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