1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * 5 * This code is derived from software contributed to Berkeley by 6 * The Mach Operating System project at Carnegie-Mellon University. 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 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 the 15 * documentation and/or other materials provided with the distribution. 16 * 3. All advertising materials mentioning features or use of this software 17 * must display the following acknowledgement: 18 * This product includes software developed by the University of 19 * California, Berkeley and its contributors. 20 * 4. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94 37 * 38 * 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 * 64 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $ 65 * $DragonFly: src/sys/vm/vm_map.c,v 1.36 2004/12/21 02:42:41 hsu Exp $ 66 */ 67 68 /* 69 * Virtual memory mapping module. 70 */ 71 72 #include <sys/param.h> 73 #include <sys/systm.h> 74 #include <sys/proc.h> 75 #include <sys/lock.h> 76 #include <sys/vmmeter.h> 77 #include <sys/mman.h> 78 #include <sys/vnode.h> 79 #include <sys/resourcevar.h> 80 #include <sys/shm.h> 81 82 #include <vm/vm.h> 83 #include <vm/vm_param.h> 84 #include <vm/pmap.h> 85 #include <vm/vm_map.h> 86 #include <vm/vm_page.h> 87 #include <vm/vm_object.h> 88 #include <vm/vm_pager.h> 89 #include <vm/vm_kern.h> 90 #include <vm/vm_extern.h> 91 #include <vm/swap_pager.h> 92 #include <vm/vm_zone.h> 93 94 #include <sys/thread2.h> 95 96 /* 97 * Virtual memory maps provide for the mapping, protection, 98 * and sharing of virtual memory objects. In addition, 99 * this module provides for an efficient virtual copy of 100 * memory from one map to another. 101 * 102 * Synchronization is required prior to most operations. 103 * 104 * Maps consist of an ordered doubly-linked list of simple 105 * entries; a single hint is used to speed up lookups. 106 * 107 * Since portions of maps are specified by start/end addresses, 108 * which may not align with existing map entries, all 109 * routines merely "clip" entries to these start/end values. 110 * [That is, an entry is split into two, bordering at a 111 * start or end value.] Note that these clippings may not 112 * always be necessary (as the two resulting entries are then 113 * not changed); however, the clipping is done for convenience. 114 * 115 * As mentioned above, virtual copy operations are performed 116 * by copying VM object references from one map to 117 * another, and then marking both regions as copy-on-write. 118 */ 119 120 /* 121 * vm_map_startup: 122 * 123 * Initialize the vm_map module. Must be called before 124 * any other vm_map routines. 125 * 126 * Map and entry structures are allocated from the general 127 * purpose memory pool with some exceptions: 128 * 129 * - The kernel map and kmem submap are allocated statically. 130 * - Kernel map entries are allocated out of a static pool. 131 * 132 * These restrictions are necessary since malloc() uses the 133 * maps and requires map entries. 134 */ 135 136 #define VMEPERCPU 2 137 138 static struct vm_zone mapentzone_store, mapzone_store; 139 static vm_zone_t mapentzone, mapzone, vmspace_zone; 140 static struct vm_object mapentobj, mapobj; 141 142 static struct vm_map_entry map_entry_init[MAX_MAPENT]; 143 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU]; 144 static struct vm_map map_init[MAX_KMAP]; 145 146 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *); 147 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *); 148 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 149 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 150 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *); 151 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t); 152 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t, 153 vm_map_entry_t); 154 static void vm_map_split (vm_map_entry_t); 155 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry, vm_offset_t start, vm_offset_t end, int *count, int flags); 156 157 void 158 vm_map_startup(void) 159 { 160 mapzone = &mapzone_store; 161 zbootinit(mapzone, "MAP", sizeof (struct vm_map), 162 map_init, MAX_KMAP); 163 mapentzone = &mapentzone_store; 164 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry), 165 map_entry_init, MAX_MAPENT); 166 } 167 168 /* 169 * Allocate a vmspace structure, including a vm_map and pmap, 170 * and initialize those structures. The refcnt is set to 1. 171 * The remaining fields must be initialized by the caller. 172 */ 173 struct vmspace * 174 vmspace_alloc(vm_offset_t min, vm_offset_t max) 175 { 176 struct vmspace *vm; 177 178 vm = zalloc(vmspace_zone); 179 vm_map_init(&vm->vm_map, min, max); 180 pmap_pinit(vmspace_pmap(vm)); 181 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */ 182 vm->vm_refcnt = 1; 183 vm->vm_shm = NULL; 184 vm->vm_exitingcnt = 0; 185 return (vm); 186 } 187 188 void 189 vm_init2(void) 190 { 191 zinitna(mapentzone, &mapentobj, NULL, 0, 0, 192 ZONE_USE_RESERVE | ZONE_SPECIAL, 1); 193 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1); 194 vmspace_zone = zinit("VMSPACE", sizeof (struct vmspace), 0, 0, 3); 195 pmap_init2(); 196 vm_object_init2(); 197 } 198 199 static __inline void 200 vmspace_dofree(struct vmspace *vm) 201 { 202 int count; 203 204 /* 205 * Make sure any SysV shm is freed, it might not have in 206 * exit1() 207 */ 208 shmexit(vm); 209 210 KKASSERT(vm->vm_upcalls == NULL); 211 212 /* 213 * Lock the map, to wait out all other references to it. 214 * Delete all of the mappings and pages they hold, then call 215 * the pmap module to reclaim anything left. 216 */ 217 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 218 vm_map_lock(&vm->vm_map); 219 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset, 220 vm->vm_map.max_offset, &count); 221 vm_map_unlock(&vm->vm_map); 222 vm_map_entry_release(count); 223 224 pmap_release(vmspace_pmap(vm)); 225 zfree(vmspace_zone, vm); 226 } 227 228 void 229 vmspace_free(struct vmspace *vm) 230 { 231 if (vm->vm_refcnt == 0) 232 panic("vmspace_free: attempt to free already freed vmspace"); 233 234 if (--vm->vm_refcnt == 0 && vm->vm_exitingcnt == 0) 235 vmspace_dofree(vm); 236 } 237 238 void 239 vmspace_exitfree(struct proc *p) 240 { 241 struct vmspace *vm; 242 243 vm = p->p_vmspace; 244 p->p_vmspace = NULL; 245 246 /* 247 * cleanup by parent process wait()ing on exiting child. vm_refcnt 248 * may not be 0 (e.g. fork() and child exits without exec()ing). 249 * exitingcnt may increment above 0 and drop back down to zero 250 * several times while vm_refcnt is held non-zero. vm_refcnt 251 * may also increment above 0 and drop back down to zero several 252 * times while vm_exitingcnt is held non-zero. 253 * 254 * The last wait on the exiting child's vmspace will clean up 255 * the remainder of the vmspace. 256 */ 257 if (--vm->vm_exitingcnt == 0 && vm->vm_refcnt == 0) 258 vmspace_dofree(vm); 259 } 260 261 /* 262 * vmspace_swap_count() - count the approximate swap useage in pages for a 263 * vmspace. 264 * 265 * Swap useage is determined by taking the proportional swap used by 266 * VM objects backing the VM map. To make up for fractional losses, 267 * if the VM object has any swap use at all the associated map entries 268 * count for at least 1 swap page. 269 */ 270 int 271 vmspace_swap_count(struct vmspace *vmspace) 272 { 273 vm_map_t map = &vmspace->vm_map; 274 vm_map_entry_t cur; 275 int count = 0; 276 277 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 278 vm_object_t object; 279 280 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 281 (object = cur->object.vm_object) != NULL && 282 object->type == OBJT_SWAP 283 ) { 284 int n = (cur->end - cur->start) / PAGE_SIZE; 285 286 if (object->un_pager.swp.swp_bcount) { 287 count += object->un_pager.swp.swp_bcount * 288 SWAP_META_PAGES * n / object->size + 1; 289 } 290 } 291 } 292 return(count); 293 } 294 295 296 /* 297 * vm_map_create: 298 * 299 * Creates and returns a new empty VM map with 300 * the given physical map structure, and having 301 * the given lower and upper address bounds. 302 */ 303 vm_map_t 304 vm_map_create(pmap_t pmap, vm_offset_t min, vm_offset_t max) 305 { 306 vm_map_t result; 307 308 result = zalloc(mapzone); 309 vm_map_init(result, min, max); 310 result->pmap = pmap; 311 return (result); 312 } 313 314 /* 315 * Initialize an existing vm_map structure 316 * such as that in the vmspace structure. 317 * The pmap is set elsewhere. 318 */ 319 void 320 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max) 321 { 322 map->header.next = map->header.prev = &map->header; 323 map->nentries = 0; 324 map->size = 0; 325 map->system_map = 0; 326 map->infork = 0; 327 map->min_offset = min; 328 map->max_offset = max; 329 map->first_free = &map->header; 330 map->hint = &map->header; 331 map->timestamp = 0; 332 lockinit(&map->lock, 0, "thrd_sleep", 0, LK_NOPAUSE); 333 } 334 335 /* 336 * vm_map_entry_reserve_cpu_init: 337 * 338 * Set an initial negative count so the first attempt to reserve 339 * space preloads a bunch of vm_map_entry's for this cpu. Also 340 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 341 * map a new page for vm_map_entry structures. SMP systems are 342 * particularly sensitive. 343 * 344 * This routine is called in early boot so we cannot just call 345 * vm_map_entry_reserve(). 346 * 347 * May be called for a gd other then mycpu, but may only be called 348 * during early boot. 349 */ 350 void 351 vm_map_entry_reserve_cpu_init(globaldata_t gd) 352 { 353 vm_map_entry_t entry; 354 int i; 355 356 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2; 357 entry = &cpu_map_entry_init[gd->gd_cpuid][0]; 358 for (i = 0; i < VMEPERCPU; ++i, ++entry) { 359 entry->next = gd->gd_vme_base; 360 gd->gd_vme_base = entry; 361 } 362 } 363 364 /* 365 * vm_map_entry_reserve: 366 * 367 * Reserves vm_map_entry structures so code later on can manipulate 368 * map_entry structures within a locked map without blocking trying 369 * to allocate a new vm_map_entry. 370 */ 371 int 372 vm_map_entry_reserve(int count) 373 { 374 struct globaldata *gd = mycpu; 375 vm_map_entry_t entry; 376 377 crit_enter(); 378 379 /* 380 * Make sure we have enough structures in gd_vme_base to handle 381 * the reservation request. 382 */ 383 while (gd->gd_vme_avail < count) { 384 entry = zalloc(mapentzone); 385 entry->next = gd->gd_vme_base; 386 gd->gd_vme_base = entry; 387 ++gd->gd_vme_avail; 388 } 389 gd->gd_vme_avail -= count; 390 crit_exit(); 391 return(count); 392 } 393 394 /* 395 * vm_map_entry_release: 396 * 397 * Releases previously reserved vm_map_entry structures that were not 398 * used. If we have too much junk in our per-cpu cache clean some of 399 * it out. 400 */ 401 void 402 vm_map_entry_release(int count) 403 { 404 struct globaldata *gd = mycpu; 405 vm_map_entry_t entry; 406 407 crit_enter(); 408 gd->gd_vme_avail += count; 409 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 410 entry = gd->gd_vme_base; 411 KKASSERT(entry != NULL); 412 gd->gd_vme_base = entry->next; 413 --gd->gd_vme_avail; 414 crit_exit(); 415 zfree(mapentzone, entry); 416 crit_enter(); 417 } 418 crit_exit(); 419 } 420 421 /* 422 * vm_map_entry_kreserve: 423 * 424 * Reserve map entry structures for use in kernel_map itself. These 425 * entries have *ALREADY* been reserved on a per-cpu basis when the map 426 * was inited. This function is used by zalloc() to avoid a recursion 427 * when zalloc() itself needs to allocate additional kernel memory. 428 * 429 * This function works like the normal reserve but does not load the 430 * vm_map_entry cache (because that would result in an infinite 431 * recursion). Note that gd_vme_avail may go negative. This is expected. 432 * 433 * Any caller of this function must be sure to renormalize after 434 * potentially eating entries to ensure that the reserve supply 435 * remains intact. 436 */ 437 int 438 vm_map_entry_kreserve(int count) 439 { 440 struct globaldata *gd = mycpu; 441 442 crit_enter(); 443 gd->gd_vme_avail -= count; 444 crit_exit(); 445 KASSERT(gd->gd_vme_base != NULL, ("no reserved entries left, gd_vme_avail = %d\n", gd->gd_vme_avail)); 446 return(count); 447 } 448 449 /* 450 * vm_map_entry_krelease: 451 * 452 * Release previously reserved map entries for kernel_map. We do not 453 * attempt to clean up like the normal release function as this would 454 * cause an unnecessary (but probably not fatal) deep procedure call. 455 */ 456 void 457 vm_map_entry_krelease(int count) 458 { 459 struct globaldata *gd = mycpu; 460 461 crit_enter(); 462 gd->gd_vme_avail += count; 463 crit_exit(); 464 } 465 466 /* 467 * vm_map_entry_create: [ internal use only ] 468 * 469 * Allocates a VM map entry for insertion. No entry fields are filled 470 * in. 471 * 472 * This routine may be called from an interrupt thread but not a FAST 473 * interrupt. This routine may recurse the map lock. 474 */ 475 static vm_map_entry_t 476 vm_map_entry_create(vm_map_t map, int *countp) 477 { 478 struct globaldata *gd = mycpu; 479 vm_map_entry_t entry; 480 481 KKASSERT(*countp > 0); 482 --*countp; 483 crit_enter(); 484 entry = gd->gd_vme_base; 485 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 486 gd->gd_vme_base = entry->next; 487 crit_exit(); 488 return(entry); 489 } 490 491 /* 492 * vm_map_entry_dispose: [ internal use only ] 493 * 494 * Dispose of a vm_map_entry that is no longer being referenced. This 495 * function may be called from an interrupt. 496 */ 497 static void 498 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 499 { 500 struct globaldata *gd = mycpu; 501 502 ++*countp; 503 crit_enter(); 504 entry->next = gd->gd_vme_base; 505 gd->gd_vme_base = entry; 506 crit_exit(); 507 } 508 509 510 /* 511 * vm_map_entry_{un,}link: 512 * 513 * Insert/remove entries from maps. 514 */ 515 static __inline void 516 vm_map_entry_link(vm_map_t map, 517 vm_map_entry_t after_where, 518 vm_map_entry_t entry) 519 { 520 map->nentries++; 521 entry->prev = after_where; 522 entry->next = after_where->next; 523 entry->next->prev = entry; 524 after_where->next = entry; 525 } 526 527 static __inline void 528 vm_map_entry_unlink(vm_map_t map, 529 vm_map_entry_t entry) 530 { 531 vm_map_entry_t prev; 532 vm_map_entry_t next; 533 534 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) 535 panic("vm_map_entry_unlink: attempt to mess with locked entry! %p", entry); 536 prev = entry->prev; 537 next = entry->next; 538 next->prev = prev; 539 prev->next = next; 540 map->nentries--; 541 } 542 543 /* 544 * SAVE_HINT: 545 * 546 * Saves the specified entry as the hint for 547 * future lookups. 548 */ 549 #define SAVE_HINT(map,value) \ 550 (map)->hint = (value); 551 552 /* 553 * vm_map_lookup_entry: [ internal use only ] 554 * 555 * Finds the map entry containing (or 556 * immediately preceding) the specified address 557 * in the given map; the entry is returned 558 * in the "entry" parameter. The boolean 559 * result indicates whether the address is 560 * actually contained in the map. 561 */ 562 boolean_t 563 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, 564 vm_map_entry_t *entry /* OUT */) 565 { 566 vm_map_entry_t cur; 567 vm_map_entry_t last; 568 569 /* 570 * Start looking either from the head of the list, or from the hint. 571 */ 572 573 cur = map->hint; 574 575 if (cur == &map->header) 576 cur = cur->next; 577 578 if (address >= cur->start) { 579 /* 580 * Go from hint to end of list. 581 * 582 * But first, make a quick check to see if we are already looking 583 * at the entry we want (which is usually the case). Note also 584 * that we don't need to save the hint here... it is the same 585 * hint (unless we are at the header, in which case the hint 586 * didn't buy us anything anyway). 587 */ 588 last = &map->header; 589 if ((cur != last) && (cur->end > address)) { 590 *entry = cur; 591 return (TRUE); 592 } 593 } else { 594 /* 595 * Go from start to hint, *inclusively* 596 */ 597 last = cur->next; 598 cur = map->header.next; 599 } 600 601 /* 602 * Search linearly 603 */ 604 605 while (cur != last) { 606 if (cur->end > address) { 607 if (address >= cur->start) { 608 /* 609 * Save this lookup for future hints, and 610 * return 611 */ 612 613 *entry = cur; 614 SAVE_HINT(map, cur); 615 return (TRUE); 616 } 617 break; 618 } 619 cur = cur->next; 620 } 621 *entry = cur->prev; 622 SAVE_HINT(map, *entry); 623 return (FALSE); 624 } 625 626 /* 627 * vm_map_insert: 628 * 629 * Inserts the given whole VM object into the target 630 * map at the specified address range. The object's 631 * size should match that of the address range. 632 * 633 * Requires that the map be locked, and leaves it so. Requires that 634 * sufficient vm_map_entry structures have been reserved and tracks 635 * the use via countp. 636 * 637 * If object is non-NULL, ref count must be bumped by caller 638 * prior to making call to account for the new entry. 639 */ 640 int 641 vm_map_insert(vm_map_t map, int *countp, 642 vm_object_t object, vm_ooffset_t offset, 643 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, 644 int cow) 645 { 646 vm_map_entry_t new_entry; 647 vm_map_entry_t prev_entry; 648 vm_map_entry_t temp_entry; 649 vm_eflags_t protoeflags; 650 651 /* 652 * Check that the start and end points are not bogus. 653 */ 654 655 if ((start < map->min_offset) || (end > map->max_offset) || 656 (start >= end)) 657 return (KERN_INVALID_ADDRESS); 658 659 /* 660 * Find the entry prior to the proposed starting address; if it's part 661 * of an existing entry, this range is bogus. 662 */ 663 664 if (vm_map_lookup_entry(map, start, &temp_entry)) 665 return (KERN_NO_SPACE); 666 667 prev_entry = temp_entry; 668 669 /* 670 * Assert that the next entry doesn't overlap the end point. 671 */ 672 673 if ((prev_entry->next != &map->header) && 674 (prev_entry->next->start < end)) 675 return (KERN_NO_SPACE); 676 677 protoeflags = 0; 678 679 if (cow & MAP_COPY_ON_WRITE) 680 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 681 682 if (cow & MAP_NOFAULT) { 683 protoeflags |= MAP_ENTRY_NOFAULT; 684 685 KASSERT(object == NULL, 686 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 687 } 688 if (cow & MAP_DISABLE_SYNCER) 689 protoeflags |= MAP_ENTRY_NOSYNC; 690 if (cow & MAP_DISABLE_COREDUMP) 691 protoeflags |= MAP_ENTRY_NOCOREDUMP; 692 693 if (object) { 694 /* 695 * When object is non-NULL, it could be shared with another 696 * process. We have to set or clear OBJ_ONEMAPPING 697 * appropriately. 698 */ 699 if ((object->ref_count > 1) || (object->shadow_count != 0)) { 700 vm_object_clear_flag(object, OBJ_ONEMAPPING); 701 } 702 } 703 else if ((prev_entry != &map->header) && 704 (prev_entry->eflags == protoeflags) && 705 (prev_entry->end == start) && 706 (prev_entry->wired_count == 0) && 707 ((prev_entry->object.vm_object == NULL) || 708 vm_object_coalesce(prev_entry->object.vm_object, 709 OFF_TO_IDX(prev_entry->offset), 710 (vm_size_t)(prev_entry->end - prev_entry->start), 711 (vm_size_t)(end - prev_entry->end)))) { 712 /* 713 * We were able to extend the object. Determine if we 714 * can extend the previous map entry to include the 715 * new range as well. 716 */ 717 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 718 (prev_entry->protection == prot) && 719 (prev_entry->max_protection == max)) { 720 map->size += (end - prev_entry->end); 721 prev_entry->end = end; 722 vm_map_simplify_entry(map, prev_entry, countp); 723 return (KERN_SUCCESS); 724 } 725 726 /* 727 * If we can extend the object but cannot extend the 728 * map entry, we have to create a new map entry. We 729 * must bump the ref count on the extended object to 730 * account for it. object may be NULL. 731 */ 732 object = prev_entry->object.vm_object; 733 offset = prev_entry->offset + 734 (prev_entry->end - prev_entry->start); 735 vm_object_reference(object); 736 } 737 738 /* 739 * NOTE: if conditionals fail, object can be NULL here. This occurs 740 * in things like the buffer map where we manage kva but do not manage 741 * backing objects. 742 */ 743 744 /* 745 * Create a new entry 746 */ 747 748 new_entry = vm_map_entry_create(map, countp); 749 new_entry->start = start; 750 new_entry->end = end; 751 752 new_entry->eflags = protoeflags; 753 new_entry->object.vm_object = object; 754 new_entry->offset = offset; 755 new_entry->avail_ssize = 0; 756 757 new_entry->inheritance = VM_INHERIT_DEFAULT; 758 new_entry->protection = prot; 759 new_entry->max_protection = max; 760 new_entry->wired_count = 0; 761 762 /* 763 * Insert the new entry into the list 764 */ 765 766 vm_map_entry_link(map, prev_entry, new_entry); 767 map->size += new_entry->end - new_entry->start; 768 769 /* 770 * Update the free space hint 771 */ 772 if ((map->first_free == prev_entry) && 773 (prev_entry->end >= new_entry->start)) { 774 map->first_free = new_entry; 775 } 776 777 #if 0 778 /* 779 * Temporarily removed to avoid MAP_STACK panic, due to 780 * MAP_STACK being a huge hack. Will be added back in 781 * when MAP_STACK (and the user stack mapping) is fixed. 782 */ 783 /* 784 * It may be possible to simplify the entry 785 */ 786 vm_map_simplify_entry(map, new_entry, countp); 787 #endif 788 789 if (cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) { 790 pmap_object_init_pt(map->pmap, start, prot, 791 object, OFF_TO_IDX(offset), end - start, 792 cow & MAP_PREFAULT_PARTIAL); 793 } 794 795 return (KERN_SUCCESS); 796 } 797 798 /* 799 * Find sufficient space for `length' bytes in the given map, starting at 800 * `start'. The map must be locked. Returns 0 on success, 1 on no space. 801 * 802 * This function will returned an arbitrarily aligned pointer. If no 803 * particular alignment is required you should pass align as 1. Note that 804 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 805 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 806 * argument. 807 * 808 * 'align' should be a power of 2 but is not required to be. 809 */ 810 int 811 vm_map_findspace( 812 vm_map_t map, 813 vm_offset_t start, 814 vm_size_t length, 815 vm_offset_t align, 816 vm_offset_t *addr) 817 { 818 vm_map_entry_t entry, next; 819 vm_offset_t end; 820 vm_offset_t align_mask; 821 822 if (start < map->min_offset) 823 start = map->min_offset; 824 if (start > map->max_offset) 825 return (1); 826 827 /* 828 * If the alignment is not a power of 2 we will have to use 829 * a mod/division, set align_mask to a special value. 830 */ 831 if ((align | (align - 1)) + 1 != (align << 1)) 832 align_mask = (vm_offset_t)-1; 833 else 834 align_mask = align - 1; 835 836 retry: 837 /* 838 * Look for the first possible address; if there's already something 839 * at this address, we have to start after it. 840 */ 841 if (start == map->min_offset) { 842 if ((entry = map->first_free) != &map->header) 843 start = entry->end; 844 } else { 845 vm_map_entry_t tmp; 846 847 if (vm_map_lookup_entry(map, start, &tmp)) 848 start = tmp->end; 849 entry = tmp; 850 } 851 852 /* 853 * Look through the rest of the map, trying to fit a new region in the 854 * gap between existing regions, or after the very last region. 855 */ 856 for (;; start = (entry = next)->end) { 857 /* 858 * Adjust the proposed start by the requested alignment, 859 * be sure that we didn't wrap the address. 860 */ 861 if (align_mask == (vm_offset_t)-1) 862 end = ((start + align - 1) / align) * align; 863 else 864 end = (start + align_mask) & ~align_mask; 865 if (end < start) 866 return (1); 867 start = end; 868 /* 869 * Find the end of the proposed new region. Be sure we didn't 870 * go beyond the end of the map, or wrap around the address. 871 * Then check to see if this is the last entry or if the 872 * proposed end fits in the gap between this and the next 873 * entry. 874 */ 875 end = start + length; 876 if (end > map->max_offset || end < start) 877 return (1); 878 next = entry->next; 879 if (next == &map->header || next->start >= end) 880 break; 881 } 882 SAVE_HINT(map, entry); 883 if (map == kernel_map) { 884 vm_offset_t ksize; 885 if ((ksize = round_page(start + length)) > kernel_vm_end) { 886 pmap_growkernel(ksize); 887 goto retry; 888 } 889 } 890 *addr = start; 891 return (0); 892 } 893 894 /* 895 * vm_map_find finds an unallocated region in the target address 896 * map with the given length. The search is defined to be 897 * first-fit from the specified address; the region found is 898 * returned in the same parameter. 899 * 900 * If object is non-NULL, ref count must be bumped by caller 901 * prior to making call to account for the new entry. 902 */ 903 int 904 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 905 vm_offset_t *addr, /* IN/OUT */ 906 vm_size_t length, boolean_t find_space, vm_prot_t prot, 907 vm_prot_t max, int cow) 908 { 909 vm_offset_t start; 910 int result; 911 int count; 912 913 start = *addr; 914 915 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 916 vm_map_lock(map); 917 if (find_space) { 918 if (vm_map_findspace(map, start, length, 1, addr)) { 919 vm_map_unlock(map); 920 vm_map_entry_release(count); 921 return (KERN_NO_SPACE); 922 } 923 start = *addr; 924 } 925 result = vm_map_insert(map, &count, object, offset, 926 start, start + length, prot, max, cow); 927 vm_map_unlock(map); 928 vm_map_entry_release(count); 929 930 return (result); 931 } 932 933 /* 934 * vm_map_simplify_entry: 935 * 936 * Simplify the given map entry by merging with either neighbor. This 937 * routine also has the ability to merge with both neighbors. 938 * 939 * The map must be locked. 940 * 941 * This routine guarentees that the passed entry remains valid (though 942 * possibly extended). When merging, this routine may delete one or 943 * both neighbors. No action is taken on entries which have their 944 * in-transition flag set. 945 */ 946 void 947 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 948 { 949 vm_map_entry_t next, prev; 950 vm_size_t prevsize, esize; 951 952 if (entry->eflags & (MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP)) { 953 ++mycpu->gd_cnt.v_intrans_coll; 954 return; 955 } 956 957 prev = entry->prev; 958 if (prev != &map->header) { 959 prevsize = prev->end - prev->start; 960 if ( (prev->end == entry->start) && 961 (prev->object.vm_object == entry->object.vm_object) && 962 (!prev->object.vm_object || 963 (prev->offset + prevsize == entry->offset)) && 964 (prev->eflags == entry->eflags) && 965 (prev->protection == entry->protection) && 966 (prev->max_protection == entry->max_protection) && 967 (prev->inheritance == entry->inheritance) && 968 (prev->wired_count == entry->wired_count)) { 969 if (map->first_free == prev) 970 map->first_free = entry; 971 if (map->hint == prev) 972 map->hint = entry; 973 vm_map_entry_unlink(map, prev); 974 entry->start = prev->start; 975 entry->offset = prev->offset; 976 if (prev->object.vm_object) 977 vm_object_deallocate(prev->object.vm_object); 978 vm_map_entry_dispose(map, prev, countp); 979 } 980 } 981 982 next = entry->next; 983 if (next != &map->header) { 984 esize = entry->end - entry->start; 985 if ((entry->end == next->start) && 986 (next->object.vm_object == entry->object.vm_object) && 987 (!entry->object.vm_object || 988 (entry->offset + esize == next->offset)) && 989 (next->eflags == entry->eflags) && 990 (next->protection == entry->protection) && 991 (next->max_protection == entry->max_protection) && 992 (next->inheritance == entry->inheritance) && 993 (next->wired_count == entry->wired_count)) { 994 if (map->first_free == next) 995 map->first_free = entry; 996 if (map->hint == next) 997 map->hint = entry; 998 vm_map_entry_unlink(map, next); 999 entry->end = next->end; 1000 if (next->object.vm_object) 1001 vm_object_deallocate(next->object.vm_object); 1002 vm_map_entry_dispose(map, next, countp); 1003 } 1004 } 1005 } 1006 /* 1007 * vm_map_clip_start: [ internal use only ] 1008 * 1009 * Asserts that the given entry begins at or after 1010 * the specified address; if necessary, 1011 * it splits the entry into two. 1012 */ 1013 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1014 { \ 1015 if (startaddr > entry->start) \ 1016 _vm_map_clip_start(map, entry, startaddr, countp); \ 1017 } 1018 1019 /* 1020 * This routine is called only when it is known that 1021 * the entry must be split. 1022 */ 1023 static void 1024 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, int *countp) 1025 { 1026 vm_map_entry_t new_entry; 1027 1028 /* 1029 * Split off the front portion -- note that we must insert the new 1030 * entry BEFORE this one, so that this entry has the specified 1031 * starting address. 1032 */ 1033 1034 vm_map_simplify_entry(map, entry, countp); 1035 1036 /* 1037 * If there is no object backing this entry, we might as well create 1038 * one now. If we defer it, an object can get created after the map 1039 * is clipped, and individual objects will be created for the split-up 1040 * map. This is a bit of a hack, but is also about the best place to 1041 * put this improvement. 1042 */ 1043 1044 if (entry->object.vm_object == NULL && !map->system_map) { 1045 vm_object_t object; 1046 object = vm_object_allocate(OBJT_DEFAULT, 1047 atop(entry->end - entry->start)); 1048 entry->object.vm_object = object; 1049 entry->offset = 0; 1050 } 1051 1052 new_entry = vm_map_entry_create(map, countp); 1053 *new_entry = *entry; 1054 1055 new_entry->end = start; 1056 entry->offset += (start - entry->start); 1057 entry->start = start; 1058 1059 vm_map_entry_link(map, entry->prev, new_entry); 1060 1061 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 1062 vm_object_reference(new_entry->object.vm_object); 1063 } 1064 } 1065 1066 /* 1067 * vm_map_clip_end: [ internal use only ] 1068 * 1069 * Asserts that the given entry ends at or before 1070 * the specified address; if necessary, 1071 * it splits the entry into two. 1072 */ 1073 1074 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1075 { \ 1076 if (endaddr < entry->end) \ 1077 _vm_map_clip_end(map, entry, endaddr, countp); \ 1078 } 1079 1080 /* 1081 * This routine is called only when it is known that 1082 * the entry must be split. 1083 */ 1084 static void 1085 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, int *countp) 1086 { 1087 vm_map_entry_t new_entry; 1088 1089 /* 1090 * If there is no object backing this entry, we might as well create 1091 * one now. If we defer it, an object can get created after the map 1092 * is clipped, and individual objects will be created for the split-up 1093 * map. This is a bit of a hack, but is also about the best place to 1094 * put this improvement. 1095 */ 1096 1097 if (entry->object.vm_object == NULL && !map->system_map) { 1098 vm_object_t object; 1099 object = vm_object_allocate(OBJT_DEFAULT, 1100 atop(entry->end - entry->start)); 1101 entry->object.vm_object = object; 1102 entry->offset = 0; 1103 } 1104 1105 /* 1106 * Create a new entry and insert it AFTER the specified entry 1107 */ 1108 1109 new_entry = vm_map_entry_create(map, countp); 1110 *new_entry = *entry; 1111 1112 new_entry->start = entry->end = end; 1113 new_entry->offset += (end - entry->start); 1114 1115 vm_map_entry_link(map, entry, new_entry); 1116 1117 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 1118 vm_object_reference(new_entry->object.vm_object); 1119 } 1120 } 1121 1122 /* 1123 * VM_MAP_RANGE_CHECK: [ internal use only ] 1124 * 1125 * Asserts that the starting and ending region 1126 * addresses fall within the valid range of the map. 1127 */ 1128 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1129 { \ 1130 if (start < vm_map_min(map)) \ 1131 start = vm_map_min(map); \ 1132 if (end > vm_map_max(map)) \ 1133 end = vm_map_max(map); \ 1134 if (start > end) \ 1135 start = end; \ 1136 } 1137 1138 /* 1139 * vm_map_transition_wait: [ kernel use only ] 1140 * 1141 * Used to block when an in-transition collison occurs. The map 1142 * is unlocked for the sleep and relocked before the return. 1143 */ 1144 static 1145 void 1146 vm_map_transition_wait(vm_map_t map) 1147 { 1148 vm_map_unlock(map); 1149 tsleep(map, 0, "vment", 0); 1150 vm_map_lock(map); 1151 } 1152 1153 /* 1154 * CLIP_CHECK_BACK 1155 * CLIP_CHECK_FWD 1156 * 1157 * When we do blocking operations with the map lock held it is 1158 * possible that a clip might have occured on our in-transit entry, 1159 * requiring an adjustment to the entry in our loop. These macros 1160 * help the pageable and clip_range code deal with the case. The 1161 * conditional costs virtually nothing if no clipping has occured. 1162 */ 1163 1164 #define CLIP_CHECK_BACK(entry, save_start) \ 1165 do { \ 1166 while (entry->start != save_start) { \ 1167 entry = entry->prev; \ 1168 KASSERT(entry != &map->header, ("bad entry clip")); \ 1169 } \ 1170 } while(0) 1171 1172 #define CLIP_CHECK_FWD(entry, save_end) \ 1173 do { \ 1174 while (entry->end != save_end) { \ 1175 entry = entry->next; \ 1176 KASSERT(entry != &map->header, ("bad entry clip")); \ 1177 } \ 1178 } while(0) 1179 1180 1181 /* 1182 * vm_map_clip_range: [ kernel use only ] 1183 * 1184 * Clip the specified range and return the base entry. The 1185 * range may cover several entries starting at the returned base 1186 * and the first and last entry in the covering sequence will be 1187 * properly clipped to the requested start and end address. 1188 * 1189 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1190 * flag. 1191 * 1192 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1193 * covered by the requested range. 1194 * 1195 * The map must be exclusively locked on entry and will remain locked 1196 * on return. If no range exists or the range contains holes and you 1197 * specified that no holes were allowed, NULL will be returned. This 1198 * routine may temporarily unlock the map in order avoid a deadlock when 1199 * sleeping. 1200 */ 1201 static 1202 vm_map_entry_t 1203 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1204 int *countp, int flags) 1205 { 1206 vm_map_entry_t start_entry; 1207 vm_map_entry_t entry; 1208 1209 /* 1210 * Locate the entry and effect initial clipping. The in-transition 1211 * case does not occur very often so do not try to optimize it. 1212 */ 1213 again: 1214 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1215 return (NULL); 1216 entry = start_entry; 1217 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1218 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1219 ++mycpu->gd_cnt.v_intrans_coll; 1220 ++mycpu->gd_cnt.v_intrans_wait; 1221 vm_map_transition_wait(map); 1222 /* 1223 * entry and/or start_entry may have been clipped while 1224 * we slept, or may have gone away entirely. We have 1225 * to restart from the lookup. 1226 */ 1227 goto again; 1228 } 1229 /* 1230 * Since we hold an exclusive map lock we do not have to restart 1231 * after clipping, even though clipping may block in zalloc. 1232 */ 1233 vm_map_clip_start(map, entry, start, countp); 1234 vm_map_clip_end(map, entry, end, countp); 1235 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1236 1237 /* 1238 * Scan entries covered by the range. When working on the next 1239 * entry a restart need only re-loop on the current entry which 1240 * we have already locked, since 'next' may have changed. Also, 1241 * even though entry is safe, it may have been clipped so we 1242 * have to iterate forwards through the clip after sleeping. 1243 */ 1244 while (entry->next != &map->header && entry->next->start < end) { 1245 vm_map_entry_t next = entry->next; 1246 1247 if (flags & MAP_CLIP_NO_HOLES) { 1248 if (next->start > entry->end) { 1249 vm_map_unclip_range(map, start_entry, 1250 start, entry->end, countp, flags); 1251 return(NULL); 1252 } 1253 } 1254 1255 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1256 vm_offset_t save_end = entry->end; 1257 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1258 ++mycpu->gd_cnt.v_intrans_coll; 1259 ++mycpu->gd_cnt.v_intrans_wait; 1260 vm_map_transition_wait(map); 1261 1262 /* 1263 * clips might have occured while we blocked. 1264 */ 1265 CLIP_CHECK_FWD(entry, save_end); 1266 CLIP_CHECK_BACK(start_entry, start); 1267 continue; 1268 } 1269 /* 1270 * No restart necessary even though clip_end may block, we 1271 * are holding the map lock. 1272 */ 1273 vm_map_clip_end(map, next, end, countp); 1274 next->eflags |= MAP_ENTRY_IN_TRANSITION; 1275 entry = next; 1276 } 1277 if (flags & MAP_CLIP_NO_HOLES) { 1278 if (entry->end != end) { 1279 vm_map_unclip_range(map, start_entry, 1280 start, entry->end, countp, flags); 1281 return(NULL); 1282 } 1283 } 1284 return(start_entry); 1285 } 1286 1287 /* 1288 * vm_map_unclip_range: [ kernel use only ] 1289 * 1290 * Undo the effect of vm_map_clip_range(). You should pass the same 1291 * flags and the same range that you passed to vm_map_clip_range(). 1292 * This code will clear the in-transition flag on the entries and 1293 * wake up anyone waiting. This code will also simplify the sequence 1294 * and attempt to merge it with entries before and after the sequence. 1295 * 1296 * The map must be locked on entry and will remain locked on return. 1297 * 1298 * Note that you should also pass the start_entry returned by 1299 * vm_map_clip_range(). However, if you block between the two calls 1300 * with the map unlocked please be aware that the start_entry may 1301 * have been clipped and you may need to scan it backwards to find 1302 * the entry corresponding with the original start address. You are 1303 * responsible for this, vm_map_unclip_range() expects the correct 1304 * start_entry to be passed to it and will KASSERT otherwise. 1305 */ 1306 static 1307 void 1308 vm_map_unclip_range( 1309 vm_map_t map, 1310 vm_map_entry_t start_entry, 1311 vm_offset_t start, 1312 vm_offset_t end, 1313 int *countp, 1314 int flags) 1315 { 1316 vm_map_entry_t entry; 1317 1318 entry = start_entry; 1319 1320 KASSERT(entry->start == start, ("unclip_range: illegal base entry")); 1321 while (entry != &map->header && entry->start < end) { 1322 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, ("in-transition flag not set during unclip on: %p", entry)); 1323 KASSERT(entry->end <= end, ("unclip_range: tail wasn't clipped")); 1324 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 1325 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 1326 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 1327 wakeup(map); 1328 } 1329 entry = entry->next; 1330 } 1331 1332 /* 1333 * Simplification does not block so there is no restart case. 1334 */ 1335 entry = start_entry; 1336 while (entry != &map->header && entry->start < end) { 1337 vm_map_simplify_entry(map, entry, countp); 1338 entry = entry->next; 1339 } 1340 } 1341 1342 /* 1343 * vm_map_submap: [ kernel use only ] 1344 * 1345 * Mark the given range as handled by a subordinate map. 1346 * 1347 * This range must have been created with vm_map_find, 1348 * and no other operations may have been performed on this 1349 * range prior to calling vm_map_submap. 1350 * 1351 * Only a limited number of operations can be performed 1352 * within this rage after calling vm_map_submap: 1353 * vm_fault 1354 * [Don't try vm_map_copy!] 1355 * 1356 * To remove a submapping, one must first remove the 1357 * range from the superior map, and then destroy the 1358 * submap (if desired). [Better yet, don't try it.] 1359 */ 1360 int 1361 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 1362 { 1363 vm_map_entry_t entry; 1364 int result = KERN_INVALID_ARGUMENT; 1365 int count; 1366 1367 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1368 vm_map_lock(map); 1369 1370 VM_MAP_RANGE_CHECK(map, start, end); 1371 1372 if (vm_map_lookup_entry(map, start, &entry)) { 1373 vm_map_clip_start(map, entry, start, &count); 1374 } else { 1375 entry = entry->next; 1376 } 1377 1378 vm_map_clip_end(map, entry, end, &count); 1379 1380 if ((entry->start == start) && (entry->end == end) && 1381 ((entry->eflags & MAP_ENTRY_COW) == 0) && 1382 (entry->object.vm_object == NULL)) { 1383 entry->object.sub_map = submap; 1384 entry->eflags |= MAP_ENTRY_IS_SUB_MAP; 1385 result = KERN_SUCCESS; 1386 } 1387 vm_map_unlock(map); 1388 vm_map_entry_release(count); 1389 1390 return (result); 1391 } 1392 1393 /* 1394 * vm_map_protect: 1395 * 1396 * Sets the protection of the specified address 1397 * region in the target map. If "set_max" is 1398 * specified, the maximum protection is to be set; 1399 * otherwise, only the current protection is affected. 1400 */ 1401 int 1402 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 1403 vm_prot_t new_prot, boolean_t set_max) 1404 { 1405 vm_map_entry_t current; 1406 vm_map_entry_t entry; 1407 int count; 1408 1409 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1410 vm_map_lock(map); 1411 1412 VM_MAP_RANGE_CHECK(map, start, end); 1413 1414 if (vm_map_lookup_entry(map, start, &entry)) { 1415 vm_map_clip_start(map, entry, start, &count); 1416 } else { 1417 entry = entry->next; 1418 } 1419 1420 /* 1421 * Make a first pass to check for protection violations. 1422 */ 1423 1424 current = entry; 1425 while ((current != &map->header) && (current->start < end)) { 1426 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { 1427 vm_map_unlock(map); 1428 vm_map_entry_release(count); 1429 return (KERN_INVALID_ARGUMENT); 1430 } 1431 if ((new_prot & current->max_protection) != new_prot) { 1432 vm_map_unlock(map); 1433 vm_map_entry_release(count); 1434 return (KERN_PROTECTION_FAILURE); 1435 } 1436 current = current->next; 1437 } 1438 1439 /* 1440 * Go back and fix up protections. [Note that clipping is not 1441 * necessary the second time.] 1442 */ 1443 current = entry; 1444 1445 while ((current != &map->header) && (current->start < end)) { 1446 vm_prot_t old_prot; 1447 1448 vm_map_clip_end(map, current, end, &count); 1449 1450 old_prot = current->protection; 1451 if (set_max) 1452 current->protection = 1453 (current->max_protection = new_prot) & 1454 old_prot; 1455 else 1456 current->protection = new_prot; 1457 1458 /* 1459 * Update physical map if necessary. Worry about copy-on-write 1460 * here -- CHECK THIS XXX 1461 */ 1462 1463 if (current->protection != old_prot) { 1464 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 1465 VM_PROT_ALL) 1466 1467 pmap_protect(map->pmap, current->start, 1468 current->end, 1469 current->protection & MASK(current)); 1470 #undef MASK 1471 } 1472 1473 vm_map_simplify_entry(map, current, &count); 1474 1475 current = current->next; 1476 } 1477 1478 vm_map_unlock(map); 1479 vm_map_entry_release(count); 1480 return (KERN_SUCCESS); 1481 } 1482 1483 /* 1484 * vm_map_madvise: 1485 * 1486 * This routine traverses a processes map handling the madvise 1487 * system call. Advisories are classified as either those effecting 1488 * the vm_map_entry structure, or those effecting the underlying 1489 * objects. 1490 */ 1491 1492 int 1493 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, int behav) 1494 { 1495 vm_map_entry_t current, entry; 1496 int modify_map = 0; 1497 int count; 1498 1499 /* 1500 * Some madvise calls directly modify the vm_map_entry, in which case 1501 * we need to use an exclusive lock on the map and we need to perform 1502 * various clipping operations. Otherwise we only need a read-lock 1503 * on the map. 1504 */ 1505 1506 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1507 1508 switch(behav) { 1509 case MADV_NORMAL: 1510 case MADV_SEQUENTIAL: 1511 case MADV_RANDOM: 1512 case MADV_NOSYNC: 1513 case MADV_AUTOSYNC: 1514 case MADV_NOCORE: 1515 case MADV_CORE: 1516 modify_map = 1; 1517 vm_map_lock(map); 1518 break; 1519 case MADV_WILLNEED: 1520 case MADV_DONTNEED: 1521 case MADV_FREE: 1522 vm_map_lock_read(map); 1523 break; 1524 default: 1525 vm_map_entry_release(count); 1526 return (KERN_INVALID_ARGUMENT); 1527 } 1528 1529 /* 1530 * Locate starting entry and clip if necessary. 1531 */ 1532 1533 VM_MAP_RANGE_CHECK(map, start, end); 1534 1535 if (vm_map_lookup_entry(map, start, &entry)) { 1536 if (modify_map) 1537 vm_map_clip_start(map, entry, start, &count); 1538 } else { 1539 entry = entry->next; 1540 } 1541 1542 if (modify_map) { 1543 /* 1544 * madvise behaviors that are implemented in the vm_map_entry. 1545 * 1546 * We clip the vm_map_entry so that behavioral changes are 1547 * limited to the specified address range. 1548 */ 1549 for (current = entry; 1550 (current != &map->header) && (current->start < end); 1551 current = current->next 1552 ) { 1553 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) 1554 continue; 1555 1556 vm_map_clip_end(map, current, end, &count); 1557 1558 switch (behav) { 1559 case MADV_NORMAL: 1560 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 1561 break; 1562 case MADV_SEQUENTIAL: 1563 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 1564 break; 1565 case MADV_RANDOM: 1566 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 1567 break; 1568 case MADV_NOSYNC: 1569 current->eflags |= MAP_ENTRY_NOSYNC; 1570 break; 1571 case MADV_AUTOSYNC: 1572 current->eflags &= ~MAP_ENTRY_NOSYNC; 1573 break; 1574 case MADV_NOCORE: 1575 current->eflags |= MAP_ENTRY_NOCOREDUMP; 1576 break; 1577 case MADV_CORE: 1578 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 1579 break; 1580 default: 1581 break; 1582 } 1583 vm_map_simplify_entry(map, current, &count); 1584 } 1585 vm_map_unlock(map); 1586 } else { 1587 vm_pindex_t pindex; 1588 int count; 1589 1590 /* 1591 * madvise behaviors that are implemented in the underlying 1592 * vm_object. 1593 * 1594 * Since we don't clip the vm_map_entry, we have to clip 1595 * the vm_object pindex and count. 1596 */ 1597 for (current = entry; 1598 (current != &map->header) && (current->start < end); 1599 current = current->next 1600 ) { 1601 vm_offset_t useStart; 1602 1603 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) 1604 continue; 1605 1606 pindex = OFF_TO_IDX(current->offset); 1607 count = atop(current->end - current->start); 1608 useStart = current->start; 1609 1610 if (current->start < start) { 1611 pindex += atop(start - current->start); 1612 count -= atop(start - current->start); 1613 useStart = start; 1614 } 1615 if (current->end > end) 1616 count -= atop(current->end - end); 1617 1618 if (count <= 0) 1619 continue; 1620 1621 vm_object_madvise(current->object.vm_object, 1622 pindex, count, behav); 1623 if (behav == MADV_WILLNEED) { 1624 pmap_object_init_pt( 1625 map->pmap, 1626 useStart, 1627 current->protection, 1628 current->object.vm_object, 1629 pindex, 1630 (count << PAGE_SHIFT), 1631 MAP_PREFAULT_MADVISE 1632 ); 1633 } 1634 } 1635 vm_map_unlock_read(map); 1636 } 1637 vm_map_entry_release(count); 1638 return(0); 1639 } 1640 1641 1642 /* 1643 * vm_map_inherit: 1644 * 1645 * Sets the inheritance of the specified address 1646 * range in the target map. Inheritance 1647 * affects how the map will be shared with 1648 * child maps at the time of vm_map_fork. 1649 */ 1650 int 1651 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 1652 vm_inherit_t new_inheritance) 1653 { 1654 vm_map_entry_t entry; 1655 vm_map_entry_t temp_entry; 1656 int count; 1657 1658 switch (new_inheritance) { 1659 case VM_INHERIT_NONE: 1660 case VM_INHERIT_COPY: 1661 case VM_INHERIT_SHARE: 1662 break; 1663 default: 1664 return (KERN_INVALID_ARGUMENT); 1665 } 1666 1667 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1668 vm_map_lock(map); 1669 1670 VM_MAP_RANGE_CHECK(map, start, end); 1671 1672 if (vm_map_lookup_entry(map, start, &temp_entry)) { 1673 entry = temp_entry; 1674 vm_map_clip_start(map, entry, start, &count); 1675 } else 1676 entry = temp_entry->next; 1677 1678 while ((entry != &map->header) && (entry->start < end)) { 1679 vm_map_clip_end(map, entry, end, &count); 1680 1681 entry->inheritance = new_inheritance; 1682 1683 vm_map_simplify_entry(map, entry, &count); 1684 1685 entry = entry->next; 1686 } 1687 vm_map_unlock(map); 1688 vm_map_entry_release(count); 1689 return (KERN_SUCCESS); 1690 } 1691 1692 /* 1693 * Implement the semantics of mlock 1694 */ 1695 int 1696 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 1697 boolean_t new_pageable) 1698 { 1699 vm_map_entry_t entry; 1700 vm_map_entry_t start_entry; 1701 vm_offset_t end; 1702 int rv = KERN_SUCCESS; 1703 int count; 1704 1705 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1706 vm_map_lock(map); 1707 VM_MAP_RANGE_CHECK(map, start, real_end); 1708 end = real_end; 1709 1710 start_entry = vm_map_clip_range(map, start, end, &count, MAP_CLIP_NO_HOLES); 1711 if (start_entry == NULL) { 1712 vm_map_unlock(map); 1713 vm_map_entry_release(count); 1714 return (KERN_INVALID_ADDRESS); 1715 } 1716 1717 if (new_pageable == 0) { 1718 entry = start_entry; 1719 while ((entry != &map->header) && (entry->start < end)) { 1720 vm_offset_t save_start; 1721 vm_offset_t save_end; 1722 1723 /* 1724 * Already user wired or hard wired (trivial cases) 1725 */ 1726 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 1727 entry = entry->next; 1728 continue; 1729 } 1730 if (entry->wired_count != 0) { 1731 entry->wired_count++; 1732 entry->eflags |= MAP_ENTRY_USER_WIRED; 1733 entry = entry->next; 1734 continue; 1735 } 1736 1737 /* 1738 * A new wiring requires instantiation of appropriate 1739 * management structures and the faulting in of the 1740 * page. 1741 */ 1742 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 1743 int copyflag = entry->eflags & MAP_ENTRY_NEEDS_COPY; 1744 if (copyflag && ((entry->protection & VM_PROT_WRITE) != 0)) { 1745 1746 vm_object_shadow(&entry->object.vm_object, 1747 &entry->offset, 1748 atop(entry->end - entry->start)); 1749 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 1750 1751 } else if (entry->object.vm_object == NULL && 1752 !map->system_map) { 1753 1754 entry->object.vm_object = 1755 vm_object_allocate(OBJT_DEFAULT, 1756 atop(entry->end - entry->start)); 1757 entry->offset = (vm_offset_t) 0; 1758 1759 } 1760 } 1761 entry->wired_count++; 1762 entry->eflags |= MAP_ENTRY_USER_WIRED; 1763 1764 /* 1765 * Now fault in the area. Note that vm_fault_wire() 1766 * may release the map lock temporarily, it will be 1767 * relocked on return. The in-transition 1768 * flag protects the entries. 1769 */ 1770 save_start = entry->start; 1771 save_end = entry->end; 1772 rv = vm_fault_wire(map, entry, TRUE); 1773 if (rv) { 1774 CLIP_CHECK_BACK(entry, save_start); 1775 for (;;) { 1776 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 1777 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 1778 entry->wired_count = 0; 1779 if (entry->end == save_end) 1780 break; 1781 entry = entry->next; 1782 KASSERT(entry != &map->header, ("bad entry clip during backout")); 1783 } 1784 end = save_start; /* unwire the rest */ 1785 break; 1786 } 1787 /* 1788 * note that even though the entry might have been 1789 * clipped, the USER_WIRED flag we set prevents 1790 * duplication so we do not have to do a 1791 * clip check. 1792 */ 1793 entry = entry->next; 1794 } 1795 1796 /* 1797 * If we failed fall through to the unwiring section to 1798 * unwire what we had wired so far. 'end' has already 1799 * been adjusted. 1800 */ 1801 if (rv) 1802 new_pageable = 1; 1803 1804 /* 1805 * start_entry might have been clipped if we unlocked the 1806 * map and blocked. No matter how clipped it has gotten 1807 * there should be a fragment that is on our start boundary. 1808 */ 1809 CLIP_CHECK_BACK(start_entry, start); 1810 } 1811 1812 /* 1813 * Deal with the unwiring case. 1814 */ 1815 if (new_pageable) { 1816 /* 1817 * This is the unwiring case. We must first ensure that the 1818 * range to be unwired is really wired down. We know there 1819 * are no holes. 1820 */ 1821 entry = start_entry; 1822 while ((entry != &map->header) && (entry->start < end)) { 1823 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 1824 rv = KERN_INVALID_ARGUMENT; 1825 goto done; 1826 } 1827 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry)); 1828 entry = entry->next; 1829 } 1830 1831 /* 1832 * Now decrement the wiring count for each region. If a region 1833 * becomes completely unwired, unwire its physical pages and 1834 * mappings. 1835 */ 1836 /* 1837 * The map entries are processed in a loop, checking to 1838 * make sure the entry is wired and asserting it has a wired 1839 * count. However, another loop was inserted more-or-less in 1840 * the middle of the unwiring path. This loop picks up the 1841 * "entry" loop variable from the first loop without first 1842 * setting it to start_entry. Naturally, the secound loop 1843 * is never entered and the pages backing the entries are 1844 * never unwired. This can lead to a leak of wired pages. 1845 */ 1846 entry = start_entry; 1847 while ((entry != &map->header) && (entry->start < end)) { 1848 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 1849 ("expected USER_WIRED on entry %p", entry)); 1850 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 1851 entry->wired_count--; 1852 if (entry->wired_count == 0) 1853 vm_fault_unwire(map, entry); 1854 entry = entry->next; 1855 } 1856 } 1857 done: 1858 vm_map_unclip_range(map, start_entry, start, real_end, &count, 1859 MAP_CLIP_NO_HOLES); 1860 map->timestamp++; 1861 vm_map_unlock(map); 1862 vm_map_entry_release(count); 1863 return (rv); 1864 } 1865 1866 /* 1867 * vm_map_wire: 1868 * 1869 * Sets the pageability of the specified address 1870 * range in the target map. Regions specified 1871 * as not pageable require locked-down physical 1872 * memory and physical page maps. 1873 * 1874 * The map must not be locked, but a reference 1875 * must remain to the map throughout the call. 1876 * 1877 * This function may be called via the zalloc path and must properly 1878 * reserve map entries for kernel_map. 1879 */ 1880 int 1881 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags) 1882 { 1883 vm_map_entry_t entry; 1884 vm_map_entry_t start_entry; 1885 vm_offset_t end; 1886 int rv = KERN_SUCCESS; 1887 int count; 1888 1889 if (kmflags & KM_KRESERVE) 1890 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 1891 else 1892 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1893 vm_map_lock(map); 1894 VM_MAP_RANGE_CHECK(map, start, real_end); 1895 end = real_end; 1896 1897 start_entry = vm_map_clip_range(map, start, end, &count, MAP_CLIP_NO_HOLES); 1898 if (start_entry == NULL) { 1899 vm_map_unlock(map); 1900 rv = KERN_INVALID_ADDRESS; 1901 goto failure; 1902 } 1903 if ((kmflags & KM_PAGEABLE) == 0) { 1904 /* 1905 * Wiring. 1906 * 1907 * 1. Holding the write lock, we create any shadow or zero-fill 1908 * objects that need to be created. Then we clip each map 1909 * entry to the region to be wired and increment its wiring 1910 * count. We create objects before clipping the map entries 1911 * to avoid object proliferation. 1912 * 1913 * 2. We downgrade to a read lock, and call vm_fault_wire to 1914 * fault in the pages for any newly wired area (wired_count is 1915 * 1). 1916 * 1917 * Downgrading to a read lock for vm_fault_wire avoids a 1918 * possible deadlock with another process that may have faulted 1919 * on one of the pages to be wired (it would mark the page busy, 1920 * blocking us, then in turn block on the map lock that we 1921 * hold). Because of problems in the recursive lock package, 1922 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 1923 * any actions that require the write lock must be done 1924 * beforehand. Because we keep the read lock on the map, the 1925 * copy-on-write status of the entries we modify here cannot 1926 * change. 1927 */ 1928 1929 entry = start_entry; 1930 while ((entry != &map->header) && (entry->start < end)) { 1931 /* 1932 * Trivial case if the entry is already wired 1933 */ 1934 if (entry->wired_count) { 1935 entry->wired_count++; 1936 entry = entry->next; 1937 continue; 1938 } 1939 1940 /* 1941 * The entry is being newly wired, we have to setup 1942 * appropriate management structures. A shadow 1943 * object is required for a copy-on-write region, 1944 * or a normal object for a zero-fill region. We 1945 * do not have to do this for entries that point to sub 1946 * maps because we won't hold the lock on the sub map. 1947 */ 1948 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 1949 int copyflag = entry->eflags & MAP_ENTRY_NEEDS_COPY; 1950 if (copyflag && 1951 ((entry->protection & VM_PROT_WRITE) != 0)) { 1952 1953 vm_object_shadow(&entry->object.vm_object, 1954 &entry->offset, 1955 atop(entry->end - entry->start)); 1956 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 1957 } else if (entry->object.vm_object == NULL && 1958 !map->system_map) { 1959 entry->object.vm_object = 1960 vm_object_allocate(OBJT_DEFAULT, 1961 atop(entry->end - entry->start)); 1962 entry->offset = (vm_offset_t) 0; 1963 } 1964 } 1965 1966 entry->wired_count++; 1967 entry = entry->next; 1968 } 1969 1970 /* 1971 * Pass 2. 1972 */ 1973 1974 /* 1975 * HACK HACK HACK HACK 1976 * 1977 * Unlock the map to avoid deadlocks. The in-transit flag 1978 * protects us from most changes but note that 1979 * clipping may still occur. To prevent clipping from 1980 * occuring after the unlock, except for when we are 1981 * blocking in vm_fault_wire, we must run in a critical 1982 * section, otherwise our accesses to entry->start and 1983 * entry->end could be corrupted. We have to enter the 1984 * critical section prior to unlocking so start_entry does 1985 * not change out from under us at the very beginning of the 1986 * loop. 1987 * 1988 * HACK HACK HACK HACK 1989 */ 1990 1991 crit_enter(); 1992 1993 entry = start_entry; 1994 while (entry != &map->header && entry->start < end) { 1995 /* 1996 * If vm_fault_wire fails for any page we need to undo 1997 * what has been done. We decrement the wiring count 1998 * for those pages which have not yet been wired (now) 1999 * and unwire those that have (later). 2000 */ 2001 vm_offset_t save_start = entry->start; 2002 vm_offset_t save_end = entry->end; 2003 2004 if (entry->wired_count == 1) 2005 rv = vm_fault_wire(map, entry, FALSE); 2006 if (rv) { 2007 CLIP_CHECK_BACK(entry, save_start); 2008 for (;;) { 2009 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly")); 2010 entry->wired_count = 0; 2011 if (entry->end == save_end) 2012 break; 2013 entry = entry->next; 2014 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2015 } 2016 end = save_start; 2017 break; 2018 } 2019 CLIP_CHECK_FWD(entry, save_end); 2020 entry = entry->next; 2021 } 2022 crit_exit(); 2023 2024 /* 2025 * If a failure occured undo everything by falling through 2026 * to the unwiring code. 'end' has already been adjusted 2027 * appropriately. 2028 */ 2029 if (rv) 2030 kmflags |= KM_PAGEABLE; 2031 2032 /* 2033 * start_entry is still IN_TRANSITION but may have been 2034 * clipped since vm_fault_wire() unlocks and relocks the 2035 * map. No matter how clipped it has gotten there should 2036 * be a fragment that is on our start boundary. 2037 */ 2038 CLIP_CHECK_BACK(start_entry, start); 2039 } 2040 2041 if (kmflags & KM_PAGEABLE) { 2042 /* 2043 * This is the unwiring case. We must first ensure that the 2044 * range to be unwired is really wired down. We know there 2045 * are no holes. 2046 */ 2047 entry = start_entry; 2048 while ((entry != &map->header) && (entry->start < end)) { 2049 if (entry->wired_count == 0) { 2050 rv = KERN_INVALID_ARGUMENT; 2051 goto done; 2052 } 2053 entry = entry->next; 2054 } 2055 2056 /* 2057 * Now decrement the wiring count for each region. If a region 2058 * becomes completely unwired, unwire its physical pages and 2059 * mappings. 2060 */ 2061 entry = start_entry; 2062 while ((entry != &map->header) && (entry->start < end)) { 2063 entry->wired_count--; 2064 if (entry->wired_count == 0) 2065 vm_fault_unwire(map, entry); 2066 entry = entry->next; 2067 } 2068 } 2069 done: 2070 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2071 MAP_CLIP_NO_HOLES); 2072 map->timestamp++; 2073 vm_map_unlock(map); 2074 failure: 2075 if (kmflags & KM_KRESERVE) 2076 vm_map_entry_krelease(count); 2077 else 2078 vm_map_entry_release(count); 2079 return (rv); 2080 } 2081 2082 /* 2083 * vm_map_set_wired_quick() 2084 * 2085 * Mark a newly allocated address range as wired but do not fault in 2086 * the pages. The caller is expected to load the pages into the object. 2087 * 2088 * The map must be locked on entry and will remain locked on return. 2089 */ 2090 void 2091 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, int *countp) 2092 { 2093 vm_map_entry_t scan; 2094 vm_map_entry_t entry; 2095 2096 entry = vm_map_clip_range(map, addr, addr + size, countp, MAP_CLIP_NO_HOLES); 2097 for (scan = entry; scan != &map->header && scan->start < addr + size; scan = scan->next) { 2098 KKASSERT(entry->wired_count == 0); 2099 entry->wired_count = 1; 2100 } 2101 vm_map_unclip_range(map, entry, addr, addr + size, countp, MAP_CLIP_NO_HOLES); 2102 } 2103 2104 /* 2105 * vm_map_clean 2106 * 2107 * Push any dirty cached pages in the address range to their pager. 2108 * If syncio is TRUE, dirty pages are written synchronously. 2109 * If invalidate is TRUE, any cached pages are freed as well. 2110 * 2111 * Returns an error if any part of the specified range is not mapped. 2112 */ 2113 int 2114 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, boolean_t syncio, 2115 boolean_t invalidate) 2116 { 2117 vm_map_entry_t current; 2118 vm_map_entry_t entry; 2119 vm_size_t size; 2120 vm_object_t object; 2121 vm_ooffset_t offset; 2122 2123 vm_map_lock_read(map); 2124 VM_MAP_RANGE_CHECK(map, start, end); 2125 if (!vm_map_lookup_entry(map, start, &entry)) { 2126 vm_map_unlock_read(map); 2127 return (KERN_INVALID_ADDRESS); 2128 } 2129 /* 2130 * Make a first pass to check for holes. 2131 */ 2132 for (current = entry; current->start < end; current = current->next) { 2133 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { 2134 vm_map_unlock_read(map); 2135 return (KERN_INVALID_ARGUMENT); 2136 } 2137 if (end > current->end && 2138 (current->next == &map->header || 2139 current->end != current->next->start)) { 2140 vm_map_unlock_read(map); 2141 return (KERN_INVALID_ADDRESS); 2142 } 2143 } 2144 2145 if (invalidate) 2146 pmap_remove(vm_map_pmap(map), start, end); 2147 /* 2148 * Make a second pass, cleaning/uncaching pages from the indicated 2149 * objects as we go. 2150 */ 2151 for (current = entry; current->start < end; current = current->next) { 2152 offset = current->offset + (start - current->start); 2153 size = (end <= current->end ? end : current->end) - start; 2154 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { 2155 vm_map_t smap; 2156 vm_map_entry_t tentry; 2157 vm_size_t tsize; 2158 2159 smap = current->object.sub_map; 2160 vm_map_lock_read(smap); 2161 (void) vm_map_lookup_entry(smap, offset, &tentry); 2162 tsize = tentry->end - offset; 2163 if (tsize < size) 2164 size = tsize; 2165 object = tentry->object.vm_object; 2166 offset = tentry->offset + (offset - tentry->start); 2167 vm_map_unlock_read(smap); 2168 } else { 2169 object = current->object.vm_object; 2170 } 2171 /* 2172 * Note that there is absolutely no sense in writing out 2173 * anonymous objects, so we track down the vnode object 2174 * to write out. 2175 * We invalidate (remove) all pages from the address space 2176 * anyway, for semantic correctness. 2177 * 2178 * note: certain anonymous maps, such as MAP_NOSYNC maps, 2179 * may start out with a NULL object. 2180 */ 2181 while (object && object->backing_object) { 2182 offset += object->backing_object_offset; 2183 object = object->backing_object; 2184 if (object->size < OFF_TO_IDX( offset + size)) 2185 size = IDX_TO_OFF(object->size) - offset; 2186 } 2187 if (object && (object->type == OBJT_VNODE) && 2188 (current->protection & VM_PROT_WRITE)) { 2189 /* 2190 * Flush pages if writing is allowed, invalidate them 2191 * if invalidation requested. Pages undergoing I/O 2192 * will be ignored by vm_object_page_remove(). 2193 * 2194 * We cannot lock the vnode and then wait for paging 2195 * to complete without deadlocking against vm_fault. 2196 * Instead we simply call vm_object_page_remove() and 2197 * allow it to block internally on a page-by-page 2198 * basis when it encounters pages undergoing async 2199 * I/O. 2200 */ 2201 int flags; 2202 2203 vm_object_reference(object); 2204 vn_lock(object->handle, 2205 LK_EXCLUSIVE | LK_RETRY, curthread); 2206 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 2207 flags |= invalidate ? OBJPC_INVAL : 0; 2208 vm_object_page_clean(object, 2209 OFF_TO_IDX(offset), 2210 OFF_TO_IDX(offset + size + PAGE_MASK), 2211 flags); 2212 VOP_UNLOCK(((struct vnode *)object->handle), 2213 0, curthread); 2214 vm_object_deallocate(object); 2215 } 2216 if (object && invalidate && 2217 ((object->type == OBJT_VNODE) || 2218 (object->type == OBJT_DEVICE))) { 2219 int clean_only = 2220 (object->type == OBJT_DEVICE) ? FALSE : TRUE; 2221 vm_object_reference(object); 2222 vm_object_page_remove(object, 2223 OFF_TO_IDX(offset), 2224 OFF_TO_IDX(offset + size + PAGE_MASK), 2225 clean_only); 2226 vm_object_deallocate(object); 2227 } 2228 start += size; 2229 } 2230 2231 vm_map_unlock_read(map); 2232 return (KERN_SUCCESS); 2233 } 2234 2235 /* 2236 * vm_map_entry_unwire: [ internal use only ] 2237 * 2238 * Make the region specified by this entry pageable. 2239 * 2240 * The map in question should be locked. 2241 * [This is the reason for this routine's existence.] 2242 */ 2243 static void 2244 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 2245 { 2246 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2247 entry->wired_count = 0; 2248 vm_fault_unwire(map, entry); 2249 } 2250 2251 /* 2252 * vm_map_entry_delete: [ internal use only ] 2253 * 2254 * Deallocate the given entry from the target map. 2255 */ 2256 static void 2257 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 2258 { 2259 vm_map_entry_unlink(map, entry); 2260 map->size -= entry->end - entry->start; 2261 2262 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2263 vm_object_deallocate(entry->object.vm_object); 2264 } 2265 2266 vm_map_entry_dispose(map, entry, countp); 2267 } 2268 2269 /* 2270 * vm_map_delete: [ internal use only ] 2271 * 2272 * Deallocates the given address range from the target 2273 * map. 2274 */ 2275 int 2276 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 2277 { 2278 vm_object_t object; 2279 vm_map_entry_t entry; 2280 vm_map_entry_t first_entry; 2281 2282 /* 2283 * Find the start of the region, and clip it 2284 */ 2285 2286 again: 2287 if (!vm_map_lookup_entry(map, start, &first_entry)) { 2288 entry = first_entry->next; 2289 } else { 2290 entry = first_entry; 2291 vm_map_clip_start(map, entry, start, countp); 2292 /* 2293 * Fix the lookup hint now, rather than each time though the 2294 * loop. 2295 */ 2296 SAVE_HINT(map, entry->prev); 2297 } 2298 2299 /* 2300 * Save the free space hint 2301 */ 2302 2303 if (entry == &map->header) { 2304 map->first_free = &map->header; 2305 } else if (map->first_free->start >= start) { 2306 map->first_free = entry->prev; 2307 } 2308 2309 /* 2310 * Step through all entries in this region 2311 */ 2312 2313 while ((entry != &map->header) && (entry->start < end)) { 2314 vm_map_entry_t next; 2315 vm_offset_t s, e; 2316 vm_pindex_t offidxstart, offidxend, count; 2317 2318 /* 2319 * If we hit an in-transition entry we have to sleep and 2320 * retry. It's easier (and not really slower) to just retry 2321 * since this case occurs so rarely and the hint is already 2322 * pointing at the right place. We have to reset the 2323 * start offset so as not to accidently delete an entry 2324 * another process just created in vacated space. 2325 */ 2326 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 2327 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2328 start = entry->start; 2329 ++mycpu->gd_cnt.v_intrans_coll; 2330 ++mycpu->gd_cnt.v_intrans_wait; 2331 vm_map_transition_wait(map); 2332 goto again; 2333 } 2334 vm_map_clip_end(map, entry, end, countp); 2335 2336 s = entry->start; 2337 e = entry->end; 2338 next = entry->next; 2339 2340 offidxstart = OFF_TO_IDX(entry->offset); 2341 count = OFF_TO_IDX(e - s); 2342 object = entry->object.vm_object; 2343 2344 /* 2345 * Unwire before removing addresses from the pmap; otherwise, 2346 * unwiring will put the entries back in the pmap. 2347 */ 2348 if (entry->wired_count != 0) 2349 vm_map_entry_unwire(map, entry); 2350 2351 offidxend = offidxstart + count; 2352 2353 if ((object == kernel_object) || (object == kmem_object)) { 2354 vm_object_page_remove(object, offidxstart, offidxend, FALSE); 2355 } else { 2356 pmap_remove(map->pmap, s, e); 2357 if (object != NULL && 2358 object->ref_count != 1 && 2359 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING && 2360 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2361 vm_object_collapse(object); 2362 vm_object_page_remove(object, offidxstart, offidxend, FALSE); 2363 if (object->type == OBJT_SWAP) { 2364 swap_pager_freespace(object, offidxstart, count); 2365 } 2366 if (offidxend >= object->size && 2367 offidxstart < object->size) { 2368 object->size = offidxstart; 2369 } 2370 } 2371 } 2372 2373 /* 2374 * Delete the entry (which may delete the object) only after 2375 * removing all pmap entries pointing to its pages. 2376 * (Otherwise, its page frames may be reallocated, and any 2377 * modify bits will be set in the wrong object!) 2378 */ 2379 vm_map_entry_delete(map, entry, countp); 2380 entry = next; 2381 } 2382 return (KERN_SUCCESS); 2383 } 2384 2385 /* 2386 * vm_map_remove: 2387 * 2388 * Remove the given address range from the target map. 2389 * This is the exported form of vm_map_delete. 2390 */ 2391 int 2392 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 2393 { 2394 int result; 2395 int count; 2396 2397 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2398 vm_map_lock(map); 2399 VM_MAP_RANGE_CHECK(map, start, end); 2400 result = vm_map_delete(map, start, end, &count); 2401 vm_map_unlock(map); 2402 vm_map_entry_release(count); 2403 2404 return (result); 2405 } 2406 2407 /* 2408 * vm_map_check_protection: 2409 * 2410 * Assert that the target map allows the specified 2411 * privilege on the entire address region given. 2412 * The entire region must be allocated. 2413 */ 2414 boolean_t 2415 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 2416 vm_prot_t protection) 2417 { 2418 vm_map_entry_t entry; 2419 vm_map_entry_t tmp_entry; 2420 2421 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 2422 return (FALSE); 2423 } 2424 entry = tmp_entry; 2425 2426 while (start < end) { 2427 if (entry == &map->header) { 2428 return (FALSE); 2429 } 2430 /* 2431 * No holes allowed! 2432 */ 2433 2434 if (start < entry->start) { 2435 return (FALSE); 2436 } 2437 /* 2438 * Check protection associated with entry. 2439 */ 2440 2441 if ((entry->protection & protection) != protection) { 2442 return (FALSE); 2443 } 2444 /* go to next entry */ 2445 2446 start = entry->end; 2447 entry = entry->next; 2448 } 2449 return (TRUE); 2450 } 2451 2452 /* 2453 * Split the pages in a map entry into a new object. This affords 2454 * easier removal of unused pages, and keeps object inheritance from 2455 * being a negative impact on memory usage. 2456 */ 2457 static void 2458 vm_map_split(vm_map_entry_t entry) 2459 { 2460 vm_page_t m; 2461 vm_object_t orig_object, new_object, source; 2462 vm_offset_t s, e; 2463 vm_pindex_t offidxstart, offidxend, idx; 2464 vm_size_t size; 2465 vm_ooffset_t offset; 2466 2467 orig_object = entry->object.vm_object; 2468 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 2469 return; 2470 if (orig_object->ref_count <= 1) 2471 return; 2472 2473 offset = entry->offset; 2474 s = entry->start; 2475 e = entry->end; 2476 2477 offidxstart = OFF_TO_IDX(offset); 2478 offidxend = offidxstart + OFF_TO_IDX(e - s); 2479 size = offidxend - offidxstart; 2480 2481 new_object = vm_pager_allocate(orig_object->type, 2482 NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL); 2483 if (new_object == NULL) 2484 return; 2485 2486 source = orig_object->backing_object; 2487 if (source != NULL) { 2488 vm_object_reference(source); /* Referenced by new_object */ 2489 LIST_INSERT_HEAD(&source->shadow_head, 2490 new_object, shadow_list); 2491 vm_object_clear_flag(source, OBJ_ONEMAPPING); 2492 new_object->backing_object_offset = 2493 orig_object->backing_object_offset + IDX_TO_OFF(offidxstart); 2494 new_object->backing_object = source; 2495 source->shadow_count++; 2496 source->generation++; 2497 } 2498 2499 for (idx = 0; idx < size; idx++) { 2500 vm_page_t m; 2501 2502 /* 2503 * A critical section is required to avoid a race between 2504 * the lookup and an interrupt/unbusy/free and our busy 2505 * check. 2506 */ 2507 crit_enter(); 2508 retry: 2509 m = vm_page_lookup(orig_object, offidxstart + idx); 2510 if (m == NULL) { 2511 crit_exit(); 2512 continue; 2513 } 2514 2515 /* 2516 * We must wait for pending I/O to complete before we can 2517 * rename the page. 2518 * 2519 * We do not have to VM_PROT_NONE the page as mappings should 2520 * not be changed by this operation. 2521 */ 2522 if (vm_page_sleep_busy(m, TRUE, "spltwt")) 2523 goto retry; 2524 vm_page_busy(m); 2525 vm_page_rename(m, new_object, idx); 2526 /* page automatically made dirty by rename and cache handled */ 2527 vm_page_busy(m); 2528 crit_exit(); 2529 } 2530 2531 if (orig_object->type == OBJT_SWAP) { 2532 vm_object_pip_add(orig_object, 1); 2533 /* 2534 * copy orig_object pages into new_object 2535 * and destroy unneeded pages in 2536 * shadow object. 2537 */ 2538 swap_pager_copy(orig_object, new_object, offidxstart, 0); 2539 vm_object_pip_wakeup(orig_object); 2540 } 2541 2542 /* 2543 * Wakeup the pages we played with. No spl protection is needed 2544 * for a simple wakeup. 2545 */ 2546 for (idx = 0; idx < size; idx++) { 2547 m = vm_page_lookup(new_object, idx); 2548 if (m) 2549 vm_page_wakeup(m); 2550 } 2551 2552 entry->object.vm_object = new_object; 2553 entry->offset = 0LL; 2554 vm_object_deallocate(orig_object); 2555 } 2556 2557 /* 2558 * vm_map_copy_entry: 2559 * 2560 * Copies the contents of the source entry to the destination 2561 * entry. The entries *must* be aligned properly. 2562 */ 2563 static void 2564 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 2565 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 2566 { 2567 vm_object_t src_object; 2568 2569 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) 2570 return; 2571 2572 if (src_entry->wired_count == 0) { 2573 2574 /* 2575 * If the source entry is marked needs_copy, it is already 2576 * write-protected. 2577 */ 2578 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 2579 pmap_protect(src_map->pmap, 2580 src_entry->start, 2581 src_entry->end, 2582 src_entry->protection & ~VM_PROT_WRITE); 2583 } 2584 2585 /* 2586 * Make a copy of the object. 2587 */ 2588 if ((src_object = src_entry->object.vm_object) != NULL) { 2589 2590 if ((src_object->handle == NULL) && 2591 (src_object->type == OBJT_DEFAULT || 2592 src_object->type == OBJT_SWAP)) { 2593 vm_object_collapse(src_object); 2594 if ((src_object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) { 2595 vm_map_split(src_entry); 2596 src_object = src_entry->object.vm_object; 2597 } 2598 } 2599 2600 vm_object_reference(src_object); 2601 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 2602 dst_entry->object.vm_object = src_object; 2603 src_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY); 2604 dst_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY); 2605 dst_entry->offset = src_entry->offset; 2606 } else { 2607 dst_entry->object.vm_object = NULL; 2608 dst_entry->offset = 0; 2609 } 2610 2611 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, 2612 dst_entry->end - dst_entry->start, src_entry->start); 2613 } else { 2614 /* 2615 * Of course, wired down pages can't be set copy-on-write. 2616 * Cause wired pages to be copied into the new map by 2617 * simulating faults (the new pages are pageable) 2618 */ 2619 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 2620 } 2621 } 2622 2623 /* 2624 * vmspace_fork: 2625 * Create a new process vmspace structure and vm_map 2626 * based on those of an existing process. The new map 2627 * is based on the old map, according to the inheritance 2628 * values on the regions in that map. 2629 * 2630 * The source map must not be locked. 2631 */ 2632 struct vmspace * 2633 vmspace_fork(struct vmspace *vm1) 2634 { 2635 struct vmspace *vm2; 2636 vm_map_t old_map = &vm1->vm_map; 2637 vm_map_t new_map; 2638 vm_map_entry_t old_entry; 2639 vm_map_entry_t new_entry; 2640 vm_object_t object; 2641 int count; 2642 2643 vm_map_lock(old_map); 2644 old_map->infork = 1; 2645 2646 /* 2647 * XXX Note: upcalls are not copied. 2648 */ 2649 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset); 2650 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 2651 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 2652 new_map = &vm2->vm_map; /* XXX */ 2653 new_map->timestamp = 1; 2654 2655 count = 0; 2656 old_entry = old_map->header.next; 2657 while (old_entry != &old_map->header) { 2658 ++count; 2659 old_entry = old_entry->next; 2660 } 2661 2662 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 2663 2664 old_entry = old_map->header.next; 2665 while (old_entry != &old_map->header) { 2666 if (old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) 2667 panic("vm_map_fork: encountered a submap"); 2668 2669 switch (old_entry->inheritance) { 2670 case VM_INHERIT_NONE: 2671 break; 2672 2673 case VM_INHERIT_SHARE: 2674 /* 2675 * Clone the entry, creating the shared object if necessary. 2676 */ 2677 object = old_entry->object.vm_object; 2678 if (object == NULL) { 2679 object = vm_object_allocate(OBJT_DEFAULT, 2680 atop(old_entry->end - old_entry->start)); 2681 old_entry->object.vm_object = object; 2682 old_entry->offset = (vm_offset_t) 0; 2683 } 2684 2685 /* 2686 * Add the reference before calling vm_object_shadow 2687 * to insure that a shadow object is created. 2688 */ 2689 vm_object_reference(object); 2690 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 2691 vm_object_shadow(&old_entry->object.vm_object, 2692 &old_entry->offset, 2693 atop(old_entry->end - old_entry->start)); 2694 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 2695 /* Transfer the second reference too. */ 2696 vm_object_reference( 2697 old_entry->object.vm_object); 2698 vm_object_deallocate(object); 2699 object = old_entry->object.vm_object; 2700 } 2701 vm_object_clear_flag(object, OBJ_ONEMAPPING); 2702 2703 /* 2704 * Clone the entry, referencing the shared object. 2705 */ 2706 new_entry = vm_map_entry_create(new_map, &count); 2707 *new_entry = *old_entry; 2708 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2709 new_entry->wired_count = 0; 2710 2711 /* 2712 * Insert the entry into the new map -- we know we're 2713 * inserting at the end of the new map. 2714 */ 2715 2716 vm_map_entry_link(new_map, new_map->header.prev, 2717 new_entry); 2718 2719 /* 2720 * Update the physical map 2721 */ 2722 2723 pmap_copy(new_map->pmap, old_map->pmap, 2724 new_entry->start, 2725 (old_entry->end - old_entry->start), 2726 old_entry->start); 2727 break; 2728 2729 case VM_INHERIT_COPY: 2730 /* 2731 * Clone the entry and link into the map. 2732 */ 2733 new_entry = vm_map_entry_create(new_map, &count); 2734 *new_entry = *old_entry; 2735 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2736 new_entry->wired_count = 0; 2737 new_entry->object.vm_object = NULL; 2738 vm_map_entry_link(new_map, new_map->header.prev, 2739 new_entry); 2740 vm_map_copy_entry(old_map, new_map, old_entry, 2741 new_entry); 2742 break; 2743 } 2744 old_entry = old_entry->next; 2745 } 2746 2747 new_map->size = old_map->size; 2748 old_map->infork = 0; 2749 vm_map_unlock(old_map); 2750 vm_map_entry_release(count); 2751 2752 return (vm2); 2753 } 2754 2755 int 2756 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 2757 vm_prot_t prot, vm_prot_t max, int cow) 2758 { 2759 vm_map_entry_t prev_entry; 2760 vm_map_entry_t new_stack_entry; 2761 vm_size_t init_ssize; 2762 int rv; 2763 int count; 2764 2765 if (VM_MIN_ADDRESS > 0 && addrbos < VM_MIN_ADDRESS) 2766 return (KERN_NO_SPACE); 2767 2768 if (max_ssize < sgrowsiz) 2769 init_ssize = max_ssize; 2770 else 2771 init_ssize = sgrowsiz; 2772 2773 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2774 vm_map_lock(map); 2775 2776 /* If addr is already mapped, no go */ 2777 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) { 2778 vm_map_unlock(map); 2779 vm_map_entry_release(count); 2780 return (KERN_NO_SPACE); 2781 } 2782 2783 /* If we would blow our VMEM resource limit, no go */ 2784 if (map->size + init_ssize > 2785 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 2786 vm_map_unlock(map); 2787 vm_map_entry_release(count); 2788 return (KERN_NO_SPACE); 2789 } 2790 2791 /* If we can't accomodate max_ssize in the current mapping, 2792 * no go. However, we need to be aware that subsequent user 2793 * mappings might map into the space we have reserved for 2794 * stack, and currently this space is not protected. 2795 * 2796 * Hopefully we will at least detect this condition 2797 * when we try to grow the stack. 2798 */ 2799 if ((prev_entry->next != &map->header) && 2800 (prev_entry->next->start < addrbos + max_ssize)) { 2801 vm_map_unlock(map); 2802 vm_map_entry_release(count); 2803 return (KERN_NO_SPACE); 2804 } 2805 2806 /* We initially map a stack of only init_ssize. We will 2807 * grow as needed later. Since this is to be a grow 2808 * down stack, we map at the top of the range. 2809 * 2810 * Note: we would normally expect prot and max to be 2811 * VM_PROT_ALL, and cow to be 0. Possibly we should 2812 * eliminate these as input parameters, and just 2813 * pass these values here in the insert call. 2814 */ 2815 rv = vm_map_insert(map, &count, 2816 NULL, 0, addrbos + max_ssize - init_ssize, 2817 addrbos + max_ssize, prot, max, cow); 2818 2819 /* Now set the avail_ssize amount */ 2820 if (rv == KERN_SUCCESS) { 2821 if (prev_entry != &map->header) 2822 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count); 2823 new_stack_entry = prev_entry->next; 2824 if (new_stack_entry->end != addrbos + max_ssize || 2825 new_stack_entry->start != addrbos + max_ssize - init_ssize) 2826 panic ("Bad entry start/end for new stack entry"); 2827 else 2828 new_stack_entry->avail_ssize = max_ssize - init_ssize; 2829 } 2830 2831 vm_map_unlock(map); 2832 vm_map_entry_release(count); 2833 return (rv); 2834 } 2835 2836 /* Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 2837 * desired address is already mapped, or if we successfully grow 2838 * the stack. Also returns KERN_SUCCESS if addr is outside the 2839 * stack range (this is strange, but preserves compatibility with 2840 * the grow function in vm_machdep.c). 2841 */ 2842 int 2843 vm_map_growstack (struct proc *p, vm_offset_t addr) 2844 { 2845 vm_map_entry_t prev_entry; 2846 vm_map_entry_t stack_entry; 2847 vm_map_entry_t new_stack_entry; 2848 struct vmspace *vm = p->p_vmspace; 2849 vm_map_t map = &vm->vm_map; 2850 vm_offset_t end; 2851 int grow_amount; 2852 int rv = KERN_SUCCESS; 2853 int is_procstack; 2854 int use_read_lock = 1; 2855 int count; 2856 2857 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2858 Retry: 2859 if (use_read_lock) 2860 vm_map_lock_read(map); 2861 else 2862 vm_map_lock(map); 2863 2864 /* If addr is already in the entry range, no need to grow.*/ 2865 if (vm_map_lookup_entry(map, addr, &prev_entry)) 2866 goto done; 2867 2868 if ((stack_entry = prev_entry->next) == &map->header) 2869 goto done; 2870 if (prev_entry == &map->header) 2871 end = stack_entry->start - stack_entry->avail_ssize; 2872 else 2873 end = prev_entry->end; 2874 2875 /* This next test mimics the old grow function in vm_machdep.c. 2876 * It really doesn't quite make sense, but we do it anyway 2877 * for compatibility. 2878 * 2879 * If not growable stack, return success. This signals the 2880 * caller to proceed as he would normally with normal vm. 2881 */ 2882 if (stack_entry->avail_ssize < 1 || 2883 addr >= stack_entry->start || 2884 addr < stack_entry->start - stack_entry->avail_ssize) { 2885 goto done; 2886 } 2887 2888 /* Find the minimum grow amount */ 2889 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE); 2890 if (grow_amount > stack_entry->avail_ssize) { 2891 rv = KERN_NO_SPACE; 2892 goto done; 2893 } 2894 2895 /* If there is no longer enough space between the entries 2896 * nogo, and adjust the available space. Note: this 2897 * should only happen if the user has mapped into the 2898 * stack area after the stack was created, and is 2899 * probably an error. 2900 * 2901 * This also effectively destroys any guard page the user 2902 * might have intended by limiting the stack size. 2903 */ 2904 if (grow_amount > stack_entry->start - end) { 2905 if (use_read_lock && vm_map_lock_upgrade(map)) { 2906 use_read_lock = 0; 2907 goto Retry; 2908 } 2909 use_read_lock = 0; 2910 stack_entry->avail_ssize = stack_entry->start - end; 2911 rv = KERN_NO_SPACE; 2912 goto done; 2913 } 2914 2915 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 2916 2917 /* If this is the main process stack, see if we're over the 2918 * stack limit. 2919 */ 2920 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 2921 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 2922 rv = KERN_NO_SPACE; 2923 goto done; 2924 } 2925 2926 /* Round up the grow amount modulo SGROWSIZ */ 2927 grow_amount = roundup (grow_amount, sgrowsiz); 2928 if (grow_amount > stack_entry->avail_ssize) { 2929 grow_amount = stack_entry->avail_ssize; 2930 } 2931 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 2932 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 2933 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - 2934 ctob(vm->vm_ssize); 2935 } 2936 2937 /* If we would blow our VMEM resource limit, no go */ 2938 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 2939 rv = KERN_NO_SPACE; 2940 goto done; 2941 } 2942 2943 if (use_read_lock && vm_map_lock_upgrade(map)) { 2944 use_read_lock = 0; 2945 goto Retry; 2946 } 2947 use_read_lock = 0; 2948 2949 /* Get the preliminary new entry start value */ 2950 addr = stack_entry->start - grow_amount; 2951 2952 /* If this puts us into the previous entry, cut back our growth 2953 * to the available space. Also, see the note above. 2954 */ 2955 if (addr < end) { 2956 stack_entry->avail_ssize = stack_entry->start - end; 2957 addr = end; 2958 } 2959 2960 rv = vm_map_insert(map, &count, 2961 NULL, 0, addr, stack_entry->start, 2962 VM_PROT_ALL, 2963 VM_PROT_ALL, 2964 0); 2965 2966 /* Adjust the available stack space by the amount we grew. */ 2967 if (rv == KERN_SUCCESS) { 2968 if (prev_entry != &map->header) 2969 vm_map_clip_end(map, prev_entry, addr, &count); 2970 new_stack_entry = prev_entry->next; 2971 if (new_stack_entry->end != stack_entry->start || 2972 new_stack_entry->start != addr) 2973 panic ("Bad stack grow start/end in new stack entry"); 2974 else { 2975 new_stack_entry->avail_ssize = stack_entry->avail_ssize - 2976 (new_stack_entry->end - 2977 new_stack_entry->start); 2978 if (is_procstack) 2979 vm->vm_ssize += btoc(new_stack_entry->end - 2980 new_stack_entry->start); 2981 } 2982 } 2983 2984 done: 2985 if (use_read_lock) 2986 vm_map_unlock_read(map); 2987 else 2988 vm_map_unlock(map); 2989 vm_map_entry_release(count); 2990 return (rv); 2991 } 2992 2993 /* 2994 * Unshare the specified VM space for exec. If other processes are 2995 * mapped to it, then create a new one. The new vmspace is null. 2996 */ 2997 2998 void 2999 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 3000 { 3001 struct vmspace *oldvmspace = p->p_vmspace; 3002 struct vmspace *newvmspace; 3003 vm_map_t map = &p->p_vmspace->vm_map; 3004 3005 /* 3006 * If we are execing a resident vmspace we fork it, otherwise 3007 * we create a new vmspace. Note that exitingcnt and upcalls 3008 * are not copied to the new vmspace. 3009 */ 3010 if (vmcopy) { 3011 newvmspace = vmspace_fork(vmcopy); 3012 } else { 3013 newvmspace = vmspace_alloc(map->min_offset, map->max_offset); 3014 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 3015 (caddr_t)&oldvmspace->vm_endcopy - 3016 (caddr_t)&oldvmspace->vm_startcopy); 3017 } 3018 3019 /* 3020 * This code is written like this for prototype purposes. The 3021 * goal is to avoid running down the vmspace here, but let the 3022 * other process's that are still using the vmspace to finally 3023 * run it down. Even though there is little or no chance of blocking 3024 * here, it is a good idea to keep this form for future mods. 3025 */ 3026 p->p_vmspace = newvmspace; 3027 pmap_pinit2(vmspace_pmap(newvmspace)); 3028 if (p == curproc) 3029 pmap_activate(p); 3030 vmspace_free(oldvmspace); 3031 } 3032 3033 /* 3034 * Unshare the specified VM space for forcing COW. This 3035 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 3036 * 3037 * The exitingcnt test is not strictly necessary but has been 3038 * included for code sanity (to make the code a bit more deterministic). 3039 */ 3040 3041 void 3042 vmspace_unshare(struct proc *p) 3043 { 3044 struct vmspace *oldvmspace = p->p_vmspace; 3045 struct vmspace *newvmspace; 3046 3047 if (oldvmspace->vm_refcnt == 1 && oldvmspace->vm_exitingcnt == 0) 3048 return; 3049 newvmspace = vmspace_fork(oldvmspace); 3050 p->p_vmspace = newvmspace; 3051 pmap_pinit2(vmspace_pmap(newvmspace)); 3052 if (p == curproc) 3053 pmap_activate(p); 3054 vmspace_free(oldvmspace); 3055 } 3056 3057 /* 3058 * vm_map_lookup: 3059 * 3060 * Finds the VM object, offset, and 3061 * protection for a given virtual address in the 3062 * specified map, assuming a page fault of the 3063 * type specified. 3064 * 3065 * Leaves the map in question locked for read; return 3066 * values are guaranteed until a vm_map_lookup_done 3067 * call is performed. Note that the map argument 3068 * is in/out; the returned map must be used in 3069 * the call to vm_map_lookup_done. 3070 * 3071 * A handle (out_entry) is returned for use in 3072 * vm_map_lookup_done, to make that fast. 3073 * 3074 * If a lookup is requested with "write protection" 3075 * specified, the map may be changed to perform virtual 3076 * copying operations, although the data referenced will 3077 * remain the same. 3078 */ 3079 int 3080 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 3081 vm_offset_t vaddr, 3082 vm_prot_t fault_typea, 3083 vm_map_entry_t *out_entry, /* OUT */ 3084 vm_object_t *object, /* OUT */ 3085 vm_pindex_t *pindex, /* OUT */ 3086 vm_prot_t *out_prot, /* OUT */ 3087 boolean_t *wired) /* OUT */ 3088 { 3089 vm_map_entry_t entry; 3090 vm_map_t map = *var_map; 3091 vm_prot_t prot; 3092 vm_prot_t fault_type = fault_typea; 3093 int use_read_lock = 1; 3094 int rv = KERN_SUCCESS; 3095 3096 RetryLookup: 3097 if (use_read_lock) 3098 vm_map_lock_read(map); 3099 else 3100 vm_map_lock(map); 3101 3102 /* 3103 * If the map has an interesting hint, try it before calling full 3104 * blown lookup routine. 3105 */ 3106 entry = map->hint; 3107 *out_entry = entry; 3108 3109 if ((entry == &map->header) || 3110 (vaddr < entry->start) || (vaddr >= entry->end)) { 3111 vm_map_entry_t tmp_entry; 3112 3113 /* 3114 * Entry was either not a valid hint, or the vaddr was not 3115 * contained in the entry, so do a full lookup. 3116 */ 3117 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 3118 rv = KERN_INVALID_ADDRESS; 3119 goto done; 3120 } 3121 3122 entry = tmp_entry; 3123 *out_entry = entry; 3124 } 3125 3126 /* 3127 * Handle submaps. 3128 */ 3129 3130 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 3131 vm_map_t old_map = map; 3132 3133 *var_map = map = entry->object.sub_map; 3134 if (use_read_lock) 3135 vm_map_unlock_read(old_map); 3136 else 3137 vm_map_unlock(old_map); 3138 use_read_lock = 1; 3139 goto RetryLookup; 3140 } 3141 3142 /* 3143 * Check whether this task is allowed to have this page. 3144 * Note the special case for MAP_ENTRY_COW 3145 * pages with an override. This is to implement a forced 3146 * COW for debuggers. 3147 */ 3148 3149 if (fault_type & VM_PROT_OVERRIDE_WRITE) 3150 prot = entry->max_protection; 3151 else 3152 prot = entry->protection; 3153 3154 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 3155 if ((fault_type & prot) != fault_type) { 3156 rv = KERN_PROTECTION_FAILURE; 3157 goto done; 3158 } 3159 3160 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 3161 (entry->eflags & MAP_ENTRY_COW) && 3162 (fault_type & VM_PROT_WRITE) && 3163 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 3164 rv = KERN_PROTECTION_FAILURE; 3165 goto done; 3166 } 3167 3168 /* 3169 * If this page is not pageable, we have to get it for all possible 3170 * accesses. 3171 */ 3172 3173 *wired = (entry->wired_count != 0); 3174 if (*wired) 3175 prot = fault_type = entry->protection; 3176 3177 /* 3178 * If the entry was copy-on-write, we either ... 3179 */ 3180 3181 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3182 /* 3183 * If we want to write the page, we may as well handle that 3184 * now since we've got the map locked. 3185 * 3186 * If we don't need to write the page, we just demote the 3187 * permissions allowed. 3188 */ 3189 3190 if (fault_type & VM_PROT_WRITE) { 3191 /* 3192 * Make a new object, and place it in the object 3193 * chain. Note that no new references have appeared 3194 * -- one just moved from the map to the new 3195 * object. 3196 */ 3197 3198 if (use_read_lock && vm_map_lock_upgrade(map)) { 3199 use_read_lock = 0; 3200 goto RetryLookup; 3201 } 3202 use_read_lock = 0; 3203 3204 vm_object_shadow( 3205 &entry->object.vm_object, 3206 &entry->offset, 3207 atop(entry->end - entry->start)); 3208 3209 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 3210 } else { 3211 /* 3212 * We're attempting to read a copy-on-write page -- 3213 * don't allow writes. 3214 */ 3215 3216 prot &= ~VM_PROT_WRITE; 3217 } 3218 } 3219 3220 /* 3221 * Create an object if necessary. 3222 */ 3223 if (entry->object.vm_object == NULL && 3224 !map->system_map) { 3225 if (use_read_lock && vm_map_lock_upgrade(map)) { 3226 use_read_lock = 0; 3227 goto RetryLookup; 3228 } 3229 use_read_lock = 0; 3230 entry->object.vm_object = vm_object_allocate(OBJT_DEFAULT, 3231 atop(entry->end - entry->start)); 3232 entry->offset = 0; 3233 } 3234 3235 /* 3236 * Return the object/offset from this entry. If the entry was 3237 * copy-on-write or empty, it has been fixed up. 3238 */ 3239 3240 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 3241 *object = entry->object.vm_object; 3242 3243 /* 3244 * Return whether this is the only map sharing this data. On 3245 * success we return with a read lock held on the map. On failure 3246 * we return with the map unlocked. 3247 */ 3248 *out_prot = prot; 3249 done: 3250 if (rv == KERN_SUCCESS) { 3251 if (use_read_lock == 0) 3252 vm_map_lock_downgrade(map); 3253 } else if (use_read_lock) { 3254 vm_map_unlock_read(map); 3255 } else { 3256 vm_map_unlock(map); 3257 } 3258 return (rv); 3259 } 3260 3261 /* 3262 * vm_map_lookup_done: 3263 * 3264 * Releases locks acquired by a vm_map_lookup 3265 * (according to the handle returned by that lookup). 3266 */ 3267 3268 void 3269 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 3270 { 3271 /* 3272 * Unlock the main-level map 3273 */ 3274 vm_map_unlock_read(map); 3275 if (count) 3276 vm_map_entry_release(count); 3277 } 3278 3279 /* 3280 * Performs the copy_on_write operations necessary to allow the virtual copies 3281 * into user space to work. This has to be called for write(2) system calls 3282 * from other processes, file unlinking, and file size shrinkage. 3283 */ 3284 void 3285 vm_freeze_copyopts(vm_object_t object, vm_pindex_t froma, vm_pindex_t toa) 3286 { 3287 int rv; 3288 vm_object_t robject; 3289 vm_pindex_t idx; 3290 3291 if ((object == NULL) || 3292 ((object->flags & OBJ_OPT) == 0)) 3293 return; 3294 3295 if (object->shadow_count > object->ref_count) 3296 panic("vm_freeze_copyopts: sc > rc"); 3297 3298 while ((robject = LIST_FIRST(&object->shadow_head)) != NULL) { 3299 vm_pindex_t bo_pindex; 3300 vm_page_t m_in, m_out; 3301 3302 bo_pindex = OFF_TO_IDX(robject->backing_object_offset); 3303 3304 vm_object_reference(robject); 3305 3306 vm_object_pip_wait(robject, "objfrz"); 3307 3308 if (robject->ref_count == 1) { 3309 vm_object_deallocate(robject); 3310 continue; 3311 } 3312 3313 vm_object_pip_add(robject, 1); 3314 3315 for (idx = 0; idx < robject->size; idx++) { 3316 3317 m_out = vm_page_grab(robject, idx, 3318 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 3319 3320 if (m_out->valid == 0) { 3321 m_in = vm_page_grab(object, bo_pindex + idx, 3322 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 3323 if (m_in->valid == 0) { 3324 rv = vm_pager_get_pages(object, &m_in, 1, 0); 3325 if (rv != VM_PAGER_OK) { 3326 printf("vm_freeze_copyopts: cannot read page from file: %lx\n", (long)m_in->pindex); 3327 continue; 3328 } 3329 vm_page_deactivate(m_in); 3330 } 3331 3332 vm_page_protect(m_in, VM_PROT_NONE); 3333 pmap_copy_page(VM_PAGE_TO_PHYS(m_in), VM_PAGE_TO_PHYS(m_out)); 3334 m_out->valid = m_in->valid; 3335 vm_page_dirty(m_out); 3336 vm_page_activate(m_out); 3337 vm_page_wakeup(m_in); 3338 } 3339 vm_page_wakeup(m_out); 3340 } 3341 3342 object->shadow_count--; 3343 object->ref_count--; 3344 LIST_REMOVE(robject, shadow_list); 3345 robject->backing_object = NULL; 3346 robject->backing_object_offset = 0; 3347 3348 vm_object_pip_wakeup(robject); 3349 vm_object_deallocate(robject); 3350 } 3351 3352 vm_object_clear_flag(object, OBJ_OPT); 3353 } 3354 3355 #include "opt_ddb.h" 3356 #ifdef DDB 3357 #include <sys/kernel.h> 3358 3359 #include <ddb/ddb.h> 3360 3361 /* 3362 * vm_map_print: [ debug ] 3363 */ 3364 DB_SHOW_COMMAND(map, vm_map_print) 3365 { 3366 static int nlines; 3367 /* XXX convert args. */ 3368 vm_map_t map = (vm_map_t)addr; 3369 boolean_t full = have_addr; 3370 3371 vm_map_entry_t entry; 3372 3373 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 3374 (void *)map, 3375 (void *)map->pmap, map->nentries, map->timestamp); 3376 nlines++; 3377 3378 if (!full && db_indent) 3379 return; 3380 3381 db_indent += 2; 3382 for (entry = map->header.next; entry != &map->header; 3383 entry = entry->next) { 3384 db_iprintf("map entry %p: start=%p, end=%p\n", 3385 (void *)entry, (void *)entry->start, (void *)entry->end); 3386 nlines++; 3387 { 3388 static char *inheritance_name[4] = 3389 {"share", "copy", "none", "donate_copy"}; 3390 3391 db_iprintf(" prot=%x/%x/%s", 3392 entry->protection, 3393 entry->max_protection, 3394 inheritance_name[(int)(unsigned char)entry->inheritance]); 3395 if (entry->wired_count != 0) 3396 db_printf(", wired"); 3397 } 3398 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 3399 /* XXX no %qd in kernel. Truncate entry->offset. */ 3400 db_printf(", share=%p, offset=0x%lx\n", 3401 (void *)entry->object.sub_map, 3402 (long)entry->offset); 3403 nlines++; 3404 if ((entry->prev == &map->header) || 3405 (entry->prev->object.sub_map != 3406 entry->object.sub_map)) { 3407 db_indent += 2; 3408 vm_map_print((db_expr_t)(intptr_t) 3409 entry->object.sub_map, 3410 full, 0, (char *)0); 3411 db_indent -= 2; 3412 } 3413 } else { 3414 /* XXX no %qd in kernel. Truncate entry->offset. */ 3415 db_printf(", object=%p, offset=0x%lx", 3416 (void *)entry->object.vm_object, 3417 (long)entry->offset); 3418 if (entry->eflags & MAP_ENTRY_COW) 3419 db_printf(", copy (%s)", 3420 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 3421 db_printf("\n"); 3422 nlines++; 3423 3424 if ((entry->prev == &map->header) || 3425 (entry->prev->object.vm_object != 3426 entry->object.vm_object)) { 3427 db_indent += 2; 3428 vm_object_print((db_expr_t)(intptr_t) 3429 entry->object.vm_object, 3430 full, 0, (char *)0); 3431 nlines += 4; 3432 db_indent -= 2; 3433 } 3434 } 3435 } 3436 db_indent -= 2; 3437 if (db_indent == 0) 3438 nlines = 0; 3439 } 3440 3441 3442 DB_SHOW_COMMAND(procvm, procvm) 3443 { 3444 struct proc *p; 3445 3446 if (have_addr) { 3447 p = (struct proc *) addr; 3448 } else { 3449 p = curproc; 3450 } 3451 3452 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 3453 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 3454 (void *)vmspace_pmap(p->p_vmspace)); 3455 3456 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 3457 } 3458 3459 #endif /* DDB */ 3460