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 1.41 2006/03/02 19:08:00 dillon 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 #include <sys/tree.h> 82 83 #include <vm/vm.h> 84 #include <vm/vm_param.h> 85 #include <vm/pmap.h> 86 #include <vm/vm_map.h> 87 #include <vm/vm_page.h> 88 #include <vm/vm_object.h> 89 #include <vm/vm_pager.h> 90 #include <vm/vm_kern.h> 91 #include <vm/vm_extern.h> 92 #include <vm/swap_pager.h> 93 #include <vm/vm_zone.h> 94 95 #include <sys/thread2.h> 96 97 /* 98 * Virtual memory maps provide for the mapping, protection, 99 * and sharing of virtual memory objects. In addition, 100 * this module provides for an efficient virtual copy of 101 * memory from one map to another. 102 * 103 * Synchronization is required prior to most operations. 104 * 105 * Maps consist of an ordered doubly-linked list of simple 106 * entries; a single hint is used to speed up lookups. 107 * 108 * Since portions of maps are specified by start/end addresses, 109 * which may not align with existing map entries, all 110 * routines merely "clip" entries to these start/end values. 111 * [That is, an entry is split into two, bordering at a 112 * start or end value.] Note that these clippings may not 113 * always be necessary (as the two resulting entries are then 114 * not changed); however, the clipping is done for convenience. 115 * 116 * As mentioned above, virtual copy operations are performed 117 * by copying VM object references from one map to 118 * another, and then marking both regions as copy-on-write. 119 */ 120 121 /* 122 * vm_map_startup: 123 * 124 * Initialize the vm_map module. Must be called before 125 * any other vm_map routines. 126 * 127 * Map and entry structures are allocated from the general 128 * purpose memory pool with some exceptions: 129 * 130 * - The kernel map and kmem submap are allocated statically. 131 * - Kernel map entries are allocated out of a static pool. 132 * 133 * These restrictions are necessary since malloc() uses the 134 * maps and requires map entries. 135 */ 136 137 #define VMEPERCPU 2 138 139 static struct vm_zone mapentzone_store, mapzone_store; 140 static vm_zone_t mapentzone, mapzone, vmspace_zone; 141 static struct vm_object mapentobj, mapobj; 142 143 static struct vm_map_entry map_entry_init[MAX_MAPENT]; 144 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU]; 145 static struct vm_map map_init[MAX_KMAP]; 146 147 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *); 148 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *); 149 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 150 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 151 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *); 152 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t); 153 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t, 154 vm_map_entry_t); 155 static void vm_map_split (vm_map_entry_t); 156 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); 157 158 void 159 vm_map_startup(void) 160 { 161 mapzone = &mapzone_store; 162 zbootinit(mapzone, "MAP", sizeof (struct vm_map), 163 map_init, MAX_KMAP); 164 mapentzone = &mapentzone_store; 165 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry), 166 map_entry_init, MAX_MAPENT); 167 } 168 169 /* 170 * Red black tree functions 171 */ 172 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b); 173 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare); 174 175 /* a->start is address, and the only field has to be initialized */ 176 static int 177 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b) 178 { 179 if (a->start < b->start) 180 return(-1); 181 else if (a->start > b->start) 182 return(1); 183 return(0); 184 } 185 186 /* 187 * Allocate a vmspace structure, including a vm_map and pmap, 188 * and initialize those structures. The refcnt is set to 1. 189 * The remaining fields must be initialized by the caller. 190 */ 191 struct vmspace * 192 vmspace_alloc(vm_offset_t min, vm_offset_t max) 193 { 194 struct vmspace *vm; 195 196 vm = zalloc(vmspace_zone); 197 vm_map_init(&vm->vm_map, min, max); 198 pmap_pinit(vmspace_pmap(vm)); 199 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */ 200 vm->vm_refcnt = 1; 201 vm->vm_shm = NULL; 202 vm->vm_exitingcnt = 0; 203 return (vm); 204 } 205 206 void 207 vm_init2(void) 208 { 209 zinitna(mapentzone, &mapentobj, NULL, 0, 0, 210 ZONE_USE_RESERVE | ZONE_SPECIAL, 1); 211 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1); 212 vmspace_zone = zinit("VMSPACE", sizeof (struct vmspace), 0, 0, 3); 213 pmap_init2(); 214 vm_object_init2(); 215 } 216 217 static __inline void 218 vmspace_dofree(struct vmspace *vm) 219 { 220 int count; 221 222 /* 223 * Make sure any SysV shm is freed, it might not have in 224 * exit1() 225 */ 226 shmexit(vm); 227 228 KKASSERT(vm->vm_upcalls == NULL); 229 230 /* 231 * Lock the map, to wait out all other references to it. 232 * Delete all of the mappings and pages they hold, then call 233 * the pmap module to reclaim anything left. 234 */ 235 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 236 vm_map_lock(&vm->vm_map); 237 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset, 238 vm->vm_map.max_offset, &count); 239 vm_map_unlock(&vm->vm_map); 240 vm_map_entry_release(count); 241 242 pmap_release(vmspace_pmap(vm)); 243 zfree(vmspace_zone, vm); 244 } 245 246 void 247 vmspace_free(struct vmspace *vm) 248 { 249 if (vm->vm_refcnt == 0) 250 panic("vmspace_free: attempt to free already freed vmspace"); 251 252 if (--vm->vm_refcnt == 0 && vm->vm_exitingcnt == 0) 253 vmspace_dofree(vm); 254 } 255 256 void 257 vmspace_exitfree(struct proc *p) 258 { 259 struct vmspace *vm; 260 261 vm = p->p_vmspace; 262 p->p_vmspace = NULL; 263 264 /* 265 * cleanup by parent process wait()ing on exiting child. vm_refcnt 266 * may not be 0 (e.g. fork() and child exits without exec()ing). 267 * exitingcnt may increment above 0 and drop back down to zero 268 * several times while vm_refcnt is held non-zero. vm_refcnt 269 * may also increment above 0 and drop back down to zero several 270 * times while vm_exitingcnt is held non-zero. 271 * 272 * The last wait on the exiting child's vmspace will clean up 273 * the remainder of the vmspace. 274 */ 275 if (--vm->vm_exitingcnt == 0 && vm->vm_refcnt == 0) 276 vmspace_dofree(vm); 277 } 278 279 /* 280 * vmspace_swap_count() - count the approximate swap useage in pages for a 281 * vmspace. 282 * 283 * Swap useage is determined by taking the proportional swap used by 284 * VM objects backing the VM map. To make up for fractional losses, 285 * if the VM object has any swap use at all the associated map entries 286 * count for at least 1 swap page. 287 */ 288 int 289 vmspace_swap_count(struct vmspace *vmspace) 290 { 291 vm_map_t map = &vmspace->vm_map; 292 vm_map_entry_t cur; 293 int count = 0; 294 295 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 296 vm_object_t object; 297 298 if ((cur->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 299 (object = cur->object.vm_object) != NULL && 300 object->type == OBJT_SWAP 301 ) { 302 int n = (cur->end - cur->start) / PAGE_SIZE; 303 304 if (object->un_pager.swp.swp_bcount) { 305 count += object->un_pager.swp.swp_bcount * 306 SWAP_META_PAGES * n / object->size + 1; 307 } 308 } 309 } 310 return(count); 311 } 312 313 314 /* 315 * vm_map_create: 316 * 317 * Creates and returns a new empty VM map with 318 * the given physical map structure, and having 319 * the given lower and upper address bounds. 320 */ 321 vm_map_t 322 vm_map_create(pmap_t pmap, vm_offset_t min, vm_offset_t max) 323 { 324 vm_map_t result; 325 326 result = zalloc(mapzone); 327 vm_map_init(result, min, max); 328 result->pmap = pmap; 329 return (result); 330 } 331 332 /* 333 * Initialize an existing vm_map structure 334 * such as that in the vmspace structure. 335 * The pmap is set elsewhere. 336 */ 337 void 338 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max) 339 { 340 map->header.next = map->header.prev = &map->header; 341 RB_INIT(&map->rb_root); 342 map->nentries = 0; 343 map->size = 0; 344 map->system_map = 0; 345 map->infork = 0; 346 map->min_offset = min; 347 map->max_offset = max; 348 map->first_free = &map->header; 349 map->hint = &map->header; 350 map->timestamp = 0; 351 lockinit(&map->lock, "thrd_sleep", 0, LK_NOPAUSE); 352 } 353 354 /* 355 * vm_map_entry_reserve_cpu_init: 356 * 357 * Set an initial negative count so the first attempt to reserve 358 * space preloads a bunch of vm_map_entry's for this cpu. Also 359 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 360 * map a new page for vm_map_entry structures. SMP systems are 361 * particularly sensitive. 362 * 363 * This routine is called in early boot so we cannot just call 364 * vm_map_entry_reserve(). 365 * 366 * May be called for a gd other then mycpu, but may only be called 367 * during early boot. 368 */ 369 void 370 vm_map_entry_reserve_cpu_init(globaldata_t gd) 371 { 372 vm_map_entry_t entry; 373 int i; 374 375 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2; 376 entry = &cpu_map_entry_init[gd->gd_cpuid][0]; 377 for (i = 0; i < VMEPERCPU; ++i, ++entry) { 378 entry->next = gd->gd_vme_base; 379 gd->gd_vme_base = entry; 380 } 381 } 382 383 /* 384 * vm_map_entry_reserve: 385 * 386 * Reserves vm_map_entry structures so code later on can manipulate 387 * map_entry structures within a locked map without blocking trying 388 * to allocate a new vm_map_entry. 389 */ 390 int 391 vm_map_entry_reserve(int count) 392 { 393 struct globaldata *gd = mycpu; 394 vm_map_entry_t entry; 395 396 crit_enter(); 397 398 /* 399 * Make sure we have enough structures in gd_vme_base to handle 400 * the reservation request. 401 */ 402 while (gd->gd_vme_avail < count) { 403 entry = zalloc(mapentzone); 404 entry->next = gd->gd_vme_base; 405 gd->gd_vme_base = entry; 406 ++gd->gd_vme_avail; 407 } 408 gd->gd_vme_avail -= count; 409 crit_exit(); 410 return(count); 411 } 412 413 /* 414 * vm_map_entry_release: 415 * 416 * Releases previously reserved vm_map_entry structures that were not 417 * used. If we have too much junk in our per-cpu cache clean some of 418 * it out. 419 */ 420 void 421 vm_map_entry_release(int count) 422 { 423 struct globaldata *gd = mycpu; 424 vm_map_entry_t entry; 425 426 crit_enter(); 427 gd->gd_vme_avail += count; 428 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 429 entry = gd->gd_vme_base; 430 KKASSERT(entry != NULL); 431 gd->gd_vme_base = entry->next; 432 --gd->gd_vme_avail; 433 crit_exit(); 434 zfree(mapentzone, entry); 435 crit_enter(); 436 } 437 crit_exit(); 438 } 439 440 /* 441 * vm_map_entry_kreserve: 442 * 443 * Reserve map entry structures for use in kernel_map itself. These 444 * entries have *ALREADY* been reserved on a per-cpu basis when the map 445 * was inited. This function is used by zalloc() to avoid a recursion 446 * when zalloc() itself needs to allocate additional kernel memory. 447 * 448 * This function works like the normal reserve but does not load the 449 * vm_map_entry cache (because that would result in an infinite 450 * recursion). Note that gd_vme_avail may go negative. This is expected. 451 * 452 * Any caller of this function must be sure to renormalize after 453 * potentially eating entries to ensure that the reserve supply 454 * remains intact. 455 */ 456 int 457 vm_map_entry_kreserve(int count) 458 { 459 struct globaldata *gd = mycpu; 460 461 crit_enter(); 462 gd->gd_vme_avail -= count; 463 crit_exit(); 464 KASSERT(gd->gd_vme_base != NULL, ("no reserved entries left, gd_vme_avail = %d\n", gd->gd_vme_avail)); 465 return(count); 466 } 467 468 /* 469 * vm_map_entry_krelease: 470 * 471 * Release previously reserved map entries for kernel_map. We do not 472 * attempt to clean up like the normal release function as this would 473 * cause an unnecessary (but probably not fatal) deep procedure call. 474 */ 475 void 476 vm_map_entry_krelease(int count) 477 { 478 struct globaldata *gd = mycpu; 479 480 crit_enter(); 481 gd->gd_vme_avail += count; 482 crit_exit(); 483 } 484 485 /* 486 * vm_map_entry_create: [ internal use only ] 487 * 488 * Allocates a VM map entry for insertion. No entry fields are filled 489 * in. 490 * 491 * This routine may be called from an interrupt thread but not a FAST 492 * interrupt. This routine may recurse the map lock. 493 */ 494 static vm_map_entry_t 495 vm_map_entry_create(vm_map_t map, int *countp) 496 { 497 struct globaldata *gd = mycpu; 498 vm_map_entry_t entry; 499 500 KKASSERT(*countp > 0); 501 --*countp; 502 crit_enter(); 503 entry = gd->gd_vme_base; 504 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 505 gd->gd_vme_base = entry->next; 506 crit_exit(); 507 return(entry); 508 } 509 510 /* 511 * vm_map_entry_dispose: [ internal use only ] 512 * 513 * Dispose of a vm_map_entry that is no longer being referenced. This 514 * function may be called from an interrupt. 515 */ 516 static void 517 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 518 { 519 struct globaldata *gd = mycpu; 520 521 KKASSERT(map->hint != entry); 522 KKASSERT(map->first_free != entry); 523 524 ++*countp; 525 crit_enter(); 526 entry->next = gd->gd_vme_base; 527 gd->gd_vme_base = entry; 528 crit_exit(); 529 } 530 531 532 /* 533 * vm_map_entry_{un,}link: 534 * 535 * Insert/remove entries from maps. 536 */ 537 static __inline void 538 vm_map_entry_link(vm_map_t map, 539 vm_map_entry_t after_where, 540 vm_map_entry_t entry) 541 { 542 map->nentries++; 543 entry->prev = after_where; 544 entry->next = after_where->next; 545 entry->next->prev = entry; 546 after_where->next = entry; 547 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry)) 548 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry); 549 } 550 551 static __inline void 552 vm_map_entry_unlink(vm_map_t map, 553 vm_map_entry_t entry) 554 { 555 vm_map_entry_t prev; 556 vm_map_entry_t next; 557 558 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) 559 panic("vm_map_entry_unlink: attempt to mess with locked entry! %p", entry); 560 prev = entry->prev; 561 next = entry->next; 562 next->prev = prev; 563 prev->next = next; 564 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry); 565 map->nentries--; 566 } 567 568 /* 569 * vm_map_lookup_entry: [ internal use only ] 570 * 571 * Finds the map entry containing (or 572 * immediately preceding) the specified address 573 * in the given map; the entry is returned 574 * in the "entry" parameter. The boolean 575 * result indicates whether the address is 576 * actually contained in the map. 577 */ 578 boolean_t 579 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, 580 vm_map_entry_t *entry /* OUT */) 581 { 582 vm_map_entry_t tmp; 583 vm_map_entry_t last; 584 585 #if 0 586 /* 587 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive 588 * the hint code with the red-black lookup meets with system crashes 589 * and lockups. We do not yet know why. 590 * 591 * It is possible that the problem is related to the setting 592 * of the hint during map_entry deletion, in the code specified 593 * at the GGG comment later on in this file. 594 */ 595 /* 596 * Quickly check the cached hint, there's a good chance of a match. 597 */ 598 if (map->hint != &map->header) { 599 tmp = map->hint; 600 if (address >= tmp->start && address < tmp->end) { 601 *entry = tmp; 602 return(TRUE); 603 } 604 } 605 #endif 606 607 /* 608 * Locate the record from the top of the tree. 'last' tracks the 609 * closest prior record and is returned if no match is found, which 610 * in binary tree terms means tracking the most recent right-branch 611 * taken. If there is no prior record, &map->header is returned. 612 */ 613 last = &map->header; 614 tmp = RB_ROOT(&map->rb_root); 615 616 while (tmp) { 617 if (address >= tmp->start) { 618 if (address < tmp->end) { 619 *entry = tmp; 620 map->hint = tmp; 621 return(TRUE); 622 } 623 last = tmp; 624 tmp = RB_RIGHT(tmp, rb_entry); 625 } else { 626 tmp = RB_LEFT(tmp, rb_entry); 627 } 628 } 629 *entry = last; 630 return (FALSE); 631 } 632 633 /* 634 * vm_map_insert: 635 * 636 * Inserts the given whole VM object into the target 637 * map at the specified address range. The object's 638 * size should match that of the address range. 639 * 640 * Requires that the map be locked, and leaves it so. Requires that 641 * sufficient vm_map_entry structures have been reserved and tracks 642 * the use via countp. 643 * 644 * If object is non-NULL, ref count must be bumped by caller 645 * prior to making call to account for the new entry. 646 */ 647 int 648 vm_map_insert(vm_map_t map, int *countp, 649 vm_object_t object, vm_ooffset_t offset, 650 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, 651 int cow) 652 { 653 vm_map_entry_t new_entry; 654 vm_map_entry_t prev_entry; 655 vm_map_entry_t temp_entry; 656 vm_eflags_t protoeflags; 657 658 /* 659 * Check that the start and end points are not bogus. 660 */ 661 662 if ((start < map->min_offset) || (end > map->max_offset) || 663 (start >= end)) 664 return (KERN_INVALID_ADDRESS); 665 666 /* 667 * Find the entry prior to the proposed starting address; if it's part 668 * of an existing entry, this range is bogus. 669 */ 670 671 if (vm_map_lookup_entry(map, start, &temp_entry)) 672 return (KERN_NO_SPACE); 673 674 prev_entry = temp_entry; 675 676 /* 677 * Assert that the next entry doesn't overlap the end point. 678 */ 679 680 if ((prev_entry->next != &map->header) && 681 (prev_entry->next->start < end)) 682 return (KERN_NO_SPACE); 683 684 protoeflags = 0; 685 686 if (cow & MAP_COPY_ON_WRITE) 687 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 688 689 if (cow & MAP_NOFAULT) { 690 protoeflags |= MAP_ENTRY_NOFAULT; 691 692 KASSERT(object == NULL, 693 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 694 } 695 if (cow & MAP_DISABLE_SYNCER) 696 protoeflags |= MAP_ENTRY_NOSYNC; 697 if (cow & MAP_DISABLE_COREDUMP) 698 protoeflags |= MAP_ENTRY_NOCOREDUMP; 699 700 if (object) { 701 /* 702 * When object is non-NULL, it could be shared with another 703 * process. We have to set or clear OBJ_ONEMAPPING 704 * appropriately. 705 */ 706 if ((object->ref_count > 1) || (object->shadow_count != 0)) { 707 vm_object_clear_flag(object, OBJ_ONEMAPPING); 708 } 709 } 710 else if ((prev_entry != &map->header) && 711 (prev_entry->eflags == protoeflags) && 712 (prev_entry->end == start) && 713 (prev_entry->wired_count == 0) && 714 ((prev_entry->object.vm_object == NULL) || 715 vm_object_coalesce(prev_entry->object.vm_object, 716 OFF_TO_IDX(prev_entry->offset), 717 (vm_size_t)(prev_entry->end - prev_entry->start), 718 (vm_size_t)(end - prev_entry->end)))) { 719 /* 720 * We were able to extend the object. Determine if we 721 * can extend the previous map entry to include the 722 * new range as well. 723 */ 724 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 725 (prev_entry->protection == prot) && 726 (prev_entry->max_protection == max)) { 727 map->size += (end - prev_entry->end); 728 prev_entry->end = end; 729 vm_map_simplify_entry(map, prev_entry, countp); 730 return (KERN_SUCCESS); 731 } 732 733 /* 734 * If we can extend the object but cannot extend the 735 * map entry, we have to create a new map entry. We 736 * must bump the ref count on the extended object to 737 * account for it. object may be NULL. 738 */ 739 object = prev_entry->object.vm_object; 740 offset = prev_entry->offset + 741 (prev_entry->end - prev_entry->start); 742 vm_object_reference(object); 743 } 744 745 /* 746 * NOTE: if conditionals fail, object can be NULL here. This occurs 747 * in things like the buffer map where we manage kva but do not manage 748 * backing objects. 749 */ 750 751 /* 752 * Create a new entry 753 */ 754 755 new_entry = vm_map_entry_create(map, countp); 756 new_entry->start = start; 757 new_entry->end = end; 758 759 new_entry->eflags = protoeflags; 760 new_entry->object.vm_object = object; 761 new_entry->offset = offset; 762 new_entry->avail_ssize = 0; 763 764 new_entry->inheritance = VM_INHERIT_DEFAULT; 765 new_entry->protection = prot; 766 new_entry->max_protection = max; 767 new_entry->wired_count = 0; 768 769 /* 770 * Insert the new entry into the list 771 */ 772 773 vm_map_entry_link(map, prev_entry, new_entry); 774 map->size += new_entry->end - new_entry->start; 775 776 /* 777 * Update the free space hint 778 */ 779 if ((map->first_free == prev_entry) && 780 (prev_entry->end >= new_entry->start)) { 781 map->first_free = new_entry; 782 } 783 784 #if 0 785 /* 786 * Temporarily removed to avoid MAP_STACK panic, due to 787 * MAP_STACK being a huge hack. Will be added back in 788 * when MAP_STACK (and the user stack mapping) is fixed. 789 */ 790 /* 791 * It may be possible to simplify the entry 792 */ 793 vm_map_simplify_entry(map, new_entry, countp); 794 #endif 795 796 if (cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) { 797 pmap_object_init_pt(map->pmap, start, prot, 798 object, OFF_TO_IDX(offset), end - start, 799 cow & MAP_PREFAULT_PARTIAL); 800 } 801 802 return (KERN_SUCCESS); 803 } 804 805 /* 806 * Find sufficient space for `length' bytes in the given map, starting at 807 * `start'. The map must be locked. Returns 0 on success, 1 on no space. 808 * 809 * This function will returned an arbitrarily aligned pointer. If no 810 * particular alignment is required you should pass align as 1. Note that 811 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 812 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 813 * argument. 814 * 815 * 'align' should be a power of 2 but is not required to be. 816 */ 817 int 818 vm_map_findspace( 819 vm_map_t map, 820 vm_offset_t start, 821 vm_size_t length, 822 vm_offset_t align, 823 vm_offset_t *addr) 824 { 825 vm_map_entry_t entry, next; 826 vm_offset_t end; 827 vm_offset_t align_mask; 828 829 if (start < map->min_offset) 830 start = map->min_offset; 831 if (start > map->max_offset) 832 return (1); 833 834 /* 835 * If the alignment is not a power of 2 we will have to use 836 * a mod/division, set align_mask to a special value. 837 */ 838 if ((align | (align - 1)) + 1 != (align << 1)) 839 align_mask = (vm_offset_t)-1; 840 else 841 align_mask = align - 1; 842 843 retry: 844 /* 845 * Look for the first possible address; if there's already something 846 * at this address, we have to start after it. 847 */ 848 if (start == map->min_offset) { 849 if ((entry = map->first_free) != &map->header) 850 start = entry->end; 851 } else { 852 vm_map_entry_t tmp; 853 854 if (vm_map_lookup_entry(map, start, &tmp)) 855 start = tmp->end; 856 entry = tmp; 857 } 858 859 /* 860 * Look through the rest of the map, trying to fit a new region in the 861 * gap between existing regions, or after the very last region. 862 */ 863 for (;; start = (entry = next)->end) { 864 /* 865 * Adjust the proposed start by the requested alignment, 866 * be sure that we didn't wrap the address. 867 */ 868 if (align_mask == (vm_offset_t)-1) 869 end = ((start + align - 1) / align) * align; 870 else 871 end = (start + align_mask) & ~align_mask; 872 if (end < start) 873 return (1); 874 start = end; 875 /* 876 * Find the end of the proposed new region. Be sure we didn't 877 * go beyond the end of the map, or wrap around the address. 878 * Then check to see if this is the last entry or if the 879 * proposed end fits in the gap between this and the next 880 * entry. 881 */ 882 end = start + length; 883 if (end > map->max_offset || end < start) 884 return (1); 885 next = entry->next; 886 if (next == &map->header || next->start >= end) 887 break; 888 } 889 map->hint = entry; 890 if (map == kernel_map) { 891 vm_offset_t ksize; 892 if ((ksize = round_page(start + length)) > kernel_vm_end) { 893 pmap_growkernel(ksize); 894 goto retry; 895 } 896 } 897 *addr = start; 898 return (0); 899 } 900 901 /* 902 * vm_map_find finds an unallocated region in the target address 903 * map with the given length. The search is defined to be 904 * first-fit from the specified address; the region found is 905 * returned in the same parameter. 906 * 907 * If object is non-NULL, ref count must be bumped by caller 908 * prior to making call to account for the new entry. 909 */ 910 int 911 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 912 vm_offset_t *addr, /* IN/OUT */ 913 vm_size_t length, boolean_t find_space, vm_prot_t prot, 914 vm_prot_t max, int cow) 915 { 916 vm_offset_t start; 917 int result; 918 int count; 919 920 start = *addr; 921 922 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 923 vm_map_lock(map); 924 if (find_space) { 925 if (vm_map_findspace(map, start, length, 1, addr)) { 926 vm_map_unlock(map); 927 vm_map_entry_release(count); 928 return (KERN_NO_SPACE); 929 } 930 start = *addr; 931 } 932 result = vm_map_insert(map, &count, object, offset, 933 start, start + length, prot, max, cow); 934 vm_map_unlock(map); 935 vm_map_entry_release(count); 936 937 return (result); 938 } 939 940 /* 941 * vm_map_simplify_entry: 942 * 943 * Simplify the given map entry by merging with either neighbor. This 944 * routine also has the ability to merge with both neighbors. 945 * 946 * The map must be locked. 947 * 948 * This routine guarentees that the passed entry remains valid (though 949 * possibly extended). When merging, this routine may delete one or 950 * both neighbors. No action is taken on entries which have their 951 * in-transition flag set. 952 */ 953 void 954 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 955 { 956 vm_map_entry_t next, prev; 957 vm_size_t prevsize, esize; 958 959 if (entry->eflags & (MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP)) { 960 ++mycpu->gd_cnt.v_intrans_coll; 961 return; 962 } 963 964 prev = entry->prev; 965 if (prev != &map->header) { 966 prevsize = prev->end - prev->start; 967 if ( (prev->end == entry->start) && 968 (prev->object.vm_object == entry->object.vm_object) && 969 (!prev->object.vm_object || 970 (prev->offset + prevsize == entry->offset)) && 971 (prev->eflags == entry->eflags) && 972 (prev->protection == entry->protection) && 973 (prev->max_protection == entry->max_protection) && 974 (prev->inheritance == entry->inheritance) && 975 (prev->wired_count == entry->wired_count)) { 976 if (map->first_free == prev) 977 map->first_free = entry; 978 if (map->hint == prev) 979 map->hint = entry; 980 vm_map_entry_unlink(map, prev); 981 entry->start = prev->start; 982 entry->offset = prev->offset; 983 if (prev->object.vm_object) 984 vm_object_deallocate(prev->object.vm_object); 985 vm_map_entry_dispose(map, prev, countp); 986 } 987 } 988 989 next = entry->next; 990 if (next != &map->header) { 991 esize = entry->end - entry->start; 992 if ((entry->end == next->start) && 993 (next->object.vm_object == entry->object.vm_object) && 994 (!entry->object.vm_object || 995 (entry->offset + esize == next->offset)) && 996 (next->eflags == entry->eflags) && 997 (next->protection == entry->protection) && 998 (next->max_protection == entry->max_protection) && 999 (next->inheritance == entry->inheritance) && 1000 (next->wired_count == entry->wired_count)) { 1001 if (map->first_free == next) 1002 map->first_free = entry; 1003 if (map->hint == next) 1004 map->hint = entry; 1005 vm_map_entry_unlink(map, next); 1006 entry->end = next->end; 1007 if (next->object.vm_object) 1008 vm_object_deallocate(next->object.vm_object); 1009 vm_map_entry_dispose(map, next, countp); 1010 } 1011 } 1012 } 1013 /* 1014 * vm_map_clip_start: [ internal use only ] 1015 * 1016 * Asserts that the given entry begins at or after 1017 * the specified address; if necessary, 1018 * it splits the entry into two. 1019 */ 1020 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1021 { \ 1022 if (startaddr > entry->start) \ 1023 _vm_map_clip_start(map, entry, startaddr, countp); \ 1024 } 1025 1026 /* 1027 * This routine is called only when it is known that 1028 * the entry must be split. 1029 */ 1030 static void 1031 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, int *countp) 1032 { 1033 vm_map_entry_t new_entry; 1034 1035 /* 1036 * Split off the front portion -- note that we must insert the new 1037 * entry BEFORE this one, so that this entry has the specified 1038 * starting address. 1039 */ 1040 1041 vm_map_simplify_entry(map, entry, countp); 1042 1043 /* 1044 * If there is no object backing this entry, we might as well create 1045 * one now. If we defer it, an object can get created after the map 1046 * is clipped, and individual objects will be created for the split-up 1047 * map. This is a bit of a hack, but is also about the best place to 1048 * put this improvement. 1049 */ 1050 1051 if (entry->object.vm_object == NULL && !map->system_map) { 1052 vm_object_t object; 1053 object = vm_object_allocate(OBJT_DEFAULT, 1054 atop(entry->end - entry->start)); 1055 entry->object.vm_object = object; 1056 entry->offset = 0; 1057 } 1058 1059 new_entry = vm_map_entry_create(map, countp); 1060 *new_entry = *entry; 1061 1062 new_entry->end = start; 1063 entry->offset += (start - entry->start); 1064 entry->start = start; 1065 1066 vm_map_entry_link(map, entry->prev, new_entry); 1067 1068 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 1069 vm_object_reference(new_entry->object.vm_object); 1070 } 1071 } 1072 1073 /* 1074 * vm_map_clip_end: [ internal use only ] 1075 * 1076 * Asserts that the given entry ends at or before 1077 * the specified address; if necessary, 1078 * it splits the entry into two. 1079 */ 1080 1081 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1082 { \ 1083 if (endaddr < entry->end) \ 1084 _vm_map_clip_end(map, entry, endaddr, countp); \ 1085 } 1086 1087 /* 1088 * This routine is called only when it is known that 1089 * the entry must be split. 1090 */ 1091 static void 1092 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, int *countp) 1093 { 1094 vm_map_entry_t new_entry; 1095 1096 /* 1097 * If there is no object backing this entry, we might as well create 1098 * one now. If we defer it, an object can get created after the map 1099 * is clipped, and individual objects will be created for the split-up 1100 * map. This is a bit of a hack, but is also about the best place to 1101 * put this improvement. 1102 */ 1103 1104 if (entry->object.vm_object == NULL && !map->system_map) { 1105 vm_object_t object; 1106 object = vm_object_allocate(OBJT_DEFAULT, 1107 atop(entry->end - entry->start)); 1108 entry->object.vm_object = object; 1109 entry->offset = 0; 1110 } 1111 1112 /* 1113 * Create a new entry and insert it AFTER the specified entry 1114 */ 1115 1116 new_entry = vm_map_entry_create(map, countp); 1117 *new_entry = *entry; 1118 1119 new_entry->start = entry->end = end; 1120 new_entry->offset += (end - entry->start); 1121 1122 vm_map_entry_link(map, entry, new_entry); 1123 1124 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 1125 vm_object_reference(new_entry->object.vm_object); 1126 } 1127 } 1128 1129 /* 1130 * VM_MAP_RANGE_CHECK: [ internal use only ] 1131 * 1132 * Asserts that the starting and ending region 1133 * addresses fall within the valid range of the map. 1134 */ 1135 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1136 { \ 1137 if (start < vm_map_min(map)) \ 1138 start = vm_map_min(map); \ 1139 if (end > vm_map_max(map)) \ 1140 end = vm_map_max(map); \ 1141 if (start > end) \ 1142 start = end; \ 1143 } 1144 1145 /* 1146 * vm_map_transition_wait: [ kernel use only ] 1147 * 1148 * Used to block when an in-transition collison occurs. The map 1149 * is unlocked for the sleep and relocked before the return. 1150 */ 1151 static 1152 void 1153 vm_map_transition_wait(vm_map_t map) 1154 { 1155 vm_map_unlock(map); 1156 tsleep(map, 0, "vment", 0); 1157 vm_map_lock(map); 1158 } 1159 1160 /* 1161 * CLIP_CHECK_BACK 1162 * CLIP_CHECK_FWD 1163 * 1164 * When we do blocking operations with the map lock held it is 1165 * possible that a clip might have occured on our in-transit entry, 1166 * requiring an adjustment to the entry in our loop. These macros 1167 * help the pageable and clip_range code deal with the case. The 1168 * conditional costs virtually nothing if no clipping has occured. 1169 */ 1170 1171 #define CLIP_CHECK_BACK(entry, save_start) \ 1172 do { \ 1173 while (entry->start != save_start) { \ 1174 entry = entry->prev; \ 1175 KASSERT(entry != &map->header, ("bad entry clip")); \ 1176 } \ 1177 } while(0) 1178 1179 #define CLIP_CHECK_FWD(entry, save_end) \ 1180 do { \ 1181 while (entry->end != save_end) { \ 1182 entry = entry->next; \ 1183 KASSERT(entry != &map->header, ("bad entry clip")); \ 1184 } \ 1185 } while(0) 1186 1187 1188 /* 1189 * vm_map_clip_range: [ kernel use only ] 1190 * 1191 * Clip the specified range and return the base entry. The 1192 * range may cover several entries starting at the returned base 1193 * and the first and last entry in the covering sequence will be 1194 * properly clipped to the requested start and end address. 1195 * 1196 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1197 * flag. 1198 * 1199 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1200 * covered by the requested range. 1201 * 1202 * The map must be exclusively locked on entry and will remain locked 1203 * on return. If no range exists or the range contains holes and you 1204 * specified that no holes were allowed, NULL will be returned. This 1205 * routine may temporarily unlock the map in order avoid a deadlock when 1206 * sleeping. 1207 */ 1208 static 1209 vm_map_entry_t 1210 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1211 int *countp, int flags) 1212 { 1213 vm_map_entry_t start_entry; 1214 vm_map_entry_t entry; 1215 1216 /* 1217 * Locate the entry and effect initial clipping. The in-transition 1218 * case does not occur very often so do not try to optimize it. 1219 */ 1220 again: 1221 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1222 return (NULL); 1223 entry = start_entry; 1224 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1225 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1226 ++mycpu->gd_cnt.v_intrans_coll; 1227 ++mycpu->gd_cnt.v_intrans_wait; 1228 vm_map_transition_wait(map); 1229 /* 1230 * entry and/or start_entry may have been clipped while 1231 * we slept, or may have gone away entirely. We have 1232 * to restart from the lookup. 1233 */ 1234 goto again; 1235 } 1236 /* 1237 * Since we hold an exclusive map lock we do not have to restart 1238 * after clipping, even though clipping may block in zalloc. 1239 */ 1240 vm_map_clip_start(map, entry, start, countp); 1241 vm_map_clip_end(map, entry, end, countp); 1242 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1243 1244 /* 1245 * Scan entries covered by the range. When working on the next 1246 * entry a restart need only re-loop on the current entry which 1247 * we have already locked, since 'next' may have changed. Also, 1248 * even though entry is safe, it may have been clipped so we 1249 * have to iterate forwards through the clip after sleeping. 1250 */ 1251 while (entry->next != &map->header && entry->next->start < end) { 1252 vm_map_entry_t next = entry->next; 1253 1254 if (flags & MAP_CLIP_NO_HOLES) { 1255 if (next->start > entry->end) { 1256 vm_map_unclip_range(map, start_entry, 1257 start, entry->end, countp, flags); 1258 return(NULL); 1259 } 1260 } 1261 1262 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1263 vm_offset_t save_end = entry->end; 1264 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1265 ++mycpu->gd_cnt.v_intrans_coll; 1266 ++mycpu->gd_cnt.v_intrans_wait; 1267 vm_map_transition_wait(map); 1268 1269 /* 1270 * clips might have occured while we blocked. 1271 */ 1272 CLIP_CHECK_FWD(entry, save_end); 1273 CLIP_CHECK_BACK(start_entry, start); 1274 continue; 1275 } 1276 /* 1277 * No restart necessary even though clip_end may block, we 1278 * are holding the map lock. 1279 */ 1280 vm_map_clip_end(map, next, end, countp); 1281 next->eflags |= MAP_ENTRY_IN_TRANSITION; 1282 entry = next; 1283 } 1284 if (flags & MAP_CLIP_NO_HOLES) { 1285 if (entry->end != end) { 1286 vm_map_unclip_range(map, start_entry, 1287 start, entry->end, countp, flags); 1288 return(NULL); 1289 } 1290 } 1291 return(start_entry); 1292 } 1293 1294 /* 1295 * vm_map_unclip_range: [ kernel use only ] 1296 * 1297 * Undo the effect of vm_map_clip_range(). You should pass the same 1298 * flags and the same range that you passed to vm_map_clip_range(). 1299 * This code will clear the in-transition flag on the entries and 1300 * wake up anyone waiting. This code will also simplify the sequence 1301 * and attempt to merge it with entries before and after the sequence. 1302 * 1303 * The map must be locked on entry and will remain locked on return. 1304 * 1305 * Note that you should also pass the start_entry returned by 1306 * vm_map_clip_range(). However, if you block between the two calls 1307 * with the map unlocked please be aware that the start_entry may 1308 * have been clipped and you may need to scan it backwards to find 1309 * the entry corresponding with the original start address. You are 1310 * responsible for this, vm_map_unclip_range() expects the correct 1311 * start_entry to be passed to it and will KASSERT otherwise. 1312 */ 1313 static 1314 void 1315 vm_map_unclip_range( 1316 vm_map_t map, 1317 vm_map_entry_t start_entry, 1318 vm_offset_t start, 1319 vm_offset_t end, 1320 int *countp, 1321 int flags) 1322 { 1323 vm_map_entry_t entry; 1324 1325 entry = start_entry; 1326 1327 KASSERT(entry->start == start, ("unclip_range: illegal base entry")); 1328 while (entry != &map->header && entry->start < end) { 1329 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, ("in-transition flag not set during unclip on: %p", entry)); 1330 KASSERT(entry->end <= end, ("unclip_range: tail wasn't clipped")); 1331 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 1332 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 1333 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 1334 wakeup(map); 1335 } 1336 entry = entry->next; 1337 } 1338 1339 /* 1340 * Simplification does not block so there is no restart case. 1341 */ 1342 entry = start_entry; 1343 while (entry != &map->header && entry->start < end) { 1344 vm_map_simplify_entry(map, entry, countp); 1345 entry = entry->next; 1346 } 1347 } 1348 1349 /* 1350 * vm_map_submap: [ kernel use only ] 1351 * 1352 * Mark the given range as handled by a subordinate map. 1353 * 1354 * This range must have been created with vm_map_find, 1355 * and no other operations may have been performed on this 1356 * range prior to calling vm_map_submap. 1357 * 1358 * Only a limited number of operations can be performed 1359 * within this rage after calling vm_map_submap: 1360 * vm_fault 1361 * [Don't try vm_map_copy!] 1362 * 1363 * To remove a submapping, one must first remove the 1364 * range from the superior map, and then destroy the 1365 * submap (if desired). [Better yet, don't try it.] 1366 */ 1367 int 1368 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 1369 { 1370 vm_map_entry_t entry; 1371 int result = KERN_INVALID_ARGUMENT; 1372 int count; 1373 1374 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1375 vm_map_lock(map); 1376 1377 VM_MAP_RANGE_CHECK(map, start, end); 1378 1379 if (vm_map_lookup_entry(map, start, &entry)) { 1380 vm_map_clip_start(map, entry, start, &count); 1381 } else { 1382 entry = entry->next; 1383 } 1384 1385 vm_map_clip_end(map, entry, end, &count); 1386 1387 if ((entry->start == start) && (entry->end == end) && 1388 ((entry->eflags & MAP_ENTRY_COW) == 0) && 1389 (entry->object.vm_object == NULL)) { 1390 entry->object.sub_map = submap; 1391 entry->eflags |= MAP_ENTRY_IS_SUB_MAP; 1392 result = KERN_SUCCESS; 1393 } 1394 vm_map_unlock(map); 1395 vm_map_entry_release(count); 1396 1397 return (result); 1398 } 1399 1400 /* 1401 * vm_map_protect: 1402 * 1403 * Sets the protection of the specified address 1404 * region in the target map. If "set_max" is 1405 * specified, the maximum protection is to be set; 1406 * otherwise, only the current protection is affected. 1407 */ 1408 int 1409 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 1410 vm_prot_t new_prot, boolean_t set_max) 1411 { 1412 vm_map_entry_t current; 1413 vm_map_entry_t entry; 1414 int count; 1415 1416 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1417 vm_map_lock(map); 1418 1419 VM_MAP_RANGE_CHECK(map, start, end); 1420 1421 if (vm_map_lookup_entry(map, start, &entry)) { 1422 vm_map_clip_start(map, entry, start, &count); 1423 } else { 1424 entry = entry->next; 1425 } 1426 1427 /* 1428 * Make a first pass to check for protection violations. 1429 */ 1430 1431 current = entry; 1432 while ((current != &map->header) && (current->start < end)) { 1433 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { 1434 vm_map_unlock(map); 1435 vm_map_entry_release(count); 1436 return (KERN_INVALID_ARGUMENT); 1437 } 1438 if ((new_prot & current->max_protection) != new_prot) { 1439 vm_map_unlock(map); 1440 vm_map_entry_release(count); 1441 return (KERN_PROTECTION_FAILURE); 1442 } 1443 current = current->next; 1444 } 1445 1446 /* 1447 * Go back and fix up protections. [Note that clipping is not 1448 * necessary the second time.] 1449 */ 1450 current = entry; 1451 1452 while ((current != &map->header) && (current->start < end)) { 1453 vm_prot_t old_prot; 1454 1455 vm_map_clip_end(map, current, end, &count); 1456 1457 old_prot = current->protection; 1458 if (set_max) 1459 current->protection = 1460 (current->max_protection = new_prot) & 1461 old_prot; 1462 else 1463 current->protection = new_prot; 1464 1465 /* 1466 * Update physical map if necessary. Worry about copy-on-write 1467 * here -- CHECK THIS XXX 1468 */ 1469 1470 if (current->protection != old_prot) { 1471 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 1472 VM_PROT_ALL) 1473 1474 pmap_protect(map->pmap, current->start, 1475 current->end, 1476 current->protection & MASK(current)); 1477 #undef MASK 1478 } 1479 1480 vm_map_simplify_entry(map, current, &count); 1481 1482 current = current->next; 1483 } 1484 1485 vm_map_unlock(map); 1486 vm_map_entry_release(count); 1487 return (KERN_SUCCESS); 1488 } 1489 1490 /* 1491 * vm_map_madvise: 1492 * 1493 * This routine traverses a processes map handling the madvise 1494 * system call. Advisories are classified as either those effecting 1495 * the vm_map_entry structure, or those effecting the underlying 1496 * objects. 1497 */ 1498 1499 int 1500 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, int behav) 1501 { 1502 vm_map_entry_t current, entry; 1503 int modify_map = 0; 1504 int count; 1505 1506 /* 1507 * Some madvise calls directly modify the vm_map_entry, in which case 1508 * we need to use an exclusive lock on the map and we need to perform 1509 * various clipping operations. Otherwise we only need a read-lock 1510 * on the map. 1511 */ 1512 1513 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1514 1515 switch(behav) { 1516 case MADV_NORMAL: 1517 case MADV_SEQUENTIAL: 1518 case MADV_RANDOM: 1519 case MADV_NOSYNC: 1520 case MADV_AUTOSYNC: 1521 case MADV_NOCORE: 1522 case MADV_CORE: 1523 modify_map = 1; 1524 vm_map_lock(map); 1525 break; 1526 case MADV_WILLNEED: 1527 case MADV_DONTNEED: 1528 case MADV_FREE: 1529 vm_map_lock_read(map); 1530 break; 1531 default: 1532 vm_map_entry_release(count); 1533 return (KERN_INVALID_ARGUMENT); 1534 } 1535 1536 /* 1537 * Locate starting entry and clip if necessary. 1538 */ 1539 1540 VM_MAP_RANGE_CHECK(map, start, end); 1541 1542 if (vm_map_lookup_entry(map, start, &entry)) { 1543 if (modify_map) 1544 vm_map_clip_start(map, entry, start, &count); 1545 } else { 1546 entry = entry->next; 1547 } 1548 1549 if (modify_map) { 1550 /* 1551 * madvise behaviors that are implemented in the vm_map_entry. 1552 * 1553 * We clip the vm_map_entry so that behavioral changes are 1554 * limited to the specified address range. 1555 */ 1556 for (current = entry; 1557 (current != &map->header) && (current->start < end); 1558 current = current->next 1559 ) { 1560 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) 1561 continue; 1562 1563 vm_map_clip_end(map, current, end, &count); 1564 1565 switch (behav) { 1566 case MADV_NORMAL: 1567 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 1568 break; 1569 case MADV_SEQUENTIAL: 1570 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 1571 break; 1572 case MADV_RANDOM: 1573 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 1574 break; 1575 case MADV_NOSYNC: 1576 current->eflags |= MAP_ENTRY_NOSYNC; 1577 break; 1578 case MADV_AUTOSYNC: 1579 current->eflags &= ~MAP_ENTRY_NOSYNC; 1580 break; 1581 case MADV_NOCORE: 1582 current->eflags |= MAP_ENTRY_NOCOREDUMP; 1583 break; 1584 case MADV_CORE: 1585 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 1586 break; 1587 default: 1588 break; 1589 } 1590 vm_map_simplify_entry(map, current, &count); 1591 } 1592 vm_map_unlock(map); 1593 } else { 1594 vm_pindex_t pindex; 1595 int count; 1596 1597 /* 1598 * madvise behaviors that are implemented in the underlying 1599 * vm_object. 1600 * 1601 * Since we don't clip the vm_map_entry, we have to clip 1602 * the vm_object pindex and count. 1603 */ 1604 for (current = entry; 1605 (current != &map->header) && (current->start < end); 1606 current = current->next 1607 ) { 1608 vm_offset_t useStart; 1609 1610 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) 1611 continue; 1612 1613 pindex = OFF_TO_IDX(current->offset); 1614 count = atop(current->end - current->start); 1615 useStart = current->start; 1616 1617 if (current->start < start) { 1618 pindex += atop(start - current->start); 1619 count -= atop(start - current->start); 1620 useStart = start; 1621 } 1622 if (current->end > end) 1623 count -= atop(current->end - end); 1624 1625 if (count <= 0) 1626 continue; 1627 1628 vm_object_madvise(current->object.vm_object, 1629 pindex, count, behav); 1630 if (behav == MADV_WILLNEED) { 1631 pmap_object_init_pt( 1632 map->pmap, 1633 useStart, 1634 current->protection, 1635 current->object.vm_object, 1636 pindex, 1637 (count << PAGE_SHIFT), 1638 MAP_PREFAULT_MADVISE 1639 ); 1640 } 1641 } 1642 vm_map_unlock_read(map); 1643 } 1644 vm_map_entry_release(count); 1645 return(0); 1646 } 1647 1648 1649 /* 1650 * vm_map_inherit: 1651 * 1652 * Sets the inheritance of the specified address 1653 * range in the target map. Inheritance 1654 * affects how the map will be shared with 1655 * child maps at the time of vm_map_fork. 1656 */ 1657 int 1658 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 1659 vm_inherit_t new_inheritance) 1660 { 1661 vm_map_entry_t entry; 1662 vm_map_entry_t temp_entry; 1663 int count; 1664 1665 switch (new_inheritance) { 1666 case VM_INHERIT_NONE: 1667 case VM_INHERIT_COPY: 1668 case VM_INHERIT_SHARE: 1669 break; 1670 default: 1671 return (KERN_INVALID_ARGUMENT); 1672 } 1673 1674 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1675 vm_map_lock(map); 1676 1677 VM_MAP_RANGE_CHECK(map, start, end); 1678 1679 if (vm_map_lookup_entry(map, start, &temp_entry)) { 1680 entry = temp_entry; 1681 vm_map_clip_start(map, entry, start, &count); 1682 } else 1683 entry = temp_entry->next; 1684 1685 while ((entry != &map->header) && (entry->start < end)) { 1686 vm_map_clip_end(map, entry, end, &count); 1687 1688 entry->inheritance = new_inheritance; 1689 1690 vm_map_simplify_entry(map, entry, &count); 1691 1692 entry = entry->next; 1693 } 1694 vm_map_unlock(map); 1695 vm_map_entry_release(count); 1696 return (KERN_SUCCESS); 1697 } 1698 1699 /* 1700 * Implement the semantics of mlock 1701 */ 1702 int 1703 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 1704 boolean_t new_pageable) 1705 { 1706 vm_map_entry_t entry; 1707 vm_map_entry_t start_entry; 1708 vm_offset_t end; 1709 int rv = KERN_SUCCESS; 1710 int count; 1711 1712 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1713 vm_map_lock(map); 1714 VM_MAP_RANGE_CHECK(map, start, real_end); 1715 end = real_end; 1716 1717 start_entry = vm_map_clip_range(map, start, end, &count, MAP_CLIP_NO_HOLES); 1718 if (start_entry == NULL) { 1719 vm_map_unlock(map); 1720 vm_map_entry_release(count); 1721 return (KERN_INVALID_ADDRESS); 1722 } 1723 1724 if (new_pageable == 0) { 1725 entry = start_entry; 1726 while ((entry != &map->header) && (entry->start < end)) { 1727 vm_offset_t save_start; 1728 vm_offset_t save_end; 1729 1730 /* 1731 * Already user wired or hard wired (trivial cases) 1732 */ 1733 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 1734 entry = entry->next; 1735 continue; 1736 } 1737 if (entry->wired_count != 0) { 1738 entry->wired_count++; 1739 entry->eflags |= MAP_ENTRY_USER_WIRED; 1740 entry = entry->next; 1741 continue; 1742 } 1743 1744 /* 1745 * A new wiring requires instantiation of appropriate 1746 * management structures and the faulting in of the 1747 * page. 1748 */ 1749 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 1750 int copyflag = entry->eflags & MAP_ENTRY_NEEDS_COPY; 1751 if (copyflag && ((entry->protection & VM_PROT_WRITE) != 0)) { 1752 1753 vm_object_shadow(&entry->object.vm_object, 1754 &entry->offset, 1755 atop(entry->end - entry->start)); 1756 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 1757 1758 } else if (entry->object.vm_object == NULL && 1759 !map->system_map) { 1760 1761 entry->object.vm_object = 1762 vm_object_allocate(OBJT_DEFAULT, 1763 atop(entry->end - entry->start)); 1764 entry->offset = (vm_offset_t) 0; 1765 1766 } 1767 } 1768 entry->wired_count++; 1769 entry->eflags |= MAP_ENTRY_USER_WIRED; 1770 1771 /* 1772 * Now fault in the area. Note that vm_fault_wire() 1773 * may release the map lock temporarily, it will be 1774 * relocked on return. The in-transition 1775 * flag protects the entries. 1776 */ 1777 save_start = entry->start; 1778 save_end = entry->end; 1779 rv = vm_fault_wire(map, entry, TRUE); 1780 if (rv) { 1781 CLIP_CHECK_BACK(entry, save_start); 1782 for (;;) { 1783 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 1784 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 1785 entry->wired_count = 0; 1786 if (entry->end == save_end) 1787 break; 1788 entry = entry->next; 1789 KASSERT(entry != &map->header, ("bad entry clip during backout")); 1790 } 1791 end = save_start; /* unwire the rest */ 1792 break; 1793 } 1794 /* 1795 * note that even though the entry might have been 1796 * clipped, the USER_WIRED flag we set prevents 1797 * duplication so we do not have to do a 1798 * clip check. 1799 */ 1800 entry = entry->next; 1801 } 1802 1803 /* 1804 * If we failed fall through to the unwiring section to 1805 * unwire what we had wired so far. 'end' has already 1806 * been adjusted. 1807 */ 1808 if (rv) 1809 new_pageable = 1; 1810 1811 /* 1812 * start_entry might have been clipped if we unlocked the 1813 * map and blocked. No matter how clipped it has gotten 1814 * there should be a fragment that is on our start boundary. 1815 */ 1816 CLIP_CHECK_BACK(start_entry, start); 1817 } 1818 1819 /* 1820 * Deal with the unwiring case. 1821 */ 1822 if (new_pageable) { 1823 /* 1824 * This is the unwiring case. We must first ensure that the 1825 * range to be unwired is really wired down. We know there 1826 * are no holes. 1827 */ 1828 entry = start_entry; 1829 while ((entry != &map->header) && (entry->start < end)) { 1830 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 1831 rv = KERN_INVALID_ARGUMENT; 1832 goto done; 1833 } 1834 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry)); 1835 entry = entry->next; 1836 } 1837 1838 /* 1839 * Now decrement the wiring count for each region. If a region 1840 * becomes completely unwired, unwire its physical pages and 1841 * mappings. 1842 */ 1843 /* 1844 * The map entries are processed in a loop, checking to 1845 * make sure the entry is wired and asserting it has a wired 1846 * count. However, another loop was inserted more-or-less in 1847 * the middle of the unwiring path. This loop picks up the 1848 * "entry" loop variable from the first loop without first 1849 * setting it to start_entry. Naturally, the secound loop 1850 * is never entered and the pages backing the entries are 1851 * never unwired. This can lead to a leak of wired pages. 1852 */ 1853 entry = start_entry; 1854 while ((entry != &map->header) && (entry->start < end)) { 1855 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 1856 ("expected USER_WIRED on entry %p", entry)); 1857 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 1858 entry->wired_count--; 1859 if (entry->wired_count == 0) 1860 vm_fault_unwire(map, entry); 1861 entry = entry->next; 1862 } 1863 } 1864 done: 1865 vm_map_unclip_range(map, start_entry, start, real_end, &count, 1866 MAP_CLIP_NO_HOLES); 1867 map->timestamp++; 1868 vm_map_unlock(map); 1869 vm_map_entry_release(count); 1870 return (rv); 1871 } 1872 1873 /* 1874 * vm_map_wire: 1875 * 1876 * Sets the pageability of the specified address 1877 * range in the target map. Regions specified 1878 * as not pageable require locked-down physical 1879 * memory and physical page maps. 1880 * 1881 * The map must not be locked, but a reference 1882 * must remain to the map throughout the call. 1883 * 1884 * This function may be called via the zalloc path and must properly 1885 * reserve map entries for kernel_map. 1886 */ 1887 int 1888 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags) 1889 { 1890 vm_map_entry_t entry; 1891 vm_map_entry_t start_entry; 1892 vm_offset_t end; 1893 int rv = KERN_SUCCESS; 1894 int count; 1895 1896 if (kmflags & KM_KRESERVE) 1897 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 1898 else 1899 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1900 vm_map_lock(map); 1901 VM_MAP_RANGE_CHECK(map, start, real_end); 1902 end = real_end; 1903 1904 start_entry = vm_map_clip_range(map, start, end, &count, MAP_CLIP_NO_HOLES); 1905 if (start_entry == NULL) { 1906 vm_map_unlock(map); 1907 rv = KERN_INVALID_ADDRESS; 1908 goto failure; 1909 } 1910 if ((kmflags & KM_PAGEABLE) == 0) { 1911 /* 1912 * Wiring. 1913 * 1914 * 1. Holding the write lock, we create any shadow or zero-fill 1915 * objects that need to be created. Then we clip each map 1916 * entry to the region to be wired and increment its wiring 1917 * count. We create objects before clipping the map entries 1918 * to avoid object proliferation. 1919 * 1920 * 2. We downgrade to a read lock, and call vm_fault_wire to 1921 * fault in the pages for any newly wired area (wired_count is 1922 * 1). 1923 * 1924 * Downgrading to a read lock for vm_fault_wire avoids a 1925 * possible deadlock with another process that may have faulted 1926 * on one of the pages to be wired (it would mark the page busy, 1927 * blocking us, then in turn block on the map lock that we 1928 * hold). Because of problems in the recursive lock package, 1929 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 1930 * any actions that require the write lock must be done 1931 * beforehand. Because we keep the read lock on the map, the 1932 * copy-on-write status of the entries we modify here cannot 1933 * change. 1934 */ 1935 1936 entry = start_entry; 1937 while ((entry != &map->header) && (entry->start < end)) { 1938 /* 1939 * Trivial case if the entry is already wired 1940 */ 1941 if (entry->wired_count) { 1942 entry->wired_count++; 1943 entry = entry->next; 1944 continue; 1945 } 1946 1947 /* 1948 * The entry is being newly wired, we have to setup 1949 * appropriate management structures. A shadow 1950 * object is required for a copy-on-write region, 1951 * or a normal object for a zero-fill region. We 1952 * do not have to do this for entries that point to sub 1953 * maps because we won't hold the lock on the sub map. 1954 */ 1955 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 1956 int copyflag = entry->eflags & MAP_ENTRY_NEEDS_COPY; 1957 if (copyflag && 1958 ((entry->protection & VM_PROT_WRITE) != 0)) { 1959 1960 vm_object_shadow(&entry->object.vm_object, 1961 &entry->offset, 1962 atop(entry->end - entry->start)); 1963 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 1964 } else if (entry->object.vm_object == NULL && 1965 !map->system_map) { 1966 entry->object.vm_object = 1967 vm_object_allocate(OBJT_DEFAULT, 1968 atop(entry->end - entry->start)); 1969 entry->offset = (vm_offset_t) 0; 1970 } 1971 } 1972 1973 entry->wired_count++; 1974 entry = entry->next; 1975 } 1976 1977 /* 1978 * Pass 2. 1979 */ 1980 1981 /* 1982 * HACK HACK HACK HACK 1983 * 1984 * Unlock the map to avoid deadlocks. The in-transit flag 1985 * protects us from most changes but note that 1986 * clipping may still occur. To prevent clipping from 1987 * occuring after the unlock, except for when we are 1988 * blocking in vm_fault_wire, we must run in a critical 1989 * section, otherwise our accesses to entry->start and 1990 * entry->end could be corrupted. We have to enter the 1991 * critical section prior to unlocking so start_entry does 1992 * not change out from under us at the very beginning of the 1993 * loop. 1994 * 1995 * HACK HACK HACK HACK 1996 */ 1997 1998 crit_enter(); 1999 2000 entry = start_entry; 2001 while (entry != &map->header && entry->start < end) { 2002 /* 2003 * If vm_fault_wire fails for any page we need to undo 2004 * what has been done. We decrement the wiring count 2005 * for those pages which have not yet been wired (now) 2006 * and unwire those that have (later). 2007 */ 2008 vm_offset_t save_start = entry->start; 2009 vm_offset_t save_end = entry->end; 2010 2011 if (entry->wired_count == 1) 2012 rv = vm_fault_wire(map, entry, FALSE); 2013 if (rv) { 2014 CLIP_CHECK_BACK(entry, save_start); 2015 for (;;) { 2016 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly")); 2017 entry->wired_count = 0; 2018 if (entry->end == save_end) 2019 break; 2020 entry = entry->next; 2021 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2022 } 2023 end = save_start; 2024 break; 2025 } 2026 CLIP_CHECK_FWD(entry, save_end); 2027 entry = entry->next; 2028 } 2029 crit_exit(); 2030 2031 /* 2032 * If a failure occured undo everything by falling through 2033 * to the unwiring code. 'end' has already been adjusted 2034 * appropriately. 2035 */ 2036 if (rv) 2037 kmflags |= KM_PAGEABLE; 2038 2039 /* 2040 * start_entry is still IN_TRANSITION but may have been 2041 * clipped since vm_fault_wire() unlocks and relocks the 2042 * map. No matter how clipped it has gotten there should 2043 * be a fragment that is on our start boundary. 2044 */ 2045 CLIP_CHECK_BACK(start_entry, start); 2046 } 2047 2048 if (kmflags & KM_PAGEABLE) { 2049 /* 2050 * This is the unwiring case. We must first ensure that the 2051 * range to be unwired is really wired down. We know there 2052 * are no holes. 2053 */ 2054 entry = start_entry; 2055 while ((entry != &map->header) && (entry->start < end)) { 2056 if (entry->wired_count == 0) { 2057 rv = KERN_INVALID_ARGUMENT; 2058 goto done; 2059 } 2060 entry = entry->next; 2061 } 2062 2063 /* 2064 * Now decrement the wiring count for each region. If a region 2065 * becomes completely unwired, unwire its physical pages and 2066 * mappings. 2067 */ 2068 entry = start_entry; 2069 while ((entry != &map->header) && (entry->start < end)) { 2070 entry->wired_count--; 2071 if (entry->wired_count == 0) 2072 vm_fault_unwire(map, entry); 2073 entry = entry->next; 2074 } 2075 } 2076 done: 2077 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2078 MAP_CLIP_NO_HOLES); 2079 map->timestamp++; 2080 vm_map_unlock(map); 2081 failure: 2082 if (kmflags & KM_KRESERVE) 2083 vm_map_entry_krelease(count); 2084 else 2085 vm_map_entry_release(count); 2086 return (rv); 2087 } 2088 2089 /* 2090 * vm_map_set_wired_quick() 2091 * 2092 * Mark a newly allocated address range as wired but do not fault in 2093 * the pages. The caller is expected to load the pages into the object. 2094 * 2095 * The map must be locked on entry and will remain locked on return. 2096 */ 2097 void 2098 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, int *countp) 2099 { 2100 vm_map_entry_t scan; 2101 vm_map_entry_t entry; 2102 2103 entry = vm_map_clip_range(map, addr, addr + size, countp, MAP_CLIP_NO_HOLES); 2104 for (scan = entry; scan != &map->header && scan->start < addr + size; scan = scan->next) { 2105 KKASSERT(entry->wired_count == 0); 2106 entry->wired_count = 1; 2107 } 2108 vm_map_unclip_range(map, entry, addr, addr + size, countp, MAP_CLIP_NO_HOLES); 2109 } 2110 2111 /* 2112 * vm_map_clean 2113 * 2114 * Push any dirty cached pages in the address range to their pager. 2115 * If syncio is TRUE, dirty pages are written synchronously. 2116 * If invalidate is TRUE, any cached pages are freed as well. 2117 * 2118 * Returns an error if any part of the specified range is not mapped. 2119 */ 2120 int 2121 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, boolean_t syncio, 2122 boolean_t invalidate) 2123 { 2124 vm_map_entry_t current; 2125 vm_map_entry_t entry; 2126 vm_size_t size; 2127 vm_object_t object; 2128 vm_ooffset_t offset; 2129 2130 vm_map_lock_read(map); 2131 VM_MAP_RANGE_CHECK(map, start, end); 2132 if (!vm_map_lookup_entry(map, start, &entry)) { 2133 vm_map_unlock_read(map); 2134 return (KERN_INVALID_ADDRESS); 2135 } 2136 /* 2137 * Make a first pass to check for holes. 2138 */ 2139 for (current = entry; current->start < end; current = current->next) { 2140 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { 2141 vm_map_unlock_read(map); 2142 return (KERN_INVALID_ARGUMENT); 2143 } 2144 if (end > current->end && 2145 (current->next == &map->header || 2146 current->end != current->next->start)) { 2147 vm_map_unlock_read(map); 2148 return (KERN_INVALID_ADDRESS); 2149 } 2150 } 2151 2152 if (invalidate) 2153 pmap_remove(vm_map_pmap(map), start, end); 2154 /* 2155 * Make a second pass, cleaning/uncaching pages from the indicated 2156 * objects as we go. 2157 */ 2158 for (current = entry; current->start < end; current = current->next) { 2159 offset = current->offset + (start - current->start); 2160 size = (end <= current->end ? end : current->end) - start; 2161 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { 2162 vm_map_t smap; 2163 vm_map_entry_t tentry; 2164 vm_size_t tsize; 2165 2166 smap = current->object.sub_map; 2167 vm_map_lock_read(smap); 2168 vm_map_lookup_entry(smap, offset, &tentry); 2169 tsize = tentry->end - offset; 2170 if (tsize < size) 2171 size = tsize; 2172 object = tentry->object.vm_object; 2173 offset = tentry->offset + (offset - tentry->start); 2174 vm_map_unlock_read(smap); 2175 } else { 2176 object = current->object.vm_object; 2177 } 2178 /* 2179 * Note that there is absolutely no sense in writing out 2180 * anonymous objects, so we track down the vnode object 2181 * to write out. 2182 * We invalidate (remove) all pages from the address space 2183 * anyway, for semantic correctness. 2184 * 2185 * note: certain anonymous maps, such as MAP_NOSYNC maps, 2186 * may start out with a NULL object. 2187 */ 2188 while (object && object->backing_object) { 2189 offset += object->backing_object_offset; 2190 object = object->backing_object; 2191 if (object->size < OFF_TO_IDX( offset + size)) 2192 size = IDX_TO_OFF(object->size) - offset; 2193 } 2194 if (object && (object->type == OBJT_VNODE) && 2195 (current->protection & VM_PROT_WRITE)) { 2196 /* 2197 * Flush pages if writing is allowed, invalidate them 2198 * if invalidation requested. Pages undergoing I/O 2199 * will be ignored by vm_object_page_remove(). 2200 * 2201 * We cannot lock the vnode and then wait for paging 2202 * to complete without deadlocking against vm_fault. 2203 * Instead we simply call vm_object_page_remove() and 2204 * allow it to block internally on a page-by-page 2205 * basis when it encounters pages undergoing async 2206 * I/O. 2207 */ 2208 int flags; 2209 2210 vm_object_reference(object); 2211 vn_lock(object->handle, 2212 LK_EXCLUSIVE | LK_RETRY, curthread); 2213 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 2214 flags |= invalidate ? OBJPC_INVAL : 0; 2215 vm_object_page_clean(object, 2216 OFF_TO_IDX(offset), 2217 OFF_TO_IDX(offset + size + PAGE_MASK), 2218 flags); 2219 VOP_UNLOCK(((struct vnode *)object->handle), 2220 0, curthread); 2221 vm_object_deallocate(object); 2222 } 2223 if (object && invalidate && 2224 ((object->type == OBJT_VNODE) || 2225 (object->type == OBJT_DEVICE))) { 2226 int clean_only = 2227 (object->type == OBJT_DEVICE) ? FALSE : TRUE; 2228 vm_object_reference(object); 2229 vm_object_page_remove(object, 2230 OFF_TO_IDX(offset), 2231 OFF_TO_IDX(offset + size + PAGE_MASK), 2232 clean_only); 2233 vm_object_deallocate(object); 2234 } 2235 start += size; 2236 } 2237 2238 vm_map_unlock_read(map); 2239 return (KERN_SUCCESS); 2240 } 2241 2242 /* 2243 * vm_map_entry_unwire: [ internal use only ] 2244 * 2245 * Make the region specified by this entry pageable. 2246 * 2247 * The map in question should be locked. 2248 * [This is the reason for this routine's existence.] 2249 */ 2250 static void 2251 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 2252 { 2253 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2254 entry->wired_count = 0; 2255 vm_fault_unwire(map, entry); 2256 } 2257 2258 /* 2259 * vm_map_entry_delete: [ internal use only ] 2260 * 2261 * Deallocate the given entry from the target map. 2262 */ 2263 static void 2264 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 2265 { 2266 vm_map_entry_unlink(map, entry); 2267 map->size -= entry->end - entry->start; 2268 2269 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2270 vm_object_deallocate(entry->object.vm_object); 2271 } 2272 2273 vm_map_entry_dispose(map, entry, countp); 2274 } 2275 2276 /* 2277 * vm_map_delete: [ internal use only ] 2278 * 2279 * Deallocates the given address range from the target 2280 * map. 2281 */ 2282 int 2283 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 2284 { 2285 vm_object_t object; 2286 vm_map_entry_t entry; 2287 vm_map_entry_t first_entry; 2288 2289 again: 2290 /* 2291 * Find the start of the region, and clip it. Set entry to point 2292 * at the first record containing the requested address or, if no 2293 * such record exists, the next record with a greater address. The 2294 * loop will run from this point until a record beyond the termination 2295 * address is encountered. 2296 * 2297 * map->hint must be adjusted to not point to anything we delete, 2298 * so set it to the entry prior to the one being deleted. 2299 * 2300 * GGG see other GGG comment. 2301 */ 2302 if (vm_map_lookup_entry(map, start, &first_entry)) { 2303 entry = first_entry; 2304 vm_map_clip_start(map, entry, start, countp); 2305 map->hint = entry->prev; /* possible problem XXX */ 2306 } else { 2307 map->hint = first_entry; /* possible problem XXX */ 2308 entry = first_entry->next; 2309 } 2310 2311 /* 2312 * If a hole opens up prior to the current first_free then 2313 * adjust first_free. As with map->hint, map->first_free 2314 * cannot be left set to anything we might delete. 2315 */ 2316 if (entry == &map->header) { 2317 map->first_free = &map->header; 2318 } else if (map->first_free->start >= start) { 2319 map->first_free = entry->prev; 2320 } 2321 2322 /* 2323 * Step through all entries in this region 2324 */ 2325 2326 while ((entry != &map->header) && (entry->start < end)) { 2327 vm_map_entry_t next; 2328 vm_offset_t s, e; 2329 vm_pindex_t offidxstart, offidxend, count; 2330 2331 /* 2332 * If we hit an in-transition entry we have to sleep and 2333 * retry. It's easier (and not really slower) to just retry 2334 * since this case occurs so rarely and the hint is already 2335 * pointing at the right place. We have to reset the 2336 * start offset so as not to accidently delete an entry 2337 * another process just created in vacated space. 2338 */ 2339 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 2340 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2341 start = entry->start; 2342 ++mycpu->gd_cnt.v_intrans_coll; 2343 ++mycpu->gd_cnt.v_intrans_wait; 2344 vm_map_transition_wait(map); 2345 goto again; 2346 } 2347 vm_map_clip_end(map, entry, end, countp); 2348 2349 s = entry->start; 2350 e = entry->end; 2351 next = entry->next; 2352 2353 offidxstart = OFF_TO_IDX(entry->offset); 2354 count = OFF_TO_IDX(e - s); 2355 object = entry->object.vm_object; 2356 2357 /* 2358 * Unwire before removing addresses from the pmap; otherwise, 2359 * unwiring will put the entries back in the pmap. 2360 */ 2361 if (entry->wired_count != 0) 2362 vm_map_entry_unwire(map, entry); 2363 2364 offidxend = offidxstart + count; 2365 2366 if ((object == kernel_object) || (object == kmem_object)) { 2367 vm_object_page_remove(object, offidxstart, offidxend, FALSE); 2368 } else { 2369 pmap_remove(map->pmap, s, e); 2370 if (object != NULL && 2371 object->ref_count != 1 && 2372 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING && 2373 (object->type == OBJT_DEFAULT || object->type == OBJT_SWAP)) { 2374 vm_object_collapse(object); 2375 vm_object_page_remove(object, offidxstart, offidxend, FALSE); 2376 if (object->type == OBJT_SWAP) { 2377 swap_pager_freespace(object, offidxstart, count); 2378 } 2379 if (offidxend >= object->size && 2380 offidxstart < object->size) { 2381 object->size = offidxstart; 2382 } 2383 } 2384 } 2385 2386 /* 2387 * Delete the entry (which may delete the object) only after 2388 * removing all pmap entries pointing to its pages. 2389 * (Otherwise, its page frames may be reallocated, and any 2390 * modify bits will be set in the wrong object!) 2391 */ 2392 vm_map_entry_delete(map, entry, countp); 2393 entry = next; 2394 } 2395 return (KERN_SUCCESS); 2396 } 2397 2398 /* 2399 * vm_map_remove: 2400 * 2401 * Remove the given address range from the target map. 2402 * This is the exported form of vm_map_delete. 2403 */ 2404 int 2405 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 2406 { 2407 int result; 2408 int count; 2409 2410 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2411 vm_map_lock(map); 2412 VM_MAP_RANGE_CHECK(map, start, end); 2413 result = vm_map_delete(map, start, end, &count); 2414 vm_map_unlock(map); 2415 vm_map_entry_release(count); 2416 2417 return (result); 2418 } 2419 2420 /* 2421 * vm_map_check_protection: 2422 * 2423 * Assert that the target map allows the specified 2424 * privilege on the entire address region given. 2425 * The entire region must be allocated. 2426 */ 2427 boolean_t 2428 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 2429 vm_prot_t protection) 2430 { 2431 vm_map_entry_t entry; 2432 vm_map_entry_t tmp_entry; 2433 2434 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 2435 return (FALSE); 2436 } 2437 entry = tmp_entry; 2438 2439 while (start < end) { 2440 if (entry == &map->header) { 2441 return (FALSE); 2442 } 2443 /* 2444 * No holes allowed! 2445 */ 2446 2447 if (start < entry->start) { 2448 return (FALSE); 2449 } 2450 /* 2451 * Check protection associated with entry. 2452 */ 2453 2454 if ((entry->protection & protection) != protection) { 2455 return (FALSE); 2456 } 2457 /* go to next entry */ 2458 2459 start = entry->end; 2460 entry = entry->next; 2461 } 2462 return (TRUE); 2463 } 2464 2465 /* 2466 * Split the pages in a map entry into a new object. This affords 2467 * easier removal of unused pages, and keeps object inheritance from 2468 * being a negative impact on memory usage. 2469 */ 2470 static void 2471 vm_map_split(vm_map_entry_t entry) 2472 { 2473 vm_page_t m; 2474 vm_object_t orig_object, new_object, source; 2475 vm_offset_t s, e; 2476 vm_pindex_t offidxstart, offidxend, idx; 2477 vm_size_t size; 2478 vm_ooffset_t offset; 2479 2480 orig_object = entry->object.vm_object; 2481 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 2482 return; 2483 if (orig_object->ref_count <= 1) 2484 return; 2485 2486 offset = entry->offset; 2487 s = entry->start; 2488 e = entry->end; 2489 2490 offidxstart = OFF_TO_IDX(offset); 2491 offidxend = offidxstart + OFF_TO_IDX(e - s); 2492 size = offidxend - offidxstart; 2493 2494 new_object = vm_pager_allocate(orig_object->type, 2495 NULL, IDX_TO_OFF(size), VM_PROT_ALL, 0LL); 2496 if (new_object == NULL) 2497 return; 2498 2499 source = orig_object->backing_object; 2500 if (source != NULL) { 2501 vm_object_reference(source); /* Referenced by new_object */ 2502 LIST_INSERT_HEAD(&source->shadow_head, 2503 new_object, shadow_list); 2504 vm_object_clear_flag(source, OBJ_ONEMAPPING); 2505 new_object->backing_object_offset = 2506 orig_object->backing_object_offset + IDX_TO_OFF(offidxstart); 2507 new_object->backing_object = source; 2508 source->shadow_count++; 2509 source->generation++; 2510 } 2511 2512 for (idx = 0; idx < size; idx++) { 2513 vm_page_t m; 2514 2515 /* 2516 * A critical section is required to avoid a race between 2517 * the lookup and an interrupt/unbusy/free and our busy 2518 * check. 2519 */ 2520 crit_enter(); 2521 retry: 2522 m = vm_page_lookup(orig_object, offidxstart + idx); 2523 if (m == NULL) { 2524 crit_exit(); 2525 continue; 2526 } 2527 2528 /* 2529 * We must wait for pending I/O to complete before we can 2530 * rename the page. 2531 * 2532 * We do not have to VM_PROT_NONE the page as mappings should 2533 * not be changed by this operation. 2534 */ 2535 if (vm_page_sleep_busy(m, TRUE, "spltwt")) 2536 goto retry; 2537 vm_page_busy(m); 2538 vm_page_rename(m, new_object, idx); 2539 /* page automatically made dirty by rename and cache handled */ 2540 vm_page_busy(m); 2541 crit_exit(); 2542 } 2543 2544 if (orig_object->type == OBJT_SWAP) { 2545 vm_object_pip_add(orig_object, 1); 2546 /* 2547 * copy orig_object pages into new_object 2548 * and destroy unneeded pages in 2549 * shadow object. 2550 */ 2551 swap_pager_copy(orig_object, new_object, offidxstart, 0); 2552 vm_object_pip_wakeup(orig_object); 2553 } 2554 2555 /* 2556 * Wakeup the pages we played with. No spl protection is needed 2557 * for a simple wakeup. 2558 */ 2559 for (idx = 0; idx < size; idx++) { 2560 m = vm_page_lookup(new_object, idx); 2561 if (m) 2562 vm_page_wakeup(m); 2563 } 2564 2565 entry->object.vm_object = new_object; 2566 entry->offset = 0LL; 2567 vm_object_deallocate(orig_object); 2568 } 2569 2570 /* 2571 * vm_map_copy_entry: 2572 * 2573 * Copies the contents of the source entry to the destination 2574 * entry. The entries *must* be aligned properly. 2575 */ 2576 static void 2577 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 2578 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 2579 { 2580 vm_object_t src_object; 2581 2582 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) 2583 return; 2584 2585 if (src_entry->wired_count == 0) { 2586 2587 /* 2588 * If the source entry is marked needs_copy, it is already 2589 * write-protected. 2590 */ 2591 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 2592 pmap_protect(src_map->pmap, 2593 src_entry->start, 2594 src_entry->end, 2595 src_entry->protection & ~VM_PROT_WRITE); 2596 } 2597 2598 /* 2599 * Make a copy of the object. 2600 */ 2601 if ((src_object = src_entry->object.vm_object) != NULL) { 2602 2603 if ((src_object->handle == NULL) && 2604 (src_object->type == OBJT_DEFAULT || 2605 src_object->type == OBJT_SWAP)) { 2606 vm_object_collapse(src_object); 2607 if ((src_object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) { 2608 vm_map_split(src_entry); 2609 src_object = src_entry->object.vm_object; 2610 } 2611 } 2612 2613 vm_object_reference(src_object); 2614 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 2615 dst_entry->object.vm_object = src_object; 2616 src_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY); 2617 dst_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY); 2618 dst_entry->offset = src_entry->offset; 2619 } else { 2620 dst_entry->object.vm_object = NULL; 2621 dst_entry->offset = 0; 2622 } 2623 2624 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, 2625 dst_entry->end - dst_entry->start, src_entry->start); 2626 } else { 2627 /* 2628 * Of course, wired down pages can't be set copy-on-write. 2629 * Cause wired pages to be copied into the new map by 2630 * simulating faults (the new pages are pageable) 2631 */ 2632 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 2633 } 2634 } 2635 2636 /* 2637 * vmspace_fork: 2638 * Create a new process vmspace structure and vm_map 2639 * based on those of an existing process. The new map 2640 * is based on the old map, according to the inheritance 2641 * values on the regions in that map. 2642 * 2643 * The source map must not be locked. 2644 */ 2645 struct vmspace * 2646 vmspace_fork(struct vmspace *vm1) 2647 { 2648 struct vmspace *vm2; 2649 vm_map_t old_map = &vm1->vm_map; 2650 vm_map_t new_map; 2651 vm_map_entry_t old_entry; 2652 vm_map_entry_t new_entry; 2653 vm_object_t object; 2654 int count; 2655 2656 vm_map_lock(old_map); 2657 old_map->infork = 1; 2658 2659 /* 2660 * XXX Note: upcalls are not copied. 2661 */ 2662 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset); 2663 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 2664 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 2665 new_map = &vm2->vm_map; /* XXX */ 2666 new_map->timestamp = 1; 2667 2668 count = 0; 2669 old_entry = old_map->header.next; 2670 while (old_entry != &old_map->header) { 2671 ++count; 2672 old_entry = old_entry->next; 2673 } 2674 2675 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 2676 2677 old_entry = old_map->header.next; 2678 while (old_entry != &old_map->header) { 2679 if (old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) 2680 panic("vm_map_fork: encountered a submap"); 2681 2682 switch (old_entry->inheritance) { 2683 case VM_INHERIT_NONE: 2684 break; 2685 2686 case VM_INHERIT_SHARE: 2687 /* 2688 * Clone the entry, creating the shared object if necessary. 2689 */ 2690 object = old_entry->object.vm_object; 2691 if (object == NULL) { 2692 object = vm_object_allocate(OBJT_DEFAULT, 2693 atop(old_entry->end - old_entry->start)); 2694 old_entry->object.vm_object = object; 2695 old_entry->offset = (vm_offset_t) 0; 2696 } 2697 2698 /* 2699 * Add the reference before calling vm_object_shadow 2700 * to insure that a shadow object is created. 2701 */ 2702 vm_object_reference(object); 2703 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 2704 vm_object_shadow(&old_entry->object.vm_object, 2705 &old_entry->offset, 2706 atop(old_entry->end - old_entry->start)); 2707 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 2708 /* Transfer the second reference too. */ 2709 vm_object_reference( 2710 old_entry->object.vm_object); 2711 vm_object_deallocate(object); 2712 object = old_entry->object.vm_object; 2713 } 2714 vm_object_clear_flag(object, OBJ_ONEMAPPING); 2715 2716 /* 2717 * Clone the entry, referencing the shared object. 2718 */ 2719 new_entry = vm_map_entry_create(new_map, &count); 2720 *new_entry = *old_entry; 2721 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2722 new_entry->wired_count = 0; 2723 2724 /* 2725 * Insert the entry into the new map -- we know we're 2726 * inserting at the end of the new map. 2727 */ 2728 2729 vm_map_entry_link(new_map, new_map->header.prev, 2730 new_entry); 2731 2732 /* 2733 * Update the physical map 2734 */ 2735 2736 pmap_copy(new_map->pmap, old_map->pmap, 2737 new_entry->start, 2738 (old_entry->end - old_entry->start), 2739 old_entry->start); 2740 break; 2741 2742 case VM_INHERIT_COPY: 2743 /* 2744 * Clone the entry and link into the map. 2745 */ 2746 new_entry = vm_map_entry_create(new_map, &count); 2747 *new_entry = *old_entry; 2748 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2749 new_entry->wired_count = 0; 2750 new_entry->object.vm_object = NULL; 2751 vm_map_entry_link(new_map, new_map->header.prev, 2752 new_entry); 2753 vm_map_copy_entry(old_map, new_map, old_entry, 2754 new_entry); 2755 break; 2756 } 2757 old_entry = old_entry->next; 2758 } 2759 2760 new_map->size = old_map->size; 2761 old_map->infork = 0; 2762 vm_map_unlock(old_map); 2763 vm_map_entry_release(count); 2764 2765 return (vm2); 2766 } 2767 2768 int 2769 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 2770 vm_prot_t prot, vm_prot_t max, int cow) 2771 { 2772 vm_map_entry_t prev_entry; 2773 vm_map_entry_t new_stack_entry; 2774 vm_size_t init_ssize; 2775 int rv; 2776 int count; 2777 2778 if (VM_MIN_ADDRESS > 0 && addrbos < VM_MIN_ADDRESS) 2779 return (KERN_NO_SPACE); 2780 2781 if (max_ssize < sgrowsiz) 2782 init_ssize = max_ssize; 2783 else 2784 init_ssize = sgrowsiz; 2785 2786 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2787 vm_map_lock(map); 2788 2789 /* If addr is already mapped, no go */ 2790 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) { 2791 vm_map_unlock(map); 2792 vm_map_entry_release(count); 2793 return (KERN_NO_SPACE); 2794 } 2795 2796 /* If we would blow our VMEM resource limit, no go */ 2797 if (map->size + init_ssize > 2798 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 2799 vm_map_unlock(map); 2800 vm_map_entry_release(count); 2801 return (KERN_NO_SPACE); 2802 } 2803 2804 /* If we can't accomodate max_ssize in the current mapping, 2805 * no go. However, we need to be aware that subsequent user 2806 * mappings might map into the space we have reserved for 2807 * stack, and currently this space is not protected. 2808 * 2809 * Hopefully we will at least detect this condition 2810 * when we try to grow the stack. 2811 */ 2812 if ((prev_entry->next != &map->header) && 2813 (prev_entry->next->start < addrbos + max_ssize)) { 2814 vm_map_unlock(map); 2815 vm_map_entry_release(count); 2816 return (KERN_NO_SPACE); 2817 } 2818 2819 /* We initially map a stack of only init_ssize. We will 2820 * grow as needed later. Since this is to be a grow 2821 * down stack, we map at the top of the range. 2822 * 2823 * Note: we would normally expect prot and max to be 2824 * VM_PROT_ALL, and cow to be 0. Possibly we should 2825 * eliminate these as input parameters, and just 2826 * pass these values here in the insert call. 2827 */ 2828 rv = vm_map_insert(map, &count, 2829 NULL, 0, addrbos + max_ssize - init_ssize, 2830 addrbos + max_ssize, prot, max, cow); 2831 2832 /* Now set the avail_ssize amount */ 2833 if (rv == KERN_SUCCESS) { 2834 if (prev_entry != &map->header) 2835 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count); 2836 new_stack_entry = prev_entry->next; 2837 if (new_stack_entry->end != addrbos + max_ssize || 2838 new_stack_entry->start != addrbos + max_ssize - init_ssize) 2839 panic ("Bad entry start/end for new stack entry"); 2840 else 2841 new_stack_entry->avail_ssize = max_ssize - init_ssize; 2842 } 2843 2844 vm_map_unlock(map); 2845 vm_map_entry_release(count); 2846 return (rv); 2847 } 2848 2849 /* Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 2850 * desired address is already mapped, or if we successfully grow 2851 * the stack. Also returns KERN_SUCCESS if addr is outside the 2852 * stack range (this is strange, but preserves compatibility with 2853 * the grow function in vm_machdep.c). 2854 */ 2855 int 2856 vm_map_growstack (struct proc *p, vm_offset_t addr) 2857 { 2858 vm_map_entry_t prev_entry; 2859 vm_map_entry_t stack_entry; 2860 vm_map_entry_t new_stack_entry; 2861 struct vmspace *vm = p->p_vmspace; 2862 vm_map_t map = &vm->vm_map; 2863 vm_offset_t end; 2864 int grow_amount; 2865 int rv = KERN_SUCCESS; 2866 int is_procstack; 2867 int use_read_lock = 1; 2868 int count; 2869 2870 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2871 Retry: 2872 if (use_read_lock) 2873 vm_map_lock_read(map); 2874 else 2875 vm_map_lock(map); 2876 2877 /* If addr is already in the entry range, no need to grow.*/ 2878 if (vm_map_lookup_entry(map, addr, &prev_entry)) 2879 goto done; 2880 2881 if ((stack_entry = prev_entry->next) == &map->header) 2882 goto done; 2883 if (prev_entry == &map->header) 2884 end = stack_entry->start - stack_entry->avail_ssize; 2885 else 2886 end = prev_entry->end; 2887 2888 /* This next test mimics the old grow function in vm_machdep.c. 2889 * It really doesn't quite make sense, but we do it anyway 2890 * for compatibility. 2891 * 2892 * If not growable stack, return success. This signals the 2893 * caller to proceed as he would normally with normal vm. 2894 */ 2895 if (stack_entry->avail_ssize < 1 || 2896 addr >= stack_entry->start || 2897 addr < stack_entry->start - stack_entry->avail_ssize) { 2898 goto done; 2899 } 2900 2901 /* Find the minimum grow amount */ 2902 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE); 2903 if (grow_amount > stack_entry->avail_ssize) { 2904 rv = KERN_NO_SPACE; 2905 goto done; 2906 } 2907 2908 /* If there is no longer enough space between the entries 2909 * nogo, and adjust the available space. Note: this 2910 * should only happen if the user has mapped into the 2911 * stack area after the stack was created, and is 2912 * probably an error. 2913 * 2914 * This also effectively destroys any guard page the user 2915 * might have intended by limiting the stack size. 2916 */ 2917 if (grow_amount > stack_entry->start - end) { 2918 if (use_read_lock && vm_map_lock_upgrade(map)) { 2919 use_read_lock = 0; 2920 goto Retry; 2921 } 2922 use_read_lock = 0; 2923 stack_entry->avail_ssize = stack_entry->start - end; 2924 rv = KERN_NO_SPACE; 2925 goto done; 2926 } 2927 2928 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 2929 2930 /* If this is the main process stack, see if we're over the 2931 * stack limit. 2932 */ 2933 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 2934 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 2935 rv = KERN_NO_SPACE; 2936 goto done; 2937 } 2938 2939 /* Round up the grow amount modulo SGROWSIZ */ 2940 grow_amount = roundup (grow_amount, sgrowsiz); 2941 if (grow_amount > stack_entry->avail_ssize) { 2942 grow_amount = stack_entry->avail_ssize; 2943 } 2944 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 2945 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 2946 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - 2947 ctob(vm->vm_ssize); 2948 } 2949 2950 /* If we would blow our VMEM resource limit, no go */ 2951 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 2952 rv = KERN_NO_SPACE; 2953 goto done; 2954 } 2955 2956 if (use_read_lock && vm_map_lock_upgrade(map)) { 2957 use_read_lock = 0; 2958 goto Retry; 2959 } 2960 use_read_lock = 0; 2961 2962 /* Get the preliminary new entry start value */ 2963 addr = stack_entry->start - grow_amount; 2964 2965 /* If this puts us into the previous entry, cut back our growth 2966 * to the available space. Also, see the note above. 2967 */ 2968 if (addr < end) { 2969 stack_entry->avail_ssize = stack_entry->start - end; 2970 addr = end; 2971 } 2972 2973 rv = vm_map_insert(map, &count, 2974 NULL, 0, addr, stack_entry->start, 2975 VM_PROT_ALL, 2976 VM_PROT_ALL, 2977 0); 2978 2979 /* Adjust the available stack space by the amount we grew. */ 2980 if (rv == KERN_SUCCESS) { 2981 if (prev_entry != &map->header) 2982 vm_map_clip_end(map, prev_entry, addr, &count); 2983 new_stack_entry = prev_entry->next; 2984 if (new_stack_entry->end != stack_entry->start || 2985 new_stack_entry->start != addr) 2986 panic ("Bad stack grow start/end in new stack entry"); 2987 else { 2988 new_stack_entry->avail_ssize = stack_entry->avail_ssize - 2989 (new_stack_entry->end - 2990 new_stack_entry->start); 2991 if (is_procstack) 2992 vm->vm_ssize += btoc(new_stack_entry->end - 2993 new_stack_entry->start); 2994 } 2995 } 2996 2997 done: 2998 if (use_read_lock) 2999 vm_map_unlock_read(map); 3000 else 3001 vm_map_unlock(map); 3002 vm_map_entry_release(count); 3003 return (rv); 3004 } 3005 3006 /* 3007 * Unshare the specified VM space for exec. If other processes are 3008 * mapped to it, then create a new one. The new vmspace is null. 3009 */ 3010 3011 void 3012 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 3013 { 3014 struct vmspace *oldvmspace = p->p_vmspace; 3015 struct vmspace *newvmspace; 3016 vm_map_t map = &p->p_vmspace->vm_map; 3017 3018 /* 3019 * If we are execing a resident vmspace we fork it, otherwise 3020 * we create a new vmspace. Note that exitingcnt and upcalls 3021 * are not copied to the new vmspace. 3022 */ 3023 if (vmcopy) { 3024 newvmspace = vmspace_fork(vmcopy); 3025 } else { 3026 newvmspace = vmspace_alloc(map->min_offset, map->max_offset); 3027 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 3028 (caddr_t)&oldvmspace->vm_endcopy - 3029 (caddr_t)&oldvmspace->vm_startcopy); 3030 } 3031 3032 /* 3033 * This code is written like this for prototype purposes. The 3034 * goal is to avoid running down the vmspace here, but let the 3035 * other process's that are still using the vmspace to finally 3036 * run it down. Even though there is little or no chance of blocking 3037 * here, it is a good idea to keep this form for future mods. 3038 */ 3039 p->p_vmspace = newvmspace; 3040 pmap_pinit2(vmspace_pmap(newvmspace)); 3041 if (p == curproc) 3042 pmap_activate(p); 3043 vmspace_free(oldvmspace); 3044 } 3045 3046 /* 3047 * Unshare the specified VM space for forcing COW. This 3048 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 3049 * 3050 * The exitingcnt test is not strictly necessary but has been 3051 * included for code sanity (to make the code a bit more deterministic). 3052 */ 3053 3054 void 3055 vmspace_unshare(struct proc *p) 3056 { 3057 struct vmspace *oldvmspace = p->p_vmspace; 3058 struct vmspace *newvmspace; 3059 3060 if (oldvmspace->vm_refcnt == 1 && oldvmspace->vm_exitingcnt == 0) 3061 return; 3062 newvmspace = vmspace_fork(oldvmspace); 3063 p->p_vmspace = newvmspace; 3064 pmap_pinit2(vmspace_pmap(newvmspace)); 3065 if (p == curproc) 3066 pmap_activate(p); 3067 vmspace_free(oldvmspace); 3068 } 3069 3070 /* 3071 * vm_map_lookup: 3072 * 3073 * Finds the VM object, offset, and 3074 * protection for a given virtual address in the 3075 * specified map, assuming a page fault of the 3076 * type specified. 3077 * 3078 * Leaves the map in question locked for read; return 3079 * values are guaranteed until a vm_map_lookup_done 3080 * call is performed. Note that the map argument 3081 * is in/out; the returned map must be used in 3082 * the call to vm_map_lookup_done. 3083 * 3084 * A handle (out_entry) is returned for use in 3085 * vm_map_lookup_done, to make that fast. 3086 * 3087 * If a lookup is requested with "write protection" 3088 * specified, the map may be changed to perform virtual 3089 * copying operations, although the data referenced will 3090 * remain the same. 3091 */ 3092 int 3093 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 3094 vm_offset_t vaddr, 3095 vm_prot_t fault_typea, 3096 vm_map_entry_t *out_entry, /* OUT */ 3097 vm_object_t *object, /* OUT */ 3098 vm_pindex_t *pindex, /* OUT */ 3099 vm_prot_t *out_prot, /* OUT */ 3100 boolean_t *wired) /* OUT */ 3101 { 3102 vm_map_entry_t entry; 3103 vm_map_t map = *var_map; 3104 vm_prot_t prot; 3105 vm_prot_t fault_type = fault_typea; 3106 int use_read_lock = 1; 3107 int rv = KERN_SUCCESS; 3108 3109 RetryLookup: 3110 if (use_read_lock) 3111 vm_map_lock_read(map); 3112 else 3113 vm_map_lock(map); 3114 3115 /* 3116 * If the map has an interesting hint, try it before calling full 3117 * blown lookup routine. 3118 */ 3119 entry = map->hint; 3120 *out_entry = entry; 3121 3122 if ((entry == &map->header) || 3123 (vaddr < entry->start) || (vaddr >= entry->end)) { 3124 vm_map_entry_t tmp_entry; 3125 3126 /* 3127 * Entry was either not a valid hint, or the vaddr was not 3128 * contained in the entry, so do a full lookup. 3129 */ 3130 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 3131 rv = KERN_INVALID_ADDRESS; 3132 goto done; 3133 } 3134 3135 entry = tmp_entry; 3136 *out_entry = entry; 3137 } 3138 3139 /* 3140 * Handle submaps. 3141 */ 3142 3143 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 3144 vm_map_t old_map = map; 3145 3146 *var_map = map = entry->object.sub_map; 3147 if (use_read_lock) 3148 vm_map_unlock_read(old_map); 3149 else 3150 vm_map_unlock(old_map); 3151 use_read_lock = 1; 3152 goto RetryLookup; 3153 } 3154 3155 /* 3156 * Check whether this task is allowed to have this page. 3157 * Note the special case for MAP_ENTRY_COW 3158 * pages with an override. This is to implement a forced 3159 * COW for debuggers. 3160 */ 3161 3162 if (fault_type & VM_PROT_OVERRIDE_WRITE) 3163 prot = entry->max_protection; 3164 else 3165 prot = entry->protection; 3166 3167 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 3168 if ((fault_type & prot) != fault_type) { 3169 rv = KERN_PROTECTION_FAILURE; 3170 goto done; 3171 } 3172 3173 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 3174 (entry->eflags & MAP_ENTRY_COW) && 3175 (fault_type & VM_PROT_WRITE) && 3176 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 3177 rv = KERN_PROTECTION_FAILURE; 3178 goto done; 3179 } 3180 3181 /* 3182 * If this page is not pageable, we have to get it for all possible 3183 * accesses. 3184 */ 3185 3186 *wired = (entry->wired_count != 0); 3187 if (*wired) 3188 prot = fault_type = entry->protection; 3189 3190 /* 3191 * If the entry was copy-on-write, we either ... 3192 */ 3193 3194 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3195 /* 3196 * If we want to write the page, we may as well handle that 3197 * now since we've got the map locked. 3198 * 3199 * If we don't need to write the page, we just demote the 3200 * permissions allowed. 3201 */ 3202 3203 if (fault_type & VM_PROT_WRITE) { 3204 /* 3205 * Make a new object, and place it in the object 3206 * chain. Note that no new references have appeared 3207 * -- one just moved from the map to the new 3208 * object. 3209 */ 3210 3211 if (use_read_lock && vm_map_lock_upgrade(map)) { 3212 use_read_lock = 0; 3213 goto RetryLookup; 3214 } 3215 use_read_lock = 0; 3216 3217 vm_object_shadow( 3218 &entry->object.vm_object, 3219 &entry->offset, 3220 atop(entry->end - entry->start)); 3221 3222 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 3223 } else { 3224 /* 3225 * We're attempting to read a copy-on-write page -- 3226 * don't allow writes. 3227 */ 3228 3229 prot &= ~VM_PROT_WRITE; 3230 } 3231 } 3232 3233 /* 3234 * Create an object if necessary. 3235 */ 3236 if (entry->object.vm_object == NULL && 3237 !map->system_map) { 3238 if (use_read_lock && vm_map_lock_upgrade(map)) { 3239 use_read_lock = 0; 3240 goto RetryLookup; 3241 } 3242 use_read_lock = 0; 3243 entry->object.vm_object = vm_object_allocate(OBJT_DEFAULT, 3244 atop(entry->end - entry->start)); 3245 entry->offset = 0; 3246 } 3247 3248 /* 3249 * Return the object/offset from this entry. If the entry was 3250 * copy-on-write or empty, it has been fixed up. 3251 */ 3252 3253 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 3254 *object = entry->object.vm_object; 3255 3256 /* 3257 * Return whether this is the only map sharing this data. On 3258 * success we return with a read lock held on the map. On failure 3259 * we return with the map unlocked. 3260 */ 3261 *out_prot = prot; 3262 done: 3263 if (rv == KERN_SUCCESS) { 3264 if (use_read_lock == 0) 3265 vm_map_lock_downgrade(map); 3266 } else if (use_read_lock) { 3267 vm_map_unlock_read(map); 3268 } else { 3269 vm_map_unlock(map); 3270 } 3271 return (rv); 3272 } 3273 3274 /* 3275 * vm_map_lookup_done: 3276 * 3277 * Releases locks acquired by a vm_map_lookup 3278 * (according to the handle returned by that lookup). 3279 */ 3280 3281 void 3282 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 3283 { 3284 /* 3285 * Unlock the main-level map 3286 */ 3287 vm_map_unlock_read(map); 3288 if (count) 3289 vm_map_entry_release(count); 3290 } 3291 3292 /* 3293 * Performs the copy_on_write operations necessary to allow the virtual copies 3294 * into user space to work. This has to be called for write(2) system calls 3295 * from other processes, file unlinking, and file size shrinkage. 3296 */ 3297 void 3298 vm_freeze_copyopts(vm_object_t object, vm_pindex_t froma, vm_pindex_t toa) 3299 { 3300 int rv; 3301 vm_object_t robject; 3302 vm_pindex_t idx; 3303 3304 if ((object == NULL) || 3305 ((object->flags & OBJ_OPT) == 0)) 3306 return; 3307 3308 if (object->shadow_count > object->ref_count) 3309 panic("vm_freeze_copyopts: sc > rc"); 3310 3311 while ((robject = LIST_FIRST(&object->shadow_head)) != NULL) { 3312 vm_pindex_t bo_pindex; 3313 vm_page_t m_in, m_out; 3314 3315 bo_pindex = OFF_TO_IDX(robject->backing_object_offset); 3316 3317 vm_object_reference(robject); 3318 3319 vm_object_pip_wait(robject, "objfrz"); 3320 3321 if (robject->ref_count == 1) { 3322 vm_object_deallocate(robject); 3323 continue; 3324 } 3325 3326 vm_object_pip_add(robject, 1); 3327 3328 for (idx = 0; idx < robject->size; idx++) { 3329 3330 m_out = vm_page_grab(robject, idx, 3331 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 3332 3333 if (m_out->valid == 0) { 3334 m_in = vm_page_grab(object, bo_pindex + idx, 3335 VM_ALLOC_NORMAL | VM_ALLOC_RETRY); 3336 if (m_in->valid == 0) { 3337 rv = vm_pager_get_pages(object, &m_in, 1, 0); 3338 if (rv != VM_PAGER_OK) { 3339 printf("vm_freeze_copyopts: cannot read page from file: %lx\n", (long)m_in->pindex); 3340 continue; 3341 } 3342 vm_page_deactivate(m_in); 3343 } 3344 3345 vm_page_protect(m_in, VM_PROT_NONE); 3346 pmap_copy_page(VM_PAGE_TO_PHYS(m_in), VM_PAGE_TO_PHYS(m_out)); 3347 m_out->valid = m_in->valid; 3348 vm_page_dirty(m_out); 3349 vm_page_activate(m_out); 3350 vm_page_wakeup(m_in); 3351 } 3352 vm_page_wakeup(m_out); 3353 } 3354 3355 object->shadow_count--; 3356 object->ref_count--; 3357 LIST_REMOVE(robject, shadow_list); 3358 robject->backing_object = NULL; 3359 robject->backing_object_offset = 0; 3360 3361 vm_object_pip_wakeup(robject); 3362 vm_object_deallocate(robject); 3363 } 3364 3365 vm_object_clear_flag(object, OBJ_OPT); 3366 } 3367 3368 #include "opt_ddb.h" 3369 #ifdef DDB 3370 #include <sys/kernel.h> 3371 3372 #include <ddb/ddb.h> 3373 3374 /* 3375 * vm_map_print: [ debug ] 3376 */ 3377 DB_SHOW_COMMAND(map, vm_map_print) 3378 { 3379 static int nlines; 3380 /* XXX convert args. */ 3381 vm_map_t map = (vm_map_t)addr; 3382 boolean_t full = have_addr; 3383 3384 vm_map_entry_t entry; 3385 3386 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 3387 (void *)map, 3388 (void *)map->pmap, map->nentries, map->timestamp); 3389 nlines++; 3390 3391 if (!full && db_indent) 3392 return; 3393 3394 db_indent += 2; 3395 for (entry = map->header.next; entry != &map->header; 3396 entry = entry->next) { 3397 db_iprintf("map entry %p: start=%p, end=%p\n", 3398 (void *)entry, (void *)entry->start, (void *)entry->end); 3399 nlines++; 3400 { 3401 static char *inheritance_name[4] = 3402 {"share", "copy", "none", "donate_copy"}; 3403 3404 db_iprintf(" prot=%x/%x/%s", 3405 entry->protection, 3406 entry->max_protection, 3407 inheritance_name[(int)(unsigned char)entry->inheritance]); 3408 if (entry->wired_count != 0) 3409 db_printf(", wired"); 3410 } 3411 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 3412 /* XXX no %qd in kernel. Truncate entry->offset. */ 3413 db_printf(", share=%p, offset=0x%lx\n", 3414 (void *)entry->object.sub_map, 3415 (long)entry->offset); 3416 nlines++; 3417 if ((entry->prev == &map->header) || 3418 (entry->prev->object.sub_map != 3419 entry->object.sub_map)) { 3420 db_indent += 2; 3421 vm_map_print((db_expr_t)(intptr_t) 3422 entry->object.sub_map, 3423 full, 0, (char *)0); 3424 db_indent -= 2; 3425 } 3426 } else { 3427 /* XXX no %qd in kernel. Truncate entry->offset. */ 3428 db_printf(", object=%p, offset=0x%lx", 3429 (void *)entry->object.vm_object, 3430 (long)entry->offset); 3431 if (entry->eflags & MAP_ENTRY_COW) 3432 db_printf(", copy (%s)", 3433 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 3434 db_printf("\n"); 3435 nlines++; 3436 3437 if ((entry->prev == &map->header) || 3438 (entry->prev->object.vm_object != 3439 entry->object.vm_object)) { 3440 db_indent += 2; 3441 vm_object_print((db_expr_t)(intptr_t) 3442 entry->object.vm_object, 3443 full, 0, (char *)0); 3444 nlines += 4; 3445 db_indent -= 2; 3446 } 3447 } 3448 } 3449 db_indent -= 2; 3450 if (db_indent == 0) 3451 nlines = 0; 3452 } 3453 3454 3455 DB_SHOW_COMMAND(procvm, procvm) 3456 { 3457 struct proc *p; 3458 3459 if (have_addr) { 3460 p = (struct proc *) addr; 3461 } else { 3462 p = curproc; 3463 } 3464 3465 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 3466 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 3467 (void *)vmspace_pmap(p->p_vmspace)); 3468 3469 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 3470 } 3471 3472 #endif /* DDB */ 3473