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 vm_page_flag_set(fs.m, PG_REFERENCED); 541 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, fs.entry); 542 mycpu->gd_cnt.v_vm_faults++; 543 if (curthread->td_lwp) 544 ++curthread->td_lwp->lwp_ru.ru_minflt; 545 546 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */ 547 KKASSERT(fs.m->flags & PG_BUSY); 548 549 /* 550 * If the page is not wired down, then put it where the pageout daemon 551 * can find it. 552 */ 553 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 554 if (fs.wired) 555 vm_page_wire(fs.m); 556 else 557 vm_page_unwire(fs.m, 1); 558 } else { 559 vm_page_activate(fs.m); 560 } 561 vm_page_wakeup(fs.m); 562 563 /* 564 * Burst in a few more pages if possible. The fs.map should still 565 * be locked. To avoid interlocking against a vnode->getblk 566 * operation we had to be sure to unbusy our primary vm_page above 567 * first. 568 */ 569 if (fault_flags & VM_FAULT_BURST) { 570 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 571 && fs.wired == 0) { 572 vm_prefault(fs.map->pmap, vaddr, 573 fs.entry, fs.prot, fault_flags); 574 } 575 } 576 if (fault_flags & VM_FAULT_BURST_QUICK) { 577 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 578 && fs.wired == 0) { 579 vm_prefault_quick(fs.map->pmap, vaddr, 580 fs.entry, fs.prot, fault_flags); 581 } 582 } 583 584 /* 585 * Unlock everything, and return 586 */ 587 unlock_things(&fs); 588 589 if (curthread->td_lwp) { 590 if (fs.hardfault) { 591 curthread->td_lwp->lwp_ru.ru_majflt++; 592 } else { 593 curthread->td_lwp->lwp_ru.ru_minflt++; 594 } 595 } 596 597 /*vm_object_deallocate(fs.first_object);*/ 598 /*fs.m = NULL; */ 599 /*fs.first_object = NULL; must still drop later */ 600 601 result = KERN_SUCCESS; 602 done: 603 if (fs.first_object) 604 vm_object_drop(fs.first_object); 605 done2: 606 lwkt_reltoken(&map->token); 607 if (lp) 608 lp->lwp_flags &= ~LWP_PAGING; 609 return (result); 610 } 611 612 /* 613 * Fault in the specified virtual address in the current process map, 614 * returning a held VM page or NULL. See vm_fault_page() for more 615 * information. 616 * 617 * No requirements. 618 */ 619 vm_page_t 620 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp) 621 { 622 struct lwp *lp = curthread->td_lwp; 623 vm_page_t m; 624 625 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 626 fault_type, VM_FAULT_NORMAL, errorp); 627 return(m); 628 } 629 630 /* 631 * Fault in the specified virtual address in the specified map, doing all 632 * necessary manipulation of the object store and all necessary I/O. Return 633 * a held VM page or NULL, and set *errorp. The related pmap is not 634 * updated. 635 * 636 * The returned page will be properly dirtied if VM_PROT_WRITE was specified, 637 * and marked PG_REFERENCED as well. 638 * 639 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 640 * error will be returned. 641 * 642 * No requirements. 643 */ 644 vm_page_t 645 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 646 int fault_flags, int *errorp) 647 { 648 vm_pindex_t first_pindex; 649 struct faultstate fs; 650 int result; 651 int retry = 0; 652 vm_prot_t orig_fault_type = fault_type; 653 654 fs.hardfault = 0; 655 fs.fault_flags = fault_flags; 656 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 657 658 lwkt_gettoken(&map->token); 659 660 RetryFault: 661 /* 662 * Find the vm_map_entry representing the backing store and resolve 663 * the top level object and page index. This may have the side 664 * effect of executing a copy-on-write on the map entry and/or 665 * creating a shadow object, but will not COW any actual VM pages. 666 * 667 * On success fs.map is left read-locked and various other fields 668 * are initialized but not otherwise referenced or locked. 669 * 670 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 671 * if the map entry is a virtual page table and also writable, 672 * so we can set the 'A'accessed bit in the virtual page table entry. 673 */ 674 fs.map = map; 675 result = vm_map_lookup(&fs.map, vaddr, fault_type, 676 &fs.entry, &fs.first_object, 677 &first_pindex, &fs.first_prot, &fs.wired); 678 679 if (result != KERN_SUCCESS) { 680 *errorp = result; 681 fs.m = NULL; 682 goto done; 683 } 684 685 /* 686 * fs.map is read-locked 687 * 688 * Misc checks. Save the map generation number to detect races. 689 */ 690 fs.map_generation = fs.map->timestamp; 691 fs.lookup_still_valid = TRUE; 692 fs.first_m = NULL; 693 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 694 fs.shared = 0; 695 fs.vp = NULL; 696 697 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 698 panic("vm_fault: fault on nofault entry, addr: %lx", 699 (u_long)vaddr); 700 } 701 702 /* 703 * A system map entry may return a NULL object. No object means 704 * no pager means an unrecoverable kernel fault. 705 */ 706 if (fs.first_object == NULL) { 707 panic("vm_fault: unrecoverable fault at %p in entry %p", 708 (void *)vaddr, fs.entry); 709 } 710 711 /* 712 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 713 * is set. 714 */ 715 if ((curthread->td_flags & TDF_NOFAULT) && 716 (retry || 717 fs.first_object->type == OBJT_VNODE || 718 fs.first_object->backing_object)) { 719 *errorp = KERN_FAILURE; 720 unlock_things(&fs); 721 goto done2; 722 } 723 724 /* 725 * Make a reference to this object to prevent its disposal while we 726 * are messing with it. Once we have the reference, the map is free 727 * to be diddled. Since objects reference their shadows (and copies), 728 * they will stay around as well. 729 * 730 * The reference should also prevent an unexpected collapse of the 731 * parent that might move pages from the current object into the 732 * parent unexpectedly, resulting in corruption. 733 * 734 * Bump the paging-in-progress count to prevent size changes (e.g. 735 * truncation operations) during I/O. This must be done after 736 * obtaining the vnode lock in order to avoid possible deadlocks. 737 */ 738 vm_object_hold(fs.first_object); 739 fs.vp = vnode_pager_lock(fs.first_object); 740 741 #if 0 742 fs.lookup_still_valid = TRUE; 743 fs.first_m = NULL; 744 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 745 fs.shared = 0; 746 #endif 747 748 /* 749 * If the entry is wired we cannot change the page protection. 750 */ 751 if (fs.wired) 752 fault_type = fs.first_prot; 753 754 /* 755 * The page we want is at (first_object, first_pindex), but if the 756 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 757 * page table to figure out the actual pindex. 758 * 759 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 760 * ONLY 761 */ 762 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 763 result = vm_fault_vpagetable(&fs, &first_pindex, 764 fs.entry->aux.master_pde, 765 fault_type, 1); 766 if (result == KERN_TRY_AGAIN) { 767 vm_object_drop(fs.first_object); 768 ++retry; 769 goto RetryFault; 770 } 771 if (result != KERN_SUCCESS) { 772 *errorp = result; 773 fs.m = NULL; 774 goto done; 775 } 776 } 777 778 /* 779 * Now we have the actual (object, pindex), fault in the page. If 780 * vm_fault_object() fails it will unlock and deallocate the FS 781 * data. If it succeeds everything remains locked and fs->object 782 * will have an additinal PIP count if it is not equal to 783 * fs->first_object 784 */ 785 fs.m = NULL; 786 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 787 788 if (result == KERN_TRY_AGAIN) { 789 vm_object_drop(fs.first_object); 790 ++retry; 791 goto RetryFault; 792 } 793 if (result != KERN_SUCCESS) { 794 *errorp = result; 795 fs.m = NULL; 796 goto done; 797 } 798 799 if ((orig_fault_type & VM_PROT_WRITE) && 800 (fs.prot & VM_PROT_WRITE) == 0) { 801 *errorp = KERN_PROTECTION_FAILURE; 802 unlock_and_deallocate(&fs); 803 fs.m = NULL; 804 goto done; 805 } 806 807 /* 808 * DO NOT UPDATE THE PMAP!!! This function may be called for 809 * a pmap unrelated to the current process pmap, in which case 810 * the current cpu core will not be listed in the pmap's pm_active 811 * mask. Thus invalidation interlocks will fail to work properly. 812 * 813 * (for example, 'ps' uses procfs to read program arguments from 814 * each process's stack). 815 * 816 * In addition to the above this function will be called to acquire 817 * a page that might already be faulted in, re-faulting it 818 * continuously is a waste of time. 819 * 820 * XXX could this have been the cause of our random seg-fault 821 * issues? procfs accesses user stacks. 822 */ 823 vm_page_flag_set(fs.m, PG_REFERENCED); 824 #if 0 825 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL); 826 mycpu->gd_cnt.v_vm_faults++; 827 if (curthread->td_lwp) 828 ++curthread->td_lwp->lwp_ru.ru_minflt; 829 #endif 830 831 /* 832 * On success vm_fault_object() does not unlock or deallocate, and fs.m 833 * will contain a busied page. So we must unlock here after having 834 * messed with the pmap. 835 */ 836 unlock_things(&fs); 837 838 /* 839 * Return a held page. We are not doing any pmap manipulation so do 840 * not set PG_MAPPED. However, adjust the page flags according to 841 * the fault type because the caller may not use a managed pmapping 842 * (so we don't want to lose the fact that the page will be dirtied 843 * if a write fault was specified). 844 */ 845 vm_page_hold(fs.m); 846 vm_page_activate(fs.m); 847 if (fault_type & VM_PROT_WRITE) 848 vm_page_dirty(fs.m); 849 850 if (curthread->td_lwp) { 851 if (fs.hardfault) { 852 curthread->td_lwp->lwp_ru.ru_majflt++; 853 } else { 854 curthread->td_lwp->lwp_ru.ru_minflt++; 855 } 856 } 857 858 /* 859 * Unlock everything, and return the held page. 860 */ 861 vm_page_wakeup(fs.m); 862 /*vm_object_deallocate(fs.first_object);*/ 863 /*fs.first_object = NULL; */ 864 *errorp = 0; 865 866 done: 867 if (fs.first_object) 868 vm_object_drop(fs.first_object); 869 done2: 870 lwkt_reltoken(&map->token); 871 return(fs.m); 872 } 873 874 /* 875 * Fault in the specified (object,offset), dirty the returned page as 876 * needed. If the requested fault_type cannot be done NULL and an 877 * error is returned. 878 * 879 * A held (but not busied) page is returned. 880 * 881 * No requirements. 882 */ 883 vm_page_t 884 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 885 vm_prot_t fault_type, int fault_flags, 886 int shared, int *errorp) 887 { 888 int result; 889 vm_pindex_t first_pindex; 890 struct faultstate fs; 891 struct vm_map_entry entry; 892 893 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 894 bzero(&entry, sizeof(entry)); 895 entry.object.vm_object = object; 896 entry.maptype = VM_MAPTYPE_NORMAL; 897 entry.protection = entry.max_protection = fault_type; 898 899 fs.hardfault = 0; 900 fs.fault_flags = fault_flags; 901 fs.map = NULL; 902 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 903 904 RetryFault: 905 906 fs.first_object = object; 907 first_pindex = OFF_TO_IDX(offset); 908 fs.entry = &entry; 909 fs.first_prot = fault_type; 910 fs.wired = 0; 911 fs.shared = shared; 912 /*fs.map_generation = 0; unused */ 913 914 /* 915 * Make a reference to this object to prevent its disposal while we 916 * are messing with it. Once we have the reference, the map is free 917 * to be diddled. Since objects reference their shadows (and copies), 918 * they will stay around as well. 919 * 920 * The reference should also prevent an unexpected collapse of the 921 * parent that might move pages from the current object into the 922 * parent unexpectedly, resulting in corruption. 923 * 924 * Bump the paging-in-progress count to prevent size changes (e.g. 925 * truncation operations) during I/O. This must be done after 926 * obtaining the vnode lock in order to avoid possible deadlocks. 927 */ 928 fs.vp = vnode_pager_lock(fs.first_object); 929 930 fs.lookup_still_valid = TRUE; 931 fs.first_m = NULL; 932 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 933 934 #if 0 935 /* XXX future - ability to operate on VM object using vpagetable */ 936 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 937 result = vm_fault_vpagetable(&fs, &first_pindex, 938 fs.entry->aux.master_pde, 939 fault_type, 0); 940 if (result == KERN_TRY_AGAIN) 941 goto RetryFault; 942 if (result != KERN_SUCCESS) { 943 *errorp = result; 944 return (NULL); 945 } 946 } 947 #endif 948 949 /* 950 * Now we have the actual (object, pindex), fault in the page. If 951 * vm_fault_object() fails it will unlock and deallocate the FS 952 * data. If it succeeds everything remains locked and fs->object 953 * will have an additinal PIP count if it is not equal to 954 * fs->first_object 955 */ 956 result = vm_fault_object(&fs, first_pindex, fault_type, 0); 957 958 if (result == KERN_TRY_AGAIN) 959 goto RetryFault; 960 if (result != KERN_SUCCESS) { 961 *errorp = result; 962 return(NULL); 963 } 964 965 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 966 *errorp = KERN_PROTECTION_FAILURE; 967 unlock_and_deallocate(&fs); 968 return(NULL); 969 } 970 971 /* 972 * On success vm_fault_object() does not unlock or deallocate, so we 973 * do it here. Note that the returned fs.m will be busied. 974 */ 975 unlock_things(&fs); 976 977 /* 978 * Return a held page. We are not doing any pmap manipulation so do 979 * not set PG_MAPPED. However, adjust the page flags according to 980 * the fault type because the caller may not use a managed pmapping 981 * (so we don't want to lose the fact that the page will be dirtied 982 * if a write fault was specified). 983 */ 984 vm_page_hold(fs.m); 985 vm_page_activate(fs.m); 986 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY)) 987 vm_page_dirty(fs.m); 988 if (fault_flags & VM_FAULT_UNSWAP) 989 swap_pager_unswapped(fs.m); 990 991 /* 992 * Indicate that the page was accessed. 993 */ 994 vm_page_flag_set(fs.m, PG_REFERENCED); 995 996 if (curthread->td_lwp) { 997 if (fs.hardfault) { 998 curthread->td_lwp->lwp_ru.ru_majflt++; 999 } else { 1000 curthread->td_lwp->lwp_ru.ru_minflt++; 1001 } 1002 } 1003 1004 /* 1005 * Unlock everything, and return the held page. 1006 */ 1007 vm_page_wakeup(fs.m); 1008 /*vm_object_deallocate(fs.first_object);*/ 1009 /*fs.first_object = NULL; */ 1010 1011 *errorp = 0; 1012 return(fs.m); 1013 } 1014 1015 /* 1016 * Translate the virtual page number (first_pindex) that is relative 1017 * to the address space into a logical page number that is relative to the 1018 * backing object. Use the virtual page table pointed to by (vpte). 1019 * 1020 * This implements an N-level page table. Any level can terminate the 1021 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 1022 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 1023 */ 1024 static 1025 int 1026 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 1027 vpte_t vpte, int fault_type, int allow_nofault) 1028 { 1029 struct lwbuf *lwb; 1030 struct lwbuf lwb_cache; 1031 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */ 1032 int result = KERN_SUCCESS; 1033 vpte_t *ptep; 1034 1035 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1036 for (;;) { 1037 /* 1038 * We cannot proceed if the vpte is not valid, not readable 1039 * for a read fault, or not writable for a write fault. 1040 */ 1041 if ((vpte & VPTE_V) == 0) { 1042 unlock_and_deallocate(fs); 1043 return (KERN_FAILURE); 1044 } 1045 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) { 1046 unlock_and_deallocate(fs); 1047 return (KERN_FAILURE); 1048 } 1049 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) { 1050 unlock_and_deallocate(fs); 1051 return (KERN_FAILURE); 1052 } 1053 if ((vpte & VPTE_PS) || vshift == 0) 1054 break; 1055 KKASSERT(vshift >= VPTE_PAGE_BITS); 1056 1057 /* 1058 * Get the page table page. Nominally we only read the page 1059 * table, but since we are actively setting VPTE_M and VPTE_A, 1060 * tell vm_fault_object() that we are writing it. 1061 * 1062 * There is currently no real need to optimize this. 1063 */ 1064 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT, 1065 VM_PROT_READ|VM_PROT_WRITE, 1066 allow_nofault); 1067 if (result != KERN_SUCCESS) 1068 return (result); 1069 1070 /* 1071 * Process the returned fs.m and look up the page table 1072 * entry in the page table page. 1073 */ 1074 vshift -= VPTE_PAGE_BITS; 1075 lwb = lwbuf_alloc(fs->m, &lwb_cache); 1076 ptep = ((vpte_t *)lwbuf_kva(lwb) + 1077 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 1078 vpte = *ptep; 1079 1080 /* 1081 * Page table write-back. If the vpte is valid for the 1082 * requested operation, do a write-back to the page table. 1083 * 1084 * XXX VPTE_M is not set properly for page directory pages. 1085 * It doesn't get set in the page directory if the page table 1086 * is modified during a read access. 1087 */ 1088 vm_page_activate(fs->m); 1089 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) && 1090 (vpte & VPTE_W)) { 1091 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) { 1092 atomic_set_long(ptep, VPTE_M | VPTE_A); 1093 vm_page_dirty(fs->m); 1094 } 1095 } 1096 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) && 1097 (vpte & VPTE_R)) { 1098 if ((vpte & VPTE_A) == 0) { 1099 atomic_set_long(ptep, VPTE_A); 1100 vm_page_dirty(fs->m); 1101 } 1102 } 1103 lwbuf_free(lwb); 1104 vm_page_flag_set(fs->m, PG_REFERENCED); 1105 vm_page_wakeup(fs->m); 1106 fs->m = NULL; 1107 cleanup_successful_fault(fs); 1108 } 1109 /* 1110 * Combine remaining address bits with the vpte. 1111 */ 1112 /* JG how many bits from each? */ 1113 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) + 1114 (*pindex & ((1L << vshift) - 1)); 1115 return (KERN_SUCCESS); 1116 } 1117 1118 1119 /* 1120 * This is the core of the vm_fault code. 1121 * 1122 * Do all operations required to fault-in (fs.first_object, pindex). Run 1123 * through the shadow chain as necessary and do required COW or virtual 1124 * copy operations. The caller has already fully resolved the vm_map_entry 1125 * and, if appropriate, has created a copy-on-write layer. All we need to 1126 * do is iterate the object chain. 1127 * 1128 * On failure (fs) is unlocked and deallocated and the caller may return or 1129 * retry depending on the failure code. On success (fs) is NOT unlocked or 1130 * deallocated, fs.m will contained a resolved, busied page, and fs.object 1131 * will have an additional PIP count if it is not equal to fs.first_object. 1132 * 1133 * fs->first_object must be held on call. 1134 */ 1135 static 1136 int 1137 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex, 1138 vm_prot_t fault_type, int allow_nofault) 1139 { 1140 vm_object_t next_object; 1141 vm_pindex_t pindex; 1142 int error; 1143 1144 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1145 fs->prot = fs->first_prot; 1146 fs->object = fs->first_object; 1147 pindex = first_pindex; 1148 1149 vm_object_chain_acquire(fs->first_object); 1150 vm_object_pip_add(fs->first_object, 1); 1151 1152 /* 1153 * If a read fault occurs we try to make the page writable if 1154 * possible. There are three cases where we cannot make the 1155 * page mapping writable: 1156 * 1157 * (1) The mapping is read-only or the VM object is read-only, 1158 * fs->prot above will simply not have VM_PROT_WRITE set. 1159 * 1160 * (2) If the mapping is a virtual page table we need to be able 1161 * to detect writes so we can set VPTE_M in the virtual page 1162 * table. 1163 * 1164 * (3) If the VM page is read-only or copy-on-write, upgrading would 1165 * just result in an unnecessary COW fault. 1166 * 1167 * VM_PROT_VPAGED is set if faulting via a virtual page table and 1168 * causes adjustments to the 'M'odify bit to also turn off write 1169 * access to force a re-fault. 1170 */ 1171 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1172 if ((fault_type & VM_PROT_WRITE) == 0) 1173 fs->prot &= ~VM_PROT_WRITE; 1174 } 1175 1176 /* vm_object_hold(fs->object); implied b/c object == first_object */ 1177 1178 for (;;) { 1179 /* 1180 * The entire backing chain from first_object to object 1181 * inclusive is chainlocked. 1182 * 1183 * If the object is dead, we stop here 1184 * 1185 * vm_shared_fault (fs->shared != 0) case: nothing special. 1186 */ 1187 if (fs->object->flags & OBJ_DEAD) { 1188 vm_object_pip_wakeup(fs->first_object); 1189 vm_object_chain_release_all(fs->first_object, 1190 fs->object); 1191 if (fs->object != fs->first_object) 1192 vm_object_drop(fs->object); 1193 unlock_and_deallocate(fs); 1194 return (KERN_PROTECTION_FAILURE); 1195 } 1196 1197 /* 1198 * See if the page is resident. Wait/Retry if the page is 1199 * busy (lots of stuff may have changed so we can't continue 1200 * in that case). 1201 * 1202 * We can theoretically allow the soft-busy case on a read 1203 * fault if the page is marked valid, but since such 1204 * pages are typically already pmap'd, putting that 1205 * special case in might be more effort then it is 1206 * worth. We cannot under any circumstances mess 1207 * around with a vm_page_t->busy page except, perhaps, 1208 * to pmap it. 1209 * 1210 * vm_shared_fault (fs->shared != 0) case: 1211 * error nothing special 1212 * fs->m relock excl if I/O needed 1213 * NULL relock excl 1214 */ 1215 fs->m = vm_page_lookup_busy_try(fs->object, pindex, 1216 TRUE, &error); 1217 if (error) { 1218 vm_object_pip_wakeup(fs->first_object); 1219 vm_object_chain_release_all(fs->first_object, 1220 fs->object); 1221 if (fs->object != fs->first_object) 1222 vm_object_drop(fs->object); 1223 unlock_things(fs); 1224 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 1225 mycpu->gd_cnt.v_intrans++; 1226 /*vm_object_deallocate(fs->first_object);*/ 1227 /*fs->first_object = NULL;*/ 1228 fs->m = NULL; 1229 return (KERN_TRY_AGAIN); 1230 } 1231 if (fs->m) { 1232 /* 1233 * The page is busied for us. 1234 * 1235 * If reactivating a page from PQ_CACHE we may have 1236 * to rate-limit. 1237 */ 1238 int queue = fs->m->queue; 1239 vm_page_unqueue_nowakeup(fs->m); 1240 1241 if ((queue - fs->m->pc) == PQ_CACHE && 1242 vm_page_count_severe()) { 1243 vm_page_activate(fs->m); 1244 vm_page_wakeup(fs->m); 1245 fs->m = NULL; 1246 vm_object_pip_wakeup(fs->first_object); 1247 vm_object_chain_release_all(fs->first_object, 1248 fs->object); 1249 if (fs->object != fs->first_object) 1250 vm_object_drop(fs->object); 1251 unlock_and_deallocate(fs); 1252 if (allow_nofault == 0 || 1253 (curthread->td_flags & TDF_NOFAULT) == 0) { 1254 vm_wait_pfault(); 1255 } 1256 return (KERN_TRY_AGAIN); 1257 } 1258 1259 /* 1260 * If it still isn't completely valid (readable), 1261 * or if a read-ahead-mark is set on the VM page, 1262 * jump to readrest, else we found the page and 1263 * can return. 1264 * 1265 * We can release the spl once we have marked the 1266 * page busy. 1267 */ 1268 if (fs->m->object != &kernel_object) { 1269 if ((fs->m->valid & VM_PAGE_BITS_ALL) != 1270 VM_PAGE_BITS_ALL) { 1271 if (fs->shared) { 1272 vm_object_drop(fs->object); 1273 vm_object_hold(fs->object); 1274 fs->shared = 0; 1275 } 1276 goto readrest; 1277 } 1278 if (fs->m->flags & PG_RAM) { 1279 if (debug_cluster) 1280 kprintf("R"); 1281 vm_page_flag_clear(fs->m, PG_RAM); 1282 if (fs->shared) { 1283 vm_object_drop(fs->object); 1284 vm_object_hold(fs->object); 1285 fs->shared = 0; 1286 } 1287 goto readrest; 1288 } 1289 } 1290 break; /* break to PAGE HAS BEEN FOUND */ 1291 } 1292 1293 if (fs->shared) { 1294 vm_object_drop(fs->object); 1295 vm_object_hold(fs->object); 1296 fs->shared = 0; 1297 } 1298 1299 /* 1300 * Page is not resident, If this is the search termination 1301 * or the pager might contain the page, allocate a new page. 1302 */ 1303 if (TRYPAGER(fs) || fs->object == fs->first_object) { 1304 /* 1305 * If the page is beyond the object size we fail 1306 */ 1307 if (pindex >= fs->object->size) { 1308 vm_object_pip_wakeup(fs->first_object); 1309 vm_object_chain_release_all(fs->first_object, 1310 fs->object); 1311 if (fs->object != fs->first_object) 1312 vm_object_drop(fs->object); 1313 unlock_and_deallocate(fs); 1314 return (KERN_PROTECTION_FAILURE); 1315 } 1316 1317 /* 1318 * Allocate a new page for this object/offset pair. 1319 * 1320 * It is possible for the allocation to race, so 1321 * handle the case. 1322 */ 1323 fs->m = NULL; 1324 if (!vm_page_count_severe()) { 1325 fs->m = vm_page_alloc(fs->object, pindex, 1326 ((fs->vp || fs->object->backing_object) ? 1327 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1328 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1329 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1330 } 1331 if (fs->m == NULL) { 1332 vm_object_pip_wakeup(fs->first_object); 1333 vm_object_chain_release_all(fs->first_object, 1334 fs->object); 1335 if (fs->object != fs->first_object) 1336 vm_object_drop(fs->object); 1337 unlock_and_deallocate(fs); 1338 if (allow_nofault == 0 || 1339 (curthread->td_flags & TDF_NOFAULT) == 0) { 1340 vm_wait_pfault(); 1341 } 1342 return (KERN_TRY_AGAIN); 1343 } 1344 1345 /* 1346 * Fall through to readrest. We have a new page which 1347 * will have to be paged (since m->valid will be 0). 1348 */ 1349 } 1350 1351 readrest: 1352 /* 1353 * We have found an invalid or partially valid page, a 1354 * page with a read-ahead mark which might be partially or 1355 * fully valid (and maybe dirty too), or we have allocated 1356 * a new page. 1357 * 1358 * Attempt to fault-in the page if there is a chance that the 1359 * pager has it, and potentially fault in additional pages 1360 * at the same time. 1361 * 1362 * If TRYPAGER is true then fs.m will be non-NULL and busied 1363 * for us. 1364 */ 1365 if (TRYPAGER(fs)) { 1366 int rv; 1367 int seqaccess; 1368 u_char behavior = vm_map_entry_behavior(fs->entry); 1369 1370 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 1371 seqaccess = 0; 1372 else 1373 seqaccess = -1; 1374 1375 #if 0 1376 /* 1377 * If sequential access is detected then attempt 1378 * to deactivate/cache pages behind the scan to 1379 * prevent resource hogging. 1380 * 1381 * Use of PG_RAM to detect sequential access 1382 * also simulates multi-zone sequential access 1383 * detection for free. 1384 * 1385 * NOTE: Partially valid dirty pages cannot be 1386 * deactivated without causing NFS picemeal 1387 * writes to barf. 1388 */ 1389 if ((fs->first_object->type != OBJT_DEVICE) && 1390 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL || 1391 (behavior != MAP_ENTRY_BEHAV_RANDOM && 1392 (fs->m->flags & PG_RAM))) 1393 ) { 1394 vm_pindex_t scan_pindex; 1395 int scan_count = 16; 1396 1397 if (first_pindex < 16) { 1398 scan_pindex = 0; 1399 scan_count = 0; 1400 } else { 1401 scan_pindex = first_pindex - 16; 1402 if (scan_pindex < 16) 1403 scan_count = scan_pindex; 1404 else 1405 scan_count = 16; 1406 } 1407 1408 while (scan_count) { 1409 vm_page_t mt; 1410 1411 mt = vm_page_lookup(fs->first_object, 1412 scan_pindex); 1413 if (mt == NULL) 1414 break; 1415 if (vm_page_busy_try(mt, TRUE)) 1416 goto skip; 1417 1418 if (mt->valid != VM_PAGE_BITS_ALL) { 1419 vm_page_wakeup(mt); 1420 break; 1421 } 1422 if ((mt->flags & 1423 (PG_FICTITIOUS | PG_UNMANAGED | 1424 PG_NEED_COMMIT)) || 1425 mt->hold_count || 1426 mt->wire_count) { 1427 vm_page_wakeup(mt); 1428 goto skip; 1429 } 1430 if (mt->dirty == 0) 1431 vm_page_test_dirty(mt); 1432 if (mt->dirty) { 1433 vm_page_protect(mt, 1434 VM_PROT_NONE); 1435 vm_page_deactivate(mt); 1436 vm_page_wakeup(mt); 1437 } else { 1438 vm_page_cache(mt); 1439 } 1440 skip: 1441 --scan_count; 1442 --scan_pindex; 1443 } 1444 1445 seqaccess = 1; 1446 } 1447 #endif 1448 1449 /* 1450 * Avoid deadlocking against the map when doing I/O. 1451 * fs.object and the page is PG_BUSY'd. 1452 * 1453 * NOTE: Once unlocked, fs->entry can become stale 1454 * so this will NULL it out. 1455 * 1456 * NOTE: fs->entry is invalid until we relock the 1457 * map and verify that the timestamp has not 1458 * changed. 1459 */ 1460 unlock_map(fs); 1461 1462 /* 1463 * Acquire the page data. We still hold a ref on 1464 * fs.object and the page has been PG_BUSY's. 1465 * 1466 * The pager may replace the page (for example, in 1467 * order to enter a fictitious page into the 1468 * object). If it does so it is responsible for 1469 * cleaning up the passed page and properly setting 1470 * the new page PG_BUSY. 1471 * 1472 * If we got here through a PG_RAM read-ahead 1473 * mark the page may be partially dirty and thus 1474 * not freeable. Don't bother checking to see 1475 * if the pager has the page because we can't free 1476 * it anyway. We have to depend on the get_page 1477 * operation filling in any gaps whether there is 1478 * backing store or not. 1479 */ 1480 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess); 1481 1482 if (rv == VM_PAGER_OK) { 1483 /* 1484 * Relookup in case pager changed page. Pager 1485 * is responsible for disposition of old page 1486 * if moved. 1487 * 1488 * XXX other code segments do relookups too. 1489 * It's a bad abstraction that needs to be 1490 * fixed/removed. 1491 */ 1492 fs->m = vm_page_lookup(fs->object, pindex); 1493 if (fs->m == NULL) { 1494 vm_object_pip_wakeup(fs->first_object); 1495 vm_object_chain_release_all( 1496 fs->first_object, fs->object); 1497 if (fs->object != fs->first_object) 1498 vm_object_drop(fs->object); 1499 unlock_and_deallocate(fs); 1500 return (KERN_TRY_AGAIN); 1501 } 1502 1503 ++fs->hardfault; 1504 break; /* break to PAGE HAS BEEN FOUND */ 1505 } 1506 1507 /* 1508 * Remove the bogus page (which does not exist at this 1509 * object/offset); before doing so, we must get back 1510 * our object lock to preserve our invariant. 1511 * 1512 * Also wake up any other process that may want to bring 1513 * in this page. 1514 * 1515 * If this is the top-level object, we must leave the 1516 * busy page to prevent another process from rushing 1517 * past us, and inserting the page in that object at 1518 * the same time that we are. 1519 */ 1520 if (rv == VM_PAGER_ERROR) { 1521 if (curproc) { 1522 kprintf("vm_fault: pager read error, " 1523 "pid %d (%s)\n", 1524 curproc->p_pid, 1525 curproc->p_comm); 1526 } else { 1527 kprintf("vm_fault: pager read error, " 1528 "thread %p (%s)\n", 1529 curthread, 1530 curproc->p_comm); 1531 } 1532 } 1533 1534 /* 1535 * Data outside the range of the pager or an I/O error 1536 * 1537 * The page may have been wired during the pagein, 1538 * e.g. by the buffer cache, and cannot simply be 1539 * freed. Call vnode_pager_freepage() to deal with it. 1540 */ 1541 /* 1542 * XXX - the check for kernel_map is a kludge to work 1543 * around having the machine panic on a kernel space 1544 * fault w/ I/O error. 1545 */ 1546 if (((fs->map != &kernel_map) && 1547 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 1548 vnode_pager_freepage(fs->m); 1549 fs->m = NULL; 1550 vm_object_pip_wakeup(fs->first_object); 1551 vm_object_chain_release_all(fs->first_object, 1552 fs->object); 1553 if (fs->object != fs->first_object) 1554 vm_object_drop(fs->object); 1555 unlock_and_deallocate(fs); 1556 if (rv == VM_PAGER_ERROR) 1557 return (KERN_FAILURE); 1558 else 1559 return (KERN_PROTECTION_FAILURE); 1560 /* NOT REACHED */ 1561 } 1562 if (fs->object != fs->first_object) { 1563 vnode_pager_freepage(fs->m); 1564 fs->m = NULL; 1565 /* 1566 * XXX - we cannot just fall out at this 1567 * point, m has been freed and is invalid! 1568 */ 1569 } 1570 } 1571 1572 /* 1573 * We get here if the object has a default pager (or unwiring) 1574 * or the pager doesn't have the page. 1575 */ 1576 if (fs->object == fs->first_object) 1577 fs->first_m = fs->m; 1578 1579 /* 1580 * Move on to the next object. The chain lock should prevent 1581 * the backing_object from getting ripped out from under us. 1582 * 1583 * vm_shared_fault case: 1584 * 1585 * If the next object is the last object and 1586 * vnode-backed (thus possibly shared), we can try a 1587 * shared object lock. There is no 'chain' for this 1588 * last object if vnode-backed (otherwise we would 1589 * need an exclusive lock). 1590 * 1591 * fs->shared mode is very fragile and only works 1592 * under certain specific conditions, and is only 1593 * handled for those conditions in our loop. Essentially 1594 * it is designed only to be able to 'dip into' the 1595 * vnode's object and extract an already-cached page. 1596 */ 1597 fs->shared = 0; 1598 if ((next_object = fs->object->backing_object) != NULL) { 1599 fs->shared = vm_object_hold_maybe_shared(next_object); 1600 vm_object_chain_acquire(next_object); 1601 KKASSERT(next_object == fs->object->backing_object); 1602 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1603 } 1604 1605 if (next_object == NULL) { 1606 /* 1607 * If there's no object left, fill the page in the top 1608 * object with zeros. 1609 */ 1610 if (fs->object != fs->first_object) { 1611 if (fs->first_object->backing_object != 1612 fs->object) { 1613 vm_object_hold(fs->first_object->backing_object); 1614 } 1615 vm_object_chain_release_all( 1616 fs->first_object->backing_object, 1617 fs->object); 1618 if (fs->first_object->backing_object != 1619 fs->object) { 1620 vm_object_drop(fs->first_object->backing_object); 1621 } 1622 vm_object_pip_wakeup(fs->object); 1623 vm_object_drop(fs->object); 1624 fs->object = fs->first_object; 1625 pindex = first_pindex; 1626 fs->m = fs->first_m; 1627 } 1628 fs->first_m = NULL; 1629 1630 /* 1631 * Zero the page if necessary and mark it valid. 1632 */ 1633 if ((fs->m->flags & PG_ZERO) == 0) { 1634 vm_page_zero_fill(fs->m); 1635 } else { 1636 #ifdef PMAP_DEBUG 1637 pmap_page_assertzero(VM_PAGE_TO_PHYS(fs->m)); 1638 #endif 1639 vm_page_flag_clear(fs->m, PG_ZERO); 1640 mycpu->gd_cnt.v_ozfod++; 1641 } 1642 mycpu->gd_cnt.v_zfod++; 1643 fs->m->valid = VM_PAGE_BITS_ALL; 1644 break; /* break to PAGE HAS BEEN FOUND */ 1645 } 1646 if (fs->object != fs->first_object) { 1647 vm_object_pip_wakeup(fs->object); 1648 vm_object_lock_swap(); 1649 vm_object_drop(fs->object); 1650 } 1651 KASSERT(fs->object != next_object, 1652 ("object loop %p", next_object)); 1653 fs->object = next_object; 1654 vm_object_pip_add(fs->object, 1); 1655 } 1656 1657 /* 1658 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1659 * is held.] 1660 * 1661 * object still held. 1662 * 1663 * If the page is being written, but isn't already owned by the 1664 * top-level object, we have to copy it into a new page owned by the 1665 * top-level object. 1666 */ 1667 KASSERT((fs->m->flags & PG_BUSY) != 0, 1668 ("vm_fault: not busy after main loop")); 1669 1670 if (fs->object != fs->first_object) { 1671 /* 1672 * We only really need to copy if we want to write it. 1673 */ 1674 if (fault_type & VM_PROT_WRITE) { 1675 /* 1676 * This allows pages to be virtually copied from a 1677 * backing_object into the first_object, where the 1678 * backing object has no other refs to it, and cannot 1679 * gain any more refs. Instead of a bcopy, we just 1680 * move the page from the backing object to the 1681 * first object. Note that we must mark the page 1682 * dirty in the first object so that it will go out 1683 * to swap when needed. 1684 */ 1685 if ( 1686 /* 1687 * Map, if present, has not changed 1688 */ 1689 (fs->map == NULL || 1690 fs->map_generation == fs->map->timestamp) && 1691 /* 1692 * Only one shadow object 1693 */ 1694 (fs->object->shadow_count == 1) && 1695 /* 1696 * No COW refs, except us 1697 */ 1698 (fs->object->ref_count == 1) && 1699 /* 1700 * No one else can look this object up 1701 */ 1702 (fs->object->handle == NULL) && 1703 /* 1704 * No other ways to look the object up 1705 */ 1706 ((fs->object->type == OBJT_DEFAULT) || 1707 (fs->object->type == OBJT_SWAP)) && 1708 /* 1709 * We don't chase down the shadow chain 1710 */ 1711 (fs->object == fs->first_object->backing_object) && 1712 1713 /* 1714 * grab the lock if we need to 1715 */ 1716 (fs->lookup_still_valid || 1717 fs->map == NULL || 1718 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 1719 ) { 1720 /* 1721 * (first_m) and (m) are both busied. We have 1722 * move (m) into (first_m)'s object/pindex 1723 * in an atomic fashion, then free (first_m). 1724 * 1725 * first_object is held so second remove 1726 * followed by the rename should wind 1727 * up being atomic. vm_page_free() might 1728 * block so we don't do it until after the 1729 * rename. 1730 */ 1731 fs->lookup_still_valid = 1; 1732 vm_page_protect(fs->first_m, VM_PROT_NONE); 1733 vm_page_remove(fs->first_m); 1734 vm_page_rename(fs->m, fs->first_object, 1735 first_pindex); 1736 vm_page_free(fs->first_m); 1737 fs->first_m = fs->m; 1738 fs->m = NULL; 1739 mycpu->gd_cnt.v_cow_optim++; 1740 } else { 1741 /* 1742 * Oh, well, lets copy it. 1743 * 1744 * Why are we unmapping the original page 1745 * here? Well, in short, not all accessors 1746 * of user memory go through the pmap. The 1747 * procfs code doesn't have access user memory 1748 * via a local pmap, so vm_fault_page*() 1749 * can't call pmap_enter(). And the umtx*() 1750 * code may modify the COW'd page via a DMAP 1751 * or kernel mapping and not via the pmap, 1752 * leaving the original page still mapped 1753 * read-only into the pmap. 1754 * 1755 * So we have to remove the page from at 1756 * least the current pmap if it is in it. 1757 * Just remove it from all pmaps. 1758 */ 1759 vm_page_copy(fs->m, fs->first_m); 1760 vm_page_protect(fs->m, VM_PROT_NONE); 1761 vm_page_event(fs->m, VMEVENT_COW); 1762 } 1763 1764 if (fs->m) { 1765 /* 1766 * We no longer need the old page or object. 1767 */ 1768 release_page(fs); 1769 } 1770 1771 /* 1772 * We intend to revert to first_object, undo the 1773 * chain lock through to that. 1774 */ 1775 if (fs->first_object->backing_object != fs->object) 1776 vm_object_hold(fs->first_object->backing_object); 1777 vm_object_chain_release_all( 1778 fs->first_object->backing_object, 1779 fs->object); 1780 if (fs->first_object->backing_object != fs->object) 1781 vm_object_drop(fs->first_object->backing_object); 1782 1783 /* 1784 * fs->object != fs->first_object due to above 1785 * conditional 1786 */ 1787 vm_object_pip_wakeup(fs->object); 1788 vm_object_drop(fs->object); 1789 1790 /* 1791 * Only use the new page below... 1792 */ 1793 1794 mycpu->gd_cnt.v_cow_faults++; 1795 fs->m = fs->first_m; 1796 fs->object = fs->first_object; 1797 pindex = first_pindex; 1798 } else { 1799 /* 1800 * If it wasn't a write fault avoid having to copy 1801 * the page by mapping it read-only. 1802 */ 1803 fs->prot &= ~VM_PROT_WRITE; 1804 } 1805 } 1806 1807 /* 1808 * Relock the map if necessary, then check the generation count. 1809 * relock_map() will update fs->timestamp to account for the 1810 * relocking if necessary. 1811 * 1812 * If the count has changed after relocking then all sorts of 1813 * crap may have happened and we have to retry. 1814 * 1815 * NOTE: The relock_map() can fail due to a deadlock against 1816 * the vm_page we are holding BUSY. 1817 */ 1818 if (fs->lookup_still_valid == FALSE && fs->map) { 1819 if (relock_map(fs) || 1820 fs->map->timestamp != fs->map_generation) { 1821 release_page(fs); 1822 vm_object_pip_wakeup(fs->first_object); 1823 vm_object_chain_release_all(fs->first_object, 1824 fs->object); 1825 if (fs->object != fs->first_object) 1826 vm_object_drop(fs->object); 1827 unlock_and_deallocate(fs); 1828 return (KERN_TRY_AGAIN); 1829 } 1830 } 1831 1832 /* 1833 * If the fault is a write, we know that this page is being 1834 * written NOW so dirty it explicitly to save on pmap_is_modified() 1835 * calls later. 1836 * 1837 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 1838 * if the page is already dirty to prevent data written with 1839 * the expectation of being synced from not being synced. 1840 * Likewise if this entry does not request NOSYNC then make 1841 * sure the page isn't marked NOSYNC. Applications sharing 1842 * data should use the same flags to avoid ping ponging. 1843 * 1844 * Also tell the backing pager, if any, that it should remove 1845 * any swap backing since the page is now dirty. 1846 */ 1847 vm_page_activate(fs->m); 1848 if (fs->prot & VM_PROT_WRITE) { 1849 vm_object_set_writeable_dirty(fs->m->object); 1850 vm_set_nosync(fs->m, fs->entry); 1851 if (fs->fault_flags & VM_FAULT_DIRTY) { 1852 vm_page_dirty(fs->m); 1853 swap_pager_unswapped(fs->m); 1854 } 1855 } 1856 1857 vm_object_pip_wakeup(fs->first_object); 1858 vm_object_chain_release_all(fs->first_object, fs->object); 1859 if (fs->object != fs->first_object) 1860 vm_object_drop(fs->object); 1861 1862 /* 1863 * Page had better still be busy. We are still locked up and 1864 * fs->object will have another PIP reference if it is not equal 1865 * to fs->first_object. 1866 */ 1867 KASSERT(fs->m->flags & PG_BUSY, 1868 ("vm_fault: page %p not busy!", fs->m)); 1869 1870 /* 1871 * Sanity check: page must be completely valid or it is not fit to 1872 * map into user space. vm_pager_get_pages() ensures this. 1873 */ 1874 if (fs->m->valid != VM_PAGE_BITS_ALL) { 1875 vm_page_zero_invalid(fs->m, TRUE); 1876 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 1877 } 1878 vm_page_flag_clear(fs->m, PG_ZERO); 1879 1880 return (KERN_SUCCESS); 1881 } 1882 1883 /* 1884 * Wire down a range of virtual addresses in a map. The entry in question 1885 * should be marked in-transition and the map must be locked. We must 1886 * release the map temporarily while faulting-in the page to avoid a 1887 * deadlock. Note that the entry may be clipped while we are blocked but 1888 * will never be freed. 1889 * 1890 * No requirements. 1891 */ 1892 int 1893 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire) 1894 { 1895 boolean_t fictitious; 1896 vm_offset_t start; 1897 vm_offset_t end; 1898 vm_offset_t va; 1899 vm_paddr_t pa; 1900 vm_page_t m; 1901 pmap_t pmap; 1902 int rv; 1903 1904 lwkt_gettoken(&map->token); 1905 1906 pmap = vm_map_pmap(map); 1907 start = entry->start; 1908 end = entry->end; 1909 fictitious = entry->object.vm_object && 1910 (entry->object.vm_object->type == OBJT_DEVICE); 1911 if (entry->eflags & MAP_ENTRY_KSTACK) 1912 start += PAGE_SIZE; 1913 map->timestamp++; 1914 vm_map_unlock(map); 1915 1916 /* 1917 * We simulate a fault to get the page and enter it in the physical 1918 * map. 1919 */ 1920 for (va = start; va < end; va += PAGE_SIZE) { 1921 if (user_wire) { 1922 rv = vm_fault(map, va, VM_PROT_READ, 1923 VM_FAULT_USER_WIRE); 1924 } else { 1925 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE, 1926 VM_FAULT_CHANGE_WIRING); 1927 } 1928 if (rv) { 1929 while (va > start) { 1930 va -= PAGE_SIZE; 1931 if ((pa = pmap_extract(pmap, va)) == 0) 1932 continue; 1933 pmap_change_wiring(pmap, va, FALSE, entry); 1934 if (!fictitious) { 1935 m = PHYS_TO_VM_PAGE(pa); 1936 vm_page_busy_wait(m, FALSE, "vmwrpg"); 1937 vm_page_unwire(m, 1); 1938 vm_page_wakeup(m); 1939 } 1940 } 1941 goto done; 1942 } 1943 } 1944 rv = KERN_SUCCESS; 1945 done: 1946 vm_map_lock(map); 1947 lwkt_reltoken(&map->token); 1948 return (rv); 1949 } 1950 1951 /* 1952 * Unwire a range of virtual addresses in a map. The map should be 1953 * locked. 1954 */ 1955 void 1956 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 1957 { 1958 boolean_t fictitious; 1959 vm_offset_t start; 1960 vm_offset_t end; 1961 vm_offset_t va; 1962 vm_paddr_t pa; 1963 vm_page_t m; 1964 pmap_t pmap; 1965 1966 lwkt_gettoken(&map->token); 1967 1968 pmap = vm_map_pmap(map); 1969 start = entry->start; 1970 end = entry->end; 1971 fictitious = entry->object.vm_object && 1972 (entry->object.vm_object->type == OBJT_DEVICE); 1973 if (entry->eflags & MAP_ENTRY_KSTACK) 1974 start += PAGE_SIZE; 1975 1976 /* 1977 * Since the pages are wired down, we must be able to get their 1978 * mappings from the physical map system. 1979 */ 1980 for (va = start; va < end; va += PAGE_SIZE) { 1981 pa = pmap_extract(pmap, va); 1982 if (pa != 0) { 1983 pmap_change_wiring(pmap, va, FALSE, entry); 1984 if (!fictitious) { 1985 m = PHYS_TO_VM_PAGE(pa); 1986 vm_page_busy_wait(m, FALSE, "vmwupg"); 1987 vm_page_unwire(m, 1); 1988 vm_page_wakeup(m); 1989 } 1990 } 1991 } 1992 lwkt_reltoken(&map->token); 1993 } 1994 1995 /* 1996 * Copy all of the pages from a wired-down map entry to another. 1997 * 1998 * The source and destination maps must be locked for write. 1999 * The source and destination maps token must be held 2000 * The source map entry must be wired down (or be a sharing map 2001 * entry corresponding to a main map entry that is wired down). 2002 * 2003 * No other requirements. 2004 * 2005 * XXX do segment optimization 2006 */ 2007 void 2008 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2009 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2010 { 2011 vm_object_t dst_object; 2012 vm_object_t src_object; 2013 vm_ooffset_t dst_offset; 2014 vm_ooffset_t src_offset; 2015 vm_prot_t prot; 2016 vm_offset_t vaddr; 2017 vm_page_t dst_m; 2018 vm_page_t src_m; 2019 2020 src_object = src_entry->object.vm_object; 2021 src_offset = src_entry->offset; 2022 2023 /* 2024 * Create the top-level object for the destination entry. (Doesn't 2025 * actually shadow anything - we copy the pages directly.) 2026 */ 2027 vm_map_entry_allocate_object(dst_entry); 2028 dst_object = dst_entry->object.vm_object; 2029 2030 prot = dst_entry->max_protection; 2031 2032 /* 2033 * Loop through all of the pages in the entry's range, copying each 2034 * one from the source object (it should be there) to the destination 2035 * object. 2036 */ 2037 vm_object_hold(src_object); 2038 vm_object_hold(dst_object); 2039 for (vaddr = dst_entry->start, dst_offset = 0; 2040 vaddr < dst_entry->end; 2041 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2042 2043 /* 2044 * Allocate a page in the destination object 2045 */ 2046 do { 2047 dst_m = vm_page_alloc(dst_object, 2048 OFF_TO_IDX(dst_offset), 2049 VM_ALLOC_NORMAL); 2050 if (dst_m == NULL) { 2051 vm_wait(0); 2052 } 2053 } while (dst_m == NULL); 2054 2055 /* 2056 * Find the page in the source object, and copy it in. 2057 * (Because the source is wired down, the page will be in 2058 * memory.) 2059 */ 2060 src_m = vm_page_lookup(src_object, 2061 OFF_TO_IDX(dst_offset + src_offset)); 2062 if (src_m == NULL) 2063 panic("vm_fault_copy_wired: page missing"); 2064 2065 vm_page_copy(src_m, dst_m); 2066 vm_page_event(src_m, VMEVENT_COW); 2067 2068 /* 2069 * Enter it in the pmap... 2070 */ 2071 2072 vm_page_flag_clear(dst_m, PG_ZERO); 2073 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2074 2075 /* 2076 * Mark it no longer busy, and put it on the active list. 2077 */ 2078 vm_page_activate(dst_m); 2079 vm_page_wakeup(dst_m); 2080 } 2081 vm_object_drop(dst_object); 2082 vm_object_drop(src_object); 2083 } 2084 2085 #if 0 2086 2087 /* 2088 * This routine checks around the requested page for other pages that 2089 * might be able to be faulted in. This routine brackets the viable 2090 * pages for the pages to be paged in. 2091 * 2092 * Inputs: 2093 * m, rbehind, rahead 2094 * 2095 * Outputs: 2096 * marray (array of vm_page_t), reqpage (index of requested page) 2097 * 2098 * Return value: 2099 * number of pages in marray 2100 */ 2101 static int 2102 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2103 vm_page_t *marray, int *reqpage) 2104 { 2105 int i,j; 2106 vm_object_t object; 2107 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2108 vm_page_t rtm; 2109 int cbehind, cahead; 2110 2111 object = m->object; 2112 pindex = m->pindex; 2113 2114 /* 2115 * we don't fault-ahead for device pager 2116 */ 2117 if (object->type == OBJT_DEVICE) { 2118 *reqpage = 0; 2119 marray[0] = m; 2120 return 1; 2121 } 2122 2123 /* 2124 * if the requested page is not available, then give up now 2125 */ 2126 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2127 *reqpage = 0; /* not used by caller, fix compiler warn */ 2128 return 0; 2129 } 2130 2131 if ((cbehind == 0) && (cahead == 0)) { 2132 *reqpage = 0; 2133 marray[0] = m; 2134 return 1; 2135 } 2136 2137 if (rahead > cahead) { 2138 rahead = cahead; 2139 } 2140 2141 if (rbehind > cbehind) { 2142 rbehind = cbehind; 2143 } 2144 2145 /* 2146 * Do not do any readahead if we have insufficient free memory. 2147 * 2148 * XXX code was broken disabled before and has instability 2149 * with this conditonal fixed, so shortcut for now. 2150 */ 2151 if (burst_fault == 0 || vm_page_count_severe()) { 2152 marray[0] = m; 2153 *reqpage = 0; 2154 return 1; 2155 } 2156 2157 /* 2158 * scan backward for the read behind pages -- in memory 2159 * 2160 * Assume that if the page is not found an interrupt will not 2161 * create it. Theoretically interrupts can only remove (busy) 2162 * pages, not create new associations. 2163 */ 2164 if (pindex > 0) { 2165 if (rbehind > pindex) { 2166 rbehind = pindex; 2167 startpindex = 0; 2168 } else { 2169 startpindex = pindex - rbehind; 2170 } 2171 2172 vm_object_hold(object); 2173 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2174 if (vm_page_lookup(object, tpindex - 1)) 2175 break; 2176 } 2177 2178 i = 0; 2179 while (tpindex < pindex) { 2180 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2181 VM_ALLOC_NULL_OK); 2182 if (rtm == NULL) { 2183 for (j = 0; j < i; j++) { 2184 vm_page_free(marray[j]); 2185 } 2186 vm_object_drop(object); 2187 marray[0] = m; 2188 *reqpage = 0; 2189 return 1; 2190 } 2191 marray[i] = rtm; 2192 ++i; 2193 ++tpindex; 2194 } 2195 vm_object_drop(object); 2196 } else { 2197 i = 0; 2198 } 2199 2200 /* 2201 * Assign requested page 2202 */ 2203 marray[i] = m; 2204 *reqpage = i; 2205 ++i; 2206 2207 /* 2208 * Scan forwards for read-ahead pages 2209 */ 2210 tpindex = pindex + 1; 2211 endpindex = tpindex + rahead; 2212 if (endpindex > object->size) 2213 endpindex = object->size; 2214 2215 vm_object_hold(object); 2216 while (tpindex < endpindex) { 2217 if (vm_page_lookup(object, tpindex)) 2218 break; 2219 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2220 VM_ALLOC_NULL_OK); 2221 if (rtm == NULL) 2222 break; 2223 marray[i] = rtm; 2224 ++i; 2225 ++tpindex; 2226 } 2227 vm_object_drop(object); 2228 2229 return (i); 2230 } 2231 2232 #endif 2233 2234 /* 2235 * vm_prefault() provides a quick way of clustering pagefaults into a 2236 * processes address space. It is a "cousin" of pmap_object_init_pt, 2237 * except it runs at page fault time instead of mmap time. 2238 * 2239 * vm.fast_fault Enables pre-faulting zero-fill pages 2240 * 2241 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2242 * prefault. Scan stops in either direction when 2243 * a page is found to already exist. 2244 * 2245 * This code used to be per-platform pmap_prefault(). It is now 2246 * machine-independent and enhanced to also pre-fault zero-fill pages 2247 * (see vm.fast_fault) as well as make them writable, which greatly 2248 * reduces the number of page faults programs incur. 2249 * 2250 * Application performance when pre-faulting zero-fill pages is heavily 2251 * dependent on the application. Very tiny applications like /bin/echo 2252 * lose a little performance while applications of any appreciable size 2253 * gain performance. Prefaulting multiple pages also reduces SMP 2254 * congestion and can improve SMP performance significantly. 2255 * 2256 * NOTE! prot may allow writing but this only applies to the top level 2257 * object. If we wind up mapping a page extracted from a backing 2258 * object we have to make sure it is read-only. 2259 * 2260 * NOTE! The caller has already handled any COW operations on the 2261 * vm_map_entry via the normal fault code. Do NOT call this 2262 * shortcut unless the normal fault code has run on this entry. 2263 * 2264 * The related map must be locked. 2265 * No other requirements. 2266 */ 2267 static int vm_prefault_pages = 8; 2268 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2269 "Maximum number of pages to pre-fault"); 2270 static int vm_fast_fault = 1; 2271 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2272 "Burst fault zero-fill regions"); 2273 2274 /* 2275 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2276 * is not already dirty by other means. This will prevent passive 2277 * filesystem syncing as well as 'sync' from writing out the page. 2278 */ 2279 static void 2280 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2281 { 2282 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2283 if (m->dirty == 0) 2284 vm_page_flag_set(m, PG_NOSYNC); 2285 } else { 2286 vm_page_flag_clear(m, PG_NOSYNC); 2287 } 2288 } 2289 2290 static void 2291 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 2292 int fault_flags) 2293 { 2294 struct lwp *lp; 2295 vm_page_t m; 2296 vm_offset_t addr; 2297 vm_pindex_t index; 2298 vm_pindex_t pindex; 2299 vm_object_t object; 2300 int pprot; 2301 int i; 2302 int noneg; 2303 int nopos; 2304 int maxpages; 2305 2306 /* 2307 * Get stable max count value, disabled if set to 0 2308 */ 2309 maxpages = vm_prefault_pages; 2310 cpu_ccfence(); 2311 if (maxpages <= 0) 2312 return; 2313 2314 /* 2315 * We do not currently prefault mappings that use virtual page 2316 * tables. We do not prefault foreign pmaps. 2317 */ 2318 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) 2319 return; 2320 lp = curthread->td_lwp; 2321 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2322 return; 2323 2324 /* 2325 * Limit pre-fault count to 1024 pages. 2326 */ 2327 if (maxpages > 1024) 2328 maxpages = 1024; 2329 2330 object = entry->object.vm_object; 2331 KKASSERT(object != NULL); 2332 KKASSERT(object == entry->object.vm_object); 2333 vm_object_hold(object); 2334 vm_object_chain_acquire(object); 2335 2336 noneg = 0; 2337 nopos = 0; 2338 for (i = 0; i < maxpages; ++i) { 2339 vm_object_t lobject; 2340 vm_object_t nobject; 2341 int allocated = 0; 2342 int error; 2343 2344 /* 2345 * This can eat a lot of time on a heavily contended 2346 * machine so yield on the tick if needed. 2347 */ 2348 if ((i & 7) == 7) 2349 lwkt_yield(); 2350 2351 /* 2352 * Calculate the page to pre-fault, stopping the scan in 2353 * each direction separately if the limit is reached. 2354 */ 2355 if (i & 1) { 2356 if (noneg) 2357 continue; 2358 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2359 } else { 2360 if (nopos) 2361 continue; 2362 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2363 } 2364 if (addr < entry->start) { 2365 noneg = 1; 2366 if (noneg && nopos) 2367 break; 2368 continue; 2369 } 2370 if (addr >= entry->end) { 2371 nopos = 1; 2372 if (noneg && nopos) 2373 break; 2374 continue; 2375 } 2376 2377 /* 2378 * Skip pages already mapped, and stop scanning in that 2379 * direction. When the scan terminates in both directions 2380 * we are done. 2381 */ 2382 if (pmap_prefault_ok(pmap, addr) == 0) { 2383 if (i & 1) 2384 noneg = 1; 2385 else 2386 nopos = 1; 2387 if (noneg && nopos) 2388 break; 2389 continue; 2390 } 2391 2392 /* 2393 * Follow the VM object chain to obtain the page to be mapped 2394 * into the pmap. 2395 * 2396 * If we reach the terminal object without finding a page 2397 * and we determine it would be advantageous, then allocate 2398 * a zero-fill page for the base object. The base object 2399 * is guaranteed to be OBJT_DEFAULT for this case. 2400 * 2401 * In order to not have to check the pager via *haspage*() 2402 * we stop if any non-default object is encountered. e.g. 2403 * a vnode or swap object would stop the loop. 2404 */ 2405 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2406 lobject = object; 2407 pindex = index; 2408 pprot = prot; 2409 2410 KKASSERT(lobject == entry->object.vm_object); 2411 /*vm_object_hold(lobject); implied */ 2412 2413 while ((m = vm_page_lookup_busy_try(lobject, pindex, 2414 TRUE, &error)) == NULL) { 2415 if (lobject->type != OBJT_DEFAULT) 2416 break; 2417 if (lobject->backing_object == NULL) { 2418 if (vm_fast_fault == 0) 2419 break; 2420 if ((prot & VM_PROT_WRITE) == 0 || 2421 vm_page_count_min(0)) { 2422 break; 2423 } 2424 2425 /* 2426 * NOTE: Allocated from base object 2427 */ 2428 m = vm_page_alloc(object, index, 2429 VM_ALLOC_NORMAL | 2430 VM_ALLOC_ZERO | 2431 VM_ALLOC_USE_GD | 2432 VM_ALLOC_NULL_OK); 2433 if (m == NULL) 2434 break; 2435 allocated = 1; 2436 pprot = prot; 2437 /* lobject = object .. not needed */ 2438 break; 2439 } 2440 if (lobject->backing_object_offset & PAGE_MASK) 2441 break; 2442 nobject = lobject->backing_object; 2443 vm_object_hold(nobject); 2444 KKASSERT(nobject == lobject->backing_object); 2445 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 2446 if (lobject != object) { 2447 vm_object_lock_swap(); 2448 vm_object_drop(lobject); 2449 } 2450 lobject = nobject; 2451 pprot &= ~VM_PROT_WRITE; 2452 vm_object_chain_acquire(lobject); 2453 } 2454 2455 /* 2456 * NOTE: A non-NULL (m) will be associated with lobject if 2457 * it was found there, otherwise it is probably a 2458 * zero-fill page associated with the base object. 2459 * 2460 * Give-up if no page is available. 2461 */ 2462 if (m == NULL) { 2463 if (lobject != object) { 2464 if (object->backing_object != lobject) 2465 vm_object_hold(object->backing_object); 2466 vm_object_chain_release_all( 2467 object->backing_object, lobject); 2468 if (object->backing_object != lobject) 2469 vm_object_drop(object->backing_object); 2470 vm_object_drop(lobject); 2471 } 2472 break; 2473 } 2474 2475 /* 2476 * The object must be marked dirty if we are mapping a 2477 * writable page. m->object is either lobject or object, 2478 * both of which are still held. Do this before we 2479 * potentially drop the object. 2480 */ 2481 if (pprot & VM_PROT_WRITE) 2482 vm_object_set_writeable_dirty(m->object); 2483 2484 /* 2485 * Do not conditionalize on PG_RAM. If pages are present in 2486 * the VM system we assume optimal caching. If caching is 2487 * not optimal the I/O gravy train will be restarted when we 2488 * hit an unavailable page. We do not want to try to restart 2489 * the gravy train now because we really don't know how much 2490 * of the object has been cached. The cost for restarting 2491 * the gravy train should be low (since accesses will likely 2492 * be I/O bound anyway). 2493 */ 2494 if (lobject != object) { 2495 if (object->backing_object != lobject) 2496 vm_object_hold(object->backing_object); 2497 vm_object_chain_release_all(object->backing_object, 2498 lobject); 2499 if (object->backing_object != lobject) 2500 vm_object_drop(object->backing_object); 2501 vm_object_drop(lobject); 2502 } 2503 2504 /* 2505 * Enter the page into the pmap if appropriate. If we had 2506 * allocated the page we have to place it on a queue. If not 2507 * we just have to make sure it isn't on the cache queue 2508 * (pages on the cache queue are not allowed to be mapped). 2509 */ 2510 if (allocated) { 2511 /* 2512 * Page must be zerod. 2513 */ 2514 if ((m->flags & PG_ZERO) == 0) { 2515 vm_page_zero_fill(m); 2516 } else { 2517 #ifdef PMAP_DEBUG 2518 pmap_page_assertzero( 2519 VM_PAGE_TO_PHYS(m)); 2520 #endif 2521 vm_page_flag_clear(m, PG_ZERO); 2522 mycpu->gd_cnt.v_ozfod++; 2523 } 2524 mycpu->gd_cnt.v_zfod++; 2525 m->valid = VM_PAGE_BITS_ALL; 2526 2527 /* 2528 * Handle dirty page case 2529 */ 2530 if (pprot & VM_PROT_WRITE) 2531 vm_set_nosync(m, entry); 2532 pmap_enter(pmap, addr, m, pprot, 0, entry); 2533 mycpu->gd_cnt.v_vm_faults++; 2534 if (curthread->td_lwp) 2535 ++curthread->td_lwp->lwp_ru.ru_minflt; 2536 vm_page_deactivate(m); 2537 if (pprot & VM_PROT_WRITE) { 2538 /*vm_object_set_writeable_dirty(m->object);*/ 2539 vm_set_nosync(m, entry); 2540 if (fault_flags & VM_FAULT_DIRTY) { 2541 vm_page_dirty(m); 2542 /*XXX*/ 2543 swap_pager_unswapped(m); 2544 } 2545 } 2546 vm_page_wakeup(m); 2547 } else if (error) { 2548 /* couldn't busy page, no wakeup */ 2549 } else if ( 2550 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2551 (m->flags & PG_FICTITIOUS) == 0) { 2552 /* 2553 * A fully valid page not undergoing soft I/O can 2554 * be immediately entered into the pmap. 2555 */ 2556 if ((m->queue - m->pc) == PQ_CACHE) 2557 vm_page_deactivate(m); 2558 if (pprot & VM_PROT_WRITE) { 2559 /*vm_object_set_writeable_dirty(m->object);*/ 2560 vm_set_nosync(m, entry); 2561 if (fault_flags & VM_FAULT_DIRTY) { 2562 vm_page_dirty(m); 2563 /*XXX*/ 2564 swap_pager_unswapped(m); 2565 } 2566 } 2567 if (pprot & VM_PROT_WRITE) 2568 vm_set_nosync(m, entry); 2569 pmap_enter(pmap, addr, m, pprot, 0, entry); 2570 mycpu->gd_cnt.v_vm_faults++; 2571 if (curthread->td_lwp) 2572 ++curthread->td_lwp->lwp_ru.ru_minflt; 2573 vm_page_wakeup(m); 2574 } else { 2575 vm_page_wakeup(m); 2576 } 2577 } 2578 vm_object_chain_release(object); 2579 vm_object_drop(object); 2580 } 2581 2582 static void 2583 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 2584 vm_map_entry_t entry, int prot, int fault_flags) 2585 { 2586 struct lwp *lp; 2587 vm_page_t m; 2588 vm_offset_t addr; 2589 vm_pindex_t pindex; 2590 vm_object_t object; 2591 int i; 2592 int noneg; 2593 int nopos; 2594 int maxpages; 2595 2596 /* 2597 * Get stable max count value, disabled if set to 0 2598 */ 2599 maxpages = vm_prefault_pages; 2600 cpu_ccfence(); 2601 if (maxpages <= 0) 2602 return; 2603 2604 /* 2605 * We do not currently prefault mappings that use virtual page 2606 * tables. We do not prefault foreign pmaps. 2607 */ 2608 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) 2609 return; 2610 lp = curthread->td_lwp; 2611 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2612 return; 2613 2614 /* 2615 * Limit pre-fault count to 1024 pages. 2616 */ 2617 if (maxpages > 1024) 2618 maxpages = 1024; 2619 2620 object = entry->object.vm_object; 2621 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2622 KKASSERT(object->backing_object == NULL); 2623 2624 noneg = 0; 2625 nopos = 0; 2626 for (i = 0; i < maxpages; ++i) { 2627 int error; 2628 2629 /* 2630 * Calculate the page to pre-fault, stopping the scan in 2631 * each direction separately if the limit is reached. 2632 */ 2633 if (i & 1) { 2634 if (noneg) 2635 continue; 2636 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2637 } else { 2638 if (nopos) 2639 continue; 2640 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2641 } 2642 if (addr < entry->start) { 2643 noneg = 1; 2644 if (noneg && nopos) 2645 break; 2646 continue; 2647 } 2648 if (addr >= entry->end) { 2649 nopos = 1; 2650 if (noneg && nopos) 2651 break; 2652 continue; 2653 } 2654 2655 /* 2656 * Skip pages already mapped, and stop scanning in that 2657 * direction. When the scan terminates in both directions 2658 * we are done. 2659 */ 2660 if (pmap_prefault_ok(pmap, addr) == 0) { 2661 if (i & 1) 2662 noneg = 1; 2663 else 2664 nopos = 1; 2665 if (noneg && nopos) 2666 break; 2667 continue; 2668 } 2669 2670 /* 2671 * Follow the VM object chain to obtain the page to be mapped 2672 * into the pmap. This version of the prefault code only 2673 * works with terminal objects. 2674 * 2675 * WARNING! We cannot call swap_pager_unswapped() with a 2676 * shared token. 2677 */ 2678 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2679 2680 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 2681 if (m == NULL || error) 2682 continue; 2683 2684 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2685 (m->flags & PG_FICTITIOUS) == 0 && 2686 ((m->flags & PG_SWAPPED) == 0 || 2687 (prot & VM_PROT_WRITE) == 0 || 2688 (fault_flags & VM_FAULT_DIRTY) == 0)) { 2689 /* 2690 * A fully valid page not undergoing soft I/O can 2691 * be immediately entered into the pmap. 2692 */ 2693 if ((m->queue - m->pc) == PQ_CACHE) 2694 vm_page_deactivate(m); 2695 if (prot & VM_PROT_WRITE) { 2696 vm_object_set_writeable_dirty(m->object); 2697 vm_set_nosync(m, entry); 2698 if (fault_flags & VM_FAULT_DIRTY) { 2699 vm_page_dirty(m); 2700 /*XXX*/ 2701 swap_pager_unswapped(m); 2702 } 2703 } 2704 pmap_enter(pmap, addr, m, prot, 0, entry); 2705 mycpu->gd_cnt.v_vm_faults++; 2706 if (curthread->td_lwp) 2707 ++curthread->td_lwp->lwp_ru.ru_minflt; 2708 } 2709 vm_page_wakeup(m); 2710 } 2711 } 2712