2128 * Mark it no longer busy, and put it on the active list. 2129 */ 2130 vm_page_activate(dst_m); 2131 vm_page_wakeup(dst_m); 2132 } 2133 vm_object_drop(dst_object); 2134 vm_object_drop(src_object); 2135 } 2136 2137 #if 0 2138 2139 /* 2140 * This routine checks around the requested page for other pages that 2141 * might be able to be faulted in. This routine brackets the viable 2142 * pages for the pages to be paged in. 2143 * 2144 * Inputs: 2145 * m, rbehind, rahead 2146 * 2147 * Outputs: 2148 * marray (array of vm_page_t), reqpage (index of requested page) 2149 * 2150 * Return value: 2151 * number of pages in marray 2152 */ 2153 static int 2154 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2155 vm_page_t *marray, int *reqpage) 2156 { 2157 int i,j; 2158 vm_object_t object; 2159 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2160 vm_page_t rtm; 2161 int cbehind, cahead; 2162 2163 object = m->object; 2164 pindex = m->pindex; 2165 2166 /* 2167 * we don't fault-ahead for device pager 2168 */ 2169 if ((object->type == OBJT_DEVICE) || 2170 (object->type == OBJT_MGTDEVICE)) { 2171 *reqpage = 0; 2172 marray[0] = m; 2173 return 1; 2174 } 2175 2176 /* 2177 * if the requested page is not available, then give up now 2178 */ 2179 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2180 *reqpage = 0; /* not used by caller, fix compiler warn */ 2181 return 0; 2182 } 2183 2184 if ((cbehind == 0) && (cahead == 0)) { 2185 *reqpage = 0; 2186 marray[0] = m; 2187 return 1; 2188 } 2189 2190 if (rahead > cahead) { 2191 rahead = cahead; 2192 } 2193 2194 if (rbehind > cbehind) { 2195 rbehind = cbehind; 2196 } 2197 2198 /* 2199 * Do not do any readahead if we have insufficient free memory. 2200 * 2201 * XXX code was broken disabled before and has instability 2202 * with this conditonal fixed, so shortcut for now. 2203 */ 2204 if (burst_fault == 0 || vm_page_count_severe()) { 2205 marray[0] = m; 2206 *reqpage = 0; 2207 return 1; 2208 } 2209 2210 /* 2211 * scan backward for the read behind pages -- in memory 2212 * 2213 * Assume that if the page is not found an interrupt will not 2214 * create it. Theoretically interrupts can only remove (busy) 2215 * pages, not create new associations. 2216 */ 2217 if (pindex > 0) { 2218 if (rbehind > pindex) { 2219 rbehind = pindex; 2220 startpindex = 0; 2221 } else { 2222 startpindex = pindex - rbehind; 2223 } 2224 2225 vm_object_hold(object); 2226 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2227 if (vm_page_lookup(object, tpindex - 1)) 2228 break; 2229 } 2230 2231 i = 0; 2232 while (tpindex < pindex) { 2233 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2234 VM_ALLOC_NULL_OK); 2235 if (rtm == NULL) { 2236 for (j = 0; j < i; j++) { 2237 vm_page_free(marray[j]); 2238 } 2239 vm_object_drop(object); 2240 marray[0] = m; 2241 *reqpage = 0; 2242 return 1; 2243 } 2244 marray[i] = rtm; 2245 ++i; 2246 ++tpindex; 2247 } 2248 vm_object_drop(object); 2249 } else { 2250 i = 0; 2251 } 2252 2253 /* 2254 * Assign requested page 2255 */ 2256 marray[i] = m; 2257 *reqpage = i; 2258 ++i; 2259 2260 /* 2261 * Scan forwards for read-ahead pages 2262 */ 2263 tpindex = pindex + 1; 2264 endpindex = tpindex + rahead; 2265 if (endpindex > object->size) 2266 endpindex = object->size; 2267 2268 vm_object_hold(object); 2269 while (tpindex < endpindex) { 2270 if (vm_page_lookup(object, tpindex)) 2271 break; 2272 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2273 VM_ALLOC_NULL_OK); 2274 if (rtm == NULL) 2275 break; 2276 marray[i] = rtm; 2277 ++i; 2278 ++tpindex; 2279 } 2280 vm_object_drop(object); 2281 2282 return (i); 2283 } 2284 2285 #endif 2286 2287 /* 2288 * vm_prefault() provides a quick way of clustering pagefaults into a 2289 * processes address space. It is a "cousin" of pmap_object_init_pt, 2290 * except it runs at page fault time instead of mmap time. 2291 * 2292 * vm.fast_fault Enables pre-faulting zero-fill pages 2293 * 2294 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2295 * prefault. Scan stops in either direction when 2296 * a page is found to already exist. 2297 * 2298 * This code used to be per-platform pmap_prefault(). It is now 2299 * machine-independent and enhanced to also pre-fault zero-fill pages 2300 * (see vm.fast_fault) as well as make them writable, which greatly 2301 * reduces the number of page faults programs incur. 2302 * 2303 * Application performance when pre-faulting zero-fill pages is heavily 2304 * dependent on the application. Very tiny applications like /bin/echo 2305 * lose a little performance while applications of any appreciable size 2306 * gain performance. Prefaulting multiple pages also reduces SMP 2307 * congestion and can improve SMP performance significantly. 2308 * 2309 * NOTE! prot may allow writing but this only applies to the top level 2310 * object. If we wind up mapping a page extracted from a backing 2311 * object we have to make sure it is read-only. 2312 * 2313 * NOTE! The caller has already handled any COW operations on the 2314 * vm_map_entry via the normal fault code. Do NOT call this 2315 * shortcut unless the normal fault code has run on this entry. 2316 * 2317 * The related map must be locked. 2318 * No other requirements. 2319 */ 2320 static int vm_prefault_pages = 8; 2321 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2322 "Maximum number of pages to pre-fault"); 2323 static int vm_fast_fault = 1; 2324 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2325 "Burst fault zero-fill regions"); 2326 2327 /* 2328 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2329 * is not already dirty by other means. This will prevent passive 2330 * filesystem syncing as well as 'sync' from writing out the page. 2331 */ 2332 static void 2333 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2334 { 2335 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2336 if (m->dirty == 0) 2337 vm_page_flag_set(m, PG_NOSYNC); 2338 } else { 2339 vm_page_flag_clear(m, PG_NOSYNC); 2340 } 2341 } 2342 2343 static void 2344 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 2345 int fault_flags) 2346 { 2347 struct lwp *lp; 2348 vm_page_t m; 2349 vm_offset_t addr; 2350 vm_pindex_t index; 2351 vm_pindex_t pindex; 2352 vm_object_t object; 2353 int pprot; 2354 int i; 2355 int noneg; 2356 int nopos; 2357 int maxpages; 2358 2359 /* 2360 * Get stable max count value, disabled if set to 0 2361 */ 2362 maxpages = vm_prefault_pages; 2363 cpu_ccfence(); 2364 if (maxpages <= 0) 2365 return; 2366 2367 /* 2368 * We do not currently prefault mappings that use virtual page 2369 * tables. We do not prefault foreign pmaps. 2370 */ 2371 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) 2372 return; 2373 lp = curthread->td_lwp; 2374 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2375 return; 2376 2377 /* 2378 * Limit pre-fault count to 1024 pages. 2379 */ 2380 if (maxpages > 1024) 2381 maxpages = 1024; 2382 2383 object = entry->object.vm_object; 2384 KKASSERT(object != NULL); 2385 KKASSERT(object == entry->object.vm_object); 2386 vm_object_hold(object); 2387 vm_object_chain_acquire(object); 2388 2389 noneg = 0; 2390 nopos = 0; 2391 for (i = 0; i < maxpages; ++i) { 2392 vm_object_t lobject; 2393 vm_object_t nobject; 2394 int allocated = 0; 2395 int error; 2396 2397 /* 2398 * This can eat a lot of time on a heavily contended 2399 * machine so yield on the tick if needed. 2400 */ 2401 if ((i & 7) == 7) 2402 lwkt_yield(); 2403 2404 /* 2405 * Calculate the page to pre-fault, stopping the scan in 2406 * each direction separately if the limit is reached. 2407 */ 2408 if (i & 1) { 2409 if (noneg) 2410 continue; 2411 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2412 } else { 2413 if (nopos) 2414 continue; 2415 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2416 } 2417 if (addr < entry->start) { 2418 noneg = 1; 2419 if (noneg && nopos) 2420 break; 2421 continue; 2422 } 2423 if (addr >= entry->end) { 2424 nopos = 1; 2425 if (noneg && nopos) 2426 break; 2427 continue; 2428 } 2429 2430 /* 2431 * Skip pages already mapped, and stop scanning in that 2432 * direction. When the scan terminates in both directions 2433 * we are done. 2434 */ 2435 if (pmap_prefault_ok(pmap, addr) == 0) { 2436 if (i & 1) 2437 noneg = 1; 2438 else 2439 nopos = 1; 2440 if (noneg && nopos) 2441 break; 2442 continue; 2443 } 2444 2445 /* 2446 * Follow the VM object chain to obtain the page to be mapped 2447 * into the pmap. 2448 * 2449 * If we reach the terminal object without finding a page 2450 * and we determine it would be advantageous, then allocate 2451 * a zero-fill page for the base object. The base object 2452 * is guaranteed to be OBJT_DEFAULT for this case. 2453 * 2454 * In order to not have to check the pager via *haspage*() 2455 * we stop if any non-default object is encountered. e.g. 2456 * a vnode or swap object would stop the loop. 2457 */ 2458 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2459 lobject = object; 2460 pindex = index; 2461 pprot = prot; 2462 2463 KKASSERT(lobject == entry->object.vm_object); 2464 /*vm_object_hold(lobject); implied */ 2465 2466 while ((m = vm_page_lookup_busy_try(lobject, pindex, 2467 TRUE, &error)) == NULL) { 2468 if (lobject->type != OBJT_DEFAULT) 2469 break; 2470 if (lobject->backing_object == NULL) { 2471 if (vm_fast_fault == 0) 2472 break; 2473 if ((prot & VM_PROT_WRITE) == 0 || 2474 vm_page_count_min(0)) { 2475 break; 2476 } 2477 2478 /* 2479 * NOTE: Allocated from base object 2480 */ 2481 m = vm_page_alloc(object, index, 2482 VM_ALLOC_NORMAL | 2483 VM_ALLOC_ZERO | 2484 VM_ALLOC_USE_GD | 2485 VM_ALLOC_NULL_OK); 2486 if (m == NULL) 2487 break; 2488 allocated = 1; 2489 pprot = prot; 2490 /* lobject = object .. not needed */ 2491 break; 2492 } 2493 if (lobject->backing_object_offset & PAGE_MASK) 2494 break; 2495 nobject = lobject->backing_object; 2496 vm_object_hold(nobject); 2497 KKASSERT(nobject == lobject->backing_object); 2498 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 2499 if (lobject != object) { 2500 vm_object_lock_swap(); 2501 vm_object_drop(lobject); 2502 } 2503 lobject = nobject; 2504 pprot &= ~VM_PROT_WRITE; 2505 vm_object_chain_acquire(lobject); 2506 } 2507 2508 /* 2509 * NOTE: A non-NULL (m) will be associated with lobject if 2510 * it was found there, otherwise it is probably a 2511 * zero-fill page associated with the base object. 2512 * 2513 * Give-up if no page is available. 2514 */ 2515 if (m == NULL) { 2516 if (lobject != object) { 2517 if (object->backing_object != lobject) 2518 vm_object_hold(object->backing_object); 2519 vm_object_chain_release_all( 2520 object->backing_object, lobject); 2521 if (object->backing_object != lobject) 2522 vm_object_drop(object->backing_object); 2523 vm_object_drop(lobject); 2524 } 2525 break; 2526 } 2527 2528 /* 2529 * The object must be marked dirty if we are mapping a 2530 * writable page. m->object is either lobject or object, 2531 * both of which are still held. Do this before we 2532 * potentially drop the object. 2533 */ 2534 if (pprot & VM_PROT_WRITE) 2535 vm_object_set_writeable_dirty(m->object); 2536 2537 /* 2538 * Do not conditionalize on PG_RAM. If pages are present in 2539 * the VM system we assume optimal caching. If caching is 2540 * not optimal the I/O gravy train will be restarted when we 2541 * hit an unavailable page. We do not want to try to restart 2542 * the gravy train now because we really don't know how much 2543 * of the object has been cached. The cost for restarting 2544 * the gravy train should be low (since accesses will likely 2545 * be I/O bound anyway). 2546 */ 2547 if (lobject != object) { 2548 if (object->backing_object != lobject) 2549 vm_object_hold(object->backing_object); 2550 vm_object_chain_release_all(object->backing_object, 2551 lobject); 2552 if (object->backing_object != lobject) 2553 vm_object_drop(object->backing_object); 2554 vm_object_drop(lobject); 2555 } 2556 2557 /* 2558 * Enter the page into the pmap if appropriate. If we had 2559 * allocated the page we have to place it on a queue. If not 2560 * we just have to make sure it isn't on the cache queue 2561 * (pages on the cache queue are not allowed to be mapped). 2562 */ 2563 if (allocated) { 2564 /* 2565 * Page must be zerod. 2566 */ 2567 if ((m->flags & PG_ZERO) == 0) { 2568 vm_page_zero_fill(m); 2569 } else { 2570 #ifdef PMAP_DEBUG 2571 pmap_page_assertzero( 2572 VM_PAGE_TO_PHYS(m)); 2573 #endif 2574 vm_page_flag_clear(m, PG_ZERO); 2575 mycpu->gd_cnt.v_ozfod++; 2576 } 2577 mycpu->gd_cnt.v_zfod++; 2578 m->valid = VM_PAGE_BITS_ALL; 2579 2580 /* 2581 * Handle dirty page case 2582 */ 2583 if (pprot & VM_PROT_WRITE) 2584 vm_set_nosync(m, entry); 2585 pmap_enter(pmap, addr, m, pprot, 0, entry); 2586 mycpu->gd_cnt.v_vm_faults++; 2587 if (curthread->td_lwp) 2588 ++curthread->td_lwp->lwp_ru.ru_minflt; 2589 vm_page_deactivate(m); 2590 if (pprot & VM_PROT_WRITE) { 2591 /*vm_object_set_writeable_dirty(m->object);*/ 2592 vm_set_nosync(m, entry); 2593 if (fault_flags & VM_FAULT_DIRTY) { 2594 vm_page_dirty(m); 2595 /*XXX*/ 2596 swap_pager_unswapped(m); 2597 } 2598 } 2599 vm_page_wakeup(m); 2600 } else if (error) { 2601 /* couldn't busy page, no wakeup */ 2602 } else if ( 2603 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2604 (m->flags & PG_FICTITIOUS) == 0) { 2605 /* 2606 * A fully valid page not undergoing soft I/O can 2607 * be immediately entered into the pmap. 2608 */ 2609 if ((m->queue - m->pc) == PQ_CACHE) 2610 vm_page_deactivate(m); 2611 if (pprot & VM_PROT_WRITE) { 2612 /*vm_object_set_writeable_dirty(m->object);*/ 2613 vm_set_nosync(m, entry); 2614 if (fault_flags & VM_FAULT_DIRTY) { 2615 vm_page_dirty(m); 2616 /*XXX*/ 2617 swap_pager_unswapped(m); 2618 } 2619 } 2620 if (pprot & VM_PROT_WRITE) 2621 vm_set_nosync(m, entry); 2622 pmap_enter(pmap, addr, m, pprot, 0, entry); 2623 mycpu->gd_cnt.v_vm_faults++; 2624 if (curthread->td_lwp) 2625 ++curthread->td_lwp->lwp_ru.ru_minflt; 2626 vm_page_wakeup(m); 2627 } else { 2628 vm_page_wakeup(m); 2629 } 2630 } 2631 vm_object_chain_release(object); 2632 vm_object_drop(object); 2633 } 2634 2635 static void 2636 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 2637 vm_map_entry_t entry, int prot, int fault_flags) 2638 { 2639 struct lwp *lp; 2640 vm_page_t m; 2641 vm_offset_t addr; 2642 vm_pindex_t pindex; 2643 vm_object_t object; 2644 int i; 2645 int noneg; 2646 int nopos; 2647 int maxpages; 2648 2649 /* 2650 * Get stable max count value, disabled if set to 0 2651 */ 2652 maxpages = vm_prefault_pages; 2653 cpu_ccfence(); 2654 if (maxpages <= 0) 2655 return; 2656 2657 /* 2658 * We do not currently prefault mappings that use virtual page 2659 * tables. We do not prefault foreign pmaps. 2660 */ 2661 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) 2662 return; 2663 lp = curthread->td_lwp; 2664 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2665 return; 2666 2667 /* 2668 * Limit pre-fault count to 1024 pages. 2669 */ 2670 if (maxpages > 1024) 2671 maxpages = 1024; 2672 2673 object = entry->object.vm_object; 2674 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2675 KKASSERT(object->backing_object == NULL); 2676 2677 noneg = 0; 2678 nopos = 0; 2679 for (i = 0; i < maxpages; ++i) { 2680 int error; 2681 2682 /* 2683 * Calculate the page to pre-fault, stopping the scan in 2684 * each direction separately if the limit is reached. 2685 */ 2686 if (i & 1) { 2687 if (noneg) 2688 continue; 2689 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2690 } else { 2691 if (nopos) 2692 continue; 2693 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2694 } 2695 if (addr < entry->start) { 2696 noneg = 1; 2697 if (noneg && nopos) 2698 break; 2699 continue; 2700 } 2701 if (addr >= entry->end) { 2702 nopos = 1; 2703 if (noneg && nopos) 2704 break; 2705 continue; 2706 } 2707 2708 /* 2709 * Skip pages already mapped, and stop scanning in that 2710 * direction. When the scan terminates in both directions 2711 * we are done. 2712 */ 2713 if (pmap_prefault_ok(pmap, addr) == 0) { 2714 if (i & 1) 2715 noneg = 1; 2716 else 2717 nopos = 1; 2718 if (noneg && nopos) 2719 break; 2720 continue; 2721 } 2722 2723 /* 2724 * Follow the VM object chain to obtain the page to be mapped 2725 * into the pmap. This version of the prefault code only 2726 * works with terminal objects. 2727 * 2728 * WARNING! We cannot call swap_pager_unswapped() with a 2729 * shared token. 2730 */ 2731 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2732 2733 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2734 if (m == NULL || error) 2735 continue; 2736 2737 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2738 (m->flags & PG_FICTITIOUS) == 0 && 2739 ((m->flags & PG_SWAPPED) == 0 || 2740 (prot & VM_PROT_WRITE) == 0 || 2741 (fault_flags & VM_FAULT_DIRTY) == 0)) { 2742 /* 2743 * A fully valid page not undergoing soft I/O can 2744 * be immediately entered into the pmap. 2745 */ 2746 if ((m->queue - m->pc) == PQ_CACHE) 2747 vm_page_deactivate(m); 2748 if (prot & VM_PROT_WRITE) { 2749 vm_object_set_writeable_dirty(m->object); 2750 vm_set_nosync(m, entry); 2751 if (fault_flags & VM_FAULT_DIRTY) { 2752 vm_page_dirty(m); 2753 /*XXX*/ 2754 swap_pager_unswapped(m); 2755 } 2756 } 2757 pmap_enter(pmap, addr, m, prot, 0, entry); 2758 mycpu->gd_cnt.v_vm_faults++; 2759 if (curthread->td_lwp) 2760 ++curthread->td_lwp->lwp_ru.ru_minflt; 2761 } 2762 vm_page_wakeup(m); 2763 } 2764 } 2765