1 /* 2 * Copyright (c) 2003-2014 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * --- 35 * 36 * Copyright (c) 1991, 1993 37 * The Regents of the University of California. All rights reserved. 38 * Copyright (c) 1994 John S. Dyson 39 * All rights reserved. 40 * Copyright (c) 1994 David Greenman 41 * All rights reserved. 42 * 43 * 44 * This code is derived from software contributed to Berkeley by 45 * The Mach Operating System project at Carnegie-Mellon University. 46 * 47 * Redistribution and use in source and binary forms, with or without 48 * modification, are permitted provided that the following conditions 49 * are met: 50 * 1. Redistributions of source code must retain the above copyright 51 * notice, this list of conditions and the following disclaimer. 52 * 2. Redistributions in binary form must reproduce the above copyright 53 * notice, this list of conditions and the following disclaimer in the 54 * documentation and/or other materials provided with the distribution. 55 * 3. Neither the name of the University nor the names of its contributors 56 * may be used to endorse or promote products derived from this software 57 * without specific prior written permission. 58 * 59 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 60 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 61 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 62 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 63 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 64 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 65 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 66 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 67 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 68 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 69 * SUCH DAMAGE. 70 * 71 * --- 72 * 73 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 74 * All rights reserved. 75 * 76 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 77 * 78 * Permission to use, copy, modify and distribute this software and 79 * its documentation is hereby granted, provided that both the copyright 80 * notice and this permission notice appear in all copies of the 81 * software, derivative works or modified versions, and any portions 82 * thereof, and that both notices appear in supporting documentation. 83 * 84 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 85 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 86 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 87 * 88 * Carnegie Mellon requests users of this software to return to 89 * 90 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 91 * School of Computer Science 92 * Carnegie Mellon University 93 * Pittsburgh PA 15213-3890 94 * 95 * any improvements or extensions that they make and grant Carnegie the 96 * rights to redistribute these changes. 97 */ 98 99 /* 100 * Page fault handling module. 101 */ 102 103 #include <sys/param.h> 104 #include <sys/systm.h> 105 #include <sys/kernel.h> 106 #include <sys/proc.h> 107 #include <sys/vnode.h> 108 #include <sys/resourcevar.h> 109 #include <sys/vmmeter.h> 110 #include <sys/vkernel.h> 111 #include <sys/lock.h> 112 #include <sys/sysctl.h> 113 114 #include <cpu/lwbuf.h> 115 116 #include <vm/vm.h> 117 #include <vm/vm_param.h> 118 #include <vm/pmap.h> 119 #include <vm/vm_map.h> 120 #include <vm/vm_object.h> 121 #include <vm/vm_page.h> 122 #include <vm/vm_pageout.h> 123 #include <vm/vm_kern.h> 124 #include <vm/vm_pager.h> 125 #include <vm/vnode_pager.h> 126 #include <vm/vm_extern.h> 127 128 #include <sys/thread2.h> 129 #include <vm/vm_page2.h> 130 131 struct faultstate { 132 vm_page_t m; 133 vm_object_t object; 134 vm_pindex_t pindex; 135 vm_prot_t prot; 136 vm_page_t first_m; 137 vm_object_t first_object; 138 vm_prot_t first_prot; 139 vm_map_t map; 140 vm_map_entry_t entry; 141 int lookup_still_valid; 142 int hardfault; 143 int fault_flags; 144 int map_generation; 145 int shared; 146 int first_shared; 147 boolean_t wired; 148 struct vnode *vp; 149 }; 150 151 static int debug_fault = 0; 152 SYSCTL_INT(_vm, OID_AUTO, debug_fault, CTLFLAG_RW, &debug_fault, 0, ""); 153 static int debug_cluster = 0; 154 SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, ""); 155 int vm_shared_fault = 1; 156 TUNABLE_INT("vm.shared_fault", &vm_shared_fault); 157 SYSCTL_INT(_vm, OID_AUTO, shared_fault, CTLFLAG_RW, &vm_shared_fault, 0, 158 "Allow shared token on vm_object"); 159 160 static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t, int); 161 static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, 162 vpte_t, int, int); 163 #if 0 164 static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *); 165 #endif 166 static void vm_set_nosync(vm_page_t m, vm_map_entry_t entry); 167 static void vm_prefault(pmap_t pmap, vm_offset_t addra, 168 vm_map_entry_t entry, int prot, int fault_flags); 169 static void vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 170 vm_map_entry_t entry, int prot, int fault_flags); 171 172 static __inline void 173 release_page(struct faultstate *fs) 174 { 175 vm_page_deactivate(fs->m); 176 vm_page_wakeup(fs->m); 177 fs->m = NULL; 178 } 179 180 /* 181 * NOTE: Once unlocked any cached fs->entry becomes invalid, any reuse 182 * requires relocking and then checking the timestamp. 183 * 184 * NOTE: vm_map_lock_read() does not bump fs->map->timestamp so we do 185 * not have to update fs->map_generation here. 186 * 187 * NOTE: This function can fail due to a deadlock against the caller's 188 * holding of a vm_page BUSY. 189 */ 190 static __inline int 191 relock_map(struct faultstate *fs) 192 { 193 int error; 194 195 if (fs->lookup_still_valid == FALSE && fs->map) { 196 error = vm_map_lock_read_to(fs->map); 197 if (error == 0) 198 fs->lookup_still_valid = TRUE; 199 } else { 200 error = 0; 201 } 202 return error; 203 } 204 205 static __inline void 206 unlock_map(struct faultstate *fs) 207 { 208 if (fs->lookup_still_valid && fs->map) { 209 vm_map_lookup_done(fs->map, fs->entry, 0); 210 fs->lookup_still_valid = FALSE; 211 } 212 } 213 214 /* 215 * Clean up after a successful call to vm_fault_object() so another call 216 * to vm_fault_object() can be made. 217 */ 218 static void 219 _cleanup_successful_fault(struct faultstate *fs, int relock) 220 { 221 /* 222 * We allocated a junk page for a COW operation that did 223 * not occur, the page must be freed. 224 */ 225 if (fs->object != fs->first_object) { 226 KKASSERT(fs->first_shared == 0); 227 vm_page_free(fs->first_m); 228 vm_object_pip_wakeup(fs->object); 229 fs->first_m = NULL; 230 } 231 232 /* 233 * Reset fs->object. 234 */ 235 fs->object = fs->first_object; 236 if (relock && fs->lookup_still_valid == FALSE) { 237 if (fs->map) 238 vm_map_lock_read(fs->map); 239 fs->lookup_still_valid = TRUE; 240 } 241 } 242 243 static void 244 _unlock_things(struct faultstate *fs, int dealloc) 245 { 246 _cleanup_successful_fault(fs, 0); 247 if (dealloc) { 248 /*vm_object_deallocate(fs->first_object);*/ 249 /*fs->first_object = NULL; drop used later on */ 250 } 251 unlock_map(fs); 252 if (fs->vp != NULL) { 253 vput(fs->vp); 254 fs->vp = NULL; 255 } 256 } 257 258 #define unlock_things(fs) _unlock_things(fs, 0) 259 #define unlock_and_deallocate(fs) _unlock_things(fs, 1) 260 #define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1) 261 262 /* 263 * TRYPAGER 264 * 265 * Determine if the pager for the current object *might* contain the page. 266 * 267 * We only need to try the pager if this is not a default object (default 268 * objects are zero-fill and have no real pager), and if we are not taking 269 * a wiring fault or if the FS entry is wired. 270 */ 271 #define TRYPAGER(fs) \ 272 (fs->object->type != OBJT_DEFAULT && \ 273 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired)) 274 275 /* 276 * vm_fault: 277 * 278 * Handle a page fault occuring at the given address, requiring the given 279 * permissions, in the map specified. If successful, the page is inserted 280 * into the associated physical map. 281 * 282 * NOTE: The given address should be truncated to the proper page address. 283 * 284 * KERN_SUCCESS is returned if the page fault is handled; otherwise, 285 * a standard error specifying why the fault is fatal is returned. 286 * 287 * The map in question must be referenced, and remains so. 288 * The caller may hold no locks. 289 * No other requirements. 290 */ 291 int 292 vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags) 293 { 294 int result; 295 vm_pindex_t first_pindex; 296 struct faultstate fs; 297 struct lwp *lp; 298 struct proc *p; 299 thread_t td; 300 struct vm_map_ilock ilock; 301 int didilock; 302 int growstack; 303 int retry = 0; 304 int inherit_prot; 305 306 inherit_prot = fault_type & VM_PROT_NOSYNC; 307 fs.hardfault = 0; 308 fs.fault_flags = fault_flags; 309 fs.vp = NULL; 310 fs.shared = vm_shared_fault; 311 fs.first_shared = vm_shared_fault; 312 growstack = 1; 313 314 /* 315 * vm_map interactions 316 */ 317 td = curthread; 318 if ((lp = td->td_lwp) != NULL) 319 lp->lwp_flags |= LWP_PAGING; 320 321 RetryFault: 322 /* 323 * Find the vm_map_entry representing the backing store and resolve 324 * the top level object and page index. This may have the side 325 * effect of executing a copy-on-write on the map entry, 326 * creating a shadow object, or splitting an anonymous entry for 327 * performance, but will not COW any actual VM pages. 328 * 329 * On success fs.map is left read-locked and various other fields 330 * are initialized but not otherwise referenced or locked. 331 * 332 * NOTE! vm_map_lookup will try to upgrade the fault_type to 333 * VM_FAULT_WRITE if the map entry is a virtual page table 334 * and also writable, so we can set the 'A'accessed bit in 335 * the virtual page table entry. 336 */ 337 fs.map = map; 338 result = vm_map_lookup(&fs.map, vaddr, fault_type, 339 &fs.entry, &fs.first_object, 340 &first_pindex, &fs.first_prot, &fs.wired); 341 342 /* 343 * If the lookup failed or the map protections are incompatible, 344 * the fault generally fails. 345 * 346 * The failure could be due to TDF_NOFAULT if vm_map_lookup() 347 * tried to do a COW fault. 348 * 349 * If the caller is trying to do a user wiring we have more work 350 * to do. 351 */ 352 if (result != KERN_SUCCESS) { 353 if (result == KERN_FAILURE_NOFAULT) { 354 result = KERN_FAILURE; 355 goto done; 356 } 357 if (result != KERN_PROTECTION_FAILURE || 358 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 359 { 360 if (result == KERN_INVALID_ADDRESS && growstack && 361 map != &kernel_map && curproc != NULL) { 362 result = vm_map_growstack(map, vaddr); 363 if (result == KERN_SUCCESS) { 364 growstack = 0; 365 ++retry; 366 goto RetryFault; 367 } 368 result = KERN_FAILURE; 369 } 370 goto done; 371 } 372 373 /* 374 * If we are user-wiring a r/w segment, and it is COW, then 375 * we need to do the COW operation. Note that we don't 376 * currently COW RO sections now, because it is NOT desirable 377 * to COW .text. We simply keep .text from ever being COW'ed 378 * and take the heat that one cannot debug wired .text sections. 379 */ 380 result = vm_map_lookup(&fs.map, vaddr, 381 VM_PROT_READ|VM_PROT_WRITE| 382 VM_PROT_OVERRIDE_WRITE, 383 &fs.entry, &fs.first_object, 384 &first_pindex, &fs.first_prot, 385 &fs.wired); 386 if (result != KERN_SUCCESS) { 387 /* could also be KERN_FAILURE_NOFAULT */ 388 result = KERN_FAILURE; 389 goto done; 390 } 391 392 /* 393 * If we don't COW now, on a user wire, the user will never 394 * be able to write to the mapping. If we don't make this 395 * restriction, the bookkeeping would be nearly impossible. 396 * 397 * XXX We have a shared lock, this will have a MP race but 398 * I don't see how it can hurt anything. 399 */ 400 if ((fs.entry->protection & VM_PROT_WRITE) == 0) { 401 atomic_clear_char(&fs.entry->max_protection, 402 VM_PROT_WRITE); 403 } 404 } 405 406 /* 407 * fs.map is read-locked 408 * 409 * Misc checks. Save the map generation number to detect races. 410 */ 411 fs.map_generation = fs.map->timestamp; 412 fs.lookup_still_valid = TRUE; 413 fs.first_m = NULL; 414 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 415 fs.prot = fs.first_prot; /* default (used by uksmap) */ 416 417 if (fs.entry->eflags & (MAP_ENTRY_NOFAULT | MAP_ENTRY_KSTACK)) { 418 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 419 panic("vm_fault: fault on nofault entry, addr: %p", 420 (void *)vaddr); 421 } 422 if ((fs.entry->eflags & MAP_ENTRY_KSTACK) && 423 vaddr >= fs.entry->start && 424 vaddr < fs.entry->start + PAGE_SIZE) { 425 panic("vm_fault: fault on stack guard, addr: %p", 426 (void *)vaddr); 427 } 428 } 429 430 /* 431 * A user-kernel shared map has no VM object and bypasses 432 * everything. We execute the uksmap function with a temporary 433 * fictitious vm_page. The address is directly mapped with no 434 * management. 435 */ 436 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 437 struct vm_page fakem; 438 439 bzero(&fakem, sizeof(fakem)); 440 fakem.pindex = first_pindex; 441 fakem.flags = PG_FICTITIOUS | PG_UNMANAGED; 442 fakem.busy_count = PBUSY_LOCKED; 443 fakem.valid = VM_PAGE_BITS_ALL; 444 fakem.pat_mode = VM_MEMATTR_DEFAULT; 445 if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) { 446 result = KERN_FAILURE; 447 unlock_things(&fs); 448 goto done2; 449 } 450 pmap_enter(fs.map->pmap, vaddr, &fakem, fs.prot | inherit_prot, 451 fs.wired, fs.entry); 452 goto done_success; 453 } 454 455 /* 456 * A system map entry may return a NULL object. No object means 457 * no pager means an unrecoverable kernel fault. 458 */ 459 if (fs.first_object == NULL) { 460 panic("vm_fault: unrecoverable fault at %p in entry %p", 461 (void *)vaddr, fs.entry); 462 } 463 464 /* 465 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 466 * is set. 467 * 468 * Unfortunately a deadlock can occur if we are forced to page-in 469 * from swap, but diving all the way into the vm_pager_get_page() 470 * function to find out is too much. Just check the object type. 471 * 472 * The deadlock is a CAM deadlock on a busy VM page when trying 473 * to finish an I/O if another process gets stuck in 474 * vop_helper_read_shortcut() due to a swap fault. 475 */ 476 if ((td->td_flags & TDF_NOFAULT) && 477 (retry || 478 fs.first_object->type == OBJT_VNODE || 479 fs.first_object->type == OBJT_SWAP || 480 fs.first_object->backing_object)) { 481 result = KERN_FAILURE; 482 unlock_things(&fs); 483 goto done2; 484 } 485 486 /* 487 * If the entry is wired we cannot change the page protection. 488 */ 489 if (fs.wired) 490 fault_type = fs.first_prot; 491 492 /* 493 * We generally want to avoid unnecessary exclusive modes on backing 494 * and terminal objects because this can seriously interfere with 495 * heavily fork()'d processes (particularly /bin/sh scripts). 496 * 497 * However, we also want to avoid unnecessary retries due to needed 498 * shared->exclusive promotion for common faults. Exclusive mode is 499 * always needed if any page insertion, rename, or free occurs in an 500 * object (and also indirectly if any I/O is done). 501 * 502 * The main issue here is going to be fs.first_shared. If the 503 * first_object has a backing object which isn't shadowed and the 504 * process is single-threaded we might as well use an exclusive 505 * lock/chain right off the bat. 506 */ 507 if (fs.first_shared && fs.first_object->backing_object && 508 LIST_EMPTY(&fs.first_object->shadow_head) && 509 td->td_proc && td->td_proc->p_nthreads == 1) { 510 fs.first_shared = 0; 511 } 512 513 /* 514 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 515 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 516 * we can try shared first. 517 */ 518 if (fault_flags & VM_FAULT_UNSWAP) { 519 fs.first_shared = 0; 520 } 521 522 /* 523 * Obtain a top-level object lock, shared or exclusive depending 524 * on fs.first_shared. If a shared lock winds up being insufficient 525 * we will retry with an exclusive lock. 526 * 527 * The vnode pager lock is always shared. 528 */ 529 if (fs.first_shared) 530 vm_object_hold_shared(fs.first_object); 531 else 532 vm_object_hold(fs.first_object); 533 if (fs.vp == NULL) 534 fs.vp = vnode_pager_lock(fs.first_object); 535 536 /* 537 * The page we want is at (first_object, first_pindex), but if the 538 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 539 * page table to figure out the actual pindex. 540 * 541 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 542 * ONLY 543 */ 544 didilock = 0; 545 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 546 vm_map_interlock(fs.map, &ilock, vaddr, vaddr + PAGE_SIZE); 547 didilock = 1; 548 result = vm_fault_vpagetable(&fs, &first_pindex, 549 fs.entry->aux.master_pde, 550 fault_type, 1); 551 if (result == KERN_TRY_AGAIN) { 552 vm_map_deinterlock(fs.map, &ilock); 553 vm_object_drop(fs.first_object); 554 ++retry; 555 goto RetryFault; 556 } 557 if (result != KERN_SUCCESS) { 558 vm_map_deinterlock(fs.map, &ilock); 559 goto done; 560 } 561 } 562 563 /* 564 * Now we have the actual (object, pindex), fault in the page. If 565 * vm_fault_object() fails it will unlock and deallocate the FS 566 * data. If it succeeds everything remains locked and fs->object 567 * will have an additional PIP count if it is not equal to 568 * fs->first_object 569 * 570 * vm_fault_object will set fs->prot for the pmap operation. It is 571 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the 572 * page can be safely written. However, it will force a read-only 573 * mapping for a read fault if the memory is managed by a virtual 574 * page table. 575 * 576 * If the fault code uses the shared object lock shortcut 577 * we must not try to burst (we can't allocate VM pages). 578 */ 579 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 580 581 if (debug_fault > 0) { 582 --debug_fault; 583 kprintf("VM_FAULT result %d addr=%jx type=%02x flags=%02x " 584 "fs.m=%p fs.prot=%02x fs.wired=%02x fs.entry=%p\n", 585 result, (intmax_t)vaddr, fault_type, fault_flags, 586 fs.m, fs.prot, fs.wired, fs.entry); 587 } 588 589 if (result == KERN_TRY_AGAIN) { 590 if (didilock) 591 vm_map_deinterlock(fs.map, &ilock); 592 vm_object_drop(fs.first_object); 593 ++retry; 594 goto RetryFault; 595 } 596 if (result != KERN_SUCCESS) { 597 if (didilock) 598 vm_map_deinterlock(fs.map, &ilock); 599 goto done; 600 } 601 602 /* 603 * On success vm_fault_object() does not unlock or deallocate, and fs.m 604 * will contain a busied page. 605 * 606 * Enter the page into the pmap and do pmap-related adjustments. 607 */ 608 KKASSERT(fs.lookup_still_valid == TRUE); 609 vm_page_flag_set(fs.m, PG_REFERENCED); 610 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot | inherit_prot, 611 fs.wired, fs.entry); 612 613 if (didilock) 614 vm_map_deinterlock(fs.map, &ilock); 615 616 /*KKASSERT(fs.m->queue == PQ_NONE); page-in op may deactivate page */ 617 KKASSERT(fs.m->busy_count & PBUSY_LOCKED); 618 619 /* 620 * If the page is not wired down, then put it where the pageout daemon 621 * can find it. 622 */ 623 if (fs.fault_flags & VM_FAULT_WIRE_MASK) { 624 if (fs.wired) 625 vm_page_wire(fs.m); 626 else 627 vm_page_unwire(fs.m, 1); 628 } else { 629 vm_page_activate(fs.m); 630 } 631 vm_page_wakeup(fs.m); 632 633 /* 634 * Burst in a few more pages if possible. The fs.map should still 635 * be locked. To avoid interlocking against a vnode->getblk 636 * operation we had to be sure to unbusy our primary vm_page above 637 * first. 638 * 639 * A normal burst can continue down backing store, only execute 640 * if we are holding an exclusive lock, otherwise the exclusive 641 * locks the burst code gets might cause excessive SMP collisions. 642 * 643 * A quick burst can be utilized when there is no backing object 644 * (i.e. a shared file mmap). 645 */ 646 if ((fault_flags & VM_FAULT_BURST) && 647 (fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 && 648 fs.wired == 0) { 649 if (fs.first_shared == 0 && fs.shared == 0) { 650 vm_prefault(fs.map->pmap, vaddr, 651 fs.entry, fs.prot, fault_flags); 652 } else { 653 vm_prefault_quick(fs.map->pmap, vaddr, 654 fs.entry, fs.prot, fault_flags); 655 } 656 } 657 658 done_success: 659 mycpu->gd_cnt.v_vm_faults++; 660 if (td->td_lwp) 661 ++td->td_lwp->lwp_ru.ru_minflt; 662 663 /* 664 * Unlock everything, and return 665 */ 666 unlock_things(&fs); 667 668 if (td->td_lwp) { 669 if (fs.hardfault) { 670 td->td_lwp->lwp_ru.ru_majflt++; 671 } else { 672 td->td_lwp->lwp_ru.ru_minflt++; 673 } 674 } 675 676 /*vm_object_deallocate(fs.first_object);*/ 677 /*fs.m = NULL; */ 678 /*fs.first_object = NULL; must still drop later */ 679 680 result = KERN_SUCCESS; 681 done: 682 if (fs.first_object) 683 vm_object_drop(fs.first_object); 684 done2: 685 if (lp) 686 lp->lwp_flags &= ~LWP_PAGING; 687 688 #if !defined(NO_SWAPPING) 689 /* 690 * Check the process RSS limit and force deactivation and 691 * (asynchronous) paging if necessary. This is a complex operation, 692 * only do it for direct user-mode faults, for now. 693 * 694 * To reduce overhead implement approximately a ~16MB hysteresis. 695 */ 696 p = td->td_proc; 697 if ((fault_flags & VM_FAULT_USERMODE) && lp && 698 p->p_limit && map->pmap && vm_pageout_memuse_mode >= 1 && 699 map != &kernel_map) { 700 vm_pindex_t limit; 701 vm_pindex_t size; 702 703 limit = OFF_TO_IDX(qmin(p->p_rlimit[RLIMIT_RSS].rlim_cur, 704 p->p_rlimit[RLIMIT_RSS].rlim_max)); 705 size = pmap_resident_tlnw_count(map->pmap); 706 if (limit >= 0 && size > 4096 && size - 4096 >= limit) { 707 vm_pageout_map_deactivate_pages(map, limit); 708 } 709 } 710 #endif 711 712 return (result); 713 } 714 715 /* 716 * Fault in the specified virtual address in the current process map, 717 * returning a held VM page or NULL. See vm_fault_page() for more 718 * information. 719 * 720 * No requirements. 721 */ 722 vm_page_t 723 vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, 724 int *errorp, int *busyp) 725 { 726 struct lwp *lp = curthread->td_lwp; 727 vm_page_t m; 728 729 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va, 730 fault_type, VM_FAULT_NORMAL, 731 errorp, busyp); 732 return(m); 733 } 734 735 /* 736 * Fault in the specified virtual address in the specified map, doing all 737 * necessary manipulation of the object store and all necessary I/O. Return 738 * a held VM page or NULL, and set *errorp. The related pmap is not 739 * updated. 740 * 741 * If busyp is not NULL then *busyp will be set to TRUE if this routine 742 * decides to return a busied page (aka VM_PROT_WRITE), or FALSE if it 743 * does not (VM_PROT_WRITE not specified or busyp is NULL). If busyp is 744 * NULL the returned page is only held. 745 * 746 * If the caller has no intention of writing to the page's contents, busyp 747 * can be passed as NULL along with VM_PROT_WRITE to force a COW operation 748 * without busying the page. 749 * 750 * The returned page will also be marked PG_REFERENCED. 751 * 752 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an 753 * error will be returned. 754 * 755 * No requirements. 756 */ 757 vm_page_t 758 vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, 759 int fault_flags, int *errorp, int *busyp) 760 { 761 vm_pindex_t first_pindex; 762 struct faultstate fs; 763 int result; 764 int retry; 765 int growstack; 766 vm_prot_t orig_fault_type = fault_type; 767 768 retry = 0; 769 fs.hardfault = 0; 770 fs.fault_flags = fault_flags; 771 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 772 773 /* 774 * Dive the pmap (concurrency possible). If we find the 775 * appropriate page we can terminate early and quickly. 776 * 777 * This works great for normal programs but will always return 778 * NULL for host lookups of vkernel maps in VMM mode. 779 * 780 * NOTE: pmap_fault_page_quick() might not busy the page. If 781 * VM_PROT_WRITE or VM_PROT_OVERRIDE_WRITE is set in 782 * fault_type and pmap_fault_page_quick() returns non-NULL, 783 * it will safely dirty the returned vm_page_t for us. We 784 * cannot safely dirty it here (it might not be busy). 785 */ 786 fs.m = pmap_fault_page_quick(map->pmap, vaddr, fault_type, busyp); 787 if (fs.m) { 788 *errorp = 0; 789 return(fs.m); 790 } 791 792 /* 793 * Otherwise take a concurrency hit and do a formal page 794 * fault. 795 */ 796 fs.vp = NULL; 797 fs.shared = vm_shared_fault; 798 fs.first_shared = vm_shared_fault; 799 growstack = 1; 800 801 /* 802 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 803 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 804 * we can try shared first. 805 */ 806 if (fault_flags & VM_FAULT_UNSWAP) { 807 fs.first_shared = 0; 808 } 809 810 RetryFault: 811 /* 812 * Find the vm_map_entry representing the backing store and resolve 813 * the top level object and page index. This may have the side 814 * effect of executing a copy-on-write on the map entry and/or 815 * creating a shadow object, but will not COW any actual VM pages. 816 * 817 * On success fs.map is left read-locked and various other fields 818 * are initialized but not otherwise referenced or locked. 819 * 820 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE 821 * if the map entry is a virtual page table and also writable, 822 * so we can set the 'A'accessed bit in the virtual page table 823 * entry. 824 */ 825 fs.map = map; 826 result = vm_map_lookup(&fs.map, vaddr, fault_type, 827 &fs.entry, &fs.first_object, 828 &first_pindex, &fs.first_prot, &fs.wired); 829 830 if (result != KERN_SUCCESS) { 831 if (result == KERN_FAILURE_NOFAULT) { 832 *errorp = KERN_FAILURE; 833 fs.m = NULL; 834 goto done; 835 } 836 if (result != KERN_PROTECTION_FAILURE || 837 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE) 838 { 839 if (result == KERN_INVALID_ADDRESS && growstack && 840 map != &kernel_map && curproc != NULL) { 841 result = vm_map_growstack(map, vaddr); 842 if (result == KERN_SUCCESS) { 843 growstack = 0; 844 ++retry; 845 goto RetryFault; 846 } 847 result = KERN_FAILURE; 848 } 849 fs.m = NULL; 850 *errorp = result; 851 goto done; 852 } 853 854 /* 855 * If we are user-wiring a r/w segment, and it is COW, then 856 * we need to do the COW operation. Note that we don't 857 * currently COW RO sections now, because it is NOT desirable 858 * to COW .text. We simply keep .text from ever being COW'ed 859 * and take the heat that one cannot debug wired .text sections. 860 */ 861 result = vm_map_lookup(&fs.map, vaddr, 862 VM_PROT_READ|VM_PROT_WRITE| 863 VM_PROT_OVERRIDE_WRITE, 864 &fs.entry, &fs.first_object, 865 &first_pindex, &fs.first_prot, 866 &fs.wired); 867 if (result != KERN_SUCCESS) { 868 /* could also be KERN_FAILURE_NOFAULT */ 869 *errorp = KERN_FAILURE; 870 fs.m = NULL; 871 goto done; 872 } 873 874 /* 875 * If we don't COW now, on a user wire, the user will never 876 * be able to write to the mapping. If we don't make this 877 * restriction, the bookkeeping would be nearly impossible. 878 * 879 * XXX We have a shared lock, this will have a MP race but 880 * I don't see how it can hurt anything. 881 */ 882 if ((fs.entry->protection & VM_PROT_WRITE) == 0) { 883 atomic_clear_char(&fs.entry->max_protection, 884 VM_PROT_WRITE); 885 } 886 } 887 888 /* 889 * fs.map is read-locked 890 * 891 * Misc checks. Save the map generation number to detect races. 892 */ 893 fs.map_generation = fs.map->timestamp; 894 fs.lookup_still_valid = TRUE; 895 fs.first_m = NULL; 896 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 897 898 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) { 899 panic("vm_fault: fault on nofault entry, addr: %lx", 900 (u_long)vaddr); 901 } 902 903 /* 904 * A user-kernel shared map has no VM object and bypasses 905 * everything. We execute the uksmap function with a temporary 906 * fictitious vm_page. The address is directly mapped with no 907 * management. 908 */ 909 if (fs.entry->maptype == VM_MAPTYPE_UKSMAP) { 910 struct vm_page fakem; 911 912 bzero(&fakem, sizeof(fakem)); 913 fakem.pindex = first_pindex; 914 fakem.flags = PG_FICTITIOUS | PG_UNMANAGED; 915 fakem.busy_count = PBUSY_LOCKED; 916 fakem.valid = VM_PAGE_BITS_ALL; 917 fakem.pat_mode = VM_MEMATTR_DEFAULT; 918 if (fs.entry->object.uksmap(fs.entry->aux.dev, &fakem)) { 919 *errorp = KERN_FAILURE; 920 fs.m = NULL; 921 unlock_things(&fs); 922 goto done2; 923 } 924 fs.m = PHYS_TO_VM_PAGE(fakem.phys_addr); 925 vm_page_hold(fs.m); 926 if (busyp) 927 *busyp = 0; /* don't need to busy R or W */ 928 unlock_things(&fs); 929 *errorp = 0; 930 goto done; 931 } 932 933 934 /* 935 * A system map entry may return a NULL object. No object means 936 * no pager means an unrecoverable kernel fault. 937 */ 938 if (fs.first_object == NULL) { 939 panic("vm_fault: unrecoverable fault at %p in entry %p", 940 (void *)vaddr, fs.entry); 941 } 942 943 /* 944 * Fail here if not a trivial anonymous page fault and TDF_NOFAULT 945 * is set. 946 * 947 * Unfortunately a deadlock can occur if we are forced to page-in 948 * from swap, but diving all the way into the vm_pager_get_page() 949 * function to find out is too much. Just check the object type. 950 */ 951 if ((curthread->td_flags & TDF_NOFAULT) && 952 (retry || 953 fs.first_object->type == OBJT_VNODE || 954 fs.first_object->type == OBJT_SWAP || 955 fs.first_object->backing_object)) { 956 *errorp = KERN_FAILURE; 957 unlock_things(&fs); 958 fs.m = NULL; 959 goto done2; 960 } 961 962 /* 963 * If the entry is wired we cannot change the page protection. 964 */ 965 if (fs.wired) 966 fault_type = fs.first_prot; 967 968 /* 969 * Make a reference to this object to prevent its disposal while we 970 * are messing with it. Once we have the reference, the map is free 971 * to be diddled. Since objects reference their shadows (and copies), 972 * they will stay around as well. 973 * 974 * The reference should also prevent an unexpected collapse of the 975 * parent that might move pages from the current object into the 976 * parent unexpectedly, resulting in corruption. 977 * 978 * Bump the paging-in-progress count to prevent size changes (e.g. 979 * truncation operations) during I/O. This must be done after 980 * obtaining the vnode lock in order to avoid possible deadlocks. 981 */ 982 if (fs.first_shared) 983 vm_object_hold_shared(fs.first_object); 984 else 985 vm_object_hold(fs.first_object); 986 if (fs.vp == NULL) 987 fs.vp = vnode_pager_lock(fs.first_object); /* shared */ 988 989 /* 990 * The page we want is at (first_object, first_pindex), but if the 991 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the 992 * page table to figure out the actual pindex. 993 * 994 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION 995 * ONLY 996 */ 997 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 998 result = vm_fault_vpagetable(&fs, &first_pindex, 999 fs.entry->aux.master_pde, 1000 fault_type, 1); 1001 if (result == KERN_TRY_AGAIN) { 1002 vm_object_drop(fs.first_object); 1003 ++retry; 1004 goto RetryFault; 1005 } 1006 if (result != KERN_SUCCESS) { 1007 *errorp = result; 1008 fs.m = NULL; 1009 goto done; 1010 } 1011 } 1012 1013 /* 1014 * Now we have the actual (object, pindex), fault in the page. If 1015 * vm_fault_object() fails it will unlock and deallocate the FS 1016 * data. If it succeeds everything remains locked and fs->object 1017 * will have an additinal PIP count if it is not equal to 1018 * fs->first_object 1019 */ 1020 fs.m = NULL; 1021 result = vm_fault_object(&fs, first_pindex, fault_type, 1); 1022 1023 if (result == KERN_TRY_AGAIN) { 1024 vm_object_drop(fs.first_object); 1025 ++retry; 1026 goto RetryFault; 1027 } 1028 if (result != KERN_SUCCESS) { 1029 *errorp = result; 1030 fs.m = NULL; 1031 goto done; 1032 } 1033 1034 if ((orig_fault_type & VM_PROT_WRITE) && 1035 (fs.prot & VM_PROT_WRITE) == 0) { 1036 *errorp = KERN_PROTECTION_FAILURE; 1037 unlock_and_deallocate(&fs); 1038 fs.m = NULL; 1039 goto done; 1040 } 1041 1042 /* 1043 * DO NOT UPDATE THE PMAP!!! This function may be called for 1044 * a pmap unrelated to the current process pmap, in which case 1045 * the current cpu core will not be listed in the pmap's pm_active 1046 * mask. Thus invalidation interlocks will fail to work properly. 1047 * 1048 * (for example, 'ps' uses procfs to read program arguments from 1049 * each process's stack). 1050 * 1051 * In addition to the above this function will be called to acquire 1052 * a page that might already be faulted in, re-faulting it 1053 * continuously is a waste of time. 1054 * 1055 * XXX could this have been the cause of our random seg-fault 1056 * issues? procfs accesses user stacks. 1057 */ 1058 vm_page_flag_set(fs.m, PG_REFERENCED); 1059 #if 0 1060 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired, NULL); 1061 mycpu->gd_cnt.v_vm_faults++; 1062 if (curthread->td_lwp) 1063 ++curthread->td_lwp->lwp_ru.ru_minflt; 1064 #endif 1065 1066 /* 1067 * On success vm_fault_object() does not unlock or deallocate, and fs.m 1068 * will contain a busied page. So we must unlock here after having 1069 * messed with the pmap. 1070 */ 1071 unlock_things(&fs); 1072 1073 /* 1074 * Return a held page. We are not doing any pmap manipulation so do 1075 * not set PG_MAPPED. However, adjust the page flags according to 1076 * the fault type because the caller may not use a managed pmapping 1077 * (so we don't want to lose the fact that the page will be dirtied 1078 * if a write fault was specified). 1079 */ 1080 if (fault_type & VM_PROT_WRITE) 1081 vm_page_dirty(fs.m); 1082 vm_page_activate(fs.m); 1083 1084 if (curthread->td_lwp) { 1085 if (fs.hardfault) { 1086 curthread->td_lwp->lwp_ru.ru_majflt++; 1087 } else { 1088 curthread->td_lwp->lwp_ru.ru_minflt++; 1089 } 1090 } 1091 1092 /* 1093 * Unlock everything, and return the held or busied page. 1094 */ 1095 if (busyp) { 1096 if (fault_type & (VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE)) { 1097 vm_page_dirty(fs.m); 1098 *busyp = 1; 1099 } else { 1100 *busyp = 0; 1101 vm_page_hold(fs.m); 1102 vm_page_wakeup(fs.m); 1103 } 1104 } else { 1105 vm_page_hold(fs.m); 1106 vm_page_wakeup(fs.m); 1107 } 1108 /*vm_object_deallocate(fs.first_object);*/ 1109 /*fs.first_object = NULL; */ 1110 *errorp = 0; 1111 1112 done: 1113 if (fs.first_object) 1114 vm_object_drop(fs.first_object); 1115 done2: 1116 return(fs.m); 1117 } 1118 1119 /* 1120 * Fault in the specified (object,offset), dirty the returned page as 1121 * needed. If the requested fault_type cannot be done NULL and an 1122 * error is returned. 1123 * 1124 * A held (but not busied) page is returned. 1125 * 1126 * The passed in object must be held as specified by the shared 1127 * argument. 1128 */ 1129 vm_page_t 1130 vm_fault_object_page(vm_object_t object, vm_ooffset_t offset, 1131 vm_prot_t fault_type, int fault_flags, 1132 int *sharedp, int *errorp) 1133 { 1134 int result; 1135 vm_pindex_t first_pindex; 1136 struct faultstate fs; 1137 struct vm_map_entry entry; 1138 1139 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1140 bzero(&entry, sizeof(entry)); 1141 entry.object.vm_object = object; 1142 entry.maptype = VM_MAPTYPE_NORMAL; 1143 entry.protection = entry.max_protection = fault_type; 1144 1145 fs.hardfault = 0; 1146 fs.fault_flags = fault_flags; 1147 fs.map = NULL; 1148 fs.shared = vm_shared_fault; 1149 fs.first_shared = *sharedp; 1150 fs.vp = NULL; 1151 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0); 1152 1153 /* 1154 * VM_FAULT_UNSWAP - swap_pager_unswapped() needs an exclusive object 1155 * VM_FAULT_DIRTY - may require swap_pager_unswapped() later, but 1156 * we can try shared first. 1157 */ 1158 if (fs.first_shared && (fault_flags & VM_FAULT_UNSWAP)) { 1159 fs.first_shared = 0; 1160 vm_object_upgrade(object); 1161 } 1162 1163 /* 1164 * Retry loop as needed (typically for shared->exclusive transitions) 1165 */ 1166 RetryFault: 1167 *sharedp = fs.first_shared; 1168 first_pindex = OFF_TO_IDX(offset); 1169 fs.first_object = object; 1170 fs.entry = &entry; 1171 fs.first_prot = fault_type; 1172 fs.wired = 0; 1173 /*fs.map_generation = 0; unused */ 1174 1175 /* 1176 * Make a reference to this object to prevent its disposal while we 1177 * are messing with it. Once we have the reference, the map is free 1178 * to be diddled. Since objects reference their shadows (and copies), 1179 * they will stay around as well. 1180 * 1181 * The reference should also prevent an unexpected collapse of the 1182 * parent that might move pages from the current object into the 1183 * parent unexpectedly, resulting in corruption. 1184 * 1185 * Bump the paging-in-progress count to prevent size changes (e.g. 1186 * truncation operations) during I/O. This must be done after 1187 * obtaining the vnode lock in order to avoid possible deadlocks. 1188 */ 1189 if (fs.vp == NULL) 1190 fs.vp = vnode_pager_lock(fs.first_object); 1191 1192 fs.lookup_still_valid = TRUE; 1193 fs.first_m = NULL; 1194 fs.object = fs.first_object; /* so unlock_and_deallocate works */ 1195 1196 #if 0 1197 /* XXX future - ability to operate on VM object using vpagetable */ 1198 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1199 result = vm_fault_vpagetable(&fs, &first_pindex, 1200 fs.entry->aux.master_pde, 1201 fault_type, 0); 1202 if (result == KERN_TRY_AGAIN) { 1203 if (fs.first_shared == 0 && *sharedp) 1204 vm_object_upgrade(object); 1205 goto RetryFault; 1206 } 1207 if (result != KERN_SUCCESS) { 1208 *errorp = result; 1209 return (NULL); 1210 } 1211 } 1212 #endif 1213 1214 /* 1215 * Now we have the actual (object, pindex), fault in the page. If 1216 * vm_fault_object() fails it will unlock and deallocate the FS 1217 * data. If it succeeds everything remains locked and fs->object 1218 * will have an additinal PIP count if it is not equal to 1219 * fs->first_object 1220 * 1221 * On KERN_TRY_AGAIN vm_fault_object() leaves fs.first_object intact. 1222 * We may have to upgrade its lock to handle the requested fault. 1223 */ 1224 result = vm_fault_object(&fs, first_pindex, fault_type, 0); 1225 1226 if (result == KERN_TRY_AGAIN) { 1227 if (fs.first_shared == 0 && *sharedp) 1228 vm_object_upgrade(object); 1229 goto RetryFault; 1230 } 1231 if (result != KERN_SUCCESS) { 1232 *errorp = result; 1233 return(NULL); 1234 } 1235 1236 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) { 1237 *errorp = KERN_PROTECTION_FAILURE; 1238 unlock_and_deallocate(&fs); 1239 return(NULL); 1240 } 1241 1242 /* 1243 * On success vm_fault_object() does not unlock or deallocate, so we 1244 * do it here. Note that the returned fs.m will be busied. 1245 */ 1246 unlock_things(&fs); 1247 1248 /* 1249 * Return a held page. We are not doing any pmap manipulation so do 1250 * not set PG_MAPPED. However, adjust the page flags according to 1251 * the fault type because the caller may not use a managed pmapping 1252 * (so we don't want to lose the fact that the page will be dirtied 1253 * if a write fault was specified). 1254 */ 1255 vm_page_hold(fs.m); 1256 vm_page_activate(fs.m); 1257 if ((fault_type & VM_PROT_WRITE) || (fault_flags & VM_FAULT_DIRTY)) 1258 vm_page_dirty(fs.m); 1259 if (fault_flags & VM_FAULT_UNSWAP) 1260 swap_pager_unswapped(fs.m); 1261 1262 /* 1263 * Indicate that the page was accessed. 1264 */ 1265 vm_page_flag_set(fs.m, PG_REFERENCED); 1266 1267 if (curthread->td_lwp) { 1268 if (fs.hardfault) { 1269 curthread->td_lwp->lwp_ru.ru_majflt++; 1270 } else { 1271 curthread->td_lwp->lwp_ru.ru_minflt++; 1272 } 1273 } 1274 1275 /* 1276 * Unlock everything, and return the held page. 1277 */ 1278 vm_page_wakeup(fs.m); 1279 /*vm_object_deallocate(fs.first_object);*/ 1280 /*fs.first_object = NULL; */ 1281 1282 *errorp = 0; 1283 return(fs.m); 1284 } 1285 1286 /* 1287 * Translate the virtual page number (first_pindex) that is relative 1288 * to the address space into a logical page number that is relative to the 1289 * backing object. Use the virtual page table pointed to by (vpte). 1290 * 1291 * Possibly downgrade the protection based on the vpte bits. 1292 * 1293 * This implements an N-level page table. Any level can terminate the 1294 * scan by setting VPTE_PS. A linear mapping is accomplished by setting 1295 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP). 1296 */ 1297 static 1298 int 1299 vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex, 1300 vpte_t vpte, int fault_type, int allow_nofault) 1301 { 1302 struct lwbuf *lwb; 1303 struct lwbuf lwb_cache; 1304 int vshift = VPTE_FRAME_END - PAGE_SHIFT; /* index bits remaining */ 1305 int result; 1306 vpte_t *ptep; 1307 1308 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1309 for (;;) { 1310 /* 1311 * We cannot proceed if the vpte is not valid, not readable 1312 * for a read fault, not writable for a write fault, or 1313 * not executable for an instruction execution fault. 1314 */ 1315 if ((vpte & VPTE_V) == 0) { 1316 unlock_and_deallocate(fs); 1317 return (KERN_FAILURE); 1318 } 1319 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW) == 0) { 1320 unlock_and_deallocate(fs); 1321 return (KERN_FAILURE); 1322 } 1323 if ((fault_type & VM_PROT_EXECUTE) && (vpte & VPTE_NX)) { 1324 unlock_and_deallocate(fs); 1325 return (KERN_FAILURE); 1326 } 1327 if ((vpte & VPTE_PS) || vshift == 0) 1328 break; 1329 1330 /* 1331 * Get the page table page. Nominally we only read the page 1332 * table, but since we are actively setting VPTE_M and VPTE_A, 1333 * tell vm_fault_object() that we are writing it. 1334 * 1335 * There is currently no real need to optimize this. 1336 */ 1337 result = vm_fault_object(fs, (vpte & VPTE_FRAME) >> PAGE_SHIFT, 1338 VM_PROT_READ|VM_PROT_WRITE, 1339 allow_nofault); 1340 if (result != KERN_SUCCESS) 1341 return (result); 1342 1343 /* 1344 * Process the returned fs.m and look up the page table 1345 * entry in the page table page. 1346 */ 1347 vshift -= VPTE_PAGE_BITS; 1348 lwb = lwbuf_alloc(fs->m, &lwb_cache); 1349 ptep = ((vpte_t *)lwbuf_kva(lwb) + 1350 ((*pindex >> vshift) & VPTE_PAGE_MASK)); 1351 vm_page_activate(fs->m); 1352 1353 /* 1354 * Page table write-back - entire operation including 1355 * validation of the pte must be atomic to avoid races 1356 * against the vkernel changing the pte. 1357 * 1358 * If the vpte is valid for the* requested operation, do 1359 * a write-back to the page table. 1360 * 1361 * XXX VPTE_M is not set properly for page directory pages. 1362 * It doesn't get set in the page directory if the page table 1363 * is modified during a read access. 1364 */ 1365 for (;;) { 1366 vpte_t nvpte; 1367 1368 /* 1369 * Reload for the cmpset, but make sure the pte is 1370 * still valid. 1371 */ 1372 vpte = *ptep; 1373 cpu_ccfence(); 1374 nvpte = vpte; 1375 1376 if ((vpte & VPTE_V) == 0) 1377 break; 1378 1379 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_RW)) 1380 nvpte |= VPTE_M | VPTE_A; 1381 if (fault_type & (VM_PROT_READ | VM_PROT_EXECUTE)) 1382 nvpte |= VPTE_A; 1383 if (vpte == nvpte) 1384 break; 1385 if (atomic_cmpset_long(ptep, vpte, nvpte)) { 1386 vm_page_dirty(fs->m); 1387 break; 1388 } 1389 } 1390 lwbuf_free(lwb); 1391 vm_page_flag_set(fs->m, PG_REFERENCED); 1392 vm_page_wakeup(fs->m); 1393 fs->m = NULL; 1394 cleanup_successful_fault(fs); 1395 } 1396 1397 /* 1398 * When the vkernel sets VPTE_RW it expects the real kernel to 1399 * reflect VPTE_M back when the page is modified via the mapping. 1400 * In order to accomplish this the real kernel must map the page 1401 * read-only for read faults and use write faults to reflect VPTE_M 1402 * back. 1403 * 1404 * Once VPTE_M has been set, the real kernel's pte allows writing. 1405 * If the vkernel clears VPTE_M the vkernel must be sure to 1406 * MADV_INVAL the real kernel's mappings to force the real kernel 1407 * to re-fault on the next write so oit can set VPTE_M again. 1408 */ 1409 if ((fault_type & VM_PROT_WRITE) == 0 && 1410 (vpte & (VPTE_RW | VPTE_M)) != (VPTE_RW | VPTE_M)) { 1411 fs->first_prot &= ~VM_PROT_WRITE; 1412 } 1413 1414 /* 1415 * Disable EXECUTE perms if NX bit is set. 1416 */ 1417 if (vpte & VPTE_NX) 1418 fs->first_prot &= ~VM_PROT_EXECUTE; 1419 1420 /* 1421 * Combine remaining address bits with the vpte. 1422 */ 1423 *pindex = ((vpte & VPTE_FRAME) >> PAGE_SHIFT) + 1424 (*pindex & ((1L << vshift) - 1)); 1425 return (KERN_SUCCESS); 1426 } 1427 1428 1429 /* 1430 * This is the core of the vm_fault code. 1431 * 1432 * Do all operations required to fault-in (fs.first_object, pindex). Run 1433 * through the shadow chain as necessary and do required COW or virtual 1434 * copy operations. The caller has already fully resolved the vm_map_entry 1435 * and, if appropriate, has created a copy-on-write layer. All we need to 1436 * do is iterate the object chain. 1437 * 1438 * On failure (fs) is unlocked and deallocated and the caller may return or 1439 * retry depending on the failure code. On success (fs) is NOT unlocked or 1440 * deallocated, fs.m will contained a resolved, busied page, and fs.object 1441 * will have an additional PIP count if it is not equal to fs.first_object. 1442 * 1443 * If locks based on fs->first_shared or fs->shared are insufficient, 1444 * clear the appropriate field(s) and return RETRY. COWs require that 1445 * first_shared be 0, while page allocations (or frees) require that 1446 * shared be 0. Renames require that both be 0. 1447 * 1448 * NOTE! fs->[first_]shared might be set with VM_FAULT_DIRTY also set. 1449 * we will have to retry with it exclusive if the vm_page is 1450 * PG_SWAPPED. 1451 * 1452 * fs->first_object must be held on call. 1453 */ 1454 static 1455 int 1456 vm_fault_object(struct faultstate *fs, vm_pindex_t first_pindex, 1457 vm_prot_t fault_type, int allow_nofault) 1458 { 1459 vm_object_t next_object; 1460 vm_pindex_t pindex; 1461 int error; 1462 1463 ASSERT_LWKT_TOKEN_HELD(vm_object_token(fs->first_object)); 1464 fs->prot = fs->first_prot; 1465 fs->object = fs->first_object; 1466 pindex = first_pindex; 1467 1468 vm_object_chain_acquire(fs->first_object, fs->shared); 1469 vm_object_pip_add(fs->first_object, 1); 1470 1471 /* 1472 * If a read fault occurs we try to upgrade the page protection 1473 * and make it also writable if possible. There are three cases 1474 * where we cannot make the page mapping writable: 1475 * 1476 * (1) The mapping is read-only or the VM object is read-only, 1477 * fs->prot above will simply not have VM_PROT_WRITE set. 1478 * 1479 * (2) If the mapping is a virtual page table fs->first_prot will 1480 * have already been properly adjusted by vm_fault_vpagetable(). 1481 * to detect writes so we can set VPTE_M in the virtual page 1482 * table. Used by vkernels. 1483 * 1484 * (3) If the VM page is read-only or copy-on-write, upgrading would 1485 * just result in an unnecessary COW fault. 1486 * 1487 * (4) If the pmap specifically requests A/M bit emulation, downgrade 1488 * here. 1489 */ 1490 #if 0 1491 /* see vpagetable code */ 1492 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) { 1493 if ((fault_type & VM_PROT_WRITE) == 0) 1494 fs->prot &= ~VM_PROT_WRITE; 1495 } 1496 #endif 1497 1498 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 1499 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 1500 if ((fault_type & VM_PROT_WRITE) == 0) 1501 fs->prot &= ~VM_PROT_WRITE; 1502 } 1503 1504 /* vm_object_hold(fs->object); implied b/c object == first_object */ 1505 1506 for (;;) { 1507 /* 1508 * The entire backing chain from first_object to object 1509 * inclusive is chainlocked. 1510 * 1511 * If the object is dead, we stop here 1512 */ 1513 if (fs->object->flags & OBJ_DEAD) { 1514 vm_object_pip_wakeup(fs->first_object); 1515 vm_object_chain_release_all(fs->first_object, 1516 fs->object); 1517 if (fs->object != fs->first_object) 1518 vm_object_drop(fs->object); 1519 unlock_and_deallocate(fs); 1520 return (KERN_PROTECTION_FAILURE); 1521 } 1522 1523 /* 1524 * See if the page is resident. Wait/Retry if the page is 1525 * busy (lots of stuff may have changed so we can't continue 1526 * in that case). 1527 * 1528 * We can theoretically allow the soft-busy case on a read 1529 * fault if the page is marked valid, but since such 1530 * pages are typically already pmap'd, putting that 1531 * special case in might be more effort then it is 1532 * worth. We cannot under any circumstances mess 1533 * around with a vm_page_t->busy page except, perhaps, 1534 * to pmap it. 1535 */ 1536 fs->m = vm_page_lookup_busy_try(fs->object, pindex, 1537 TRUE, &error); 1538 if (error) { 1539 vm_object_pip_wakeup(fs->first_object); 1540 vm_object_chain_release_all(fs->first_object, 1541 fs->object); 1542 if (fs->object != fs->first_object) 1543 vm_object_drop(fs->object); 1544 unlock_things(fs); 1545 vm_page_sleep_busy(fs->m, TRUE, "vmpfw"); 1546 mycpu->gd_cnt.v_intrans++; 1547 /*vm_object_deallocate(fs->first_object);*/ 1548 /*fs->first_object = NULL;*/ 1549 fs->m = NULL; 1550 return (KERN_TRY_AGAIN); 1551 } 1552 if (fs->m) { 1553 /* 1554 * The page is busied for us. 1555 * 1556 * If reactivating a page from PQ_CACHE we may have 1557 * to rate-limit. 1558 */ 1559 int queue = fs->m->queue; 1560 vm_page_unqueue_nowakeup(fs->m); 1561 1562 if ((queue - fs->m->pc) == PQ_CACHE && 1563 vm_page_count_severe()) { 1564 vm_page_activate(fs->m); 1565 vm_page_wakeup(fs->m); 1566 fs->m = NULL; 1567 vm_object_pip_wakeup(fs->first_object); 1568 vm_object_chain_release_all(fs->first_object, 1569 fs->object); 1570 if (fs->object != fs->first_object) 1571 vm_object_drop(fs->object); 1572 unlock_and_deallocate(fs); 1573 if (allow_nofault == 0 || 1574 (curthread->td_flags & TDF_NOFAULT) == 0) { 1575 thread_t td; 1576 1577 vm_wait_pfault(); 1578 td = curthread; 1579 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1580 return (KERN_PROTECTION_FAILURE); 1581 } 1582 return (KERN_TRY_AGAIN); 1583 } 1584 1585 /* 1586 * If it still isn't completely valid (readable), 1587 * or if a read-ahead-mark is set on the VM page, 1588 * jump to readrest, else we found the page and 1589 * can return. 1590 * 1591 * We can release the spl once we have marked the 1592 * page busy. 1593 */ 1594 if (fs->m->object != &kernel_object) { 1595 if ((fs->m->valid & VM_PAGE_BITS_ALL) != 1596 VM_PAGE_BITS_ALL) { 1597 goto readrest; 1598 } 1599 if (fs->m->flags & PG_RAM) { 1600 if (debug_cluster) 1601 kprintf("R"); 1602 vm_page_flag_clear(fs->m, PG_RAM); 1603 goto readrest; 1604 } 1605 } 1606 break; /* break to PAGE HAS BEEN FOUND */ 1607 } 1608 1609 /* 1610 * Page is not resident, If this is the search termination 1611 * or the pager might contain the page, allocate a new page. 1612 */ 1613 if (TRYPAGER(fs) || fs->object == fs->first_object) { 1614 /* 1615 * Allocating, must be exclusive. 1616 */ 1617 if (fs->object == fs->first_object && 1618 fs->first_shared) { 1619 fs->first_shared = 0; 1620 vm_object_pip_wakeup(fs->first_object); 1621 vm_object_chain_release_all(fs->first_object, 1622 fs->object); 1623 if (fs->object != fs->first_object) 1624 vm_object_drop(fs->object); 1625 unlock_and_deallocate(fs); 1626 return (KERN_TRY_AGAIN); 1627 } 1628 if (fs->object != fs->first_object && 1629 fs->shared) { 1630 fs->first_shared = 0; 1631 fs->shared = 0; 1632 vm_object_pip_wakeup(fs->first_object); 1633 vm_object_chain_release_all(fs->first_object, 1634 fs->object); 1635 if (fs->object != fs->first_object) 1636 vm_object_drop(fs->object); 1637 unlock_and_deallocate(fs); 1638 return (KERN_TRY_AGAIN); 1639 } 1640 1641 /* 1642 * If the page is beyond the object size we fail 1643 */ 1644 if (pindex >= fs->object->size) { 1645 vm_object_pip_wakeup(fs->first_object); 1646 vm_object_chain_release_all(fs->first_object, 1647 fs->object); 1648 if (fs->object != fs->first_object) 1649 vm_object_drop(fs->object); 1650 unlock_and_deallocate(fs); 1651 return (KERN_PROTECTION_FAILURE); 1652 } 1653 1654 /* 1655 * Allocate a new page for this object/offset pair. 1656 * 1657 * It is possible for the allocation to race, so 1658 * handle the case. 1659 */ 1660 fs->m = NULL; 1661 if (!vm_page_count_severe()) { 1662 fs->m = vm_page_alloc(fs->object, pindex, 1663 ((fs->vp || fs->object->backing_object) ? 1664 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL : 1665 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL | 1666 VM_ALLOC_USE_GD | VM_ALLOC_ZERO)); 1667 } 1668 if (fs->m == NULL) { 1669 vm_object_pip_wakeup(fs->first_object); 1670 vm_object_chain_release_all(fs->first_object, 1671 fs->object); 1672 if (fs->object != fs->first_object) 1673 vm_object_drop(fs->object); 1674 unlock_and_deallocate(fs); 1675 if (allow_nofault == 0 || 1676 (curthread->td_flags & TDF_NOFAULT) == 0) { 1677 thread_t td; 1678 1679 vm_wait_pfault(); 1680 td = curthread; 1681 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL)) 1682 return (KERN_PROTECTION_FAILURE); 1683 } 1684 return (KERN_TRY_AGAIN); 1685 } 1686 1687 /* 1688 * Fall through to readrest. We have a new page which 1689 * will have to be paged (since m->valid will be 0). 1690 */ 1691 } 1692 1693 readrest: 1694 /* 1695 * We have found an invalid or partially valid page, a 1696 * page with a read-ahead mark which might be partially or 1697 * fully valid (and maybe dirty too), or we have allocated 1698 * a new page. 1699 * 1700 * Attempt to fault-in the page if there is a chance that the 1701 * pager has it, and potentially fault in additional pages 1702 * at the same time. 1703 * 1704 * If TRYPAGER is true then fs.m will be non-NULL and busied 1705 * for us. 1706 */ 1707 if (TRYPAGER(fs)) { 1708 int rv; 1709 int seqaccess; 1710 u_char behavior = vm_map_entry_behavior(fs->entry); 1711 1712 if (behavior == MAP_ENTRY_BEHAV_RANDOM) 1713 seqaccess = 0; 1714 else 1715 seqaccess = -1; 1716 1717 /* 1718 * Doing I/O may synchronously insert additional 1719 * pages so we can't be shared at this point either. 1720 * 1721 * NOTE: We can't free fs->m here in the allocated 1722 * case (fs->object != fs->first_object) as 1723 * this would require an exclusively locked 1724 * VM object. 1725 */ 1726 if (fs->object == fs->first_object && 1727 fs->first_shared) { 1728 vm_page_deactivate(fs->m); 1729 vm_page_wakeup(fs->m); 1730 fs->m = NULL; 1731 fs->first_shared = 0; 1732 vm_object_pip_wakeup(fs->first_object); 1733 vm_object_chain_release_all(fs->first_object, 1734 fs->object); 1735 if (fs->object != fs->first_object) 1736 vm_object_drop(fs->object); 1737 unlock_and_deallocate(fs); 1738 return (KERN_TRY_AGAIN); 1739 } 1740 if (fs->object != fs->first_object && 1741 fs->shared) { 1742 vm_page_deactivate(fs->m); 1743 vm_page_wakeup(fs->m); 1744 fs->m = NULL; 1745 fs->first_shared = 0; 1746 fs->shared = 0; 1747 vm_object_pip_wakeup(fs->first_object); 1748 vm_object_chain_release_all(fs->first_object, 1749 fs->object); 1750 if (fs->object != fs->first_object) 1751 vm_object_drop(fs->object); 1752 unlock_and_deallocate(fs); 1753 return (KERN_TRY_AGAIN); 1754 } 1755 1756 /* 1757 * Avoid deadlocking against the map when doing I/O. 1758 * fs.object and the page is BUSY'd. 1759 * 1760 * NOTE: Once unlocked, fs->entry can become stale 1761 * so this will NULL it out. 1762 * 1763 * NOTE: fs->entry is invalid until we relock the 1764 * map and verify that the timestamp has not 1765 * changed. 1766 */ 1767 unlock_map(fs); 1768 1769 /* 1770 * Acquire the page data. We still hold a ref on 1771 * fs.object and the page has been BUSY's. 1772 * 1773 * The pager may replace the page (for example, in 1774 * order to enter a fictitious page into the 1775 * object). If it does so it is responsible for 1776 * cleaning up the passed page and properly setting 1777 * the new page BUSY. 1778 * 1779 * If we got here through a PG_RAM read-ahead 1780 * mark the page may be partially dirty and thus 1781 * not freeable. Don't bother checking to see 1782 * if the pager has the page because we can't free 1783 * it anyway. We have to depend on the get_page 1784 * operation filling in any gaps whether there is 1785 * backing store or not. 1786 */ 1787 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess); 1788 1789 if (rv == VM_PAGER_OK) { 1790 /* 1791 * Relookup in case pager changed page. Pager 1792 * is responsible for disposition of old page 1793 * if moved. 1794 * 1795 * XXX other code segments do relookups too. 1796 * It's a bad abstraction that needs to be 1797 * fixed/removed. 1798 */ 1799 fs->m = vm_page_lookup(fs->object, pindex); 1800 if (fs->m == NULL) { 1801 vm_object_pip_wakeup(fs->first_object); 1802 vm_object_chain_release_all( 1803 fs->first_object, fs->object); 1804 if (fs->object != fs->first_object) 1805 vm_object_drop(fs->object); 1806 unlock_and_deallocate(fs); 1807 return (KERN_TRY_AGAIN); 1808 } 1809 ++fs->hardfault; 1810 break; /* break to PAGE HAS BEEN FOUND */ 1811 } 1812 1813 /* 1814 * Remove the bogus page (which does not exist at this 1815 * object/offset); before doing so, we must get back 1816 * our object lock to preserve our invariant. 1817 * 1818 * Also wake up any other process that may want to bring 1819 * in this page. 1820 * 1821 * If this is the top-level object, we must leave the 1822 * busy page to prevent another process from rushing 1823 * past us, and inserting the page in that object at 1824 * the same time that we are. 1825 */ 1826 if (rv == VM_PAGER_ERROR) { 1827 if (curproc) { 1828 kprintf("vm_fault: pager read error, " 1829 "pid %d (%s)\n", 1830 curproc->p_pid, 1831 curproc->p_comm); 1832 } else { 1833 kprintf("vm_fault: pager read error, " 1834 "thread %p (%s)\n", 1835 curthread, 1836 curproc->p_comm); 1837 } 1838 } 1839 1840 /* 1841 * Data outside the range of the pager or an I/O error 1842 * 1843 * The page may have been wired during the pagein, 1844 * e.g. by the buffer cache, and cannot simply be 1845 * freed. Call vnode_pager_freepage() to deal with it. 1846 * 1847 * Also note that we cannot free the page if we are 1848 * holding the related object shared. XXX not sure 1849 * what to do in that case. 1850 */ 1851 if (fs->object != fs->first_object) { 1852 /* 1853 * Scrap the page. Check to see if the 1854 * vm_pager_get_page() call has already 1855 * dealt with it. 1856 */ 1857 if (fs->m) { 1858 vnode_pager_freepage(fs->m); 1859 fs->m = NULL; 1860 } 1861 1862 /* 1863 * XXX - we cannot just fall out at this 1864 * point, m has been freed and is invalid! 1865 */ 1866 } 1867 /* 1868 * XXX - the check for kernel_map is a kludge to work 1869 * around having the machine panic on a kernel space 1870 * fault w/ I/O error. 1871 */ 1872 if (((fs->map != &kernel_map) && 1873 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) { 1874 if (fs->m) { 1875 if (fs->first_shared) { 1876 vm_page_deactivate(fs->m); 1877 vm_page_wakeup(fs->m); 1878 } else { 1879 vnode_pager_freepage(fs->m); 1880 } 1881 fs->m = NULL; 1882 } 1883 vm_object_pip_wakeup(fs->first_object); 1884 vm_object_chain_release_all(fs->first_object, 1885 fs->object); 1886 if (fs->object != fs->first_object) 1887 vm_object_drop(fs->object); 1888 unlock_and_deallocate(fs); 1889 if (rv == VM_PAGER_ERROR) 1890 return (KERN_FAILURE); 1891 else 1892 return (KERN_PROTECTION_FAILURE); 1893 /* NOT REACHED */ 1894 } 1895 } 1896 1897 /* 1898 * We get here if the object has a default pager (or unwiring) 1899 * or the pager doesn't have the page. 1900 * 1901 * fs->first_m will be used for the COW unless we find a 1902 * deeper page to be mapped read-only, in which case the 1903 * unlock*(fs) will free first_m. 1904 */ 1905 if (fs->object == fs->first_object) 1906 fs->first_m = fs->m; 1907 1908 /* 1909 * Move on to the next object. The chain lock should prevent 1910 * the backing_object from getting ripped out from under us. 1911 * 1912 * The object lock for the next object is governed by 1913 * fs->shared. 1914 */ 1915 if ((next_object = fs->object->backing_object) != NULL) { 1916 if (fs->shared) 1917 vm_object_hold_shared(next_object); 1918 else 1919 vm_object_hold(next_object); 1920 vm_object_chain_acquire(next_object, fs->shared); 1921 KKASSERT(next_object == fs->object->backing_object); 1922 pindex += OFF_TO_IDX(fs->object->backing_object_offset); 1923 } 1924 1925 if (next_object == NULL) { 1926 /* 1927 * If there's no object left, fill the page in the top 1928 * object with zeros. 1929 */ 1930 if (fs->object != fs->first_object) { 1931 #if 0 1932 if (fs->first_object->backing_object != 1933 fs->object) { 1934 vm_object_hold(fs->first_object->backing_object); 1935 } 1936 #endif 1937 vm_object_chain_release_all( 1938 fs->first_object->backing_object, 1939 fs->object); 1940 #if 0 1941 if (fs->first_object->backing_object != 1942 fs->object) { 1943 vm_object_drop(fs->first_object->backing_object); 1944 } 1945 #endif 1946 vm_object_pip_wakeup(fs->object); 1947 vm_object_drop(fs->object); 1948 fs->object = fs->first_object; 1949 pindex = first_pindex; 1950 fs->m = fs->first_m; 1951 } 1952 fs->first_m = NULL; 1953 1954 /* 1955 * Zero the page and mark it valid. 1956 */ 1957 vm_page_zero_fill(fs->m); 1958 mycpu->gd_cnt.v_zfod++; 1959 fs->m->valid = VM_PAGE_BITS_ALL; 1960 break; /* break to PAGE HAS BEEN FOUND */ 1961 } 1962 if (fs->object != fs->first_object) { 1963 vm_object_pip_wakeup(fs->object); 1964 vm_object_lock_swap(); 1965 vm_object_drop(fs->object); 1966 } 1967 KASSERT(fs->object != next_object, 1968 ("object loop %p", next_object)); 1969 fs->object = next_object; 1970 vm_object_pip_add(fs->object, 1); 1971 } 1972 1973 /* 1974 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock 1975 * is held.] 1976 * 1977 * object still held. 1978 * 1979 * local shared variable may be different from fs->shared. 1980 * 1981 * If the page is being written, but isn't already owned by the 1982 * top-level object, we have to copy it into a new page owned by the 1983 * top-level object. 1984 */ 1985 KASSERT((fs->m->busy_count & PBUSY_LOCKED) != 0, 1986 ("vm_fault: not busy after main loop")); 1987 1988 if (fs->object != fs->first_object) { 1989 /* 1990 * We only really need to copy if we want to write it. 1991 */ 1992 if (fault_type & VM_PROT_WRITE) { 1993 /* 1994 * This allows pages to be virtually copied from a 1995 * backing_object into the first_object, where the 1996 * backing object has no other refs to it, and cannot 1997 * gain any more refs. Instead of a bcopy, we just 1998 * move the page from the backing object to the 1999 * first object. Note that we must mark the page 2000 * dirty in the first object so that it will go out 2001 * to swap when needed. 2002 */ 2003 if ( 2004 /* 2005 * Must be holding exclusive locks 2006 */ 2007 fs->first_shared == 0 && 2008 fs->shared == 0 && 2009 /* 2010 * Map, if present, has not changed 2011 */ 2012 (fs->map == NULL || 2013 fs->map_generation == fs->map->timestamp) && 2014 /* 2015 * Only one shadow object 2016 */ 2017 (fs->object->shadow_count == 1) && 2018 /* 2019 * No COW refs, except us 2020 */ 2021 (fs->object->ref_count == 1) && 2022 /* 2023 * No one else can look this object up 2024 */ 2025 (fs->object->handle == NULL) && 2026 /* 2027 * No other ways to look the object up 2028 */ 2029 ((fs->object->type == OBJT_DEFAULT) || 2030 (fs->object->type == OBJT_SWAP)) && 2031 /* 2032 * We don't chase down the shadow chain 2033 */ 2034 (fs->object == fs->first_object->backing_object) && 2035 2036 /* 2037 * grab the lock if we need to 2038 */ 2039 (fs->lookup_still_valid || 2040 fs->map == NULL || 2041 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0) 2042 ) { 2043 /* 2044 * (first_m) and (m) are both busied. We have 2045 * move (m) into (first_m)'s object/pindex 2046 * in an atomic fashion, then free (first_m). 2047 * 2048 * first_object is held so second remove 2049 * followed by the rename should wind 2050 * up being atomic. vm_page_free() might 2051 * block so we don't do it until after the 2052 * rename. 2053 */ 2054 fs->lookup_still_valid = 1; 2055 vm_page_protect(fs->first_m, VM_PROT_NONE); 2056 vm_page_remove(fs->first_m); 2057 vm_page_rename(fs->m, fs->first_object, 2058 first_pindex); 2059 vm_page_free(fs->first_m); 2060 fs->first_m = fs->m; 2061 fs->m = NULL; 2062 mycpu->gd_cnt.v_cow_optim++; 2063 } else { 2064 /* 2065 * Oh, well, lets copy it. 2066 * 2067 * Why are we unmapping the original page 2068 * here? Well, in short, not all accessors 2069 * of user memory go through the pmap. The 2070 * procfs code doesn't have access user memory 2071 * via a local pmap, so vm_fault_page*() 2072 * can't call pmap_enter(). And the umtx*() 2073 * code may modify the COW'd page via a DMAP 2074 * or kernel mapping and not via the pmap, 2075 * leaving the original page still mapped 2076 * read-only into the pmap. 2077 * 2078 * So we have to remove the page from at 2079 * least the current pmap if it is in it. 2080 * 2081 * We used to just remove it from all pmaps 2082 * but that creates inefficiencies on SMP, 2083 * particularly for COW program & library 2084 * mappings that are concurrently exec'd. 2085 * Only remove the page from the current 2086 * pmap. 2087 */ 2088 KKASSERT(fs->first_shared == 0); 2089 vm_page_copy(fs->m, fs->first_m); 2090 /*vm_page_protect(fs->m, VM_PROT_NONE);*/ 2091 pmap_remove_specific( 2092 &curthread->td_lwp->lwp_vmspace->vm_pmap, 2093 fs->m); 2094 } 2095 2096 /* 2097 * We no longer need the old page or object. 2098 */ 2099 if (fs->m) 2100 release_page(fs); 2101 2102 /* 2103 * We intend to revert to first_object, undo the 2104 * chain lock through to that. 2105 */ 2106 #if 0 2107 if (fs->first_object->backing_object != fs->object) 2108 vm_object_hold(fs->first_object->backing_object); 2109 #endif 2110 vm_object_chain_release_all( 2111 fs->first_object->backing_object, 2112 fs->object); 2113 #if 0 2114 if (fs->first_object->backing_object != fs->object) 2115 vm_object_drop(fs->first_object->backing_object); 2116 #endif 2117 2118 /* 2119 * fs->object != fs->first_object due to above 2120 * conditional 2121 */ 2122 vm_object_pip_wakeup(fs->object); 2123 vm_object_drop(fs->object); 2124 2125 /* 2126 * Only use the new page below... 2127 */ 2128 mycpu->gd_cnt.v_cow_faults++; 2129 fs->m = fs->first_m; 2130 fs->object = fs->first_object; 2131 pindex = first_pindex; 2132 } else { 2133 /* 2134 * If it wasn't a write fault avoid having to copy 2135 * the page by mapping it read-only. 2136 */ 2137 fs->prot &= ~VM_PROT_WRITE; 2138 } 2139 } 2140 2141 /* 2142 * Relock the map if necessary, then check the generation count. 2143 * relock_map() will update fs->timestamp to account for the 2144 * relocking if necessary. 2145 * 2146 * If the count has changed after relocking then all sorts of 2147 * crap may have happened and we have to retry. 2148 * 2149 * NOTE: The relock_map() can fail due to a deadlock against 2150 * the vm_page we are holding BUSY. 2151 */ 2152 if (fs->lookup_still_valid == FALSE && fs->map) { 2153 if (relock_map(fs) || 2154 fs->map->timestamp != fs->map_generation) { 2155 release_page(fs); 2156 vm_object_pip_wakeup(fs->first_object); 2157 vm_object_chain_release_all(fs->first_object, 2158 fs->object); 2159 if (fs->object != fs->first_object) 2160 vm_object_drop(fs->object); 2161 unlock_and_deallocate(fs); 2162 return (KERN_TRY_AGAIN); 2163 } 2164 } 2165 2166 /* 2167 * If the fault is a write, we know that this page is being 2168 * written NOW so dirty it explicitly to save on pmap_is_modified() 2169 * calls later. 2170 * 2171 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC 2172 * if the page is already dirty to prevent data written with 2173 * the expectation of being synced from not being synced. 2174 * Likewise if this entry does not request NOSYNC then make 2175 * sure the page isn't marked NOSYNC. Applications sharing 2176 * data should use the same flags to avoid ping ponging. 2177 * 2178 * Also tell the backing pager, if any, that it should remove 2179 * any swap backing since the page is now dirty. 2180 */ 2181 vm_page_activate(fs->m); 2182 if (fs->prot & VM_PROT_WRITE) { 2183 vm_object_set_writeable_dirty(fs->m->object); 2184 vm_set_nosync(fs->m, fs->entry); 2185 if (fs->fault_flags & VM_FAULT_DIRTY) { 2186 vm_page_dirty(fs->m); 2187 if (fs->m->flags & PG_SWAPPED) { 2188 /* 2189 * If the page is swapped out we have to call 2190 * swap_pager_unswapped() which requires an 2191 * exclusive object lock. If we are shared, 2192 * we must clear the shared flag and retry. 2193 */ 2194 if ((fs->object == fs->first_object && 2195 fs->first_shared) || 2196 (fs->object != fs->first_object && 2197 fs->shared)) { 2198 vm_page_wakeup(fs->m); 2199 fs->m = NULL; 2200 if (fs->object == fs->first_object) 2201 fs->first_shared = 0; 2202 else 2203 fs->shared = 0; 2204 vm_object_pip_wakeup(fs->first_object); 2205 vm_object_chain_release_all( 2206 fs->first_object, fs->object); 2207 if (fs->object != fs->first_object) 2208 vm_object_drop(fs->object); 2209 unlock_and_deallocate(fs); 2210 return (KERN_TRY_AGAIN); 2211 } 2212 swap_pager_unswapped(fs->m); 2213 } 2214 } 2215 } 2216 2217 vm_object_pip_wakeup(fs->first_object); 2218 vm_object_chain_release_all(fs->first_object, fs->object); 2219 if (fs->object != fs->first_object) 2220 vm_object_drop(fs->object); 2221 2222 /* 2223 * Page had better still be busy. We are still locked up and 2224 * fs->object will have another PIP reference if it is not equal 2225 * to fs->first_object. 2226 */ 2227 KASSERT(fs->m->busy_count & PBUSY_LOCKED, 2228 ("vm_fault: page %p not busy!", fs->m)); 2229 2230 /* 2231 * Sanity check: page must be completely valid or it is not fit to 2232 * map into user space. vm_pager_get_pages() ensures this. 2233 */ 2234 if (fs->m->valid != VM_PAGE_BITS_ALL) { 2235 vm_page_zero_invalid(fs->m, TRUE); 2236 kprintf("Warning: page %p partially invalid on fault\n", fs->m); 2237 } 2238 2239 return (KERN_SUCCESS); 2240 } 2241 2242 /* 2243 * Wire down a range of virtual addresses in a map. The entry in question 2244 * should be marked in-transition and the map must be locked. We must 2245 * release the map temporarily while faulting-in the page to avoid a 2246 * deadlock. Note that the entry may be clipped while we are blocked but 2247 * will never be freed. 2248 * 2249 * No requirements. 2250 */ 2251 int 2252 vm_fault_wire(vm_map_t map, vm_map_entry_t entry, 2253 boolean_t user_wire, int kmflags) 2254 { 2255 boolean_t fictitious; 2256 vm_offset_t start; 2257 vm_offset_t end; 2258 vm_offset_t va; 2259 pmap_t pmap; 2260 int rv; 2261 int wire_prot; 2262 int fault_flags; 2263 vm_page_t m; 2264 2265 if (user_wire) { 2266 wire_prot = VM_PROT_READ; 2267 fault_flags = VM_FAULT_USER_WIRE; 2268 } else { 2269 wire_prot = VM_PROT_READ | VM_PROT_WRITE; 2270 fault_flags = VM_FAULT_CHANGE_WIRING; 2271 } 2272 if (kmflags & KM_NOTLBSYNC) 2273 wire_prot |= VM_PROT_NOSYNC; 2274 2275 pmap = vm_map_pmap(map); 2276 start = entry->start; 2277 end = entry->end; 2278 2279 switch(entry->maptype) { 2280 case VM_MAPTYPE_NORMAL: 2281 case VM_MAPTYPE_VPAGETABLE: 2282 fictitious = entry->object.vm_object && 2283 ((entry->object.vm_object->type == OBJT_DEVICE) || 2284 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2285 break; 2286 case VM_MAPTYPE_UKSMAP: 2287 fictitious = TRUE; 2288 break; 2289 default: 2290 fictitious = FALSE; 2291 break; 2292 } 2293 2294 if (entry->eflags & MAP_ENTRY_KSTACK) 2295 start += PAGE_SIZE; 2296 map->timestamp++; 2297 vm_map_unlock(map); 2298 2299 /* 2300 * We simulate a fault to get the page and enter it in the physical 2301 * map. 2302 */ 2303 for (va = start; va < end; va += PAGE_SIZE) { 2304 rv = vm_fault(map, va, wire_prot, fault_flags); 2305 if (rv) { 2306 while (va > start) { 2307 va -= PAGE_SIZE; 2308 m = pmap_unwire(pmap, va); 2309 if (m && !fictitious) { 2310 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2311 vm_page_unwire(m, 1); 2312 vm_page_wakeup(m); 2313 } 2314 } 2315 goto done; 2316 } 2317 } 2318 rv = KERN_SUCCESS; 2319 done: 2320 vm_map_lock(map); 2321 2322 return (rv); 2323 } 2324 2325 /* 2326 * Unwire a range of virtual addresses in a map. The map should be 2327 * locked. 2328 */ 2329 void 2330 vm_fault_unwire(vm_map_t map, vm_map_entry_t entry) 2331 { 2332 boolean_t fictitious; 2333 vm_offset_t start; 2334 vm_offset_t end; 2335 vm_offset_t va; 2336 pmap_t pmap; 2337 vm_page_t m; 2338 2339 pmap = vm_map_pmap(map); 2340 start = entry->start; 2341 end = entry->end; 2342 fictitious = entry->object.vm_object && 2343 ((entry->object.vm_object->type == OBJT_DEVICE) || 2344 (entry->object.vm_object->type == OBJT_MGTDEVICE)); 2345 if (entry->eflags & MAP_ENTRY_KSTACK) 2346 start += PAGE_SIZE; 2347 2348 /* 2349 * Since the pages are wired down, we must be able to get their 2350 * mappings from the physical map system. 2351 */ 2352 for (va = start; va < end; va += PAGE_SIZE) { 2353 m = pmap_unwire(pmap, va); 2354 if (m && !fictitious) { 2355 vm_page_busy_wait(m, FALSE, "vmwrpg"); 2356 vm_page_unwire(m, 1); 2357 vm_page_wakeup(m); 2358 } 2359 } 2360 } 2361 2362 /* 2363 * Copy all of the pages from a wired-down map entry to another. 2364 * 2365 * The source and destination maps must be locked for write. 2366 * The source and destination maps token must be held 2367 * The source map entry must be wired down (or be a sharing map 2368 * entry corresponding to a main map entry that is wired down). 2369 * 2370 * No other requirements. 2371 * 2372 * XXX do segment optimization 2373 */ 2374 void 2375 vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map, 2376 vm_map_entry_t dst_entry, vm_map_entry_t src_entry) 2377 { 2378 vm_object_t dst_object; 2379 vm_object_t src_object; 2380 vm_ooffset_t dst_offset; 2381 vm_ooffset_t src_offset; 2382 vm_prot_t prot; 2383 vm_offset_t vaddr; 2384 vm_page_t dst_m; 2385 vm_page_t src_m; 2386 2387 src_object = src_entry->object.vm_object; 2388 src_offset = src_entry->offset; 2389 2390 /* 2391 * Create the top-level object for the destination entry. (Doesn't 2392 * actually shadow anything - we copy the pages directly.) 2393 */ 2394 vm_map_entry_allocate_object(dst_entry); 2395 dst_object = dst_entry->object.vm_object; 2396 2397 prot = dst_entry->max_protection; 2398 2399 /* 2400 * Loop through all of the pages in the entry's range, copying each 2401 * one from the source object (it should be there) to the destination 2402 * object. 2403 */ 2404 vm_object_hold(src_object); 2405 vm_object_hold(dst_object); 2406 for (vaddr = dst_entry->start, dst_offset = 0; 2407 vaddr < dst_entry->end; 2408 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) { 2409 2410 /* 2411 * Allocate a page in the destination object 2412 */ 2413 do { 2414 dst_m = vm_page_alloc(dst_object, 2415 OFF_TO_IDX(dst_offset), 2416 VM_ALLOC_NORMAL); 2417 if (dst_m == NULL) { 2418 vm_wait(0); 2419 } 2420 } while (dst_m == NULL); 2421 2422 /* 2423 * Find the page in the source object, and copy it in. 2424 * (Because the source is wired down, the page will be in 2425 * memory.) 2426 */ 2427 src_m = vm_page_lookup(src_object, 2428 OFF_TO_IDX(dst_offset + src_offset)); 2429 if (src_m == NULL) 2430 panic("vm_fault_copy_wired: page missing"); 2431 2432 vm_page_copy(src_m, dst_m); 2433 2434 /* 2435 * Enter it in the pmap... 2436 */ 2437 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE, dst_entry); 2438 2439 /* 2440 * Mark it no longer busy, and put it on the active list. 2441 */ 2442 vm_page_activate(dst_m); 2443 vm_page_wakeup(dst_m); 2444 } 2445 vm_object_drop(dst_object); 2446 vm_object_drop(src_object); 2447 } 2448 2449 #if 0 2450 2451 /* 2452 * This routine checks around the requested page for other pages that 2453 * might be able to be faulted in. This routine brackets the viable 2454 * pages for the pages to be paged in. 2455 * 2456 * Inputs: 2457 * m, rbehind, rahead 2458 * 2459 * Outputs: 2460 * marray (array of vm_page_t), reqpage (index of requested page) 2461 * 2462 * Return value: 2463 * number of pages in marray 2464 */ 2465 static int 2466 vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead, 2467 vm_page_t *marray, int *reqpage) 2468 { 2469 int i,j; 2470 vm_object_t object; 2471 vm_pindex_t pindex, startpindex, endpindex, tpindex; 2472 vm_page_t rtm; 2473 int cbehind, cahead; 2474 2475 object = m->object; 2476 pindex = m->pindex; 2477 2478 /* 2479 * we don't fault-ahead for device pager 2480 */ 2481 if ((object->type == OBJT_DEVICE) || 2482 (object->type == OBJT_MGTDEVICE)) { 2483 *reqpage = 0; 2484 marray[0] = m; 2485 return 1; 2486 } 2487 2488 /* 2489 * if the requested page is not available, then give up now 2490 */ 2491 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) { 2492 *reqpage = 0; /* not used by caller, fix compiler warn */ 2493 return 0; 2494 } 2495 2496 if ((cbehind == 0) && (cahead == 0)) { 2497 *reqpage = 0; 2498 marray[0] = m; 2499 return 1; 2500 } 2501 2502 if (rahead > cahead) { 2503 rahead = cahead; 2504 } 2505 2506 if (rbehind > cbehind) { 2507 rbehind = cbehind; 2508 } 2509 2510 /* 2511 * Do not do any readahead if we have insufficient free memory. 2512 * 2513 * XXX code was broken disabled before and has instability 2514 * with this conditonal fixed, so shortcut for now. 2515 */ 2516 if (burst_fault == 0 || vm_page_count_severe()) { 2517 marray[0] = m; 2518 *reqpage = 0; 2519 return 1; 2520 } 2521 2522 /* 2523 * scan backward for the read behind pages -- in memory 2524 * 2525 * Assume that if the page is not found an interrupt will not 2526 * create it. Theoretically interrupts can only remove (busy) 2527 * pages, not create new associations. 2528 */ 2529 if (pindex > 0) { 2530 if (rbehind > pindex) { 2531 rbehind = pindex; 2532 startpindex = 0; 2533 } else { 2534 startpindex = pindex - rbehind; 2535 } 2536 2537 vm_object_hold(object); 2538 for (tpindex = pindex; tpindex > startpindex; --tpindex) { 2539 if (vm_page_lookup(object, tpindex - 1)) 2540 break; 2541 } 2542 2543 i = 0; 2544 while (tpindex < pindex) { 2545 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2546 VM_ALLOC_NULL_OK); 2547 if (rtm == NULL) { 2548 for (j = 0; j < i; j++) { 2549 vm_page_free(marray[j]); 2550 } 2551 vm_object_drop(object); 2552 marray[0] = m; 2553 *reqpage = 0; 2554 return 1; 2555 } 2556 marray[i] = rtm; 2557 ++i; 2558 ++tpindex; 2559 } 2560 vm_object_drop(object); 2561 } else { 2562 i = 0; 2563 } 2564 2565 /* 2566 * Assign requested page 2567 */ 2568 marray[i] = m; 2569 *reqpage = i; 2570 ++i; 2571 2572 /* 2573 * Scan forwards for read-ahead pages 2574 */ 2575 tpindex = pindex + 1; 2576 endpindex = tpindex + rahead; 2577 if (endpindex > object->size) 2578 endpindex = object->size; 2579 2580 vm_object_hold(object); 2581 while (tpindex < endpindex) { 2582 if (vm_page_lookup(object, tpindex)) 2583 break; 2584 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM | 2585 VM_ALLOC_NULL_OK); 2586 if (rtm == NULL) 2587 break; 2588 marray[i] = rtm; 2589 ++i; 2590 ++tpindex; 2591 } 2592 vm_object_drop(object); 2593 2594 return (i); 2595 } 2596 2597 #endif 2598 2599 /* 2600 * vm_prefault() provides a quick way of clustering pagefaults into a 2601 * processes address space. It is a "cousin" of pmap_object_init_pt, 2602 * except it runs at page fault time instead of mmap time. 2603 * 2604 * vm.fast_fault Enables pre-faulting zero-fill pages 2605 * 2606 * vm.prefault_pages Number of pages (1/2 negative, 1/2 positive) to 2607 * prefault. Scan stops in either direction when 2608 * a page is found to already exist. 2609 * 2610 * This code used to be per-platform pmap_prefault(). It is now 2611 * machine-independent and enhanced to also pre-fault zero-fill pages 2612 * (see vm.fast_fault) as well as make them writable, which greatly 2613 * reduces the number of page faults programs incur. 2614 * 2615 * Application performance when pre-faulting zero-fill pages is heavily 2616 * dependent on the application. Very tiny applications like /bin/echo 2617 * lose a little performance while applications of any appreciable size 2618 * gain performance. Prefaulting multiple pages also reduces SMP 2619 * congestion and can improve SMP performance significantly. 2620 * 2621 * NOTE! prot may allow writing but this only applies to the top level 2622 * object. If we wind up mapping a page extracted from a backing 2623 * object we have to make sure it is read-only. 2624 * 2625 * NOTE! The caller has already handled any COW operations on the 2626 * vm_map_entry via the normal fault code. Do NOT call this 2627 * shortcut unless the normal fault code has run on this entry. 2628 * 2629 * The related map must be locked. 2630 * No other requirements. 2631 */ 2632 static int vm_prefault_pages = 8; 2633 SYSCTL_INT(_vm, OID_AUTO, prefault_pages, CTLFLAG_RW, &vm_prefault_pages, 0, 2634 "Maximum number of pages to pre-fault"); 2635 static int vm_fast_fault = 1; 2636 SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, 2637 "Burst fault zero-fill regions"); 2638 2639 /* 2640 * Set PG_NOSYNC if the map entry indicates so, but only if the page 2641 * is not already dirty by other means. This will prevent passive 2642 * filesystem syncing as well as 'sync' from writing out the page. 2643 */ 2644 static void 2645 vm_set_nosync(vm_page_t m, vm_map_entry_t entry) 2646 { 2647 if (entry->eflags & MAP_ENTRY_NOSYNC) { 2648 if (m->dirty == 0) 2649 vm_page_flag_set(m, PG_NOSYNC); 2650 } else { 2651 vm_page_flag_clear(m, PG_NOSYNC); 2652 } 2653 } 2654 2655 static void 2656 vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot, 2657 int fault_flags) 2658 { 2659 struct lwp *lp; 2660 vm_page_t m; 2661 vm_offset_t addr; 2662 vm_pindex_t index; 2663 vm_pindex_t pindex; 2664 vm_object_t object; 2665 int pprot; 2666 int i; 2667 int noneg; 2668 int nopos; 2669 int maxpages; 2670 2671 /* 2672 * Get stable max count value, disabled if set to 0 2673 */ 2674 maxpages = vm_prefault_pages; 2675 cpu_ccfence(); 2676 if (maxpages <= 0) 2677 return; 2678 2679 /* 2680 * We do not currently prefault mappings that use virtual page 2681 * tables. We do not prefault foreign pmaps. 2682 */ 2683 if (entry->maptype != VM_MAPTYPE_NORMAL) 2684 return; 2685 lp = curthread->td_lwp; 2686 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2687 return; 2688 2689 /* 2690 * Limit pre-fault count to 1024 pages. 2691 */ 2692 if (maxpages > 1024) 2693 maxpages = 1024; 2694 2695 object = entry->object.vm_object; 2696 KKASSERT(object != NULL); 2697 KKASSERT(object == entry->object.vm_object); 2698 2699 /* 2700 * NOTE: VM_FAULT_DIRTY allowed later so must hold object exclusively 2701 * now (or do something more complex XXX). 2702 */ 2703 vm_object_hold(object); 2704 vm_object_chain_acquire(object, 0); 2705 2706 noneg = 0; 2707 nopos = 0; 2708 for (i = 0; i < maxpages; ++i) { 2709 vm_object_t lobject; 2710 vm_object_t nobject; 2711 int allocated = 0; 2712 int error; 2713 2714 /* 2715 * This can eat a lot of time on a heavily contended 2716 * machine so yield on the tick if needed. 2717 */ 2718 if ((i & 7) == 7) 2719 lwkt_yield(); 2720 2721 /* 2722 * Calculate the page to pre-fault, stopping the scan in 2723 * each direction separately if the limit is reached. 2724 */ 2725 if (i & 1) { 2726 if (noneg) 2727 continue; 2728 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 2729 } else { 2730 if (nopos) 2731 continue; 2732 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 2733 } 2734 if (addr < entry->start) { 2735 noneg = 1; 2736 if (noneg && nopos) 2737 break; 2738 continue; 2739 } 2740 if (addr >= entry->end) { 2741 nopos = 1; 2742 if (noneg && nopos) 2743 break; 2744 continue; 2745 } 2746 2747 /* 2748 * Skip pages already mapped, and stop scanning in that 2749 * direction. When the scan terminates in both directions 2750 * we are done. 2751 */ 2752 if (pmap_prefault_ok(pmap, addr) == 0) { 2753 if (i & 1) 2754 noneg = 1; 2755 else 2756 nopos = 1; 2757 if (noneg && nopos) 2758 break; 2759 continue; 2760 } 2761 2762 /* 2763 * Follow the VM object chain to obtain the page to be mapped 2764 * into the pmap. 2765 * 2766 * If we reach the terminal object without finding a page 2767 * and we determine it would be advantageous, then allocate 2768 * a zero-fill page for the base object. The base object 2769 * is guaranteed to be OBJT_DEFAULT for this case. 2770 * 2771 * In order to not have to check the pager via *haspage*() 2772 * we stop if any non-default object is encountered. e.g. 2773 * a vnode or swap object would stop the loop. 2774 */ 2775 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 2776 lobject = object; 2777 pindex = index; 2778 pprot = prot; 2779 2780 KKASSERT(lobject == entry->object.vm_object); 2781 /*vm_object_hold(lobject); implied */ 2782 2783 while ((m = vm_page_lookup_busy_try(lobject, pindex, 2784 TRUE, &error)) == NULL) { 2785 if (lobject->type != OBJT_DEFAULT) 2786 break; 2787 if (lobject->backing_object == NULL) { 2788 if (vm_fast_fault == 0) 2789 break; 2790 if ((prot & VM_PROT_WRITE) == 0 || 2791 vm_page_count_min(0)) { 2792 break; 2793 } 2794 2795 /* 2796 * NOTE: Allocated from base object 2797 */ 2798 m = vm_page_alloc(object, index, 2799 VM_ALLOC_NORMAL | 2800 VM_ALLOC_ZERO | 2801 VM_ALLOC_USE_GD | 2802 VM_ALLOC_NULL_OK); 2803 if (m == NULL) 2804 break; 2805 allocated = 1; 2806 pprot = prot; 2807 /* lobject = object .. not needed */ 2808 break; 2809 } 2810 if (lobject->backing_object_offset & PAGE_MASK) 2811 break; 2812 nobject = lobject->backing_object; 2813 vm_object_hold(nobject); 2814 KKASSERT(nobject == lobject->backing_object); 2815 pindex += lobject->backing_object_offset >> PAGE_SHIFT; 2816 if (lobject != object) { 2817 vm_object_lock_swap(); 2818 vm_object_drop(lobject); 2819 } 2820 lobject = nobject; 2821 pprot &= ~VM_PROT_WRITE; 2822 vm_object_chain_acquire(lobject, 0); 2823 } 2824 2825 /* 2826 * NOTE: A non-NULL (m) will be associated with lobject if 2827 * it was found there, otherwise it is probably a 2828 * zero-fill page associated with the base object. 2829 * 2830 * Give-up if no page is available. 2831 */ 2832 if (m == NULL) { 2833 if (lobject != object) { 2834 #if 0 2835 if (object->backing_object != lobject) 2836 vm_object_hold(object->backing_object); 2837 #endif 2838 vm_object_chain_release_all( 2839 object->backing_object, lobject); 2840 #if 0 2841 if (object->backing_object != lobject) 2842 vm_object_drop(object->backing_object); 2843 #endif 2844 vm_object_drop(lobject); 2845 } 2846 break; 2847 } 2848 2849 /* 2850 * The object must be marked dirty if we are mapping a 2851 * writable page. m->object is either lobject or object, 2852 * both of which are still held. Do this before we 2853 * potentially drop the object. 2854 */ 2855 if (pprot & VM_PROT_WRITE) 2856 vm_object_set_writeable_dirty(m->object); 2857 2858 /* 2859 * Do not conditionalize on PG_RAM. If pages are present in 2860 * the VM system we assume optimal caching. If caching is 2861 * not optimal the I/O gravy train will be restarted when we 2862 * hit an unavailable page. We do not want to try to restart 2863 * the gravy train now because we really don't know how much 2864 * of the object has been cached. The cost for restarting 2865 * the gravy train should be low (since accesses will likely 2866 * be I/O bound anyway). 2867 */ 2868 if (lobject != object) { 2869 #if 0 2870 if (object->backing_object != lobject) 2871 vm_object_hold(object->backing_object); 2872 #endif 2873 vm_object_chain_release_all(object->backing_object, 2874 lobject); 2875 #if 0 2876 if (object->backing_object != lobject) 2877 vm_object_drop(object->backing_object); 2878 #endif 2879 vm_object_drop(lobject); 2880 } 2881 2882 /* 2883 * Enter the page into the pmap if appropriate. If we had 2884 * allocated the page we have to place it on a queue. If not 2885 * we just have to make sure it isn't on the cache queue 2886 * (pages on the cache queue are not allowed to be mapped). 2887 */ 2888 if (allocated) { 2889 /* 2890 * Page must be zerod. 2891 */ 2892 vm_page_zero_fill(m); 2893 mycpu->gd_cnt.v_zfod++; 2894 m->valid = VM_PAGE_BITS_ALL; 2895 2896 /* 2897 * Handle dirty page case 2898 */ 2899 if (pprot & VM_PROT_WRITE) 2900 vm_set_nosync(m, entry); 2901 pmap_enter(pmap, addr, m, pprot, 0, entry); 2902 mycpu->gd_cnt.v_vm_faults++; 2903 if (curthread->td_lwp) 2904 ++curthread->td_lwp->lwp_ru.ru_minflt; 2905 vm_page_deactivate(m); 2906 if (pprot & VM_PROT_WRITE) { 2907 /*vm_object_set_writeable_dirty(m->object);*/ 2908 vm_set_nosync(m, entry); 2909 if (fault_flags & VM_FAULT_DIRTY) { 2910 vm_page_dirty(m); 2911 /*XXX*/ 2912 swap_pager_unswapped(m); 2913 } 2914 } 2915 vm_page_wakeup(m); 2916 } else if (error) { 2917 /* couldn't busy page, no wakeup */ 2918 } else if ( 2919 ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) && 2920 (m->flags & PG_FICTITIOUS) == 0) { 2921 /* 2922 * A fully valid page not undergoing soft I/O can 2923 * be immediately entered into the pmap. 2924 */ 2925 if ((m->queue - m->pc) == PQ_CACHE) 2926 vm_page_deactivate(m); 2927 if (pprot & VM_PROT_WRITE) { 2928 /*vm_object_set_writeable_dirty(m->object);*/ 2929 vm_set_nosync(m, entry); 2930 if (fault_flags & VM_FAULT_DIRTY) { 2931 vm_page_dirty(m); 2932 /*XXX*/ 2933 swap_pager_unswapped(m); 2934 } 2935 } 2936 if (pprot & VM_PROT_WRITE) 2937 vm_set_nosync(m, entry); 2938 pmap_enter(pmap, addr, m, pprot, 0, entry); 2939 mycpu->gd_cnt.v_vm_faults++; 2940 if (curthread->td_lwp) 2941 ++curthread->td_lwp->lwp_ru.ru_minflt; 2942 vm_page_wakeup(m); 2943 } else { 2944 vm_page_wakeup(m); 2945 } 2946 } 2947 vm_object_chain_release(object); 2948 vm_object_drop(object); 2949 } 2950 2951 /* 2952 * Object can be held shared 2953 */ 2954 static void 2955 vm_prefault_quick(pmap_t pmap, vm_offset_t addra, 2956 vm_map_entry_t entry, int prot, int fault_flags) 2957 { 2958 struct lwp *lp; 2959 vm_page_t m; 2960 vm_offset_t addr; 2961 vm_pindex_t pindex; 2962 vm_object_t object; 2963 int i; 2964 int noneg; 2965 int nopos; 2966 int maxpages; 2967 2968 /* 2969 * Get stable max count value, disabled if set to 0 2970 */ 2971 maxpages = vm_prefault_pages; 2972 cpu_ccfence(); 2973 if (maxpages <= 0) 2974 return; 2975 2976 /* 2977 * We do not currently prefault mappings that use virtual page 2978 * tables. We do not prefault foreign pmaps. 2979 */ 2980 if (entry->maptype != VM_MAPTYPE_NORMAL) 2981 return; 2982 lp = curthread->td_lwp; 2983 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace))) 2984 return; 2985 object = entry->object.vm_object; 2986 if (object->backing_object != NULL) 2987 return; 2988 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 2989 2990 /* 2991 * Limit pre-fault count to 1024 pages. 2992 */ 2993 if (maxpages > 1024) 2994 maxpages = 1024; 2995 2996 noneg = 0; 2997 nopos = 0; 2998 for (i = 0; i < maxpages; ++i) { 2999 int error; 3000 3001 /* 3002 * Calculate the page to pre-fault, stopping the scan in 3003 * each direction separately if the limit is reached. 3004 */ 3005 if (i & 1) { 3006 if (noneg) 3007 continue; 3008 addr = addra - ((i + 1) >> 1) * PAGE_SIZE; 3009 } else { 3010 if (nopos) 3011 continue; 3012 addr = addra + ((i + 2) >> 1) * PAGE_SIZE; 3013 } 3014 if (addr < entry->start) { 3015 noneg = 1; 3016 if (noneg && nopos) 3017 break; 3018 continue; 3019 } 3020 if (addr >= entry->end) { 3021 nopos = 1; 3022 if (noneg && nopos) 3023 break; 3024 continue; 3025 } 3026 3027 /* 3028 * Follow the VM object chain to obtain the page to be mapped 3029 * into the pmap. This version of the prefault code only 3030 * works with terminal objects. 3031 * 3032 * The page must already exist. If we encounter a problem 3033 * we stop here. 3034 * 3035 * WARNING! We cannot call swap_pager_unswapped() or insert 3036 * a new vm_page with a shared token. 3037 */ 3038 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT; 3039 3040 /* 3041 * Skip pages already mapped, and stop scanning in that 3042 * direction. When the scan terminates in both directions 3043 * we are done. 3044 */ 3045 if (pmap_prefault_ok(pmap, addr) == 0) { 3046 if (i & 1) 3047 noneg = 1; 3048 else 3049 nopos = 1; 3050 if (noneg && nopos) 3051 break; 3052 continue; 3053 } 3054 3055 /* 3056 * Shortcut the read-only mapping case using the far more 3057 * efficient vm_page_lookup_sbusy_try() function. This 3058 * allows us to acquire the page soft-busied only which 3059 * is especially nice for concurrent execs of the same 3060 * program. 3061 * 3062 * The lookup function also validates page suitability 3063 * (all valid bits set, and not fictitious). 3064 * 3065 * If the page is in PQ_CACHE we have to fall-through 3066 * and hard-busy it so we can move it out of PQ_CACHE. 3067 */ 3068 if ((prot & (VM_PROT_WRITE|VM_PROT_OVERRIDE_WRITE)) == 0) { 3069 m = vm_page_lookup_sbusy_try(object, pindex, 3070 0, PAGE_SIZE); 3071 if (m == NULL) 3072 break; 3073 if ((m->queue - m->pc) != PQ_CACHE) { 3074 pmap_enter(pmap, addr, m, prot, 0, entry); 3075 mycpu->gd_cnt.v_vm_faults++; 3076 if (curthread->td_lwp) 3077 ++curthread->td_lwp->lwp_ru.ru_minflt; 3078 vm_page_sbusy_drop(m); 3079 continue; 3080 } 3081 vm_page_sbusy_drop(m); 3082 } 3083 3084 /* 3085 * Fallback to normal vm_page lookup code. This code 3086 * hard-busies the page. Not only that, but the page 3087 * can remain in that state for a significant period 3088 * time due to pmap_enter()'s overhead. 3089 */ 3090 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error); 3091 if (m == NULL || error) 3092 break; 3093 3094 /* 3095 * Stop if the page cannot be trivially entered into the 3096 * pmap. 3097 */ 3098 if (((m->valid & VM_PAGE_BITS_ALL) != VM_PAGE_BITS_ALL) || 3099 (m->flags & PG_FICTITIOUS) || 3100 ((m->flags & PG_SWAPPED) && 3101 (prot & VM_PROT_WRITE) && 3102 (fault_flags & VM_FAULT_DIRTY))) { 3103 vm_page_wakeup(m); 3104 break; 3105 } 3106 3107 /* 3108 * Enter the page into the pmap. The object might be held 3109 * shared so we can't do any (serious) modifying operation 3110 * on it. 3111 */ 3112 if ((m->queue - m->pc) == PQ_CACHE) 3113 vm_page_deactivate(m); 3114 if (prot & VM_PROT_WRITE) { 3115 vm_object_set_writeable_dirty(m->object); 3116 vm_set_nosync(m, entry); 3117 if (fault_flags & VM_FAULT_DIRTY) { 3118 vm_page_dirty(m); 3119 /* can't happeen due to conditional above */ 3120 /* swap_pager_unswapped(m); */ 3121 } 3122 } 3123 pmap_enter(pmap, addr, m, prot, 0, entry); 3124 mycpu->gd_cnt.v_vm_faults++; 3125 if (curthread->td_lwp) 3126 ++curthread->td_lwp->lwp_ru.ru_minflt; 3127 vm_page_wakeup(m); 3128 } 3129 } 3130