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