1 /* 2 * (MPSAFE) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94 35 * 36 * 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 * 62 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $ 63 */ 64 65 /* 66 * Virtual memory mapping module. 67 */ 68 69 #include <sys/param.h> 70 #include <sys/systm.h> 71 #include <sys/kernel.h> 72 #include <sys/proc.h> 73 #include <sys/serialize.h> 74 #include <sys/lock.h> 75 #include <sys/vmmeter.h> 76 #include <sys/mman.h> 77 #include <sys/vnode.h> 78 #include <sys/resourcevar.h> 79 #include <sys/shm.h> 80 #include <sys/tree.h> 81 #include <sys/malloc.h> 82 #include <sys/objcache.h> 83 84 #include <vm/vm.h> 85 #include <vm/vm_param.h> 86 #include <vm/pmap.h> 87 #include <vm/vm_map.h> 88 #include <vm/vm_page.h> 89 #include <vm/vm_object.h> 90 #include <vm/vm_pager.h> 91 #include <vm/vm_kern.h> 92 #include <vm/vm_extern.h> 93 #include <vm/swap_pager.h> 94 #include <vm/vm_zone.h> 95 96 #include <sys/thread2.h> 97 #include <sys/random.h> 98 #include <sys/sysctl.h> 99 100 /* 101 * Virtual memory maps provide for the mapping, protection, and sharing 102 * of virtual memory objects. In addition, this module provides for an 103 * efficient virtual copy of memory from one map to another. 104 * 105 * Synchronization is required prior to most operations. 106 * 107 * Maps consist of an ordered doubly-linked list of simple entries. 108 * A hint and a RB tree is used to speed-up lookups. 109 * 110 * Callers looking to modify maps specify start/end addresses which cause 111 * the related map entry to be clipped if necessary, and then later 112 * recombined if the pieces remained compatible. 113 * 114 * Virtual copy operations are performed by copying VM object references 115 * from one map to another, and then marking both regions as copy-on-write. 116 */ 117 static __boolean_t vmspace_ctor(void *obj, void *privdata, int ocflags); 118 static void vmspace_dtor(void *obj, void *privdata); 119 static void vmspace_terminate(struct vmspace *vm, int final); 120 121 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore"); 122 static struct objcache *vmspace_cache; 123 124 /* 125 * per-cpu page table cross mappings are initialized in early boot 126 * and might require a considerable number of vm_map_entry structures. 127 */ 128 #define MAPENTRYBSP_CACHE (MAXCPU+1) 129 #define MAPENTRYAP_CACHE 8 130 131 static struct vm_zone mapentzone_store, mapzone_store; 132 static vm_zone_t mapentzone, mapzone; 133 static struct vm_object mapentobj, mapobj; 134 135 static struct vm_map_entry map_entry_init[MAX_MAPENT]; 136 static struct vm_map_entry cpu_map_entry_init_bsp[MAPENTRYBSP_CACHE]; 137 static struct vm_map_entry cpu_map_entry_init_ap[MAXCPU][MAPENTRYAP_CACHE]; 138 static struct vm_map map_init[MAX_KMAP]; 139 140 static int randomize_mmap; 141 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0, 142 "Randomize mmap offsets"); 143 static int vm_map_relock_enable = 1; 144 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW, 145 &vm_map_relock_enable, 0, "Randomize mmap offsets"); 146 147 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref); 148 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *); 149 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *); 150 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 151 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 152 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *); 153 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t); 154 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t, 155 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 /* 159 * Initialize the vm_map module. Must be called before any other vm_map 160 * routines. 161 * 162 * Map and entry structures are allocated from the general purpose 163 * memory pool with some exceptions: 164 * 165 * - The kernel map is allocated statically. 166 * - Initial kernel map entries are allocated out of a static pool. 167 * - We must set ZONE_SPECIAL here or the early boot code can get 168 * stuck if there are >63 cores. 169 * 170 * These restrictions are necessary since malloc() uses the 171 * maps and requires map entries. 172 * 173 * Called from the low level boot code only. 174 */ 175 void 176 vm_map_startup(void) 177 { 178 mapzone = &mapzone_store; 179 zbootinit(mapzone, "MAP", sizeof (struct vm_map), 180 map_init, MAX_KMAP); 181 mapentzone = &mapentzone_store; 182 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry), 183 map_entry_init, MAX_MAPENT); 184 mapentzone_store.zflags |= ZONE_SPECIAL; 185 } 186 187 /* 188 * Called prior to any vmspace allocations. 189 * 190 * Called from the low level boot code only. 191 */ 192 void 193 vm_init2(void) 194 { 195 vmspace_cache = objcache_create_mbacked(M_VMSPACE, 196 sizeof(struct vmspace), 197 0, ncpus * 4, 198 vmspace_ctor, vmspace_dtor, 199 NULL); 200 zinitna(mapentzone, &mapentobj, NULL, 0, 0, 201 ZONE_USE_RESERVE | ZONE_SPECIAL, 1); 202 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1); 203 pmap_init2(); 204 vm_object_init2(); 205 } 206 207 /* 208 * objcache support. We leave the pmap root cached as long as possible 209 * for performance reasons. 210 */ 211 static 212 __boolean_t 213 vmspace_ctor(void *obj, void *privdata, int ocflags) 214 { 215 struct vmspace *vm = obj; 216 217 bzero(vm, sizeof(*vm)); 218 vm->vm_refcnt = (u_int)-1; 219 220 return 1; 221 } 222 223 static 224 void 225 vmspace_dtor(void *obj, void *privdata) 226 { 227 struct vmspace *vm = obj; 228 229 KKASSERT(vm->vm_refcnt == (u_int)-1); 230 pmap_puninit(vmspace_pmap(vm)); 231 } 232 233 /* 234 * Red black tree functions 235 * 236 * The caller must hold the related map lock. 237 */ 238 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b); 239 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare); 240 241 /* a->start is address, and the only field has to be initialized */ 242 static int 243 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b) 244 { 245 if (a->start < b->start) 246 return(-1); 247 else if (a->start > b->start) 248 return(1); 249 return(0); 250 } 251 252 /* 253 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for 254 * every refcnt. 255 */ 256 void 257 vmspace_initrefs(struct vmspace *vm) 258 { 259 vm->vm_refcnt = 1; 260 vm->vm_holdcnt = 1; 261 } 262 263 /* 264 * Allocate a vmspace structure, including a vm_map and pmap. 265 * Initialize numerous fields. While the initial allocation is zerod, 266 * subsequence reuse from the objcache leaves elements of the structure 267 * intact (particularly the pmap), so portions must be zerod. 268 * 269 * Returns a referenced vmspace. 270 * 271 * No requirements. 272 */ 273 struct vmspace * 274 vmspace_alloc(vm_offset_t min, vm_offset_t max) 275 { 276 struct vmspace *vm; 277 278 vm = objcache_get(vmspace_cache, M_WAITOK); 279 280 bzero(&vm->vm_startcopy, 281 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy); 282 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */ 283 284 /* 285 * NOTE: hold to acquires token for safety. 286 * 287 * On return vmspace is referenced (refs=1, hold=1). That is, 288 * each refcnt also has a holdcnt. There can be additional holds 289 * (holdcnt) above and beyond the refcnt. Finalization is handled in 290 * two stages, one on refs 1->0, and the the second on hold 1->0. 291 */ 292 KKASSERT(vm->vm_holdcnt == 0); 293 KKASSERT(vm->vm_refcnt == (u_int)-1); 294 vmspace_initrefs(vm); 295 vmspace_hold(vm); 296 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */ 297 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */ 298 vm->vm_shm = NULL; 299 vm->vm_flags = 0; 300 cpu_vmspace_alloc(vm); 301 vmspace_drop(vm); 302 303 return (vm); 304 } 305 306 /* 307 * NOTE: Can return -1 if the vmspace is exiting. 308 */ 309 int 310 vmspace_getrefs(struct vmspace *vm) 311 { 312 return ((int)vm->vm_refcnt); 313 } 314 315 /* 316 * A vmspace object must already have a non-zero hold to be able to gain 317 * further holds on it. 318 */ 319 static void 320 vmspace_hold_notoken(struct vmspace *vm) 321 { 322 KKASSERT(vm->vm_holdcnt != 0); 323 refcount_acquire(&vm->vm_holdcnt); 324 } 325 326 static void 327 vmspace_drop_notoken(struct vmspace *vm) 328 { 329 if (refcount_release(&vm->vm_holdcnt)) { 330 if (vm->vm_refcnt == (u_int)-1) { 331 vmspace_terminate(vm, 1); 332 } 333 } 334 } 335 336 void 337 vmspace_hold(struct vmspace *vm) 338 { 339 vmspace_hold_notoken(vm); 340 lwkt_gettoken(&vm->vm_map.token); 341 } 342 343 void 344 vmspace_drop(struct vmspace *vm) 345 { 346 lwkt_reltoken(&vm->vm_map.token); 347 vmspace_drop_notoken(vm); 348 } 349 350 /* 351 * A vmspace object must not be in a terminated state to be able to obtain 352 * additional refs on it. 353 * 354 * Ref'ing a vmspace object also increments its hold count. 355 */ 356 void 357 vmspace_ref(struct vmspace *vm) 358 { 359 KKASSERT((int)vm->vm_refcnt >= 0); 360 vmspace_hold_notoken(vm); 361 refcount_acquire(&vm->vm_refcnt); 362 } 363 364 /* 365 * Release a ref on the vmspace. On the 1->0 transition we do stage-1 366 * termination of the vmspace. Then, on the final drop of the hold we 367 * will do stage-2 final termination. 368 */ 369 void 370 vmspace_rel(struct vmspace *vm) 371 { 372 if (refcount_release(&vm->vm_refcnt)) { 373 vm->vm_refcnt = (u_int)-1; /* no other refs possible */ 374 vmspace_terminate(vm, 0); 375 } 376 vmspace_drop_notoken(vm); 377 } 378 379 /* 380 * This is called during exit indicating that the vmspace is no 381 * longer in used by an exiting process, but the process has not yet 382 * been reaped. 383 * 384 * We release the refcnt but not the associated holdcnt. 385 * 386 * No requirements. 387 */ 388 void 389 vmspace_relexit(struct vmspace *vm) 390 { 391 if (refcount_release(&vm->vm_refcnt)) { 392 vm->vm_refcnt = (u_int)-1; /* no other refs possible */ 393 vmspace_terminate(vm, 0); 394 } 395 } 396 397 /* 398 * Called during reap to disconnect the remainder of the vmspace from 399 * the process. On the hold drop the vmspace termination is finalized. 400 * 401 * No requirements. 402 */ 403 void 404 vmspace_exitfree(struct proc *p) 405 { 406 struct vmspace *vm; 407 408 vm = p->p_vmspace; 409 p->p_vmspace = NULL; 410 vmspace_drop_notoken(vm); 411 } 412 413 /* 414 * Called in two cases: 415 * 416 * (1) When the last refcnt is dropped and the vmspace becomes inactive, 417 * called with final == 0. refcnt will be (u_int)-1 at this point, 418 * and holdcnt will still be non-zero. 419 * 420 * (2) When holdcnt becomes 0, called with final == 1. There should no 421 * longer be anyone with access to the vmspace. 422 * 423 * VMSPACE_EXIT1 flags the primary deactivation 424 * VMSPACE_EXIT2 flags the last reap 425 */ 426 static void 427 vmspace_terminate(struct vmspace *vm, int final) 428 { 429 int count; 430 431 lwkt_gettoken(&vm->vm_map.token); 432 if (final == 0) { 433 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0); 434 435 /* 436 * Get rid of most of the resources. Leave the kernel pmap 437 * intact. 438 */ 439 vm->vm_flags |= VMSPACE_EXIT1; 440 shmexit(vm); 441 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 442 VM_MAX_USER_ADDRESS); 443 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 444 VM_MAX_USER_ADDRESS); 445 lwkt_reltoken(&vm->vm_map.token); 446 } else { 447 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0); 448 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0); 449 450 /* 451 * Get rid of remaining basic resources. 452 */ 453 vm->vm_flags |= VMSPACE_EXIT2; 454 cpu_vmspace_free(vm); 455 shmexit(vm); 456 457 /* 458 * Lock the map, to wait out all other references to it. 459 * Delete all of the mappings and pages they hold, then call 460 * the pmap module to reclaim anything left. 461 */ 462 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 463 vm_map_lock(&vm->vm_map); 464 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset, 465 vm->vm_map.max_offset, &count); 466 vm_map_unlock(&vm->vm_map); 467 vm_map_entry_release(count); 468 469 lwkt_gettoken(&vmspace_pmap(vm)->pm_token); 470 pmap_release(vmspace_pmap(vm)); 471 lwkt_reltoken(&vmspace_pmap(vm)->pm_token); 472 lwkt_reltoken(&vm->vm_map.token); 473 objcache_put(vmspace_cache, vm); 474 } 475 } 476 477 /* 478 * Swap useage is determined by taking the proportional swap used by 479 * VM objects backing the VM map. To make up for fractional losses, 480 * if the VM object has any swap use at all the associated map entries 481 * count for at least 1 swap page. 482 * 483 * No requirements. 484 */ 485 int 486 vmspace_swap_count(struct vmspace *vm) 487 { 488 vm_map_t map = &vm->vm_map; 489 vm_map_entry_t cur; 490 vm_object_t object; 491 int count = 0; 492 int n; 493 494 vmspace_hold(vm); 495 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 496 switch(cur->maptype) { 497 case VM_MAPTYPE_NORMAL: 498 case VM_MAPTYPE_VPAGETABLE: 499 if ((object = cur->object.vm_object) == NULL) 500 break; 501 if (object->swblock_count) { 502 n = (cur->end - cur->start) / PAGE_SIZE; 503 count += object->swblock_count * 504 SWAP_META_PAGES * n / object->size + 1; 505 } 506 break; 507 default: 508 break; 509 } 510 } 511 vmspace_drop(vm); 512 513 return(count); 514 } 515 516 /* 517 * Calculate the approximate number of anonymous pages in use by 518 * this vmspace. To make up for fractional losses, we count each 519 * VM object as having at least 1 anonymous page. 520 * 521 * No requirements. 522 */ 523 int 524 vmspace_anonymous_count(struct vmspace *vm) 525 { 526 vm_map_t map = &vm->vm_map; 527 vm_map_entry_t cur; 528 vm_object_t object; 529 int count = 0; 530 531 vmspace_hold(vm); 532 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 533 switch(cur->maptype) { 534 case VM_MAPTYPE_NORMAL: 535 case VM_MAPTYPE_VPAGETABLE: 536 if ((object = cur->object.vm_object) == NULL) 537 break; 538 if (object->type != OBJT_DEFAULT && 539 object->type != OBJT_SWAP) { 540 break; 541 } 542 count += object->resident_page_count; 543 break; 544 default: 545 break; 546 } 547 } 548 vmspace_drop(vm); 549 550 return(count); 551 } 552 553 /* 554 * Creates and returns a new empty VM map with the given physical map 555 * structure, and having the given lower and upper address bounds. 556 * 557 * No requirements. 558 */ 559 vm_map_t 560 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max) 561 { 562 if (result == NULL) 563 result = zalloc(mapzone); 564 vm_map_init(result, min, max, pmap); 565 return (result); 566 } 567 568 /* 569 * Initialize an existing vm_map structure such as that in the vmspace 570 * structure. The pmap is initialized elsewhere. 571 * 572 * No requirements. 573 */ 574 void 575 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap) 576 { 577 map->header.next = map->header.prev = &map->header; 578 RB_INIT(&map->rb_root); 579 map->nentries = 0; 580 map->size = 0; 581 map->system_map = 0; 582 map->min_offset = min; 583 map->max_offset = max; 584 map->pmap = pmap; 585 map->first_free = &map->header; 586 map->hint = &map->header; 587 map->timestamp = 0; 588 map->flags = 0; 589 lwkt_token_init(&map->token, "vm_map"); 590 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0); 591 } 592 593 /* 594 * Shadow the vm_map_entry's object. This typically needs to be done when 595 * a write fault is taken on an entry which had previously been cloned by 596 * fork(). The shared object (which might be NULL) must become private so 597 * we add a shadow layer above it. 598 * 599 * Object allocation for anonymous mappings is defered as long as possible. 600 * When creating a shadow, however, the underlying object must be instantiated 601 * so it can be shared. 602 * 603 * If the map segment is governed by a virtual page table then it is 604 * possible to address offsets beyond the mapped area. Just allocate 605 * a maximally sized object for this case. 606 * 607 * The vm_map must be exclusively locked. 608 * No other requirements. 609 */ 610 static 611 void 612 vm_map_entry_shadow(vm_map_entry_t entry, int addref) 613 { 614 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 615 vm_object_shadow(&entry->object.vm_object, &entry->offset, 616 0x7FFFFFFF, addref); /* XXX */ 617 } else { 618 vm_object_shadow(&entry->object.vm_object, &entry->offset, 619 atop(entry->end - entry->start), addref); 620 } 621 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 622 } 623 624 /* 625 * Allocate an object for a vm_map_entry. 626 * 627 * Object allocation for anonymous mappings is defered as long as possible. 628 * This function is called when we can defer no longer, generally when a map 629 * entry might be split or forked or takes a page fault. 630 * 631 * If the map segment is governed by a virtual page table then it is 632 * possible to address offsets beyond the mapped area. Just allocate 633 * a maximally sized object for this case. 634 * 635 * The vm_map must be exclusively locked. 636 * No other requirements. 637 */ 638 void 639 vm_map_entry_allocate_object(vm_map_entry_t entry) 640 { 641 vm_object_t obj; 642 643 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 644 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */ 645 } else { 646 obj = vm_object_allocate(OBJT_DEFAULT, 647 atop(entry->end - entry->start)); 648 } 649 entry->object.vm_object = obj; 650 entry->offset = 0; 651 } 652 653 /* 654 * Set an initial negative count so the first attempt to reserve 655 * space preloads a bunch of vm_map_entry's for this cpu. Also 656 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 657 * map a new page for vm_map_entry structures. SMP systems are 658 * particularly sensitive. 659 * 660 * This routine is called in early boot so we cannot just call 661 * vm_map_entry_reserve(). 662 * 663 * Called from the low level boot code only (for each cpu) 664 * 665 * WARNING! Take care not to have too-big a static/BSS structure here 666 * as MAXCPU can be 256+, otherwise the loader's 64MB heap 667 * can get blown out by the kernel plus the initrd image. 668 */ 669 void 670 vm_map_entry_reserve_cpu_init(globaldata_t gd) 671 { 672 vm_map_entry_t entry; 673 int count; 674 int i; 675 676 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2; 677 if (gd->gd_cpuid == 0) { 678 entry = &cpu_map_entry_init_bsp[0]; 679 count = MAPENTRYBSP_CACHE; 680 } else { 681 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0]; 682 count = MAPENTRYAP_CACHE; 683 } 684 for (i = 0; i < count; ++i, ++entry) { 685 entry->next = gd->gd_vme_base; 686 gd->gd_vme_base = entry; 687 } 688 } 689 690 /* 691 * Reserves vm_map_entry structures so code later on can manipulate 692 * map_entry structures within a locked map without blocking trying 693 * to allocate a new vm_map_entry. 694 * 695 * No requirements. 696 */ 697 int 698 vm_map_entry_reserve(int count) 699 { 700 struct globaldata *gd = mycpu; 701 vm_map_entry_t entry; 702 703 /* 704 * Make sure we have enough structures in gd_vme_base to handle 705 * the reservation request. 706 * 707 * The critical section protects access to the per-cpu gd. 708 */ 709 crit_enter(); 710 while (gd->gd_vme_avail < count) { 711 entry = zalloc(mapentzone); 712 entry->next = gd->gd_vme_base; 713 gd->gd_vme_base = entry; 714 ++gd->gd_vme_avail; 715 } 716 gd->gd_vme_avail -= count; 717 crit_exit(); 718 719 return(count); 720 } 721 722 /* 723 * Releases previously reserved vm_map_entry structures that were not 724 * used. If we have too much junk in our per-cpu cache clean some of 725 * it out. 726 * 727 * No requirements. 728 */ 729 void 730 vm_map_entry_release(int count) 731 { 732 struct globaldata *gd = mycpu; 733 vm_map_entry_t entry; 734 735 crit_enter(); 736 gd->gd_vme_avail += count; 737 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 738 entry = gd->gd_vme_base; 739 KKASSERT(entry != NULL); 740 gd->gd_vme_base = entry->next; 741 --gd->gd_vme_avail; 742 crit_exit(); 743 zfree(mapentzone, entry); 744 crit_enter(); 745 } 746 crit_exit(); 747 } 748 749 /* 750 * Reserve map entry structures for use in kernel_map itself. These 751 * entries have *ALREADY* been reserved on a per-cpu basis when the map 752 * was inited. This function is used by zalloc() to avoid a recursion 753 * when zalloc() itself needs to allocate additional kernel memory. 754 * 755 * This function works like the normal reserve but does not load the 756 * vm_map_entry cache (because that would result in an infinite 757 * recursion). Note that gd_vme_avail may go negative. This is expected. 758 * 759 * Any caller of this function must be sure to renormalize after 760 * potentially eating entries to ensure that the reserve supply 761 * remains intact. 762 * 763 * No requirements. 764 */ 765 int 766 vm_map_entry_kreserve(int count) 767 { 768 struct globaldata *gd = mycpu; 769 770 crit_enter(); 771 gd->gd_vme_avail -= count; 772 crit_exit(); 773 KASSERT(gd->gd_vme_base != NULL, 774 ("no reserved entries left, gd_vme_avail = %d", 775 gd->gd_vme_avail)); 776 return(count); 777 } 778 779 /* 780 * Release previously reserved map entries for kernel_map. We do not 781 * attempt to clean up like the normal release function as this would 782 * cause an unnecessary (but probably not fatal) deep procedure call. 783 * 784 * No requirements. 785 */ 786 void 787 vm_map_entry_krelease(int count) 788 { 789 struct globaldata *gd = mycpu; 790 791 crit_enter(); 792 gd->gd_vme_avail += count; 793 crit_exit(); 794 } 795 796 /* 797 * Allocates a VM map entry for insertion. No entry fields are filled in. 798 * 799 * The entries should have previously been reserved. The reservation count 800 * is tracked in (*countp). 801 * 802 * No requirements. 803 */ 804 static vm_map_entry_t 805 vm_map_entry_create(vm_map_t map, int *countp) 806 { 807 struct globaldata *gd = mycpu; 808 vm_map_entry_t entry; 809 810 KKASSERT(*countp > 0); 811 --*countp; 812 crit_enter(); 813 entry = gd->gd_vme_base; 814 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 815 gd->gd_vme_base = entry->next; 816 crit_exit(); 817 818 return(entry); 819 } 820 821 /* 822 * Dispose of a vm_map_entry that is no longer being referenced. 823 * 824 * No requirements. 825 */ 826 static void 827 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 828 { 829 struct globaldata *gd = mycpu; 830 831 KKASSERT(map->hint != entry); 832 KKASSERT(map->first_free != entry); 833 834 ++*countp; 835 crit_enter(); 836 entry->next = gd->gd_vme_base; 837 gd->gd_vme_base = entry; 838 crit_exit(); 839 } 840 841 842 /* 843 * Insert/remove entries from maps. 844 * 845 * The related map must be exclusively locked. 846 * The caller must hold map->token 847 * No other requirements. 848 */ 849 static __inline void 850 vm_map_entry_link(vm_map_t map, 851 vm_map_entry_t after_where, 852 vm_map_entry_t entry) 853 { 854 ASSERT_VM_MAP_LOCKED(map); 855 856 map->nentries++; 857 entry->prev = after_where; 858 entry->next = after_where->next; 859 entry->next->prev = entry; 860 after_where->next = entry; 861 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry)) 862 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry); 863 } 864 865 static __inline void 866 vm_map_entry_unlink(vm_map_t map, 867 vm_map_entry_t entry) 868 { 869 vm_map_entry_t prev; 870 vm_map_entry_t next; 871 872 ASSERT_VM_MAP_LOCKED(map); 873 874 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 875 panic("vm_map_entry_unlink: attempt to mess with " 876 "locked entry! %p", entry); 877 } 878 prev = entry->prev; 879 next = entry->next; 880 next->prev = prev; 881 prev->next = next; 882 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry); 883 map->nentries--; 884 } 885 886 /* 887 * Finds the map entry containing (or immediately preceding) the specified 888 * address in the given map. The entry is returned in (*entry). 889 * 890 * The boolean result indicates whether the address is actually contained 891 * in the map. 892 * 893 * The related map must be locked. 894 * No other requirements. 895 */ 896 boolean_t 897 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry) 898 { 899 vm_map_entry_t tmp; 900 vm_map_entry_t last; 901 902 ASSERT_VM_MAP_LOCKED(map); 903 #if 0 904 /* 905 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive 906 * the hint code with the red-black lookup meets with system crashes 907 * and lockups. We do not yet know why. 908 * 909 * It is possible that the problem is related to the setting 910 * of the hint during map_entry deletion, in the code specified 911 * at the GGG comment later on in this file. 912 * 913 * YYY More likely it's because this function can be called with 914 * a shared lock on the map, resulting in map->hint updates possibly 915 * racing. Fixed now but untested. 916 */ 917 /* 918 * Quickly check the cached hint, there's a good chance of a match. 919 */ 920 tmp = map->hint; 921 cpu_ccfence(); 922 if (tmp != &map->header) { 923 if (address >= tmp->start && address < tmp->end) { 924 *entry = tmp; 925 return(TRUE); 926 } 927 } 928 #endif 929 930 /* 931 * Locate the record from the top of the tree. 'last' tracks the 932 * closest prior record and is returned if no match is found, which 933 * in binary tree terms means tracking the most recent right-branch 934 * taken. If there is no prior record, &map->header is returned. 935 */ 936 last = &map->header; 937 tmp = RB_ROOT(&map->rb_root); 938 939 while (tmp) { 940 if (address >= tmp->start) { 941 if (address < tmp->end) { 942 *entry = tmp; 943 map->hint = tmp; 944 return(TRUE); 945 } 946 last = tmp; 947 tmp = RB_RIGHT(tmp, rb_entry); 948 } else { 949 tmp = RB_LEFT(tmp, rb_entry); 950 } 951 } 952 *entry = last; 953 return (FALSE); 954 } 955 956 /* 957 * Inserts the given whole VM object into the target map at the specified 958 * address range. The object's size should match that of the address range. 959 * 960 * The map must be exclusively locked. 961 * The object must be held. 962 * The caller must have reserved sufficient vm_map_entry structures. 963 * 964 * If object is non-NULL, ref count must be bumped by caller prior to 965 * making call to account for the new entry. 966 */ 967 int 968 vm_map_insert(vm_map_t map, int *countp, 969 vm_object_t object, vm_ooffset_t offset, 970 vm_offset_t start, vm_offset_t end, 971 vm_maptype_t maptype, 972 vm_prot_t prot, vm_prot_t max, 973 int cow) 974 { 975 vm_map_entry_t new_entry; 976 vm_map_entry_t prev_entry; 977 vm_map_entry_t temp_entry; 978 vm_eflags_t protoeflags; 979 int must_drop = 0; 980 981 ASSERT_VM_MAP_LOCKED(map); 982 if (object) 983 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 984 985 /* 986 * Check that the start and end points are not bogus. 987 */ 988 if ((start < map->min_offset) || (end > map->max_offset) || 989 (start >= end)) 990 return (KERN_INVALID_ADDRESS); 991 992 /* 993 * Find the entry prior to the proposed starting address; if it's part 994 * of an existing entry, this range is bogus. 995 */ 996 if (vm_map_lookup_entry(map, start, &temp_entry)) 997 return (KERN_NO_SPACE); 998 999 prev_entry = temp_entry; 1000 1001 /* 1002 * Assert that the next entry doesn't overlap the end point. 1003 */ 1004 1005 if ((prev_entry->next != &map->header) && 1006 (prev_entry->next->start < end)) 1007 return (KERN_NO_SPACE); 1008 1009 protoeflags = 0; 1010 1011 if (cow & MAP_COPY_ON_WRITE) 1012 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 1013 1014 if (cow & MAP_NOFAULT) { 1015 protoeflags |= MAP_ENTRY_NOFAULT; 1016 1017 KASSERT(object == NULL, 1018 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 1019 } 1020 if (cow & MAP_DISABLE_SYNCER) 1021 protoeflags |= MAP_ENTRY_NOSYNC; 1022 if (cow & MAP_DISABLE_COREDUMP) 1023 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1024 if (cow & MAP_IS_STACK) 1025 protoeflags |= MAP_ENTRY_STACK; 1026 if (cow & MAP_IS_KSTACK) 1027 protoeflags |= MAP_ENTRY_KSTACK; 1028 1029 lwkt_gettoken(&map->token); 1030 1031 if (object) { 1032 /* 1033 * When object is non-NULL, it could be shared with another 1034 * process. We have to set or clear OBJ_ONEMAPPING 1035 * appropriately. 1036 * 1037 * NOTE: This flag is only applicable to DEFAULT and SWAP 1038 * objects and will already be clear in other types 1039 * of objects, so a shared object lock is ok for 1040 * VNODE objects. 1041 */ 1042 if ((object->ref_count > 1) || (object->shadow_count != 0)) { 1043 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1044 } 1045 } 1046 else if ((prev_entry != &map->header) && 1047 (prev_entry->eflags == protoeflags) && 1048 (prev_entry->end == start) && 1049 (prev_entry->wired_count == 0) && 1050 prev_entry->maptype == maptype && 1051 ((prev_entry->object.vm_object == NULL) || 1052 vm_object_coalesce(prev_entry->object.vm_object, 1053 OFF_TO_IDX(prev_entry->offset), 1054 (vm_size_t)(prev_entry->end - prev_entry->start), 1055 (vm_size_t)(end - prev_entry->end)))) { 1056 /* 1057 * We were able to extend the object. Determine if we 1058 * can extend the previous map entry to include the 1059 * new range as well. 1060 */ 1061 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 1062 (prev_entry->protection == prot) && 1063 (prev_entry->max_protection == max)) { 1064 map->size += (end - prev_entry->end); 1065 prev_entry->end = end; 1066 vm_map_simplify_entry(map, prev_entry, countp); 1067 lwkt_reltoken(&map->token); 1068 return (KERN_SUCCESS); 1069 } 1070 1071 /* 1072 * If we can extend the object but cannot extend the 1073 * map entry, we have to create a new map entry. We 1074 * must bump the ref count on the extended object to 1075 * account for it. object may be NULL. 1076 */ 1077 object = prev_entry->object.vm_object; 1078 offset = prev_entry->offset + 1079 (prev_entry->end - prev_entry->start); 1080 if (object) { 1081 vm_object_hold(object); 1082 vm_object_chain_wait(object, 0); 1083 vm_object_reference_locked(object); 1084 must_drop = 1; 1085 } 1086 } 1087 1088 /* 1089 * NOTE: if conditionals fail, object can be NULL here. This occurs 1090 * in things like the buffer map where we manage kva but do not manage 1091 * backing objects. 1092 */ 1093 1094 /* 1095 * Create a new entry 1096 */ 1097 1098 new_entry = vm_map_entry_create(map, countp); 1099 new_entry->start = start; 1100 new_entry->end = end; 1101 1102 new_entry->maptype = maptype; 1103 new_entry->eflags = protoeflags; 1104 new_entry->object.vm_object = object; 1105 new_entry->offset = offset; 1106 new_entry->aux.master_pde = 0; 1107 1108 new_entry->inheritance = VM_INHERIT_DEFAULT; 1109 new_entry->protection = prot; 1110 new_entry->max_protection = max; 1111 new_entry->wired_count = 0; 1112 1113 /* 1114 * Insert the new entry into the list 1115 */ 1116 1117 vm_map_entry_link(map, prev_entry, new_entry); 1118 map->size += new_entry->end - new_entry->start; 1119 1120 /* 1121 * Update the free space hint. Entries cannot overlap. 1122 * An exact comparison is needed to avoid matching 1123 * against the map->header. 1124 */ 1125 if ((map->first_free == prev_entry) && 1126 (prev_entry->end == new_entry->start)) { 1127 map->first_free = new_entry; 1128 } 1129 1130 #if 0 1131 /* 1132 * Temporarily removed to avoid MAP_STACK panic, due to 1133 * MAP_STACK being a huge hack. Will be added back in 1134 * when MAP_STACK (and the user stack mapping) is fixed. 1135 */ 1136 /* 1137 * It may be possible to simplify the entry 1138 */ 1139 vm_map_simplify_entry(map, new_entry, countp); 1140 #endif 1141 1142 /* 1143 * Try to pre-populate the page table. Mappings governed by virtual 1144 * page tables cannot be prepopulated without a lot of work, so 1145 * don't try. 1146 */ 1147 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) && 1148 maptype != VM_MAPTYPE_VPAGETABLE) { 1149 int dorelock = 0; 1150 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) { 1151 dorelock = 1; 1152 vm_object_lock_swap(); 1153 vm_object_drop(object); 1154 } 1155 pmap_object_init_pt(map->pmap, start, prot, 1156 object, OFF_TO_IDX(offset), end - start, 1157 cow & MAP_PREFAULT_PARTIAL); 1158 if (dorelock) { 1159 vm_object_hold(object); 1160 vm_object_lock_swap(); 1161 } 1162 } 1163 if (must_drop) 1164 vm_object_drop(object); 1165 1166 lwkt_reltoken(&map->token); 1167 return (KERN_SUCCESS); 1168 } 1169 1170 /* 1171 * Find sufficient space for `length' bytes in the given map, starting at 1172 * `start'. Returns 0 on success, 1 on no space. 1173 * 1174 * This function will returned an arbitrarily aligned pointer. If no 1175 * particular alignment is required you should pass align as 1. Note that 1176 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 1177 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 1178 * argument. 1179 * 1180 * 'align' should be a power of 2 but is not required to be. 1181 * 1182 * The map must be exclusively locked. 1183 * No other requirements. 1184 */ 1185 int 1186 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length, 1187 vm_size_t align, int flags, vm_offset_t *addr) 1188 { 1189 vm_map_entry_t entry, next; 1190 vm_offset_t end; 1191 vm_offset_t align_mask; 1192 1193 if (start < map->min_offset) 1194 start = map->min_offset; 1195 if (start > map->max_offset) 1196 return (1); 1197 1198 /* 1199 * If the alignment is not a power of 2 we will have to use 1200 * a mod/division, set align_mask to a special value. 1201 */ 1202 if ((align | (align - 1)) + 1 != (align << 1)) 1203 align_mask = (vm_offset_t)-1; 1204 else 1205 align_mask = align - 1; 1206 1207 /* 1208 * Look for the first possible address; if there's already something 1209 * at this address, we have to start after it. 1210 */ 1211 if (start == map->min_offset) { 1212 if ((entry = map->first_free) != &map->header) 1213 start = entry->end; 1214 } else { 1215 vm_map_entry_t tmp; 1216 1217 if (vm_map_lookup_entry(map, start, &tmp)) 1218 start = tmp->end; 1219 entry = tmp; 1220 } 1221 1222 /* 1223 * Look through the rest of the map, trying to fit a new region in the 1224 * gap between existing regions, or after the very last region. 1225 */ 1226 for (;; start = (entry = next)->end) { 1227 /* 1228 * Adjust the proposed start by the requested alignment, 1229 * be sure that we didn't wrap the address. 1230 */ 1231 if (align_mask == (vm_offset_t)-1) 1232 end = ((start + align - 1) / align) * align; 1233 else 1234 end = (start + align_mask) & ~align_mask; 1235 if (end < start) 1236 return (1); 1237 start = end; 1238 /* 1239 * Find the end of the proposed new region. Be sure we didn't 1240 * go beyond the end of the map, or wrap around the address. 1241 * Then check to see if this is the last entry or if the 1242 * proposed end fits in the gap between this and the next 1243 * entry. 1244 */ 1245 end = start + length; 1246 if (end > map->max_offset || end < start) 1247 return (1); 1248 next = entry->next; 1249 1250 /* 1251 * If the next entry's start address is beyond the desired 1252 * end address we may have found a good entry. 1253 * 1254 * If the next entry is a stack mapping we do not map into 1255 * the stack's reserved space. 1256 * 1257 * XXX continue to allow mapping into the stack's reserved 1258 * space if doing a MAP_STACK mapping inside a MAP_STACK 1259 * mapping, for backwards compatibility. But the caller 1260 * really should use MAP_STACK | MAP_TRYFIXED if they 1261 * want to do that. 1262 */ 1263 if (next == &map->header) 1264 break; 1265 if (next->start >= end) { 1266 if ((next->eflags & MAP_ENTRY_STACK) == 0) 1267 break; 1268 if (flags & MAP_STACK) 1269 break; 1270 if (next->start - next->aux.avail_ssize >= end) 1271 break; 1272 } 1273 } 1274 map->hint = entry; 1275 1276 /* 1277 * Grow the kernel_map if necessary. pmap_growkernel() will panic 1278 * if it fails. The kernel_map is locked and nothing can steal 1279 * our address space if pmap_growkernel() blocks. 1280 * 1281 * NOTE: This may be unconditionally called for kldload areas on 1282 * x86_64 because these do not bump kernel_vm_end (which would 1283 * fill 128G worth of page tables!). Therefore we must not 1284 * retry. 1285 */ 1286 if (map == &kernel_map) { 1287 vm_offset_t kstop; 1288 1289 kstop = round_page(start + length); 1290 if (kstop > kernel_vm_end) 1291 pmap_growkernel(start, kstop); 1292 } 1293 *addr = start; 1294 return (0); 1295 } 1296 1297 /* 1298 * vm_map_find finds an unallocated region in the target address map with 1299 * the given length and allocates it. The search is defined to be first-fit 1300 * from the specified address; the region found is returned in the same 1301 * parameter. 1302 * 1303 * If object is non-NULL, ref count must be bumped by caller 1304 * prior to making call to account for the new entry. 1305 * 1306 * No requirements. This function will lock the map temporarily. 1307 */ 1308 int 1309 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1310 vm_offset_t *addr, vm_size_t length, vm_size_t align, 1311 boolean_t fitit, 1312 vm_maptype_t maptype, 1313 vm_prot_t prot, vm_prot_t max, 1314 int cow) 1315 { 1316 vm_offset_t start; 1317 int result; 1318 int count; 1319 1320 start = *addr; 1321 1322 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1323 vm_map_lock(map); 1324 if (object) 1325 vm_object_hold_shared(object); 1326 if (fitit) { 1327 if (vm_map_findspace(map, start, length, align, 0, addr)) { 1328 if (object) 1329 vm_object_drop(object); 1330 vm_map_unlock(map); 1331 vm_map_entry_release(count); 1332 return (KERN_NO_SPACE); 1333 } 1334 start = *addr; 1335 } 1336 result = vm_map_insert(map, &count, object, offset, 1337 start, start + length, 1338 maptype, 1339 prot, max, 1340 cow); 1341 if (object) 1342 vm_object_drop(object); 1343 vm_map_unlock(map); 1344 vm_map_entry_release(count); 1345 1346 return (result); 1347 } 1348 1349 /* 1350 * Simplify the given map entry by merging with either neighbor. This 1351 * routine also has the ability to merge with both neighbors. 1352 * 1353 * This routine guarentees that the passed entry remains valid (though 1354 * possibly extended). When merging, this routine may delete one or 1355 * both neighbors. No action is taken on entries which have their 1356 * in-transition flag set. 1357 * 1358 * The map must be exclusively locked. 1359 */ 1360 void 1361 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 1362 { 1363 vm_map_entry_t next, prev; 1364 vm_size_t prevsize, esize; 1365 1366 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1367 ++mycpu->gd_cnt.v_intrans_coll; 1368 return; 1369 } 1370 1371 if (entry->maptype == VM_MAPTYPE_SUBMAP) 1372 return; 1373 1374 prev = entry->prev; 1375 if (prev != &map->header) { 1376 prevsize = prev->end - prev->start; 1377 if ( (prev->end == entry->start) && 1378 (prev->maptype == entry->maptype) && 1379 (prev->object.vm_object == entry->object.vm_object) && 1380 (!prev->object.vm_object || 1381 (prev->offset + prevsize == entry->offset)) && 1382 (prev->eflags == entry->eflags) && 1383 (prev->protection == entry->protection) && 1384 (prev->max_protection == entry->max_protection) && 1385 (prev->inheritance == entry->inheritance) && 1386 (prev->wired_count == entry->wired_count)) { 1387 if (map->first_free == prev) 1388 map->first_free = entry; 1389 if (map->hint == prev) 1390 map->hint = entry; 1391 vm_map_entry_unlink(map, prev); 1392 entry->start = prev->start; 1393 entry->offset = prev->offset; 1394 if (prev->object.vm_object) 1395 vm_object_deallocate(prev->object.vm_object); 1396 vm_map_entry_dispose(map, prev, countp); 1397 } 1398 } 1399 1400 next = entry->next; 1401 if (next != &map->header) { 1402 esize = entry->end - entry->start; 1403 if ((entry->end == next->start) && 1404 (next->maptype == entry->maptype) && 1405 (next->object.vm_object == entry->object.vm_object) && 1406 (!entry->object.vm_object || 1407 (entry->offset + esize == next->offset)) && 1408 (next->eflags == entry->eflags) && 1409 (next->protection == entry->protection) && 1410 (next->max_protection == entry->max_protection) && 1411 (next->inheritance == entry->inheritance) && 1412 (next->wired_count == entry->wired_count)) { 1413 if (map->first_free == next) 1414 map->first_free = entry; 1415 if (map->hint == next) 1416 map->hint = entry; 1417 vm_map_entry_unlink(map, next); 1418 entry->end = next->end; 1419 if (next->object.vm_object) 1420 vm_object_deallocate(next->object.vm_object); 1421 vm_map_entry_dispose(map, next, countp); 1422 } 1423 } 1424 } 1425 1426 /* 1427 * Asserts that the given entry begins at or after the specified address. 1428 * If necessary, it splits the entry into two. 1429 */ 1430 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1431 { \ 1432 if (startaddr > entry->start) \ 1433 _vm_map_clip_start(map, entry, startaddr, countp); \ 1434 } 1435 1436 /* 1437 * This routine is called only when it is known that the entry must be split. 1438 * 1439 * The map must be exclusively locked. 1440 */ 1441 static void 1442 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, 1443 int *countp) 1444 { 1445 vm_map_entry_t new_entry; 1446 1447 /* 1448 * Split off the front portion -- note that we must insert the new 1449 * entry BEFORE this one, so that this entry has the specified 1450 * starting address. 1451 */ 1452 1453 vm_map_simplify_entry(map, entry, countp); 1454 1455 /* 1456 * If there is no object backing this entry, we might as well create 1457 * one now. If we defer it, an object can get created after the map 1458 * is clipped, and individual objects will be created for the split-up 1459 * map. This is a bit of a hack, but is also about the best place to 1460 * put this improvement. 1461 */ 1462 if (entry->object.vm_object == NULL && !map->system_map) { 1463 vm_map_entry_allocate_object(entry); 1464 } 1465 1466 new_entry = vm_map_entry_create(map, countp); 1467 *new_entry = *entry; 1468 1469 new_entry->end = start; 1470 entry->offset += (start - entry->start); 1471 entry->start = start; 1472 1473 vm_map_entry_link(map, entry->prev, new_entry); 1474 1475 switch(entry->maptype) { 1476 case VM_MAPTYPE_NORMAL: 1477 case VM_MAPTYPE_VPAGETABLE: 1478 if (new_entry->object.vm_object) { 1479 vm_object_hold(new_entry->object.vm_object); 1480 vm_object_chain_wait(new_entry->object.vm_object, 0); 1481 vm_object_reference_locked(new_entry->object.vm_object); 1482 vm_object_drop(new_entry->object.vm_object); 1483 } 1484 break; 1485 default: 1486 break; 1487 } 1488 } 1489 1490 /* 1491 * Asserts that the given entry ends at or before the specified address. 1492 * If necessary, it splits the entry into two. 1493 * 1494 * The map must be exclusively locked. 1495 */ 1496 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1497 { \ 1498 if (endaddr < entry->end) \ 1499 _vm_map_clip_end(map, entry, endaddr, countp); \ 1500 } 1501 1502 /* 1503 * This routine is called only when it is known that the entry must be split. 1504 * 1505 * The map must be exclusively locked. 1506 */ 1507 static void 1508 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, 1509 int *countp) 1510 { 1511 vm_map_entry_t new_entry; 1512 1513 /* 1514 * If there is no object backing this entry, we might as well create 1515 * one now. If we defer it, an object can get created after the map 1516 * is clipped, and individual objects will be created for the split-up 1517 * map. This is a bit of a hack, but is also about the best place to 1518 * put this improvement. 1519 */ 1520 1521 if (entry->object.vm_object == NULL && !map->system_map) { 1522 vm_map_entry_allocate_object(entry); 1523 } 1524 1525 /* 1526 * Create a new entry and insert it AFTER the specified entry 1527 */ 1528 1529 new_entry = vm_map_entry_create(map, countp); 1530 *new_entry = *entry; 1531 1532 new_entry->start = entry->end = end; 1533 new_entry->offset += (end - entry->start); 1534 1535 vm_map_entry_link(map, entry, new_entry); 1536 1537 switch(entry->maptype) { 1538 case VM_MAPTYPE_NORMAL: 1539 case VM_MAPTYPE_VPAGETABLE: 1540 if (new_entry->object.vm_object) { 1541 vm_object_hold(new_entry->object.vm_object); 1542 vm_object_chain_wait(new_entry->object.vm_object, 0); 1543 vm_object_reference_locked(new_entry->object.vm_object); 1544 vm_object_drop(new_entry->object.vm_object); 1545 } 1546 break; 1547 default: 1548 break; 1549 } 1550 } 1551 1552 /* 1553 * Asserts that the starting and ending region addresses fall within the 1554 * valid range for the map. 1555 */ 1556 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1557 { \ 1558 if (start < vm_map_min(map)) \ 1559 start = vm_map_min(map); \ 1560 if (end > vm_map_max(map)) \ 1561 end = vm_map_max(map); \ 1562 if (start > end) \ 1563 start = end; \ 1564 } 1565 1566 /* 1567 * Used to block when an in-transition collison occurs. The map 1568 * is unlocked for the sleep and relocked before the return. 1569 */ 1570 void 1571 vm_map_transition_wait(vm_map_t map) 1572 { 1573 tsleep_interlock(map, 0); 1574 vm_map_unlock(map); 1575 tsleep(map, PINTERLOCKED, "vment", 0); 1576 vm_map_lock(map); 1577 } 1578 1579 /* 1580 * When we do blocking operations with the map lock held it is 1581 * possible that a clip might have occured on our in-transit entry, 1582 * requiring an adjustment to the entry in our loop. These macros 1583 * help the pageable and clip_range code deal with the case. The 1584 * conditional costs virtually nothing if no clipping has occured. 1585 */ 1586 1587 #define CLIP_CHECK_BACK(entry, save_start) \ 1588 do { \ 1589 while (entry->start != save_start) { \ 1590 entry = entry->prev; \ 1591 KASSERT(entry != &map->header, ("bad entry clip")); \ 1592 } \ 1593 } while(0) 1594 1595 #define CLIP_CHECK_FWD(entry, save_end) \ 1596 do { \ 1597 while (entry->end != save_end) { \ 1598 entry = entry->next; \ 1599 KASSERT(entry != &map->header, ("bad entry clip")); \ 1600 } \ 1601 } while(0) 1602 1603 1604 /* 1605 * Clip the specified range and return the base entry. The 1606 * range may cover several entries starting at the returned base 1607 * and the first and last entry in the covering sequence will be 1608 * properly clipped to the requested start and end address. 1609 * 1610 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1611 * flag. 1612 * 1613 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1614 * covered by the requested range. 1615 * 1616 * The map must be exclusively locked on entry and will remain locked 1617 * on return. If no range exists or the range contains holes and you 1618 * specified that no holes were allowed, NULL will be returned. This 1619 * routine may temporarily unlock the map in order avoid a deadlock when 1620 * sleeping. 1621 */ 1622 static 1623 vm_map_entry_t 1624 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1625 int *countp, int flags) 1626 { 1627 vm_map_entry_t start_entry; 1628 vm_map_entry_t entry; 1629 1630 /* 1631 * Locate the entry and effect initial clipping. The in-transition 1632 * case does not occur very often so do not try to optimize it. 1633 */ 1634 again: 1635 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1636 return (NULL); 1637 entry = start_entry; 1638 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1639 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1640 ++mycpu->gd_cnt.v_intrans_coll; 1641 ++mycpu->gd_cnt.v_intrans_wait; 1642 vm_map_transition_wait(map); 1643 /* 1644 * entry and/or start_entry may have been clipped while 1645 * we slept, or may have gone away entirely. We have 1646 * to restart from the lookup. 1647 */ 1648 goto again; 1649 } 1650 1651 /* 1652 * Since we hold an exclusive map lock we do not have to restart 1653 * after clipping, even though clipping may block in zalloc. 1654 */ 1655 vm_map_clip_start(map, entry, start, countp); 1656 vm_map_clip_end(map, entry, end, countp); 1657 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1658 1659 /* 1660 * Scan entries covered by the range. When working on the next 1661 * entry a restart need only re-loop on the current entry which 1662 * we have already locked, since 'next' may have changed. Also, 1663 * even though entry is safe, it may have been clipped so we 1664 * have to iterate forwards through the clip after sleeping. 1665 */ 1666 while (entry->next != &map->header && entry->next->start < end) { 1667 vm_map_entry_t next = entry->next; 1668 1669 if (flags & MAP_CLIP_NO_HOLES) { 1670 if (next->start > entry->end) { 1671 vm_map_unclip_range(map, start_entry, 1672 start, entry->end, countp, flags); 1673 return(NULL); 1674 } 1675 } 1676 1677 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1678 vm_offset_t save_end = entry->end; 1679 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1680 ++mycpu->gd_cnt.v_intrans_coll; 1681 ++mycpu->gd_cnt.v_intrans_wait; 1682 vm_map_transition_wait(map); 1683 1684 /* 1685 * clips might have occured while we blocked. 1686 */ 1687 CLIP_CHECK_FWD(entry, save_end); 1688 CLIP_CHECK_BACK(start_entry, start); 1689 continue; 1690 } 1691 /* 1692 * No restart necessary even though clip_end may block, we 1693 * are holding the map lock. 1694 */ 1695 vm_map_clip_end(map, next, end, countp); 1696 next->eflags |= MAP_ENTRY_IN_TRANSITION; 1697 entry = next; 1698 } 1699 if (flags & MAP_CLIP_NO_HOLES) { 1700 if (entry->end != end) { 1701 vm_map_unclip_range(map, start_entry, 1702 start, entry->end, countp, flags); 1703 return(NULL); 1704 } 1705 } 1706 return(start_entry); 1707 } 1708 1709 /* 1710 * Undo the effect of vm_map_clip_range(). You should pass the same 1711 * flags and the same range that you passed to vm_map_clip_range(). 1712 * This code will clear the in-transition flag on the entries and 1713 * wake up anyone waiting. This code will also simplify the sequence 1714 * and attempt to merge it with entries before and after the sequence. 1715 * 1716 * The map must be locked on entry and will remain locked on return. 1717 * 1718 * Note that you should also pass the start_entry returned by 1719 * vm_map_clip_range(). However, if you block between the two calls 1720 * with the map unlocked please be aware that the start_entry may 1721 * have been clipped and you may need to scan it backwards to find 1722 * the entry corresponding with the original start address. You are 1723 * responsible for this, vm_map_unclip_range() expects the correct 1724 * start_entry to be passed to it and will KASSERT otherwise. 1725 */ 1726 static 1727 void 1728 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry, 1729 vm_offset_t start, vm_offset_t end, 1730 int *countp, int flags) 1731 { 1732 vm_map_entry_t entry; 1733 1734 entry = start_entry; 1735 1736 KASSERT(entry->start == start, ("unclip_range: illegal base entry")); 1737 while (entry != &map->header && entry->start < end) { 1738 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, 1739 ("in-transition flag not set during unclip on: %p", 1740 entry)); 1741 KASSERT(entry->end <= end, 1742 ("unclip_range: tail wasn't clipped")); 1743 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 1744 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 1745 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 1746 wakeup(map); 1747 } 1748 entry = entry->next; 1749 } 1750 1751 /* 1752 * Simplification does not block so there is no restart case. 1753 */ 1754 entry = start_entry; 1755 while (entry != &map->header && entry->start < end) { 1756 vm_map_simplify_entry(map, entry, countp); 1757 entry = entry->next; 1758 } 1759 } 1760 1761 /* 1762 * Mark the given range as handled by a subordinate map. 1763 * 1764 * This range must have been created with vm_map_find(), and no other 1765 * operations may have been performed on this range prior to calling 1766 * vm_map_submap(). 1767 * 1768 * Submappings cannot be removed. 1769 * 1770 * No requirements. 1771 */ 1772 int 1773 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 1774 { 1775 vm_map_entry_t entry; 1776 int result = KERN_INVALID_ARGUMENT; 1777 int count; 1778 1779 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1780 vm_map_lock(map); 1781 1782 VM_MAP_RANGE_CHECK(map, start, end); 1783 1784 if (vm_map_lookup_entry(map, start, &entry)) { 1785 vm_map_clip_start(map, entry, start, &count); 1786 } else { 1787 entry = entry->next; 1788 } 1789 1790 vm_map_clip_end(map, entry, end, &count); 1791 1792 if ((entry->start == start) && (entry->end == end) && 1793 ((entry->eflags & MAP_ENTRY_COW) == 0) && 1794 (entry->object.vm_object == NULL)) { 1795 entry->object.sub_map = submap; 1796 entry->maptype = VM_MAPTYPE_SUBMAP; 1797 result = KERN_SUCCESS; 1798 } 1799 vm_map_unlock(map); 1800 vm_map_entry_release(count); 1801 1802 return (result); 1803 } 1804 1805 /* 1806 * Sets the protection of the specified address region in the target map. 1807 * If "set_max" is specified, the maximum protection is to be set; 1808 * otherwise, only the current protection is affected. 1809 * 1810 * The protection is not applicable to submaps, but is applicable to normal 1811 * maps and maps governed by virtual page tables. For example, when operating 1812 * on a virtual page table our protection basically controls how COW occurs 1813 * on the backing object, whereas the virtual page table abstraction itself 1814 * is an abstraction for userland. 1815 * 1816 * No requirements. 1817 */ 1818 int 1819 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 1820 vm_prot_t new_prot, boolean_t set_max) 1821 { 1822 vm_map_entry_t current; 1823 vm_map_entry_t entry; 1824 int count; 1825 1826 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1827 vm_map_lock(map); 1828 1829 VM_MAP_RANGE_CHECK(map, start, end); 1830 1831 if (vm_map_lookup_entry(map, start, &entry)) { 1832 vm_map_clip_start(map, entry, start, &count); 1833 } else { 1834 entry = entry->next; 1835 } 1836 1837 /* 1838 * Make a first pass to check for protection violations. 1839 */ 1840 current = entry; 1841 while ((current != &map->header) && (current->start < end)) { 1842 if (current->maptype == VM_MAPTYPE_SUBMAP) { 1843 vm_map_unlock(map); 1844 vm_map_entry_release(count); 1845 return (KERN_INVALID_ARGUMENT); 1846 } 1847 if ((new_prot & current->max_protection) != new_prot) { 1848 vm_map_unlock(map); 1849 vm_map_entry_release(count); 1850 return (KERN_PROTECTION_FAILURE); 1851 } 1852 current = current->next; 1853 } 1854 1855 /* 1856 * Go back and fix up protections. [Note that clipping is not 1857 * necessary the second time.] 1858 */ 1859 current = entry; 1860 1861 while ((current != &map->header) && (current->start < end)) { 1862 vm_prot_t old_prot; 1863 1864 vm_map_clip_end(map, current, end, &count); 1865 1866 old_prot = current->protection; 1867 if (set_max) { 1868 current->protection = 1869 (current->max_protection = new_prot) & 1870 old_prot; 1871 } else { 1872 current->protection = new_prot; 1873 } 1874 1875 /* 1876 * Update physical map if necessary. Worry about copy-on-write 1877 * here -- CHECK THIS XXX 1878 */ 1879 1880 if (current->protection != old_prot) { 1881 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 1882 VM_PROT_ALL) 1883 1884 pmap_protect(map->pmap, current->start, 1885 current->end, 1886 current->protection & MASK(current)); 1887 #undef MASK 1888 } 1889 1890 vm_map_simplify_entry(map, current, &count); 1891 1892 current = current->next; 1893 } 1894 1895 vm_map_unlock(map); 1896 vm_map_entry_release(count); 1897 return (KERN_SUCCESS); 1898 } 1899 1900 /* 1901 * This routine traverses a processes map handling the madvise 1902 * system call. Advisories are classified as either those effecting 1903 * the vm_map_entry structure, or those effecting the underlying 1904 * objects. 1905 * 1906 * The <value> argument is used for extended madvise calls. 1907 * 1908 * No requirements. 1909 */ 1910 int 1911 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, 1912 int behav, off_t value) 1913 { 1914 vm_map_entry_t current, entry; 1915 int modify_map = 0; 1916 int error = 0; 1917 int count; 1918 1919 /* 1920 * Some madvise calls directly modify the vm_map_entry, in which case 1921 * we need to use an exclusive lock on the map and we need to perform 1922 * various clipping operations. Otherwise we only need a read-lock 1923 * on the map. 1924 */ 1925 1926 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1927 1928 switch(behav) { 1929 case MADV_NORMAL: 1930 case MADV_SEQUENTIAL: 1931 case MADV_RANDOM: 1932 case MADV_NOSYNC: 1933 case MADV_AUTOSYNC: 1934 case MADV_NOCORE: 1935 case MADV_CORE: 1936 case MADV_SETMAP: 1937 case MADV_INVAL: 1938 modify_map = 1; 1939 vm_map_lock(map); 1940 break; 1941 case MADV_WILLNEED: 1942 case MADV_DONTNEED: 1943 case MADV_FREE: 1944 vm_map_lock_read(map); 1945 break; 1946 default: 1947 vm_map_entry_release(count); 1948 return (EINVAL); 1949 } 1950 1951 /* 1952 * Locate starting entry and clip if necessary. 1953 */ 1954 1955 VM_MAP_RANGE_CHECK(map, start, end); 1956 1957 if (vm_map_lookup_entry(map, start, &entry)) { 1958 if (modify_map) 1959 vm_map_clip_start(map, entry, start, &count); 1960 } else { 1961 entry = entry->next; 1962 } 1963 1964 if (modify_map) { 1965 /* 1966 * madvise behaviors that are implemented in the vm_map_entry. 1967 * 1968 * We clip the vm_map_entry so that behavioral changes are 1969 * limited to the specified address range. 1970 */ 1971 for (current = entry; 1972 (current != &map->header) && (current->start < end); 1973 current = current->next 1974 ) { 1975 if (current->maptype == VM_MAPTYPE_SUBMAP) 1976 continue; 1977 1978 vm_map_clip_end(map, current, end, &count); 1979 1980 switch (behav) { 1981 case MADV_NORMAL: 1982 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 1983 break; 1984 case MADV_SEQUENTIAL: 1985 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 1986 break; 1987 case MADV_RANDOM: 1988 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 1989 break; 1990 case MADV_NOSYNC: 1991 current->eflags |= MAP_ENTRY_NOSYNC; 1992 break; 1993 case MADV_AUTOSYNC: 1994 current->eflags &= ~MAP_ENTRY_NOSYNC; 1995 break; 1996 case MADV_NOCORE: 1997 current->eflags |= MAP_ENTRY_NOCOREDUMP; 1998 break; 1999 case MADV_CORE: 2000 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 2001 break; 2002 case MADV_INVAL: 2003 /* 2004 * Invalidate the related pmap entries, used 2005 * to flush portions of the real kernel's 2006 * pmap when the caller has removed or 2007 * modified existing mappings in a virtual 2008 * page table. 2009 */ 2010 pmap_remove(map->pmap, 2011 current->start, current->end); 2012 break; 2013 case MADV_SETMAP: 2014 /* 2015 * Set the page directory page for a map 2016 * governed by a virtual page table. Mark 2017 * the entry as being governed by a virtual 2018 * page table if it is not. 2019 * 2020 * XXX the page directory page is stored 2021 * in the avail_ssize field if the map_entry. 2022 * 2023 * XXX the map simplification code does not 2024 * compare this field so weird things may 2025 * happen if you do not apply this function 2026 * to the entire mapping governed by the 2027 * virtual page table. 2028 */ 2029 if (current->maptype != VM_MAPTYPE_VPAGETABLE) { 2030 error = EINVAL; 2031 break; 2032 } 2033 current->aux.master_pde = value; 2034 pmap_remove(map->pmap, 2035 current->start, current->end); 2036 break; 2037 default: 2038 error = EINVAL; 2039 break; 2040 } 2041 vm_map_simplify_entry(map, current, &count); 2042 } 2043 vm_map_unlock(map); 2044 } else { 2045 vm_pindex_t pindex; 2046 int count; 2047 2048 /* 2049 * madvise behaviors that are implemented in the underlying 2050 * vm_object. 2051 * 2052 * Since we don't clip the vm_map_entry, we have to clip 2053 * the vm_object pindex and count. 2054 * 2055 * NOTE! We currently do not support these functions on 2056 * virtual page tables. 2057 */ 2058 for (current = entry; 2059 (current != &map->header) && (current->start < end); 2060 current = current->next 2061 ) { 2062 vm_offset_t useStart; 2063 2064 if (current->maptype != VM_MAPTYPE_NORMAL) 2065 continue; 2066 2067 pindex = OFF_TO_IDX(current->offset); 2068 count = atop(current->end - current->start); 2069 useStart = current->start; 2070 2071 if (current->start < start) { 2072 pindex += atop(start - current->start); 2073 count -= atop(start - current->start); 2074 useStart = start; 2075 } 2076 if (current->end > end) 2077 count -= atop(current->end - end); 2078 2079 if (count <= 0) 2080 continue; 2081 2082 vm_object_madvise(current->object.vm_object, 2083 pindex, count, behav); 2084 2085 /* 2086 * Try to populate the page table. Mappings governed 2087 * by virtual page tables cannot be pre-populated 2088 * without a lot of work so don't try. 2089 */ 2090 if (behav == MADV_WILLNEED && 2091 current->maptype != VM_MAPTYPE_VPAGETABLE) { 2092 pmap_object_init_pt( 2093 map->pmap, 2094 useStart, 2095 current->protection, 2096 current->object.vm_object, 2097 pindex, 2098 (count << PAGE_SHIFT), 2099 MAP_PREFAULT_MADVISE 2100 ); 2101 } 2102 } 2103 vm_map_unlock_read(map); 2104 } 2105 vm_map_entry_release(count); 2106 return(error); 2107 } 2108 2109 2110 /* 2111 * Sets the inheritance of the specified address range in the target map. 2112 * Inheritance affects how the map will be shared with child maps at the 2113 * time of vm_map_fork. 2114 */ 2115 int 2116 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 2117 vm_inherit_t new_inheritance) 2118 { 2119 vm_map_entry_t entry; 2120 vm_map_entry_t temp_entry; 2121 int count; 2122 2123 switch (new_inheritance) { 2124 case VM_INHERIT_NONE: 2125 case VM_INHERIT_COPY: 2126 case VM_INHERIT_SHARE: 2127 break; 2128 default: 2129 return (KERN_INVALID_ARGUMENT); 2130 } 2131 2132 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2133 vm_map_lock(map); 2134 2135 VM_MAP_RANGE_CHECK(map, start, end); 2136 2137 if (vm_map_lookup_entry(map, start, &temp_entry)) { 2138 entry = temp_entry; 2139 vm_map_clip_start(map, entry, start, &count); 2140 } else 2141 entry = temp_entry->next; 2142 2143 while ((entry != &map->header) && (entry->start < end)) { 2144 vm_map_clip_end(map, entry, end, &count); 2145 2146 entry->inheritance = new_inheritance; 2147 2148 vm_map_simplify_entry(map, entry, &count); 2149 2150 entry = entry->next; 2151 } 2152 vm_map_unlock(map); 2153 vm_map_entry_release(count); 2154 return (KERN_SUCCESS); 2155 } 2156 2157 /* 2158 * Implement the semantics of mlock 2159 */ 2160 int 2161 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 2162 boolean_t new_pageable) 2163 { 2164 vm_map_entry_t entry; 2165 vm_map_entry_t start_entry; 2166 vm_offset_t end; 2167 int rv = KERN_SUCCESS; 2168 int count; 2169 2170 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2171 vm_map_lock(map); 2172 VM_MAP_RANGE_CHECK(map, start, real_end); 2173 end = real_end; 2174 2175 start_entry = vm_map_clip_range(map, start, end, &count, 2176 MAP_CLIP_NO_HOLES); 2177 if (start_entry == NULL) { 2178 vm_map_unlock(map); 2179 vm_map_entry_release(count); 2180 return (KERN_INVALID_ADDRESS); 2181 } 2182 2183 if (new_pageable == 0) { 2184 entry = start_entry; 2185 while ((entry != &map->header) && (entry->start < end)) { 2186 vm_offset_t save_start; 2187 vm_offset_t save_end; 2188 2189 /* 2190 * Already user wired or hard wired (trivial cases) 2191 */ 2192 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 2193 entry = entry->next; 2194 continue; 2195 } 2196 if (entry->wired_count != 0) { 2197 entry->wired_count++; 2198 entry->eflags |= MAP_ENTRY_USER_WIRED; 2199 entry = entry->next; 2200 continue; 2201 } 2202 2203 /* 2204 * A new wiring requires instantiation of appropriate 2205 * management structures and the faulting in of the 2206 * page. 2207 */ 2208 if (entry->maptype != VM_MAPTYPE_SUBMAP) { 2209 int copyflag = entry->eflags & 2210 MAP_ENTRY_NEEDS_COPY; 2211 if (copyflag && ((entry->protection & 2212 VM_PROT_WRITE) != 0)) { 2213 vm_map_entry_shadow(entry, 0); 2214 } else if (entry->object.vm_object == NULL && 2215 !map->system_map) { 2216 vm_map_entry_allocate_object(entry); 2217 } 2218 } 2219 entry->wired_count++; 2220 entry->eflags |= MAP_ENTRY_USER_WIRED; 2221 2222 /* 2223 * Now fault in the area. Note that vm_fault_wire() 2224 * may release the map lock temporarily, it will be 2225 * relocked on return. The in-transition 2226 * flag protects the entries. 2227 */ 2228 save_start = entry->start; 2229 save_end = entry->end; 2230 rv = vm_fault_wire(map, entry, TRUE, 0); 2231 if (rv) { 2232 CLIP_CHECK_BACK(entry, save_start); 2233 for (;;) { 2234 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 2235 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2236 entry->wired_count = 0; 2237 if (entry->end == save_end) 2238 break; 2239 entry = entry->next; 2240 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2241 } 2242 end = save_start; /* unwire the rest */ 2243 break; 2244 } 2245 /* 2246 * note that even though the entry might have been 2247 * clipped, the USER_WIRED flag we set prevents 2248 * duplication so we do not have to do a 2249 * clip check. 2250 */ 2251 entry = entry->next; 2252 } 2253 2254 /* 2255 * If we failed fall through to the unwiring section to 2256 * unwire what we had wired so far. 'end' has already 2257 * been adjusted. 2258 */ 2259 if (rv) 2260 new_pageable = 1; 2261 2262 /* 2263 * start_entry might have been clipped if we unlocked the 2264 * map and blocked. No matter how clipped it has gotten 2265 * there should be a fragment that is on our start boundary. 2266 */ 2267 CLIP_CHECK_BACK(start_entry, start); 2268 } 2269 2270 /* 2271 * Deal with the unwiring case. 2272 */ 2273 if (new_pageable) { 2274 /* 2275 * This is the unwiring case. We must first ensure that the 2276 * range to be unwired is really wired down. We know there 2277 * are no holes. 2278 */ 2279 entry = start_entry; 2280 while ((entry != &map->header) && (entry->start < end)) { 2281 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 2282 rv = KERN_INVALID_ARGUMENT; 2283 goto done; 2284 } 2285 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry)); 2286 entry = entry->next; 2287 } 2288 2289 /* 2290 * Now decrement the wiring count for each region. If a region 2291 * becomes completely unwired, unwire its physical pages and 2292 * mappings. 2293 */ 2294 /* 2295 * The map entries are processed in a loop, checking to 2296 * make sure the entry is wired and asserting it has a wired 2297 * count. However, another loop was inserted more-or-less in 2298 * the middle of the unwiring path. This loop picks up the 2299 * "entry" loop variable from the first loop without first 2300 * setting it to start_entry. Naturally, the secound loop 2301 * is never entered and the pages backing the entries are 2302 * never unwired. This can lead to a leak of wired pages. 2303 */ 2304 entry = start_entry; 2305 while ((entry != &map->header) && (entry->start < end)) { 2306 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 2307 ("expected USER_WIRED on entry %p", entry)); 2308 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2309 entry->wired_count--; 2310 if (entry->wired_count == 0) 2311 vm_fault_unwire(map, entry); 2312 entry = entry->next; 2313 } 2314 } 2315 done: 2316 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2317 MAP_CLIP_NO_HOLES); 2318 map->timestamp++; 2319 vm_map_unlock(map); 2320 vm_map_entry_release(count); 2321 return (rv); 2322 } 2323 2324 /* 2325 * Sets the pageability of the specified address range in the target map. 2326 * Regions specified as not pageable require locked-down physical 2327 * memory and physical page maps. 2328 * 2329 * The map must not be locked, but a reference must remain to the map 2330 * throughout the call. 2331 * 2332 * This function may be called via the zalloc path and must properly 2333 * reserve map entries for kernel_map. 2334 * 2335 * No requirements. 2336 */ 2337 int 2338 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags) 2339 { 2340 vm_map_entry_t entry; 2341 vm_map_entry_t start_entry; 2342 vm_offset_t end; 2343 int rv = KERN_SUCCESS; 2344 int count; 2345 2346 if (kmflags & KM_KRESERVE) 2347 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 2348 else 2349 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2350 vm_map_lock(map); 2351 VM_MAP_RANGE_CHECK(map, start, real_end); 2352 end = real_end; 2353 2354 start_entry = vm_map_clip_range(map, start, end, &count, 2355 MAP_CLIP_NO_HOLES); 2356 if (start_entry == NULL) { 2357 vm_map_unlock(map); 2358 rv = KERN_INVALID_ADDRESS; 2359 goto failure; 2360 } 2361 if ((kmflags & KM_PAGEABLE) == 0) { 2362 /* 2363 * Wiring. 2364 * 2365 * 1. Holding the write lock, we create any shadow or zero-fill 2366 * objects that need to be created. Then we clip each map 2367 * entry to the region to be wired and increment its wiring 2368 * count. We create objects before clipping the map entries 2369 * to avoid object proliferation. 2370 * 2371 * 2. We downgrade to a read lock, and call vm_fault_wire to 2372 * fault in the pages for any newly wired area (wired_count is 2373 * 1). 2374 * 2375 * Downgrading to a read lock for vm_fault_wire avoids a 2376 * possible deadlock with another process that may have faulted 2377 * on one of the pages to be wired (it would mark the page busy, 2378 * blocking us, then in turn block on the map lock that we 2379 * hold). Because of problems in the recursive lock package, 2380 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 2381 * any actions that require the write lock must be done 2382 * beforehand. Because we keep the read lock on the map, the 2383 * copy-on-write status of the entries we modify here cannot 2384 * change. 2385 */ 2386 entry = start_entry; 2387 while ((entry != &map->header) && (entry->start < end)) { 2388 /* 2389 * Trivial case if the entry is already wired 2390 */ 2391 if (entry->wired_count) { 2392 entry->wired_count++; 2393 entry = entry->next; 2394 continue; 2395 } 2396 2397 /* 2398 * The entry is being newly wired, we have to setup 2399 * appropriate management structures. A shadow 2400 * object is required for a copy-on-write region, 2401 * or a normal object for a zero-fill region. We 2402 * do not have to do this for entries that point to sub 2403 * maps because we won't hold the lock on the sub map. 2404 */ 2405 if (entry->maptype != VM_MAPTYPE_SUBMAP) { 2406 int copyflag = entry->eflags & 2407 MAP_ENTRY_NEEDS_COPY; 2408 if (copyflag && ((entry->protection & 2409 VM_PROT_WRITE) != 0)) { 2410 vm_map_entry_shadow(entry, 0); 2411 } else if (entry->object.vm_object == NULL && 2412 !map->system_map) { 2413 vm_map_entry_allocate_object(entry); 2414 } 2415 } 2416 2417 entry->wired_count++; 2418 entry = entry->next; 2419 } 2420 2421 /* 2422 * Pass 2. 2423 */ 2424 2425 /* 2426 * HACK HACK HACK HACK 2427 * 2428 * vm_fault_wire() temporarily unlocks the map to avoid 2429 * deadlocks. The in-transition flag from vm_map_clip_range 2430 * call should protect us from changes while the map is 2431 * unlocked. T 2432 * 2433 * NOTE: Previously this comment stated that clipping might 2434 * still occur while the entry is unlocked, but from 2435 * what I can tell it actually cannot. 2436 * 2437 * It is unclear whether the CLIP_CHECK_*() calls 2438 * are still needed but we keep them in anyway. 2439 * 2440 * HACK HACK HACK HACK 2441 */ 2442 2443 entry = start_entry; 2444 while (entry != &map->header && entry->start < end) { 2445 /* 2446 * If vm_fault_wire fails for any page we need to undo 2447 * what has been done. We decrement the wiring count 2448 * for those pages which have not yet been wired (now) 2449 * and unwire those that have (later). 2450 */ 2451 vm_offset_t save_start = entry->start; 2452 vm_offset_t save_end = entry->end; 2453 2454 if (entry->wired_count == 1) 2455 rv = vm_fault_wire(map, entry, FALSE, kmflags); 2456 if (rv) { 2457 CLIP_CHECK_BACK(entry, save_start); 2458 for (;;) { 2459 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly")); 2460 entry->wired_count = 0; 2461 if (entry->end == save_end) 2462 break; 2463 entry = entry->next; 2464 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2465 } 2466 end = save_start; 2467 break; 2468 } 2469 CLIP_CHECK_FWD(entry, save_end); 2470 entry = entry->next; 2471 } 2472 2473 /* 2474 * If a failure occured undo everything by falling through 2475 * to the unwiring code. 'end' has already been adjusted 2476 * appropriately. 2477 */ 2478 if (rv) 2479 kmflags |= KM_PAGEABLE; 2480 2481 /* 2482 * start_entry is still IN_TRANSITION but may have been 2483 * clipped since vm_fault_wire() unlocks and relocks the 2484 * map. No matter how clipped it has gotten there should 2485 * be a fragment that is on our start boundary. 2486 */ 2487 CLIP_CHECK_BACK(start_entry, start); 2488 } 2489 2490 if (kmflags & KM_PAGEABLE) { 2491 /* 2492 * This is the unwiring case. We must first ensure that the 2493 * range to be unwired is really wired down. We know there 2494 * are no holes. 2495 */ 2496 entry = start_entry; 2497 while ((entry != &map->header) && (entry->start < end)) { 2498 if (entry->wired_count == 0) { 2499 rv = KERN_INVALID_ARGUMENT; 2500 goto done; 2501 } 2502 entry = entry->next; 2503 } 2504 2505 /* 2506 * Now decrement the wiring count for each region. If a region 2507 * becomes completely unwired, unwire its physical pages and 2508 * mappings. 2509 */ 2510 entry = start_entry; 2511 while ((entry != &map->header) && (entry->start < end)) { 2512 entry->wired_count--; 2513 if (entry->wired_count == 0) 2514 vm_fault_unwire(map, entry); 2515 entry = entry->next; 2516 } 2517 } 2518 done: 2519 vm_map_unclip_range(map, start_entry, start, real_end, 2520 &count, MAP_CLIP_NO_HOLES); 2521 map->timestamp++; 2522 vm_map_unlock(map); 2523 failure: 2524 if (kmflags & KM_KRESERVE) 2525 vm_map_entry_krelease(count); 2526 else 2527 vm_map_entry_release(count); 2528 return (rv); 2529 } 2530 2531 /* 2532 * Mark a newly allocated address range as wired but do not fault in 2533 * the pages. The caller is expected to load the pages into the object. 2534 * 2535 * The map must be locked on entry and will remain locked on return. 2536 * No other requirements. 2537 */ 2538 void 2539 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, 2540 int *countp) 2541 { 2542 vm_map_entry_t scan; 2543 vm_map_entry_t entry; 2544 2545 entry = vm_map_clip_range(map, addr, addr + size, 2546 countp, MAP_CLIP_NO_HOLES); 2547 for (scan = entry; 2548 scan != &map->header && scan->start < addr + size; 2549 scan = scan->next) { 2550 KKASSERT(scan->wired_count == 0); 2551 scan->wired_count = 1; 2552 } 2553 vm_map_unclip_range(map, entry, addr, addr + size, 2554 countp, MAP_CLIP_NO_HOLES); 2555 } 2556 2557 /* 2558 * Push any dirty cached pages in the address range to their pager. 2559 * If syncio is TRUE, dirty pages are written synchronously. 2560 * If invalidate is TRUE, any cached pages are freed as well. 2561 * 2562 * This routine is called by sys_msync() 2563 * 2564 * Returns an error if any part of the specified range is not mapped. 2565 * 2566 * No requirements. 2567 */ 2568 int 2569 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, 2570 boolean_t syncio, boolean_t invalidate) 2571 { 2572 vm_map_entry_t current; 2573 vm_map_entry_t entry; 2574 vm_size_t size; 2575 vm_object_t object; 2576 vm_object_t tobj; 2577 vm_ooffset_t offset; 2578 2579 vm_map_lock_read(map); 2580 VM_MAP_RANGE_CHECK(map, start, end); 2581 if (!vm_map_lookup_entry(map, start, &entry)) { 2582 vm_map_unlock_read(map); 2583 return (KERN_INVALID_ADDRESS); 2584 } 2585 lwkt_gettoken(&map->token); 2586 2587 /* 2588 * Make a first pass to check for holes. 2589 */ 2590 for (current = entry; current->start < end; current = current->next) { 2591 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2592 lwkt_reltoken(&map->token); 2593 vm_map_unlock_read(map); 2594 return (KERN_INVALID_ARGUMENT); 2595 } 2596 if (end > current->end && 2597 (current->next == &map->header || 2598 current->end != current->next->start)) { 2599 lwkt_reltoken(&map->token); 2600 vm_map_unlock_read(map); 2601 return (KERN_INVALID_ADDRESS); 2602 } 2603 } 2604 2605 if (invalidate) 2606 pmap_remove(vm_map_pmap(map), start, end); 2607 2608 /* 2609 * Make a second pass, cleaning/uncaching pages from the indicated 2610 * objects as we go. 2611 */ 2612 for (current = entry; current->start < end; current = current->next) { 2613 offset = current->offset + (start - current->start); 2614 size = (end <= current->end ? end : current->end) - start; 2615 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2616 vm_map_t smap; 2617 vm_map_entry_t tentry; 2618 vm_size_t tsize; 2619 2620 smap = current->object.sub_map; 2621 vm_map_lock_read(smap); 2622 vm_map_lookup_entry(smap, offset, &tentry); 2623 tsize = tentry->end - offset; 2624 if (tsize < size) 2625 size = tsize; 2626 object = tentry->object.vm_object; 2627 offset = tentry->offset + (offset - tentry->start); 2628 vm_map_unlock_read(smap); 2629 } else { 2630 object = current->object.vm_object; 2631 } 2632 2633 if (object) 2634 vm_object_hold(object); 2635 2636 /* 2637 * Note that there is absolutely no sense in writing out 2638 * anonymous objects, so we track down the vnode object 2639 * to write out. 2640 * We invalidate (remove) all pages from the address space 2641 * anyway, for semantic correctness. 2642 * 2643 * note: certain anonymous maps, such as MAP_NOSYNC maps, 2644 * may start out with a NULL object. 2645 */ 2646 while (object && (tobj = object->backing_object) != NULL) { 2647 vm_object_hold(tobj); 2648 if (tobj == object->backing_object) { 2649 vm_object_lock_swap(); 2650 offset += object->backing_object_offset; 2651 vm_object_drop(object); 2652 object = tobj; 2653 if (object->size < OFF_TO_IDX(offset + size)) 2654 size = IDX_TO_OFF(object->size) - 2655 offset; 2656 break; 2657 } 2658 vm_object_drop(tobj); 2659 } 2660 if (object && (object->type == OBJT_VNODE) && 2661 (current->protection & VM_PROT_WRITE) && 2662 (object->flags & OBJ_NOMSYNC) == 0) { 2663 /* 2664 * Flush pages if writing is allowed, invalidate them 2665 * if invalidation requested. Pages undergoing I/O 2666 * will be ignored by vm_object_page_remove(). 2667 * 2668 * We cannot lock the vnode and then wait for paging 2669 * to complete without deadlocking against vm_fault. 2670 * Instead we simply call vm_object_page_remove() and 2671 * allow it to block internally on a page-by-page 2672 * basis when it encounters pages undergoing async 2673 * I/O. 2674 */ 2675 int flags; 2676 2677 /* no chain wait needed for vnode objects */ 2678 vm_object_reference_locked(object); 2679 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY); 2680 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 2681 flags |= invalidate ? OBJPC_INVAL : 0; 2682 2683 /* 2684 * When operating on a virtual page table just 2685 * flush the whole object. XXX we probably ought 2686 * to 2687 */ 2688 switch(current->maptype) { 2689 case VM_MAPTYPE_NORMAL: 2690 vm_object_page_clean(object, 2691 OFF_TO_IDX(offset), 2692 OFF_TO_IDX(offset + size + PAGE_MASK), 2693 flags); 2694 break; 2695 case VM_MAPTYPE_VPAGETABLE: 2696 vm_object_page_clean(object, 0, 0, flags); 2697 break; 2698 } 2699 vn_unlock(((struct vnode *)object->handle)); 2700 vm_object_deallocate_locked(object); 2701 } 2702 if (object && invalidate && 2703 ((object->type == OBJT_VNODE) || 2704 (object->type == OBJT_DEVICE) || 2705 (object->type == OBJT_MGTDEVICE))) { 2706 int clean_only = 2707 ((object->type == OBJT_DEVICE) || 2708 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE; 2709 /* no chain wait needed for vnode/device objects */ 2710 vm_object_reference_locked(object); 2711 switch(current->maptype) { 2712 case VM_MAPTYPE_NORMAL: 2713 vm_object_page_remove(object, 2714 OFF_TO_IDX(offset), 2715 OFF_TO_IDX(offset + size + PAGE_MASK), 2716 clean_only); 2717 break; 2718 case VM_MAPTYPE_VPAGETABLE: 2719 vm_object_page_remove(object, 0, 0, clean_only); 2720 break; 2721 } 2722 vm_object_deallocate_locked(object); 2723 } 2724 start += size; 2725 if (object) 2726 vm_object_drop(object); 2727 } 2728 2729 lwkt_reltoken(&map->token); 2730 vm_map_unlock_read(map); 2731 2732 return (KERN_SUCCESS); 2733 } 2734 2735 /* 2736 * Make the region specified by this entry pageable. 2737 * 2738 * The vm_map must be exclusively locked. 2739 */ 2740 static void 2741 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 2742 { 2743 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2744 entry->wired_count = 0; 2745 vm_fault_unwire(map, entry); 2746 } 2747 2748 /* 2749 * Deallocate the given entry from the target map. 2750 * 2751 * The vm_map must be exclusively locked. 2752 */ 2753 static void 2754 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 2755 { 2756 vm_map_entry_unlink(map, entry); 2757 map->size -= entry->end - entry->start; 2758 2759 switch(entry->maptype) { 2760 case VM_MAPTYPE_NORMAL: 2761 case VM_MAPTYPE_VPAGETABLE: 2762 vm_object_deallocate(entry->object.vm_object); 2763 break; 2764 default: 2765 break; 2766 } 2767 2768 vm_map_entry_dispose(map, entry, countp); 2769 } 2770 2771 /* 2772 * Deallocates the given address range from the target map. 2773 * 2774 * The vm_map must be exclusively locked. 2775 */ 2776 int 2777 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 2778 { 2779 vm_object_t object; 2780 vm_map_entry_t entry; 2781 vm_map_entry_t first_entry; 2782 2783 ASSERT_VM_MAP_LOCKED(map); 2784 lwkt_gettoken(&map->token); 2785 again: 2786 /* 2787 * Find the start of the region, and clip it. Set entry to point 2788 * at the first record containing the requested address or, if no 2789 * such record exists, the next record with a greater address. The 2790 * loop will run from this point until a record beyond the termination 2791 * address is encountered. 2792 * 2793 * map->hint must be adjusted to not point to anything we delete, 2794 * so set it to the entry prior to the one being deleted. 2795 * 2796 * GGG see other GGG comment. 2797 */ 2798 if (vm_map_lookup_entry(map, start, &first_entry)) { 2799 entry = first_entry; 2800 vm_map_clip_start(map, entry, start, countp); 2801 map->hint = entry->prev; /* possible problem XXX */ 2802 } else { 2803 map->hint = first_entry; /* possible problem XXX */ 2804 entry = first_entry->next; 2805 } 2806 2807 /* 2808 * If a hole opens up prior to the current first_free then 2809 * adjust first_free. As with map->hint, map->first_free 2810 * cannot be left set to anything we might delete. 2811 */ 2812 if (entry == &map->header) { 2813 map->first_free = &map->header; 2814 } else if (map->first_free->start >= start) { 2815 map->first_free = entry->prev; 2816 } 2817 2818 /* 2819 * Step through all entries in this region 2820 */ 2821 while ((entry != &map->header) && (entry->start < end)) { 2822 vm_map_entry_t next; 2823 vm_offset_t s, e; 2824 vm_pindex_t offidxstart, offidxend, count; 2825 2826 /* 2827 * If we hit an in-transition entry we have to sleep and 2828 * retry. It's easier (and not really slower) to just retry 2829 * since this case occurs so rarely and the hint is already 2830 * pointing at the right place. We have to reset the 2831 * start offset so as not to accidently delete an entry 2832 * another process just created in vacated space. 2833 */ 2834 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 2835 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2836 start = entry->start; 2837 ++mycpu->gd_cnt.v_intrans_coll; 2838 ++mycpu->gd_cnt.v_intrans_wait; 2839 vm_map_transition_wait(map); 2840 goto again; 2841 } 2842 vm_map_clip_end(map, entry, end, countp); 2843 2844 s = entry->start; 2845 e = entry->end; 2846 next = entry->next; 2847 2848 offidxstart = OFF_TO_IDX(entry->offset); 2849 count = OFF_TO_IDX(e - s); 2850 object = entry->object.vm_object; 2851 2852 /* 2853 * Unwire before removing addresses from the pmap; otherwise, 2854 * unwiring will put the entries back in the pmap. 2855 */ 2856 if (entry->wired_count != 0) 2857 vm_map_entry_unwire(map, entry); 2858 2859 offidxend = offidxstart + count; 2860 2861 if (object == &kernel_object) { 2862 vm_object_hold(object); 2863 vm_object_page_remove(object, offidxstart, 2864 offidxend, FALSE); 2865 vm_object_drop(object); 2866 } else if (object && object->type != OBJT_DEFAULT && 2867 object->type != OBJT_SWAP) { 2868 /* 2869 * vnode object routines cannot be chain-locked, 2870 * but since we aren't removing pages from the 2871 * object here we can use a shared hold. 2872 */ 2873 vm_object_hold_shared(object); 2874 pmap_remove(map->pmap, s, e); 2875 vm_object_drop(object); 2876 } else if (object) { 2877 vm_object_hold(object); 2878 vm_object_chain_acquire(object, 0); 2879 pmap_remove(map->pmap, s, e); 2880 2881 if (object != NULL && 2882 object->ref_count != 1 && 2883 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == 2884 OBJ_ONEMAPPING && 2885 (object->type == OBJT_DEFAULT || 2886 object->type == OBJT_SWAP)) { 2887 vm_object_collapse(object, NULL); 2888 vm_object_page_remove(object, offidxstart, 2889 offidxend, FALSE); 2890 if (object->type == OBJT_SWAP) { 2891 swap_pager_freespace(object, 2892 offidxstart, 2893 count); 2894 } 2895 if (offidxend >= object->size && 2896 offidxstart < object->size) { 2897 object->size = offidxstart; 2898 } 2899 } 2900 vm_object_chain_release(object); 2901 vm_object_drop(object); 2902 } 2903 2904 /* 2905 * Delete the entry (which may delete the object) only after 2906 * removing all pmap entries pointing to its pages. 2907 * (Otherwise, its page frames may be reallocated, and any 2908 * modify bits will be set in the wrong object!) 2909 */ 2910 vm_map_entry_delete(map, entry, countp); 2911 entry = next; 2912 } 2913 lwkt_reltoken(&map->token); 2914 return (KERN_SUCCESS); 2915 } 2916 2917 /* 2918 * Remove the given address range from the target map. 2919 * This is the exported form of vm_map_delete. 2920 * 2921 * No requirements. 2922 */ 2923 int 2924 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 2925 { 2926 int result; 2927 int count; 2928 2929 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2930 vm_map_lock(map); 2931 VM_MAP_RANGE_CHECK(map, start, end); 2932 result = vm_map_delete(map, start, end, &count); 2933 vm_map_unlock(map); 2934 vm_map_entry_release(count); 2935 2936 return (result); 2937 } 2938 2939 /* 2940 * Assert that the target map allows the specified privilege on the 2941 * entire address region given. The entire region must be allocated. 2942 * 2943 * The caller must specify whether the vm_map is already locked or not. 2944 */ 2945 boolean_t 2946 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 2947 vm_prot_t protection, boolean_t have_lock) 2948 { 2949 vm_map_entry_t entry; 2950 vm_map_entry_t tmp_entry; 2951 boolean_t result; 2952 2953 if (have_lock == FALSE) 2954 vm_map_lock_read(map); 2955 2956 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 2957 if (have_lock == FALSE) 2958 vm_map_unlock_read(map); 2959 return (FALSE); 2960 } 2961 entry = tmp_entry; 2962 2963 result = TRUE; 2964 while (start < end) { 2965 if (entry == &map->header) { 2966 result = FALSE; 2967 break; 2968 } 2969 /* 2970 * No holes allowed! 2971 */ 2972 2973 if (start < entry->start) { 2974 result = FALSE; 2975 break; 2976 } 2977 /* 2978 * Check protection associated with entry. 2979 */ 2980 2981 if ((entry->protection & protection) != protection) { 2982 result = FALSE; 2983 break; 2984 } 2985 /* go to next entry */ 2986 2987 start = entry->end; 2988 entry = entry->next; 2989 } 2990 if (have_lock == FALSE) 2991 vm_map_unlock_read(map); 2992 return (result); 2993 } 2994 2995 /* 2996 * If appropriate this function shadows the original object with a new object 2997 * and moves the VM pages from the original object to the new object. 2998 * The original object will also be collapsed, if possible. 2999 * 3000 * We can only do this for normal memory objects with a single mapping, and 3001 * it only makes sense to do it if there are 2 or more refs on the original 3002 * object. i.e. typically a memory object that has been extended into 3003 * multiple vm_map_entry's with non-overlapping ranges. 3004 * 3005 * This makes it easier to remove unused pages and keeps object inheritance 3006 * from being a negative impact on memory usage. 3007 * 3008 * On return the (possibly new) entry->object.vm_object will have an 3009 * additional ref on it for the caller to dispose of (usually by cloning 3010 * the vm_map_entry). The additional ref had to be done in this routine 3011 * to avoid racing a collapse. The object's ONEMAPPING flag will also be 3012 * cleared. 3013 * 3014 * The vm_map must be locked and its token held. 3015 */ 3016 static void 3017 vm_map_split(vm_map_entry_t entry) 3018 { 3019 /* OPTIMIZED */ 3020 vm_object_t oobject, nobject, bobject; 3021 vm_offset_t s, e; 3022 vm_page_t m; 3023 vm_pindex_t offidxstart, offidxend, idx; 3024 vm_size_t size; 3025 vm_ooffset_t offset; 3026 int useshadowlist; 3027 3028 /* 3029 * Optimize away object locks for vnode objects. Important exit/exec 3030 * critical path. 3031 * 3032 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag 3033 * anyway. 3034 */ 3035 oobject = entry->object.vm_object; 3036 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) { 3037 vm_object_reference_quick(oobject); 3038 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3039 return; 3040 } 3041 3042 /* 3043 * Setup. Chain lock the original object throughout the entire 3044 * routine to prevent new page faults from occuring. 3045 * 3046 * XXX can madvise WILLNEED interfere with us too? 3047 */ 3048 vm_object_hold(oobject); 3049 vm_object_chain_acquire(oobject, 0); 3050 3051 /* 3052 * Original object cannot be split? Might have also changed state. 3053 */ 3054 if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT && 3055 oobject->type != OBJT_SWAP)) { 3056 vm_object_chain_release(oobject); 3057 vm_object_reference_locked(oobject); 3058 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3059 vm_object_drop(oobject); 3060 return; 3061 } 3062 3063 /* 3064 * Collapse original object with its backing store as an 3065 * optimization to reduce chain lengths when possible. 3066 * 3067 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's 3068 * for oobject, so there's no point collapsing it. 3069 * 3070 * Then re-check whether the object can be split. 3071 */ 3072 vm_object_collapse(oobject, NULL); 3073 3074 if (oobject->ref_count <= 1 || 3075 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) || 3076 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) { 3077 vm_object_chain_release(oobject); 3078 vm_object_reference_locked(oobject); 3079 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3080 vm_object_drop(oobject); 3081 return; 3082 } 3083 3084 /* 3085 * Acquire the chain lock on the backing object. 3086 * 3087 * Give bobject an additional ref count for when it will be shadowed 3088 * by nobject. 3089 */ 3090 useshadowlist = 0; 3091 if ((bobject = oobject->backing_object) != NULL) { 3092 if (bobject->type != OBJT_VNODE) { 3093 useshadowlist = 1; 3094 vm_object_hold(bobject); 3095 vm_object_chain_wait(bobject, 0); 3096 vm_object_reference_locked(bobject); 3097 vm_object_chain_acquire(bobject, 0); 3098 KKASSERT(bobject->backing_object == bobject); 3099 KKASSERT((bobject->flags & OBJ_DEAD) == 0); 3100 } else { 3101 vm_object_reference_quick(bobject); 3102 } 3103 } 3104 3105 /* 3106 * Calculate the object page range and allocate the new object. 3107 */ 3108 offset = entry->offset; 3109 s = entry->start; 3110 e = entry->end; 3111 3112 offidxstart = OFF_TO_IDX(offset); 3113 offidxend = offidxstart + OFF_TO_IDX(e - s); 3114 size = offidxend - offidxstart; 3115 3116 switch(oobject->type) { 3117 case OBJT_DEFAULT: 3118 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size), 3119 VM_PROT_ALL, 0); 3120 break; 3121 case OBJT_SWAP: 3122 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size), 3123 VM_PROT_ALL, 0); 3124 break; 3125 default: 3126 /* not reached */ 3127 nobject = NULL; 3128 KKASSERT(0); 3129 } 3130 3131 if (nobject == NULL) { 3132 if (bobject) { 3133 if (useshadowlist) { 3134 vm_object_chain_release(bobject); 3135 vm_object_deallocate(bobject); 3136 vm_object_drop(bobject); 3137 } else { 3138 vm_object_deallocate(bobject); 3139 } 3140 } 3141 vm_object_chain_release(oobject); 3142 vm_object_reference_locked(oobject); 3143 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3144 vm_object_drop(oobject); 3145 return; 3146 } 3147 3148 /* 3149 * The new object will replace entry->object.vm_object so it needs 3150 * a second reference (the caller expects an additional ref). 3151 */ 3152 vm_object_hold(nobject); 3153 vm_object_reference_locked(nobject); 3154 vm_object_chain_acquire(nobject, 0); 3155 3156 /* 3157 * nobject shadows bobject (oobject already shadows bobject). 3158 */ 3159 if (bobject) { 3160 nobject->backing_object_offset = 3161 oobject->backing_object_offset + IDX_TO_OFF(offidxstart); 3162 nobject->backing_object = bobject; 3163 if (useshadowlist) { 3164 bobject->shadow_count++; 3165 bobject->generation++; 3166 LIST_INSERT_HEAD(&bobject->shadow_head, 3167 nobject, shadow_list); 3168 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/ 3169 vm_object_chain_release(bobject); 3170 vm_object_drop(bobject); 3171 vm_object_set_flag(nobject, OBJ_ONSHADOW); 3172 } 3173 } 3174 3175 /* 3176 * Move the VM pages from oobject to nobject 3177 */ 3178 for (idx = 0; idx < size; idx++) { 3179 vm_page_t m; 3180 3181 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx, 3182 TRUE, "vmpg"); 3183 if (m == NULL) 3184 continue; 3185 3186 /* 3187 * We must wait for pending I/O to complete before we can 3188 * rename the page. 3189 * 3190 * We do not have to VM_PROT_NONE the page as mappings should 3191 * not be changed by this operation. 3192 * 3193 * NOTE: The act of renaming a page updates chaingen for both 3194 * objects. 3195 */ 3196 vm_page_rename(m, nobject, idx); 3197 /* page automatically made dirty by rename and cache handled */ 3198 /* page remains busy */ 3199 } 3200 3201 if (oobject->type == OBJT_SWAP) { 3202 vm_object_pip_add(oobject, 1); 3203 /* 3204 * copy oobject pages into nobject and destroy unneeded 3205 * pages in shadow object. 3206 */ 3207 swap_pager_copy(oobject, nobject, offidxstart, 0); 3208 vm_object_pip_wakeup(oobject); 3209 } 3210 3211 /* 3212 * Wakeup the pages we played with. No spl protection is needed 3213 * for a simple wakeup. 3214 */ 3215 for (idx = 0; idx < size; idx++) { 3216 m = vm_page_lookup(nobject, idx); 3217 if (m) { 3218 KKASSERT(m->flags & PG_BUSY); 3219 vm_page_wakeup(m); 3220 } 3221 } 3222 entry->object.vm_object = nobject; 3223 entry->offset = 0LL; 3224 3225 /* 3226 * Cleanup 3227 * 3228 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the 3229 * related pages were moved and are no longer applicable to the 3230 * original object. 3231 * 3232 * NOTE: Deallocate oobject (due to its entry->object.vm_object being 3233 * replaced by nobject). 3234 */ 3235 vm_object_chain_release(nobject); 3236 vm_object_drop(nobject); 3237 if (bobject && useshadowlist) { 3238 vm_object_chain_release(bobject); 3239 vm_object_drop(bobject); 3240 } 3241 vm_object_chain_release(oobject); 3242 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/ 3243 vm_object_deallocate_locked(oobject); 3244 vm_object_drop(oobject); 3245 } 3246 3247 /* 3248 * Copies the contents of the source entry to the destination 3249 * entry. The entries *must* be aligned properly. 3250 * 3251 * The vm_maps must be exclusively locked. 3252 * The vm_map's token must be held. 3253 * 3254 * Because the maps are locked no faults can be in progress during the 3255 * operation. 3256 */ 3257 static void 3258 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 3259 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 3260 { 3261 vm_object_t src_object; 3262 3263 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP) 3264 return; 3265 if (src_entry->maptype == VM_MAPTYPE_SUBMAP) 3266 return; 3267 3268 if (src_entry->wired_count == 0) { 3269 /* 3270 * If the source entry is marked needs_copy, it is already 3271 * write-protected. 3272 */ 3273 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 3274 pmap_protect(src_map->pmap, 3275 src_entry->start, 3276 src_entry->end, 3277 src_entry->protection & ~VM_PROT_WRITE); 3278 } 3279 3280 /* 3281 * Make a copy of the object. 3282 * 3283 * The object must be locked prior to checking the object type 3284 * and for the call to vm_object_collapse() and vm_map_split(). 3285 * We cannot use *_hold() here because the split code will 3286 * probably try to destroy the object. The lock is a pool 3287 * token and doesn't care. 3288 * 3289 * We must bump src_map->timestamp when setting 3290 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault 3291 * to retry, otherwise the concurrent fault might improperly 3292 * install a RW pte when its supposed to be a RO(COW) pte. 3293 * This race can occur because a vnode-backed fault may have 3294 * to temporarily release the map lock. 3295 */ 3296 if (src_entry->object.vm_object != NULL) { 3297 vm_map_split(src_entry); 3298 src_object = src_entry->object.vm_object; 3299 dst_entry->object.vm_object = src_object; 3300 src_entry->eflags |= (MAP_ENTRY_COW | 3301 MAP_ENTRY_NEEDS_COPY); 3302 dst_entry->eflags |= (MAP_ENTRY_COW | 3303 MAP_ENTRY_NEEDS_COPY); 3304 dst_entry->offset = src_entry->offset; 3305 ++src_map->timestamp; 3306 } else { 3307 dst_entry->object.vm_object = NULL; 3308 dst_entry->offset = 0; 3309 } 3310 3311 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, 3312 dst_entry->end - dst_entry->start, src_entry->start); 3313 } else { 3314 /* 3315 * Of course, wired down pages can't be set copy-on-write. 3316 * Cause wired pages to be copied into the new map by 3317 * simulating faults (the new pages are pageable) 3318 */ 3319 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 3320 } 3321 } 3322 3323 /* 3324 * vmspace_fork: 3325 * Create a new process vmspace structure and vm_map 3326 * based on those of an existing process. The new map 3327 * is based on the old map, according to the inheritance 3328 * values on the regions in that map. 3329 * 3330 * The source map must not be locked. 3331 * No requirements. 3332 */ 3333 struct vmspace * 3334 vmspace_fork(struct vmspace *vm1) 3335 { 3336 struct vmspace *vm2; 3337 vm_map_t old_map = &vm1->vm_map; 3338 vm_map_t new_map; 3339 vm_map_entry_t old_entry; 3340 vm_map_entry_t new_entry; 3341 vm_object_t object; 3342 int count; 3343 3344 lwkt_gettoken(&vm1->vm_map.token); 3345 vm_map_lock(old_map); 3346 3347 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset); 3348 lwkt_gettoken(&vm2->vm_map.token); 3349 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 3350 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 3351 new_map = &vm2->vm_map; /* XXX */ 3352 new_map->timestamp = 1; 3353 3354 vm_map_lock(new_map); 3355 3356 count = 0; 3357 old_entry = old_map->header.next; 3358 while (old_entry != &old_map->header) { 3359 ++count; 3360 old_entry = old_entry->next; 3361 } 3362 3363 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 3364 3365 old_entry = old_map->header.next; 3366 while (old_entry != &old_map->header) { 3367 if (old_entry->maptype == VM_MAPTYPE_SUBMAP) 3368 panic("vm_map_fork: encountered a submap"); 3369 3370 switch (old_entry->inheritance) { 3371 case VM_INHERIT_NONE: 3372 break; 3373 case VM_INHERIT_SHARE: 3374 /* 3375 * Clone the entry, creating the shared object if 3376 * necessary. 3377 */ 3378 if (old_entry->object.vm_object == NULL) 3379 vm_map_entry_allocate_object(old_entry); 3380 3381 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3382 /* 3383 * Shadow a map_entry which needs a copy, 3384 * replacing its object with a new object 3385 * that points to the old one. Ask the 3386 * shadow code to automatically add an 3387 * additional ref. We can't do it afterwords 3388 * because we might race a collapse. The call 3389 * to vm_map_entry_shadow() will also clear 3390 * OBJ_ONEMAPPING. 3391 */ 3392 vm_map_entry_shadow(old_entry, 1); 3393 } else if (old_entry->object.vm_object) { 3394 /* 3395 * We will make a shared copy of the object, 3396 * and must clear OBJ_ONEMAPPING. 3397 * 3398 * Optimize vnode objects. OBJ_ONEMAPPING 3399 * is non-applicable but clear it anyway, 3400 * and its terminal so we don'th ave to deal 3401 * with chains. Reduces SMP conflicts. 3402 * 3403 * XXX assert that object.vm_object != NULL 3404 * since we allocate it above. 3405 */ 3406 object = old_entry->object.vm_object; 3407 if (object->type == OBJT_VNODE) { 3408 vm_object_reference_quick(object); 3409 vm_object_clear_flag(object, 3410 OBJ_ONEMAPPING); 3411 } else { 3412 vm_object_hold(object); 3413 vm_object_chain_wait(object, 0); 3414 vm_object_reference_locked(object); 3415 vm_object_clear_flag(object, 3416 OBJ_ONEMAPPING); 3417 vm_object_drop(object); 3418 } 3419 } 3420 3421 /* 3422 * Clone the entry. We've already bumped the ref on 3423 * any vm_object. 3424 */ 3425 new_entry = vm_map_entry_create(new_map, &count); 3426 *new_entry = *old_entry; 3427 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3428 new_entry->wired_count = 0; 3429 3430 /* 3431 * Insert the entry into the new map -- we know we're 3432 * inserting at the end of the new map. 3433 */ 3434 3435 vm_map_entry_link(new_map, new_map->header.prev, 3436 new_entry); 3437 3438 /* 3439 * Update the physical map 3440 */ 3441 pmap_copy(new_map->pmap, old_map->pmap, 3442 new_entry->start, 3443 (old_entry->end - old_entry->start), 3444 old_entry->start); 3445 break; 3446 case VM_INHERIT_COPY: 3447 /* 3448 * Clone the entry and link into the map. 3449 */ 3450 new_entry = vm_map_entry_create(new_map, &count); 3451 *new_entry = *old_entry; 3452 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3453 new_entry->wired_count = 0; 3454 new_entry->object.vm_object = NULL; 3455 vm_map_entry_link(new_map, new_map->header.prev, 3456 new_entry); 3457 vm_map_copy_entry(old_map, new_map, old_entry, 3458 new_entry); 3459 break; 3460 } 3461 old_entry = old_entry->next; 3462 } 3463 3464 new_map->size = old_map->size; 3465 vm_map_unlock(old_map); 3466 vm_map_unlock(new_map); 3467 vm_map_entry_release(count); 3468 3469 lwkt_reltoken(&vm2->vm_map.token); 3470 lwkt_reltoken(&vm1->vm_map.token); 3471 3472 return (vm2); 3473 } 3474 3475 /* 3476 * Create an auto-grow stack entry 3477 * 3478 * No requirements. 3479 */ 3480 int 3481 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 3482 int flags, vm_prot_t prot, vm_prot_t max, int cow) 3483 { 3484 vm_map_entry_t prev_entry; 3485 vm_map_entry_t new_stack_entry; 3486 vm_size_t init_ssize; 3487 int rv; 3488 int count; 3489 vm_offset_t tmpaddr; 3490 3491 cow |= MAP_IS_STACK; 3492 3493 if (max_ssize < sgrowsiz) 3494 init_ssize = max_ssize; 3495 else 3496 init_ssize = sgrowsiz; 3497 3498 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3499 vm_map_lock(map); 3500 3501 /* 3502 * Find space for the mapping 3503 */ 3504 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) { 3505 if (vm_map_findspace(map, addrbos, max_ssize, 1, 3506 flags, &tmpaddr)) { 3507 vm_map_unlock(map); 3508 vm_map_entry_release(count); 3509 return (KERN_NO_SPACE); 3510 } 3511 addrbos = tmpaddr; 3512 } 3513 3514 /* If addr is already mapped, no go */ 3515 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) { 3516 vm_map_unlock(map); 3517 vm_map_entry_release(count); 3518 return (KERN_NO_SPACE); 3519 } 3520 3521 #if 0 3522 /* XXX already handled by kern_mmap() */ 3523 /* If we would blow our VMEM resource limit, no go */ 3524 if (map->size + init_ssize > 3525 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3526 vm_map_unlock(map); 3527 vm_map_entry_release(count); 3528 return (KERN_NO_SPACE); 3529 } 3530 #endif 3531 3532 /* 3533 * If we can't accomodate max_ssize in the current mapping, 3534 * no go. However, we need to be aware that subsequent user 3535 * mappings might map into the space we have reserved for 3536 * stack, and currently this space is not protected. 3537 * 3538 * Hopefully we will at least detect this condition 3539 * when we try to grow the stack. 3540 */ 3541 if ((prev_entry->next != &map->header) && 3542 (prev_entry->next->start < addrbos + max_ssize)) { 3543 vm_map_unlock(map); 3544 vm_map_entry_release(count); 3545 return (KERN_NO_SPACE); 3546 } 3547 3548 /* 3549 * We initially map a stack of only init_ssize. We will 3550 * grow as needed later. Since this is to be a grow 3551 * down stack, we map at the top of the range. 3552 * 3553 * Note: we would normally expect prot and max to be 3554 * VM_PROT_ALL, and cow to be 0. Possibly we should 3555 * eliminate these as input parameters, and just 3556 * pass these values here in the insert call. 3557 */ 3558 rv = vm_map_insert(map, &count, 3559 NULL, 0, addrbos + max_ssize - init_ssize, 3560 addrbos + max_ssize, 3561 VM_MAPTYPE_NORMAL, 3562 prot, max, 3563 cow); 3564 3565 /* Now set the avail_ssize amount */ 3566 if (rv == KERN_SUCCESS) { 3567 if (prev_entry != &map->header) 3568 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count); 3569 new_stack_entry = prev_entry->next; 3570 if (new_stack_entry->end != addrbos + max_ssize || 3571 new_stack_entry->start != addrbos + max_ssize - init_ssize) 3572 panic ("Bad entry start/end for new stack entry"); 3573 else 3574 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize; 3575 } 3576 3577 vm_map_unlock(map); 3578 vm_map_entry_release(count); 3579 return (rv); 3580 } 3581 3582 /* 3583 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 3584 * desired address is already mapped, or if we successfully grow 3585 * the stack. Also returns KERN_SUCCESS if addr is outside the 3586 * stack range (this is strange, but preserves compatibility with 3587 * the grow function in vm_machdep.c). 3588 * 3589 * No requirements. 3590 */ 3591 int 3592 vm_map_growstack (struct proc *p, vm_offset_t addr) 3593 { 3594 vm_map_entry_t prev_entry; 3595 vm_map_entry_t stack_entry; 3596 vm_map_entry_t new_stack_entry; 3597 struct vmspace *vm = p->p_vmspace; 3598 vm_map_t map = &vm->vm_map; 3599 vm_offset_t end; 3600 int grow_amount; 3601 int rv = KERN_SUCCESS; 3602 int is_procstack; 3603 int use_read_lock = 1; 3604 int count; 3605 3606 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3607 Retry: 3608 if (use_read_lock) 3609 vm_map_lock_read(map); 3610 else 3611 vm_map_lock(map); 3612 3613 /* If addr is already in the entry range, no need to grow.*/ 3614 if (vm_map_lookup_entry(map, addr, &prev_entry)) 3615 goto done; 3616 3617 if ((stack_entry = prev_entry->next) == &map->header) 3618 goto done; 3619 if (prev_entry == &map->header) 3620 end = stack_entry->start - stack_entry->aux.avail_ssize; 3621 else 3622 end = prev_entry->end; 3623 3624 /* 3625 * This next test mimics the old grow function in vm_machdep.c. 3626 * It really doesn't quite make sense, but we do it anyway 3627 * for compatibility. 3628 * 3629 * If not growable stack, return success. This signals the 3630 * caller to proceed as he would normally with normal vm. 3631 */ 3632 if (stack_entry->aux.avail_ssize < 1 || 3633 addr >= stack_entry->start || 3634 addr < stack_entry->start - stack_entry->aux.avail_ssize) { 3635 goto done; 3636 } 3637 3638 /* Find the minimum grow amount */ 3639 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE); 3640 if (grow_amount > stack_entry->aux.avail_ssize) { 3641 rv = KERN_NO_SPACE; 3642 goto done; 3643 } 3644 3645 /* 3646 * If there is no longer enough space between the entries 3647 * nogo, and adjust the available space. Note: this 3648 * should only happen if the user has mapped into the 3649 * stack area after the stack was created, and is 3650 * probably an error. 3651 * 3652 * This also effectively destroys any guard page the user 3653 * might have intended by limiting the stack size. 3654 */ 3655 if (grow_amount > stack_entry->start - end) { 3656 if (use_read_lock && vm_map_lock_upgrade(map)) { 3657 /* lost lock */ 3658 use_read_lock = 0; 3659 goto Retry; 3660 } 3661 use_read_lock = 0; 3662 stack_entry->aux.avail_ssize = stack_entry->start - end; 3663 rv = KERN_NO_SPACE; 3664 goto done; 3665 } 3666 3667 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 3668 3669 /* If this is the main process stack, see if we're over the 3670 * stack limit. 3671 */ 3672 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3673 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3674 rv = KERN_NO_SPACE; 3675 goto done; 3676 } 3677 3678 /* Round up the grow amount modulo SGROWSIZ */ 3679 grow_amount = roundup (grow_amount, sgrowsiz); 3680 if (grow_amount > stack_entry->aux.avail_ssize) { 3681 grow_amount = stack_entry->aux.avail_ssize; 3682 } 3683 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3684 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3685 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - 3686 ctob(vm->vm_ssize); 3687 } 3688 3689 /* If we would blow our VMEM resource limit, no go */ 3690 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3691 rv = KERN_NO_SPACE; 3692 goto done; 3693 } 3694 3695 if (use_read_lock && vm_map_lock_upgrade(map)) { 3696 /* lost lock */ 3697 use_read_lock = 0; 3698 goto Retry; 3699 } 3700 use_read_lock = 0; 3701 3702 /* Get the preliminary new entry start value */ 3703 addr = stack_entry->start - grow_amount; 3704 3705 /* If this puts us into the previous entry, cut back our growth 3706 * to the available space. Also, see the note above. 3707 */ 3708 if (addr < end) { 3709 stack_entry->aux.avail_ssize = stack_entry->start - end; 3710 addr = end; 3711 } 3712 3713 rv = vm_map_insert(map, &count, 3714 NULL, 0, addr, stack_entry->start, 3715 VM_MAPTYPE_NORMAL, 3716 VM_PROT_ALL, VM_PROT_ALL, 3717 0); 3718 3719 /* Adjust the available stack space by the amount we grew. */ 3720 if (rv == KERN_SUCCESS) { 3721 if (prev_entry != &map->header) 3722 vm_map_clip_end(map, prev_entry, addr, &count); 3723 new_stack_entry = prev_entry->next; 3724 if (new_stack_entry->end != stack_entry->start || 3725 new_stack_entry->start != addr) 3726 panic ("Bad stack grow start/end in new stack entry"); 3727 else { 3728 new_stack_entry->aux.avail_ssize = 3729 stack_entry->aux.avail_ssize - 3730 (new_stack_entry->end - new_stack_entry->start); 3731 if (is_procstack) 3732 vm->vm_ssize += btoc(new_stack_entry->end - 3733 new_stack_entry->start); 3734 } 3735 3736 if (map->flags & MAP_WIREFUTURE) 3737 vm_map_unwire(map, new_stack_entry->start, 3738 new_stack_entry->end, FALSE); 3739 } 3740 3741 done: 3742 if (use_read_lock) 3743 vm_map_unlock_read(map); 3744 else 3745 vm_map_unlock(map); 3746 vm_map_entry_release(count); 3747 return (rv); 3748 } 3749 3750 /* 3751 * Unshare the specified VM space for exec. If other processes are 3752 * mapped to it, then create a new one. The new vmspace is null. 3753 * 3754 * No requirements. 3755 */ 3756 void 3757 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 3758 { 3759 struct vmspace *oldvmspace = p->p_vmspace; 3760 struct vmspace *newvmspace; 3761 vm_map_t map = &p->p_vmspace->vm_map; 3762 3763 /* 3764 * If we are execing a resident vmspace we fork it, otherwise 3765 * we create a new vmspace. Note that exitingcnt is not 3766 * copied to the new vmspace. 3767 */ 3768 lwkt_gettoken(&oldvmspace->vm_map.token); 3769 if (vmcopy) { 3770 newvmspace = vmspace_fork(vmcopy); 3771 lwkt_gettoken(&newvmspace->vm_map.token); 3772 } else { 3773 newvmspace = vmspace_alloc(map->min_offset, map->max_offset); 3774 lwkt_gettoken(&newvmspace->vm_map.token); 3775 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 3776 (caddr_t)&oldvmspace->vm_endcopy - 3777 (caddr_t)&oldvmspace->vm_startcopy); 3778 } 3779 3780 /* 3781 * Finish initializing the vmspace before assigning it 3782 * to the process. The vmspace will become the current vmspace 3783 * if p == curproc. 3784 */ 3785 pmap_pinit2(vmspace_pmap(newvmspace)); 3786 pmap_replacevm(p, newvmspace, 0); 3787 lwkt_reltoken(&newvmspace->vm_map.token); 3788 lwkt_reltoken(&oldvmspace->vm_map.token); 3789 vmspace_rel(oldvmspace); 3790 } 3791 3792 /* 3793 * Unshare the specified VM space for forcing COW. This 3794 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 3795 */ 3796 void 3797 vmspace_unshare(struct proc *p) 3798 { 3799 struct vmspace *oldvmspace = p->p_vmspace; 3800 struct vmspace *newvmspace; 3801 3802 lwkt_gettoken(&oldvmspace->vm_map.token); 3803 if (vmspace_getrefs(oldvmspace) == 1) { 3804 lwkt_reltoken(&oldvmspace->vm_map.token); 3805 return; 3806 } 3807 newvmspace = vmspace_fork(oldvmspace); 3808 lwkt_gettoken(&newvmspace->vm_map.token); 3809 pmap_pinit2(vmspace_pmap(newvmspace)); 3810 pmap_replacevm(p, newvmspace, 0); 3811 lwkt_reltoken(&newvmspace->vm_map.token); 3812 lwkt_reltoken(&oldvmspace->vm_map.token); 3813 vmspace_rel(oldvmspace); 3814 } 3815 3816 /* 3817 * vm_map_hint: return the beginning of the best area suitable for 3818 * creating a new mapping with "prot" protection. 3819 * 3820 * No requirements. 3821 */ 3822 vm_offset_t 3823 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot) 3824 { 3825 struct vmspace *vms = p->p_vmspace; 3826 3827 if (!randomize_mmap || addr != 0) { 3828 /* 3829 * Set a reasonable start point for the hint if it was 3830 * not specified or if it falls within the heap space. 3831 * Hinted mmap()s do not allocate out of the heap space. 3832 */ 3833 if (addr == 0 || 3834 (addr >= round_page((vm_offset_t)vms->vm_taddr) && 3835 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) { 3836 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz); 3837 } 3838 3839 return addr; 3840 } 3841 3842 #ifdef notyet 3843 #ifdef __i386__ 3844 /* 3845 * If executable skip first two pages, otherwise start 3846 * after data + heap region. 3847 */ 3848 if ((prot & VM_PROT_EXECUTE) && 3849 ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) { 3850 addr = (PAGE_SIZE * 2) + 3851 (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1)); 3852 return (round_page(addr)); 3853 } 3854 #endif /* __i386__ */ 3855 #endif /* notyet */ 3856 3857 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ; 3858 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1); 3859 3860 return (round_page(addr)); 3861 } 3862 3863 /* 3864 * Finds the VM object, offset, and protection for a given virtual address 3865 * in the specified map, assuming a page fault of the type specified. 3866 * 3867 * Leaves the map in question locked for read; return values are guaranteed 3868 * until a vm_map_lookup_done call is performed. Note that the map argument 3869 * is in/out; the returned map must be used in the call to vm_map_lookup_done. 3870 * 3871 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make 3872 * that fast. 3873 * 3874 * If a lookup is requested with "write protection" specified, the map may 3875 * be changed to perform virtual copying operations, although the data 3876 * referenced will remain the same. 3877 * 3878 * No requirements. 3879 */ 3880 int 3881 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 3882 vm_offset_t vaddr, 3883 vm_prot_t fault_typea, 3884 vm_map_entry_t *out_entry, /* OUT */ 3885 vm_object_t *object, /* OUT */ 3886 vm_pindex_t *pindex, /* OUT */ 3887 vm_prot_t *out_prot, /* OUT */ 3888 boolean_t *wired) /* OUT */ 3889 { 3890 vm_map_entry_t entry; 3891 vm_map_t map = *var_map; 3892 vm_prot_t prot; 3893 vm_prot_t fault_type = fault_typea; 3894 int use_read_lock = 1; 3895 int rv = KERN_SUCCESS; 3896 3897 RetryLookup: 3898 if (use_read_lock) 3899 vm_map_lock_read(map); 3900 else 3901 vm_map_lock(map); 3902 3903 /* 3904 * If the map has an interesting hint, try it before calling full 3905 * blown lookup routine. 3906 */ 3907 entry = map->hint; 3908 cpu_ccfence(); 3909 *out_entry = entry; 3910 *object = NULL; 3911 3912 if ((entry == &map->header) || 3913 (vaddr < entry->start) || (vaddr >= entry->end)) { 3914 vm_map_entry_t tmp_entry; 3915 3916 /* 3917 * Entry was either not a valid hint, or the vaddr was not 3918 * contained in the entry, so do a full lookup. 3919 */ 3920 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 3921 rv = KERN_INVALID_ADDRESS; 3922 goto done; 3923 } 3924 3925 entry = tmp_entry; 3926 *out_entry = entry; 3927 } 3928 3929 /* 3930 * Handle submaps. 3931 */ 3932 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 3933 vm_map_t old_map = map; 3934 3935 *var_map = map = entry->object.sub_map; 3936 if (use_read_lock) 3937 vm_map_unlock_read(old_map); 3938 else 3939 vm_map_unlock(old_map); 3940 use_read_lock = 1; 3941 goto RetryLookup; 3942 } 3943 3944 /* 3945 * Check whether this task is allowed to have this page. 3946 * Note the special case for MAP_ENTRY_COW 3947 * pages with an override. This is to implement a forced 3948 * COW for debuggers. 3949 */ 3950 3951 if (fault_type & VM_PROT_OVERRIDE_WRITE) 3952 prot = entry->max_protection; 3953 else 3954 prot = entry->protection; 3955 3956 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 3957 if ((fault_type & prot) != fault_type) { 3958 rv = KERN_PROTECTION_FAILURE; 3959 goto done; 3960 } 3961 3962 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 3963 (entry->eflags & MAP_ENTRY_COW) && 3964 (fault_type & VM_PROT_WRITE) && 3965 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 3966 rv = KERN_PROTECTION_FAILURE; 3967 goto done; 3968 } 3969 3970 /* 3971 * If this page is not pageable, we have to get it for all possible 3972 * accesses. 3973 */ 3974 *wired = (entry->wired_count != 0); 3975 if (*wired) 3976 prot = fault_type = entry->protection; 3977 3978 /* 3979 * Virtual page tables may need to update the accessed (A) bit 3980 * in a page table entry. Upgrade the fault to a write fault for 3981 * that case if the map will support it. If the map does not support 3982 * it the page table entry simply will not be updated. 3983 */ 3984 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 3985 if (prot & VM_PROT_WRITE) 3986 fault_type |= VM_PROT_WRITE; 3987 } 3988 3989 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 3990 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 3991 if ((prot & VM_PROT_WRITE) == 0) 3992 fault_type |= VM_PROT_WRITE; 3993 } 3994 3995 /* 3996 * If the entry was copy-on-write, we either ... 3997 */ 3998 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3999 /* 4000 * If we want to write the page, we may as well handle that 4001 * now since we've got the map locked. 4002 * 4003 * If we don't need to write the page, we just demote the 4004 * permissions allowed. 4005 */ 4006 4007 if (fault_type & VM_PROT_WRITE) { 4008 /* 4009 * Make a new object, and place it in the object 4010 * chain. Note that no new references have appeared 4011 * -- one just moved from the map to the new 4012 * object. 4013 */ 4014 4015 if (use_read_lock && vm_map_lock_upgrade(map)) { 4016 /* lost lock */ 4017 use_read_lock = 0; 4018 goto RetryLookup; 4019 } 4020 use_read_lock = 0; 4021 4022 vm_map_entry_shadow(entry, 0); 4023 } else { 4024 /* 4025 * We're attempting to read a copy-on-write page -- 4026 * don't allow writes. 4027 */ 4028 4029 prot &= ~VM_PROT_WRITE; 4030 } 4031 } 4032 4033 /* 4034 * Create an object if necessary. 4035 */ 4036 if (entry->object.vm_object == NULL && !map->system_map) { 4037 if (use_read_lock && vm_map_lock_upgrade(map)) { 4038 /* lost lock */ 4039 use_read_lock = 0; 4040 goto RetryLookup; 4041 } 4042 use_read_lock = 0; 4043 vm_map_entry_allocate_object(entry); 4044 } 4045 4046 /* 4047 * Return the object/offset from this entry. If the entry was 4048 * copy-on-write or empty, it has been fixed up. 4049 */ 4050 4051 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 4052 *object = entry->object.vm_object; 4053 4054 /* 4055 * Return whether this is the only map sharing this data. On 4056 * success we return with a read lock held on the map. On failure 4057 * we return with the map unlocked. 4058 */ 4059 *out_prot = prot; 4060 done: 4061 if (rv == KERN_SUCCESS) { 4062 if (use_read_lock == 0) 4063 vm_map_lock_downgrade(map); 4064 } else if (use_read_lock) { 4065 vm_map_unlock_read(map); 4066 } else { 4067 vm_map_unlock(map); 4068 } 4069 return (rv); 4070 } 4071 4072 /* 4073 * Releases locks acquired by a vm_map_lookup() 4074 * (according to the handle returned by that lookup). 4075 * 4076 * No other requirements. 4077 */ 4078 void 4079 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 4080 { 4081 /* 4082 * Unlock the main-level map 4083 */ 4084 vm_map_unlock_read(map); 4085 if (count) 4086 vm_map_entry_release(count); 4087 } 4088 4089 #include "opt_ddb.h" 4090 #ifdef DDB 4091 #include <sys/kernel.h> 4092 4093 #include <ddb/ddb.h> 4094 4095 /* 4096 * Debugging only 4097 */ 4098 DB_SHOW_COMMAND(map, vm_map_print) 4099 { 4100 static int nlines; 4101 /* XXX convert args. */ 4102 vm_map_t map = (vm_map_t)addr; 4103 boolean_t full = have_addr; 4104 4105 vm_map_entry_t entry; 4106 4107 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 4108 (void *)map, 4109 (void *)map->pmap, map->nentries, map->timestamp); 4110 nlines++; 4111 4112 if (!full && db_indent) 4113 return; 4114 4115 db_indent += 2; 4116 for (entry = map->header.next; entry != &map->header; 4117 entry = entry->next) { 4118 db_iprintf("map entry %p: start=%p, end=%p\n", 4119 (void *)entry, (void *)entry->start, (void *)entry->end); 4120 nlines++; 4121 { 4122 static char *inheritance_name[4] = 4123 {"share", "copy", "none", "donate_copy"}; 4124 4125 db_iprintf(" prot=%x/%x/%s", 4126 entry->protection, 4127 entry->max_protection, 4128 inheritance_name[(int)(unsigned char)entry->inheritance]); 4129 if (entry->wired_count != 0) 4130 db_printf(", wired"); 4131 } 4132 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 4133 /* XXX no %qd in kernel. Truncate entry->offset. */ 4134 db_printf(", share=%p, offset=0x%lx\n", 4135 (void *)entry->object.sub_map, 4136 (long)entry->offset); 4137 nlines++; 4138 if ((entry->prev == &map->header) || 4139 (entry->prev->object.sub_map != 4140 entry->object.sub_map)) { 4141 db_indent += 2; 4142 vm_map_print((db_expr_t)(intptr_t) 4143 entry->object.sub_map, 4144 full, 0, NULL); 4145 db_indent -= 2; 4146 } 4147 } else { 4148 /* XXX no %qd in kernel. Truncate entry->offset. */ 4149 db_printf(", object=%p, offset=0x%lx", 4150 (void *)entry->object.vm_object, 4151 (long)entry->offset); 4152 if (entry->eflags & MAP_ENTRY_COW) 4153 db_printf(", copy (%s)", 4154 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4155 db_printf("\n"); 4156 nlines++; 4157 4158 if ((entry->prev == &map->header) || 4159 (entry->prev->object.vm_object != 4160 entry->object.vm_object)) { 4161 db_indent += 2; 4162 vm_object_print((db_expr_t)(intptr_t) 4163 entry->object.vm_object, 4164 full, 0, NULL); 4165 nlines += 4; 4166 db_indent -= 2; 4167 } 4168 } 4169 } 4170 db_indent -= 2; 4171 if (db_indent == 0) 4172 nlines = 0; 4173 } 4174 4175 /* 4176 * Debugging only 4177 */ 4178 DB_SHOW_COMMAND(procvm, procvm) 4179 { 4180 struct proc *p; 4181 4182 if (have_addr) { 4183 p = (struct proc *) addr; 4184 } else { 4185 p = curproc; 4186 } 4187 4188 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 4189 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 4190 (void *)vmspace_pmap(p->p_vmspace)); 4191 4192 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 4193 } 4194 4195 #endif /* DDB */ 4196