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 * If the pmap does not contain wired pages we can bulk-delete 440 * the pmap as a performance optimization before removing the related mappings. 441 * 442 * If the pmap contains wired pages we cannot do this pre-optimization 443 * because currently vm_fault_unwire() expects the pmap pages to exist 444 * and will not decrement p->wire_count if they do not. 445 */ 446 vm->vm_flags |= VMSPACE_EXIT1; 447 shmexit(vm); 448 if (vmspace_pmap(vm)->pm_stats.wired_count) { 449 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 450 VM_MAX_USER_ADDRESS); 451 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 452 VM_MAX_USER_ADDRESS); 453 } else { 454 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 455 VM_MAX_USER_ADDRESS); 456 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 457 VM_MAX_USER_ADDRESS); 458 } 459 lwkt_reltoken(&vm->vm_map.token); 460 } else { 461 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0); 462 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0); 463 464 /* 465 * Get rid of remaining basic resources. 466 */ 467 vm->vm_flags |= VMSPACE_EXIT2; 468 cpu_vmspace_free(vm); 469 shmexit(vm); 470 471 /* 472 * Lock the map, to wait out all other references to it. 473 * Delete all of the mappings and pages they hold, then call 474 * the pmap module to reclaim anything left. 475 */ 476 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 477 vm_map_lock(&vm->vm_map); 478 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset, 479 vm->vm_map.max_offset, &count); 480 vm_map_unlock(&vm->vm_map); 481 vm_map_entry_release(count); 482 483 lwkt_gettoken(&vmspace_pmap(vm)->pm_token); 484 pmap_release(vmspace_pmap(vm)); 485 lwkt_reltoken(&vmspace_pmap(vm)->pm_token); 486 lwkt_reltoken(&vm->vm_map.token); 487 objcache_put(vmspace_cache, vm); 488 } 489 } 490 491 /* 492 * Swap useage is determined by taking the proportional swap used by 493 * VM objects backing the VM map. To make up for fractional losses, 494 * if the VM object has any swap use at all the associated map entries 495 * count for at least 1 swap page. 496 * 497 * No requirements. 498 */ 499 int 500 vmspace_swap_count(struct vmspace *vm) 501 { 502 vm_map_t map = &vm->vm_map; 503 vm_map_entry_t cur; 504 vm_object_t object; 505 int count = 0; 506 int n; 507 508 vmspace_hold(vm); 509 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 510 switch(cur->maptype) { 511 case VM_MAPTYPE_NORMAL: 512 case VM_MAPTYPE_VPAGETABLE: 513 if ((object = cur->object.vm_object) == NULL) 514 break; 515 if (object->swblock_count) { 516 n = (cur->end - cur->start) / PAGE_SIZE; 517 count += object->swblock_count * 518 SWAP_META_PAGES * n / object->size + 1; 519 } 520 break; 521 default: 522 break; 523 } 524 } 525 vmspace_drop(vm); 526 527 return(count); 528 } 529 530 /* 531 * Calculate the approximate number of anonymous pages in use by 532 * this vmspace. To make up for fractional losses, we count each 533 * VM object as having at least 1 anonymous page. 534 * 535 * No requirements. 536 */ 537 int 538 vmspace_anonymous_count(struct vmspace *vm) 539 { 540 vm_map_t map = &vm->vm_map; 541 vm_map_entry_t cur; 542 vm_object_t object; 543 int count = 0; 544 545 vmspace_hold(vm); 546 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 547 switch(cur->maptype) { 548 case VM_MAPTYPE_NORMAL: 549 case VM_MAPTYPE_VPAGETABLE: 550 if ((object = cur->object.vm_object) == NULL) 551 break; 552 if (object->type != OBJT_DEFAULT && 553 object->type != OBJT_SWAP) { 554 break; 555 } 556 count += object->resident_page_count; 557 break; 558 default: 559 break; 560 } 561 } 562 vmspace_drop(vm); 563 564 return(count); 565 } 566 567 /* 568 * Creates and returns a new empty VM map with the given physical map 569 * structure, and having the given lower and upper address bounds. 570 * 571 * No requirements. 572 */ 573 vm_map_t 574 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max) 575 { 576 if (result == NULL) 577 result = zalloc(mapzone); 578 vm_map_init(result, min, max, pmap); 579 return (result); 580 } 581 582 /* 583 * Initialize an existing vm_map structure such as that in the vmspace 584 * structure. The pmap is initialized elsewhere. 585 * 586 * No requirements. 587 */ 588 void 589 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap) 590 { 591 map->header.next = map->header.prev = &map->header; 592 RB_INIT(&map->rb_root); 593 map->nentries = 0; 594 map->size = 0; 595 map->system_map = 0; 596 map->min_offset = min; 597 map->max_offset = max; 598 map->pmap = pmap; 599 map->first_free = &map->header; 600 map->hint = &map->header; 601 map->timestamp = 0; 602 map->flags = 0; 603 lwkt_token_init(&map->token, "vm_map"); 604 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0); 605 } 606 607 /* 608 * Shadow the vm_map_entry's object. This typically needs to be done when 609 * a write fault is taken on an entry which had previously been cloned by 610 * fork(). The shared object (which might be NULL) must become private so 611 * we add a shadow layer above it. 612 * 613 * Object allocation for anonymous mappings is defered as long as possible. 614 * When creating a shadow, however, the underlying object must be instantiated 615 * so it can be shared. 616 * 617 * If the map segment is governed by a virtual page table then it is 618 * possible to address offsets beyond the mapped area. Just allocate 619 * a maximally sized object for this case. 620 * 621 * If addref is non-zero an additional reference is added to the returned 622 * entry. This mechanic exists because the additional reference might have 623 * to be added atomically and not after return to prevent a premature 624 * collapse. 625 * 626 * The vm_map must be exclusively locked. 627 * No other requirements. 628 */ 629 static 630 void 631 vm_map_entry_shadow(vm_map_entry_t entry, int addref) 632 { 633 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 634 vm_object_shadow(&entry->object.vm_object, &entry->offset, 635 0x7FFFFFFF, addref); /* XXX */ 636 } else { 637 vm_object_shadow(&entry->object.vm_object, &entry->offset, 638 atop(entry->end - entry->start), addref); 639 } 640 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 641 } 642 643 /* 644 * Allocate an object for a vm_map_entry. 645 * 646 * Object allocation for anonymous mappings is defered as long as possible. 647 * This function is called when we can defer no longer, generally when a map 648 * entry might be split or forked or takes a page fault. 649 * 650 * If the map segment is governed by a virtual page table then it is 651 * possible to address offsets beyond the mapped area. Just allocate 652 * a maximally sized object for this case. 653 * 654 * The vm_map must be exclusively locked. 655 * No other requirements. 656 */ 657 void 658 vm_map_entry_allocate_object(vm_map_entry_t entry) 659 { 660 vm_object_t obj; 661 662 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 663 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */ 664 } else { 665 obj = vm_object_allocate(OBJT_DEFAULT, 666 atop(entry->end - entry->start)); 667 } 668 entry->object.vm_object = obj; 669 entry->offset = 0; 670 } 671 672 /* 673 * Set an initial negative count so the first attempt to reserve 674 * space preloads a bunch of vm_map_entry's for this cpu. Also 675 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 676 * map a new page for vm_map_entry structures. SMP systems are 677 * particularly sensitive. 678 * 679 * This routine is called in early boot so we cannot just call 680 * vm_map_entry_reserve(). 681 * 682 * Called from the low level boot code only (for each cpu) 683 * 684 * WARNING! Take care not to have too-big a static/BSS structure here 685 * as MAXCPU can be 256+, otherwise the loader's 64MB heap 686 * can get blown out by the kernel plus the initrd image. 687 */ 688 void 689 vm_map_entry_reserve_cpu_init(globaldata_t gd) 690 { 691 vm_map_entry_t entry; 692 int count; 693 int i; 694 695 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2; 696 if (gd->gd_cpuid == 0) { 697 entry = &cpu_map_entry_init_bsp[0]; 698 count = MAPENTRYBSP_CACHE; 699 } else { 700 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0]; 701 count = MAPENTRYAP_CACHE; 702 } 703 for (i = 0; i < count; ++i, ++entry) { 704 entry->next = gd->gd_vme_base; 705 gd->gd_vme_base = entry; 706 } 707 } 708 709 /* 710 * Reserves vm_map_entry structures so code later on can manipulate 711 * map_entry structures within a locked map without blocking trying 712 * to allocate a new vm_map_entry. 713 * 714 * No requirements. 715 */ 716 int 717 vm_map_entry_reserve(int count) 718 { 719 struct globaldata *gd = mycpu; 720 vm_map_entry_t entry; 721 722 /* 723 * Make sure we have enough structures in gd_vme_base to handle 724 * the reservation request. 725 * 726 * The critical section protects access to the per-cpu gd. 727 */ 728 crit_enter(); 729 while (gd->gd_vme_avail < count) { 730 entry = zalloc(mapentzone); 731 entry->next = gd->gd_vme_base; 732 gd->gd_vme_base = entry; 733 ++gd->gd_vme_avail; 734 } 735 gd->gd_vme_avail -= count; 736 crit_exit(); 737 738 return(count); 739 } 740 741 /* 742 * Releases previously reserved vm_map_entry structures that were not 743 * used. If we have too much junk in our per-cpu cache clean some of 744 * it out. 745 * 746 * No requirements. 747 */ 748 void 749 vm_map_entry_release(int count) 750 { 751 struct globaldata *gd = mycpu; 752 vm_map_entry_t entry; 753 754 crit_enter(); 755 gd->gd_vme_avail += count; 756 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 757 entry = gd->gd_vme_base; 758 KKASSERT(entry != NULL); 759 gd->gd_vme_base = entry->next; 760 --gd->gd_vme_avail; 761 crit_exit(); 762 zfree(mapentzone, entry); 763 crit_enter(); 764 } 765 crit_exit(); 766 } 767 768 /* 769 * Reserve map entry structures for use in kernel_map itself. These 770 * entries have *ALREADY* been reserved on a per-cpu basis when the map 771 * was inited. This function is used by zalloc() to avoid a recursion 772 * when zalloc() itself needs to allocate additional kernel memory. 773 * 774 * This function works like the normal reserve but does not load the 775 * vm_map_entry cache (because that would result in an infinite 776 * recursion). Note that gd_vme_avail may go negative. This is expected. 777 * 778 * Any caller of this function must be sure to renormalize after 779 * potentially eating entries to ensure that the reserve supply 780 * remains intact. 781 * 782 * No requirements. 783 */ 784 int 785 vm_map_entry_kreserve(int count) 786 { 787 struct globaldata *gd = mycpu; 788 789 crit_enter(); 790 gd->gd_vme_avail -= count; 791 crit_exit(); 792 KASSERT(gd->gd_vme_base != NULL, 793 ("no reserved entries left, gd_vme_avail = %d", 794 gd->gd_vme_avail)); 795 return(count); 796 } 797 798 /* 799 * Release previously reserved map entries for kernel_map. We do not 800 * attempt to clean up like the normal release function as this would 801 * cause an unnecessary (but probably not fatal) deep procedure call. 802 * 803 * No requirements. 804 */ 805 void 806 vm_map_entry_krelease(int count) 807 { 808 struct globaldata *gd = mycpu; 809 810 crit_enter(); 811 gd->gd_vme_avail += count; 812 crit_exit(); 813 } 814 815 /* 816 * Allocates a VM map entry for insertion. No entry fields are filled in. 817 * 818 * The entries should have previously been reserved. The reservation count 819 * is tracked in (*countp). 820 * 821 * No requirements. 822 */ 823 static vm_map_entry_t 824 vm_map_entry_create(vm_map_t map, int *countp) 825 { 826 struct globaldata *gd = mycpu; 827 vm_map_entry_t entry; 828 829 KKASSERT(*countp > 0); 830 --*countp; 831 crit_enter(); 832 entry = gd->gd_vme_base; 833 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 834 gd->gd_vme_base = entry->next; 835 crit_exit(); 836 837 return(entry); 838 } 839 840 /* 841 * Dispose of a vm_map_entry that is no longer being referenced. 842 * 843 * No requirements. 844 */ 845 static void 846 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 847 { 848 struct globaldata *gd = mycpu; 849 850 KKASSERT(map->hint != entry); 851 KKASSERT(map->first_free != entry); 852 853 ++*countp; 854 crit_enter(); 855 entry->next = gd->gd_vme_base; 856 gd->gd_vme_base = entry; 857 crit_exit(); 858 } 859 860 861 /* 862 * Insert/remove entries from maps. 863 * 864 * The related map must be exclusively locked. 865 * The caller must hold map->token 866 * No other requirements. 867 */ 868 static __inline void 869 vm_map_entry_link(vm_map_t map, 870 vm_map_entry_t after_where, 871 vm_map_entry_t entry) 872 { 873 ASSERT_VM_MAP_LOCKED(map); 874 875 map->nentries++; 876 entry->prev = after_where; 877 entry->next = after_where->next; 878 entry->next->prev = entry; 879 after_where->next = entry; 880 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry)) 881 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry); 882 } 883 884 static __inline void 885 vm_map_entry_unlink(vm_map_t map, 886 vm_map_entry_t entry) 887 { 888 vm_map_entry_t prev; 889 vm_map_entry_t next; 890 891 ASSERT_VM_MAP_LOCKED(map); 892 893 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 894 panic("vm_map_entry_unlink: attempt to mess with " 895 "locked entry! %p", entry); 896 } 897 prev = entry->prev; 898 next = entry->next; 899 next->prev = prev; 900 prev->next = next; 901 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry); 902 map->nentries--; 903 } 904 905 /* 906 * Finds the map entry containing (or immediately preceding) the specified 907 * address in the given map. The entry is returned in (*entry). 908 * 909 * The boolean result indicates whether the address is actually contained 910 * in the map. 911 * 912 * The related map must be locked. 913 * No other requirements. 914 */ 915 boolean_t 916 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry) 917 { 918 vm_map_entry_t tmp; 919 vm_map_entry_t last; 920 921 ASSERT_VM_MAP_LOCKED(map); 922 #if 0 923 /* 924 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive 925 * the hint code with the red-black lookup meets with system crashes 926 * and lockups. We do not yet know why. 927 * 928 * It is possible that the problem is related to the setting 929 * of the hint during map_entry deletion, in the code specified 930 * at the GGG comment later on in this file. 931 * 932 * YYY More likely it's because this function can be called with 933 * a shared lock on the map, resulting in map->hint updates possibly 934 * racing. Fixed now but untested. 935 */ 936 /* 937 * Quickly check the cached hint, there's a good chance of a match. 938 */ 939 tmp = map->hint; 940 cpu_ccfence(); 941 if (tmp != &map->header) { 942 if (address >= tmp->start && address < tmp->end) { 943 *entry = tmp; 944 return(TRUE); 945 } 946 } 947 #endif 948 949 /* 950 * Locate the record from the top of the tree. 'last' tracks the 951 * closest prior record and is returned if no match is found, which 952 * in binary tree terms means tracking the most recent right-branch 953 * taken. If there is no prior record, &map->header is returned. 954 */ 955 last = &map->header; 956 tmp = RB_ROOT(&map->rb_root); 957 958 while (tmp) { 959 if (address >= tmp->start) { 960 if (address < tmp->end) { 961 *entry = tmp; 962 map->hint = tmp; 963 return(TRUE); 964 } 965 last = tmp; 966 tmp = RB_RIGHT(tmp, rb_entry); 967 } else { 968 tmp = RB_LEFT(tmp, rb_entry); 969 } 970 } 971 *entry = last; 972 return (FALSE); 973 } 974 975 /* 976 * Inserts the given whole VM object into the target map at the specified 977 * address range. The object's size should match that of the address range. 978 * 979 * The map must be exclusively locked. 980 * The object must be held. 981 * The caller must have reserved sufficient vm_map_entry structures. 982 * 983 * If object is non-NULL, ref count must be bumped by caller prior to 984 * making call to account for the new entry. 985 */ 986 int 987 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux, 988 vm_ooffset_t offset, vm_offset_t start, vm_offset_t end, 989 vm_maptype_t maptype, 990 vm_prot_t prot, vm_prot_t max, int cow) 991 { 992 vm_map_entry_t new_entry; 993 vm_map_entry_t prev_entry; 994 vm_map_entry_t temp_entry; 995 vm_eflags_t protoeflags; 996 int must_drop = 0; 997 vm_object_t object; 998 999 if (maptype == VM_MAPTYPE_UKSMAP) 1000 object = NULL; 1001 else 1002 object = map_object; 1003 1004 ASSERT_VM_MAP_LOCKED(map); 1005 if (object) 1006 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1007 1008 /* 1009 * Check that the start and end points are not bogus. 1010 */ 1011 if ((start < map->min_offset) || (end > map->max_offset) || 1012 (start >= end)) 1013 return (KERN_INVALID_ADDRESS); 1014 1015 /* 1016 * Find the entry prior to the proposed starting address; if it's part 1017 * of an existing entry, this range is bogus. 1018 */ 1019 if (vm_map_lookup_entry(map, start, &temp_entry)) 1020 return (KERN_NO_SPACE); 1021 1022 prev_entry = temp_entry; 1023 1024 /* 1025 * Assert that the next entry doesn't overlap the end point. 1026 */ 1027 1028 if ((prev_entry->next != &map->header) && 1029 (prev_entry->next->start < end)) 1030 return (KERN_NO_SPACE); 1031 1032 protoeflags = 0; 1033 1034 if (cow & MAP_COPY_ON_WRITE) 1035 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 1036 1037 if (cow & MAP_NOFAULT) { 1038 protoeflags |= MAP_ENTRY_NOFAULT; 1039 1040 KASSERT(object == NULL, 1041 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 1042 } 1043 if (cow & MAP_DISABLE_SYNCER) 1044 protoeflags |= MAP_ENTRY_NOSYNC; 1045 if (cow & MAP_DISABLE_COREDUMP) 1046 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1047 if (cow & MAP_IS_STACK) 1048 protoeflags |= MAP_ENTRY_STACK; 1049 if (cow & MAP_IS_KSTACK) 1050 protoeflags |= MAP_ENTRY_KSTACK; 1051 1052 lwkt_gettoken(&map->token); 1053 1054 if (object) { 1055 /* 1056 * When object is non-NULL, it could be shared with another 1057 * process. We have to set or clear OBJ_ONEMAPPING 1058 * appropriately. 1059 * 1060 * NOTE: This flag is only applicable to DEFAULT and SWAP 1061 * objects and will already be clear in other types 1062 * of objects, so a shared object lock is ok for 1063 * VNODE objects. 1064 */ 1065 if ((object->ref_count > 1) || (object->shadow_count != 0)) { 1066 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1067 } 1068 } 1069 else if ((prev_entry != &map->header) && 1070 (prev_entry->eflags == protoeflags) && 1071 (prev_entry->end == start) && 1072 (prev_entry->wired_count == 0) && 1073 prev_entry->maptype == maptype && 1074 maptype == VM_MAPTYPE_NORMAL && 1075 ((prev_entry->object.vm_object == NULL) || 1076 vm_object_coalesce(prev_entry->object.vm_object, 1077 OFF_TO_IDX(prev_entry->offset), 1078 (vm_size_t)(prev_entry->end - prev_entry->start), 1079 (vm_size_t)(end - prev_entry->end)))) { 1080 /* 1081 * We were able to extend the object. Determine if we 1082 * can extend the previous map entry to include the 1083 * new range as well. 1084 */ 1085 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 1086 (prev_entry->protection == prot) && 1087 (prev_entry->max_protection == max)) { 1088 map->size += (end - prev_entry->end); 1089 prev_entry->end = end; 1090 vm_map_simplify_entry(map, prev_entry, countp); 1091 lwkt_reltoken(&map->token); 1092 return (KERN_SUCCESS); 1093 } 1094 1095 /* 1096 * If we can extend the object but cannot extend the 1097 * map entry, we have to create a new map entry. We 1098 * must bump the ref count on the extended object to 1099 * account for it. object may be NULL. 1100 * 1101 * XXX if object is NULL should we set offset to 0 here ? 1102 */ 1103 object = prev_entry->object.vm_object; 1104 offset = prev_entry->offset + 1105 (prev_entry->end - prev_entry->start); 1106 if (object) { 1107 vm_object_hold(object); 1108 vm_object_chain_wait(object, 0); 1109 vm_object_reference_locked(object); 1110 must_drop = 1; 1111 map_object = object; 1112 } 1113 } 1114 1115 /* 1116 * NOTE: if conditionals fail, object can be NULL here. This occurs 1117 * in things like the buffer map where we manage kva but do not manage 1118 * backing objects. 1119 */ 1120 1121 /* 1122 * Create a new entry 1123 */ 1124 1125 new_entry = vm_map_entry_create(map, countp); 1126 new_entry->start = start; 1127 new_entry->end = end; 1128 1129 new_entry->maptype = maptype; 1130 new_entry->eflags = protoeflags; 1131 new_entry->object.map_object = map_object; 1132 new_entry->aux.master_pde = 0; /* in case size is different */ 1133 new_entry->aux.map_aux = map_aux; 1134 new_entry->offset = offset; 1135 1136 new_entry->inheritance = VM_INHERIT_DEFAULT; 1137 new_entry->protection = prot; 1138 new_entry->max_protection = max; 1139 new_entry->wired_count = 0; 1140 1141 /* 1142 * Insert the new entry into the list 1143 */ 1144 1145 vm_map_entry_link(map, prev_entry, new_entry); 1146 map->size += new_entry->end - new_entry->start; 1147 1148 /* 1149 * Update the free space hint. Entries cannot overlap. 1150 * An exact comparison is needed to avoid matching 1151 * against the map->header. 1152 */ 1153 if ((map->first_free == prev_entry) && 1154 (prev_entry->end == new_entry->start)) { 1155 map->first_free = new_entry; 1156 } 1157 1158 #if 0 1159 /* 1160 * Temporarily removed to avoid MAP_STACK panic, due to 1161 * MAP_STACK being a huge hack. Will be added back in 1162 * when MAP_STACK (and the user stack mapping) is fixed. 1163 */ 1164 /* 1165 * It may be possible to simplify the entry 1166 */ 1167 vm_map_simplify_entry(map, new_entry, countp); 1168 #endif 1169 1170 /* 1171 * Try to pre-populate the page table. Mappings governed by virtual 1172 * page tables cannot be prepopulated without a lot of work, so 1173 * don't try. 1174 */ 1175 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) && 1176 maptype != VM_MAPTYPE_VPAGETABLE && 1177 maptype != VM_MAPTYPE_UKSMAP) { 1178 int dorelock = 0; 1179 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) { 1180 dorelock = 1; 1181 vm_object_lock_swap(); 1182 vm_object_drop(object); 1183 } 1184 pmap_object_init_pt(map->pmap, start, prot, 1185 object, OFF_TO_IDX(offset), end - start, 1186 cow & MAP_PREFAULT_PARTIAL); 1187 if (dorelock) { 1188 vm_object_hold(object); 1189 vm_object_lock_swap(); 1190 } 1191 } 1192 if (must_drop) 1193 vm_object_drop(object); 1194 1195 lwkt_reltoken(&map->token); 1196 return (KERN_SUCCESS); 1197 } 1198 1199 /* 1200 * Find sufficient space for `length' bytes in the given map, starting at 1201 * `start'. Returns 0 on success, 1 on no space. 1202 * 1203 * This function will returned an arbitrarily aligned pointer. If no 1204 * particular alignment is required you should pass align as 1. Note that 1205 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 1206 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 1207 * argument. 1208 * 1209 * 'align' should be a power of 2 but is not required to be. 1210 * 1211 * The map must be exclusively locked. 1212 * No other requirements. 1213 */ 1214 int 1215 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length, 1216 vm_size_t align, int flags, vm_offset_t *addr) 1217 { 1218 vm_map_entry_t entry, next; 1219 vm_offset_t end; 1220 vm_offset_t align_mask; 1221 1222 if (start < map->min_offset) 1223 start = map->min_offset; 1224 if (start > map->max_offset) 1225 return (1); 1226 1227 /* 1228 * If the alignment is not a power of 2 we will have to use 1229 * a mod/division, set align_mask to a special value. 1230 */ 1231 if ((align | (align - 1)) + 1 != (align << 1)) 1232 align_mask = (vm_offset_t)-1; 1233 else 1234 align_mask = align - 1; 1235 1236 /* 1237 * Look for the first possible address; if there's already something 1238 * at this address, we have to start after it. 1239 */ 1240 if (start == map->min_offset) { 1241 if ((entry = map->first_free) != &map->header) 1242 start = entry->end; 1243 } else { 1244 vm_map_entry_t tmp; 1245 1246 if (vm_map_lookup_entry(map, start, &tmp)) 1247 start = tmp->end; 1248 entry = tmp; 1249 } 1250 1251 /* 1252 * Look through the rest of the map, trying to fit a new region in the 1253 * gap between existing regions, or after the very last region. 1254 */ 1255 for (;; start = (entry = next)->end) { 1256 /* 1257 * Adjust the proposed start by the requested alignment, 1258 * be sure that we didn't wrap the address. 1259 */ 1260 if (align_mask == (vm_offset_t)-1) 1261 end = roundup(start, align); 1262 else 1263 end = (start + align_mask) & ~align_mask; 1264 if (end < start) 1265 return (1); 1266 start = end; 1267 /* 1268 * Find the end of the proposed new region. Be sure we didn't 1269 * go beyond the end of the map, or wrap around the address. 1270 * Then check to see if this is the last entry or if the 1271 * proposed end fits in the gap between this and the next 1272 * entry. 1273 */ 1274 end = start + length; 1275 if (end > map->max_offset || end < start) 1276 return (1); 1277 next = entry->next; 1278 1279 /* 1280 * If the next entry's start address is beyond the desired 1281 * end address we may have found a good entry. 1282 * 1283 * If the next entry is a stack mapping we do not map into 1284 * the stack's reserved space. 1285 * 1286 * XXX continue to allow mapping into the stack's reserved 1287 * space if doing a MAP_STACK mapping inside a MAP_STACK 1288 * mapping, for backwards compatibility. But the caller 1289 * really should use MAP_STACK | MAP_TRYFIXED if they 1290 * want to do that. 1291 */ 1292 if (next == &map->header) 1293 break; 1294 if (next->start >= end) { 1295 if ((next->eflags & MAP_ENTRY_STACK) == 0) 1296 break; 1297 if (flags & MAP_STACK) 1298 break; 1299 if (next->start - next->aux.avail_ssize >= end) 1300 break; 1301 } 1302 } 1303 map->hint = entry; 1304 1305 /* 1306 * Grow the kernel_map if necessary. pmap_growkernel() will panic 1307 * if it fails. The kernel_map is locked and nothing can steal 1308 * our address space if pmap_growkernel() blocks. 1309 * 1310 * NOTE: This may be unconditionally called for kldload areas on 1311 * x86_64 because these do not bump kernel_vm_end (which would 1312 * fill 128G worth of page tables!). Therefore we must not 1313 * retry. 1314 */ 1315 if (map == &kernel_map) { 1316 vm_offset_t kstop; 1317 1318 kstop = round_page(start + length); 1319 if (kstop > kernel_vm_end) 1320 pmap_growkernel(start, kstop); 1321 } 1322 *addr = start; 1323 return (0); 1324 } 1325 1326 /* 1327 * vm_map_find finds an unallocated region in the target address map with 1328 * the given length and allocates it. The search is defined to be first-fit 1329 * from the specified address; the region found is returned in the same 1330 * parameter. 1331 * 1332 * If object is non-NULL, ref count must be bumped by caller 1333 * prior to making call to account for the new entry. 1334 * 1335 * No requirements. This function will lock the map temporarily. 1336 */ 1337 int 1338 vm_map_find(vm_map_t map, void *map_object, void *map_aux, 1339 vm_ooffset_t offset, vm_offset_t *addr, 1340 vm_size_t length, vm_size_t align, 1341 boolean_t fitit, 1342 vm_maptype_t maptype, 1343 vm_prot_t prot, vm_prot_t max, 1344 int cow) 1345 { 1346 vm_offset_t start; 1347 vm_object_t object; 1348 int result; 1349 int count; 1350 1351 if (maptype == VM_MAPTYPE_UKSMAP) 1352 object = NULL; 1353 else 1354 object = map_object; 1355 1356 start = *addr; 1357 1358 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1359 vm_map_lock(map); 1360 if (object) 1361 vm_object_hold_shared(object); 1362 if (fitit) { 1363 if (vm_map_findspace(map, start, length, align, 0, addr)) { 1364 if (object) 1365 vm_object_drop(object); 1366 vm_map_unlock(map); 1367 vm_map_entry_release(count); 1368 return (KERN_NO_SPACE); 1369 } 1370 start = *addr; 1371 } 1372 result = vm_map_insert(map, &count, map_object, map_aux, 1373 offset, start, start + length, 1374 maptype, prot, max, cow); 1375 if (object) 1376 vm_object_drop(object); 1377 vm_map_unlock(map); 1378 vm_map_entry_release(count); 1379 1380 return (result); 1381 } 1382 1383 /* 1384 * Simplify the given map entry by merging with either neighbor. This 1385 * routine also has the ability to merge with both neighbors. 1386 * 1387 * This routine guarentees that the passed entry remains valid (though 1388 * possibly extended). When merging, this routine may delete one or 1389 * both neighbors. No action is taken on entries which have their 1390 * in-transition flag set. 1391 * 1392 * The map must be exclusively locked. 1393 */ 1394 void 1395 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 1396 { 1397 vm_map_entry_t next, prev; 1398 vm_size_t prevsize, esize; 1399 1400 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1401 ++mycpu->gd_cnt.v_intrans_coll; 1402 return; 1403 } 1404 1405 if (entry->maptype == VM_MAPTYPE_SUBMAP) 1406 return; 1407 if (entry->maptype == VM_MAPTYPE_UKSMAP) 1408 return; 1409 1410 prev = entry->prev; 1411 if (prev != &map->header) { 1412 prevsize = prev->end - prev->start; 1413 if ( (prev->end == entry->start) && 1414 (prev->maptype == entry->maptype) && 1415 (prev->object.vm_object == entry->object.vm_object) && 1416 (!prev->object.vm_object || 1417 (prev->offset + prevsize == entry->offset)) && 1418 (prev->eflags == entry->eflags) && 1419 (prev->protection == entry->protection) && 1420 (prev->max_protection == entry->max_protection) && 1421 (prev->inheritance == entry->inheritance) && 1422 (prev->wired_count == entry->wired_count)) { 1423 if (map->first_free == prev) 1424 map->first_free = entry; 1425 if (map->hint == prev) 1426 map->hint = entry; 1427 vm_map_entry_unlink(map, prev); 1428 entry->start = prev->start; 1429 entry->offset = prev->offset; 1430 if (prev->object.vm_object) 1431 vm_object_deallocate(prev->object.vm_object); 1432 vm_map_entry_dispose(map, prev, countp); 1433 } 1434 } 1435 1436 next = entry->next; 1437 if (next != &map->header) { 1438 esize = entry->end - entry->start; 1439 if ((entry->end == next->start) && 1440 (next->maptype == entry->maptype) && 1441 (next->object.vm_object == entry->object.vm_object) && 1442 (!entry->object.vm_object || 1443 (entry->offset + esize == next->offset)) && 1444 (next->eflags == entry->eflags) && 1445 (next->protection == entry->protection) && 1446 (next->max_protection == entry->max_protection) && 1447 (next->inheritance == entry->inheritance) && 1448 (next->wired_count == entry->wired_count)) { 1449 if (map->first_free == next) 1450 map->first_free = entry; 1451 if (map->hint == next) 1452 map->hint = entry; 1453 vm_map_entry_unlink(map, next); 1454 entry->end = next->end; 1455 if (next->object.vm_object) 1456 vm_object_deallocate(next->object.vm_object); 1457 vm_map_entry_dispose(map, next, countp); 1458 } 1459 } 1460 } 1461 1462 /* 1463 * Asserts that the given entry begins at or after the specified address. 1464 * If necessary, it splits the entry into two. 1465 */ 1466 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1467 { \ 1468 if (startaddr > entry->start) \ 1469 _vm_map_clip_start(map, entry, startaddr, countp); \ 1470 } 1471 1472 /* 1473 * This routine is called only when it is known that the entry must be split. 1474 * 1475 * The map must be exclusively locked. 1476 */ 1477 static void 1478 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, 1479 int *countp) 1480 { 1481 vm_map_entry_t new_entry; 1482 1483 /* 1484 * Split off the front portion -- note that we must insert the new 1485 * entry BEFORE this one, so that this entry has the specified 1486 * starting address. 1487 */ 1488 1489 vm_map_simplify_entry(map, entry, countp); 1490 1491 /* 1492 * If there is no object backing this entry, we might as well create 1493 * one now. If we defer it, an object can get created after the map 1494 * is clipped, and individual objects will be created for the split-up 1495 * map. This is a bit of a hack, but is also about the best place to 1496 * put this improvement. 1497 */ 1498 if (entry->object.vm_object == NULL && !map->system_map) { 1499 vm_map_entry_allocate_object(entry); 1500 } 1501 1502 new_entry = vm_map_entry_create(map, countp); 1503 *new_entry = *entry; 1504 1505 new_entry->end = start; 1506 entry->offset += (start - entry->start); 1507 entry->start = start; 1508 1509 vm_map_entry_link(map, entry->prev, new_entry); 1510 1511 switch(entry->maptype) { 1512 case VM_MAPTYPE_NORMAL: 1513 case VM_MAPTYPE_VPAGETABLE: 1514 if (new_entry->object.vm_object) { 1515 vm_object_hold(new_entry->object.vm_object); 1516 vm_object_chain_wait(new_entry->object.vm_object, 0); 1517 vm_object_reference_locked(new_entry->object.vm_object); 1518 vm_object_drop(new_entry->object.vm_object); 1519 } 1520 break; 1521 default: 1522 break; 1523 } 1524 } 1525 1526 /* 1527 * Asserts that the given entry ends at or before the specified address. 1528 * If necessary, it splits the entry into two. 1529 * 1530 * The map must be exclusively locked. 1531 */ 1532 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1533 { \ 1534 if (endaddr < entry->end) \ 1535 _vm_map_clip_end(map, entry, endaddr, countp); \ 1536 } 1537 1538 /* 1539 * This routine is called only when it is known that the entry must be split. 1540 * 1541 * The map must be exclusively locked. 1542 */ 1543 static void 1544 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, 1545 int *countp) 1546 { 1547 vm_map_entry_t new_entry; 1548 1549 /* 1550 * If there is no object backing this entry, we might as well create 1551 * one now. If we defer it, an object can get created after the map 1552 * is clipped, and individual objects will be created for the split-up 1553 * map. This is a bit of a hack, but is also about the best place to 1554 * put this improvement. 1555 */ 1556 1557 if (entry->object.vm_object == NULL && !map->system_map) { 1558 vm_map_entry_allocate_object(entry); 1559 } 1560 1561 /* 1562 * Create a new entry and insert it AFTER the specified entry 1563 */ 1564 1565 new_entry = vm_map_entry_create(map, countp); 1566 *new_entry = *entry; 1567 1568 new_entry->start = entry->end = end; 1569 new_entry->offset += (end - entry->start); 1570 1571 vm_map_entry_link(map, entry, new_entry); 1572 1573 switch(entry->maptype) { 1574 case VM_MAPTYPE_NORMAL: 1575 case VM_MAPTYPE_VPAGETABLE: 1576 if (new_entry->object.vm_object) { 1577 vm_object_hold(new_entry->object.vm_object); 1578 vm_object_chain_wait(new_entry->object.vm_object, 0); 1579 vm_object_reference_locked(new_entry->object.vm_object); 1580 vm_object_drop(new_entry->object.vm_object); 1581 } 1582 break; 1583 default: 1584 break; 1585 } 1586 } 1587 1588 /* 1589 * Asserts that the starting and ending region addresses fall within the 1590 * valid range for the map. 1591 */ 1592 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1593 { \ 1594 if (start < vm_map_min(map)) \ 1595 start = vm_map_min(map); \ 1596 if (end > vm_map_max(map)) \ 1597 end = vm_map_max(map); \ 1598 if (start > end) \ 1599 start = end; \ 1600 } 1601 1602 /* 1603 * Used to block when an in-transition collison occurs. The map 1604 * is unlocked for the sleep and relocked before the return. 1605 */ 1606 void 1607 vm_map_transition_wait(vm_map_t map) 1608 { 1609 tsleep_interlock(map, 0); 1610 vm_map_unlock(map); 1611 tsleep(map, PINTERLOCKED, "vment", 0); 1612 vm_map_lock(map); 1613 } 1614 1615 /* 1616 * When we do blocking operations with the map lock held it is 1617 * possible that a clip might have occured on our in-transit entry, 1618 * requiring an adjustment to the entry in our loop. These macros 1619 * help the pageable and clip_range code deal with the case. The 1620 * conditional costs virtually nothing if no clipping has occured. 1621 */ 1622 1623 #define CLIP_CHECK_BACK(entry, save_start) \ 1624 do { \ 1625 while (entry->start != save_start) { \ 1626 entry = entry->prev; \ 1627 KASSERT(entry != &map->header, ("bad entry clip")); \ 1628 } \ 1629 } while(0) 1630 1631 #define CLIP_CHECK_FWD(entry, save_end) \ 1632 do { \ 1633 while (entry->end != save_end) { \ 1634 entry = entry->next; \ 1635 KASSERT(entry != &map->header, ("bad entry clip")); \ 1636 } \ 1637 } while(0) 1638 1639 1640 /* 1641 * Clip the specified range and return the base entry. The 1642 * range may cover several entries starting at the returned base 1643 * and the first and last entry in the covering sequence will be 1644 * properly clipped to the requested start and end address. 1645 * 1646 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1647 * flag. 1648 * 1649 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1650 * covered by the requested range. 1651 * 1652 * The map must be exclusively locked on entry and will remain locked 1653 * on return. If no range exists or the range contains holes and you 1654 * specified that no holes were allowed, NULL will be returned. This 1655 * routine may temporarily unlock the map in order avoid a deadlock when 1656 * sleeping. 1657 */ 1658 static 1659 vm_map_entry_t 1660 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1661 int *countp, int flags) 1662 { 1663 vm_map_entry_t start_entry; 1664 vm_map_entry_t entry; 1665 1666 /* 1667 * Locate the entry and effect initial clipping. The in-transition 1668 * case does not occur very often so do not try to optimize it. 1669 */ 1670 again: 1671 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1672 return (NULL); 1673 entry = start_entry; 1674 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1675 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1676 ++mycpu->gd_cnt.v_intrans_coll; 1677 ++mycpu->gd_cnt.v_intrans_wait; 1678 vm_map_transition_wait(map); 1679 /* 1680 * entry and/or start_entry may have been clipped while 1681 * we slept, or may have gone away entirely. We have 1682 * to restart from the lookup. 1683 */ 1684 goto again; 1685 } 1686 1687 /* 1688 * Since we hold an exclusive map lock we do not have to restart 1689 * after clipping, even though clipping may block in zalloc. 1690 */ 1691 vm_map_clip_start(map, entry, start, countp); 1692 vm_map_clip_end(map, entry, end, countp); 1693 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1694 1695 /* 1696 * Scan entries covered by the range. When working on the next 1697 * entry a restart need only re-loop on the current entry which 1698 * we have already locked, since 'next' may have changed. Also, 1699 * even though entry is safe, it may have been clipped so we 1700 * have to iterate forwards through the clip after sleeping. 1701 */ 1702 while (entry->next != &map->header && entry->next->start < end) { 1703 vm_map_entry_t next = entry->next; 1704 1705 if (flags & MAP_CLIP_NO_HOLES) { 1706 if (next->start > entry->end) { 1707 vm_map_unclip_range(map, start_entry, 1708 start, entry->end, countp, flags); 1709 return(NULL); 1710 } 1711 } 1712 1713 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1714 vm_offset_t save_end = entry->end; 1715 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1716 ++mycpu->gd_cnt.v_intrans_coll; 1717 ++mycpu->gd_cnt.v_intrans_wait; 1718 vm_map_transition_wait(map); 1719 1720 /* 1721 * clips might have occured while we blocked. 1722 */ 1723 CLIP_CHECK_FWD(entry, save_end); 1724 CLIP_CHECK_BACK(start_entry, start); 1725 continue; 1726 } 1727 /* 1728 * No restart necessary even though clip_end may block, we 1729 * are holding the map lock. 1730 */ 1731 vm_map_clip_end(map, next, end, countp); 1732 next->eflags |= MAP_ENTRY_IN_TRANSITION; 1733 entry = next; 1734 } 1735 if (flags & MAP_CLIP_NO_HOLES) { 1736 if (entry->end != end) { 1737 vm_map_unclip_range(map, start_entry, 1738 start, entry->end, countp, flags); 1739 return(NULL); 1740 } 1741 } 1742 return(start_entry); 1743 } 1744 1745 /* 1746 * Undo the effect of vm_map_clip_range(). You should pass the same 1747 * flags and the same range that you passed to vm_map_clip_range(). 1748 * This code will clear the in-transition flag on the entries and 1749 * wake up anyone waiting. This code will also simplify the sequence 1750 * and attempt to merge it with entries before and after the sequence. 1751 * 1752 * The map must be locked on entry and will remain locked on return. 1753 * 1754 * Note that you should also pass the start_entry returned by 1755 * vm_map_clip_range(). However, if you block between the two calls 1756 * with the map unlocked please be aware that the start_entry may 1757 * have been clipped and you may need to scan it backwards to find 1758 * the entry corresponding with the original start address. You are 1759 * responsible for this, vm_map_unclip_range() expects the correct 1760 * start_entry to be passed to it and will KASSERT otherwise. 1761 */ 1762 static 1763 void 1764 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry, 1765 vm_offset_t start, vm_offset_t end, 1766 int *countp, int flags) 1767 { 1768 vm_map_entry_t entry; 1769 1770 entry = start_entry; 1771 1772 KASSERT(entry->start == start, ("unclip_range: illegal base entry")); 1773 while (entry != &map->header && entry->start < end) { 1774 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, 1775 ("in-transition flag not set during unclip on: %p", 1776 entry)); 1777 KASSERT(entry->end <= end, 1778 ("unclip_range: tail wasn't clipped")); 1779 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 1780 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 1781 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 1782 wakeup(map); 1783 } 1784 entry = entry->next; 1785 } 1786 1787 /* 1788 * Simplification does not block so there is no restart case. 1789 */ 1790 entry = start_entry; 1791 while (entry != &map->header && entry->start < end) { 1792 vm_map_simplify_entry(map, entry, countp); 1793 entry = entry->next; 1794 } 1795 } 1796 1797 /* 1798 * Mark the given range as handled by a subordinate map. 1799 * 1800 * This range must have been created with vm_map_find(), and no other 1801 * operations may have been performed on this range prior to calling 1802 * vm_map_submap(). 1803 * 1804 * Submappings cannot be removed. 1805 * 1806 * No requirements. 1807 */ 1808 int 1809 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 1810 { 1811 vm_map_entry_t entry; 1812 int result = KERN_INVALID_ARGUMENT; 1813 int count; 1814 1815 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1816 vm_map_lock(map); 1817 1818 VM_MAP_RANGE_CHECK(map, start, end); 1819 1820 if (vm_map_lookup_entry(map, start, &entry)) { 1821 vm_map_clip_start(map, entry, start, &count); 1822 } else { 1823 entry = entry->next; 1824 } 1825 1826 vm_map_clip_end(map, entry, end, &count); 1827 1828 if ((entry->start == start) && (entry->end == end) && 1829 ((entry->eflags & MAP_ENTRY_COW) == 0) && 1830 (entry->object.vm_object == NULL)) { 1831 entry->object.sub_map = submap; 1832 entry->maptype = VM_MAPTYPE_SUBMAP; 1833 result = KERN_SUCCESS; 1834 } 1835 vm_map_unlock(map); 1836 vm_map_entry_release(count); 1837 1838 return (result); 1839 } 1840 1841 /* 1842 * Sets the protection of the specified address region in the target map. 1843 * If "set_max" is specified, the maximum protection is to be set; 1844 * otherwise, only the current protection is affected. 1845 * 1846 * The protection is not applicable to submaps, but is applicable to normal 1847 * maps and maps governed by virtual page tables. For example, when operating 1848 * on a virtual page table our protection basically controls how COW occurs 1849 * on the backing object, whereas the virtual page table abstraction itself 1850 * is an abstraction for userland. 1851 * 1852 * No requirements. 1853 */ 1854 int 1855 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 1856 vm_prot_t new_prot, boolean_t set_max) 1857 { 1858 vm_map_entry_t current; 1859 vm_map_entry_t entry; 1860 int count; 1861 1862 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1863 vm_map_lock(map); 1864 1865 VM_MAP_RANGE_CHECK(map, start, end); 1866 1867 if (vm_map_lookup_entry(map, start, &entry)) { 1868 vm_map_clip_start(map, entry, start, &count); 1869 } else { 1870 entry = entry->next; 1871 } 1872 1873 /* 1874 * Make a first pass to check for protection violations. 1875 */ 1876 current = entry; 1877 while ((current != &map->header) && (current->start < end)) { 1878 if (current->maptype == VM_MAPTYPE_SUBMAP) { 1879 vm_map_unlock(map); 1880 vm_map_entry_release(count); 1881 return (KERN_INVALID_ARGUMENT); 1882 } 1883 if ((new_prot & current->max_protection) != new_prot) { 1884 vm_map_unlock(map); 1885 vm_map_entry_release(count); 1886 return (KERN_PROTECTION_FAILURE); 1887 } 1888 current = current->next; 1889 } 1890 1891 /* 1892 * Go back and fix up protections. [Note that clipping is not 1893 * necessary the second time.] 1894 */ 1895 current = entry; 1896 1897 while ((current != &map->header) && (current->start < end)) { 1898 vm_prot_t old_prot; 1899 1900 vm_map_clip_end(map, current, end, &count); 1901 1902 old_prot = current->protection; 1903 if (set_max) { 1904 current->protection = 1905 (current->max_protection = new_prot) & 1906 old_prot; 1907 } else { 1908 current->protection = new_prot; 1909 } 1910 1911 /* 1912 * Update physical map if necessary. Worry about copy-on-write 1913 * here -- CHECK THIS XXX 1914 */ 1915 1916 if (current->protection != old_prot) { 1917 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 1918 VM_PROT_ALL) 1919 1920 pmap_protect(map->pmap, current->start, 1921 current->end, 1922 current->protection & MASK(current)); 1923 #undef MASK 1924 } 1925 1926 vm_map_simplify_entry(map, current, &count); 1927 1928 current = current->next; 1929 } 1930 1931 vm_map_unlock(map); 1932 vm_map_entry_release(count); 1933 return (KERN_SUCCESS); 1934 } 1935 1936 /* 1937 * This routine traverses a processes map handling the madvise 1938 * system call. Advisories are classified as either those effecting 1939 * the vm_map_entry structure, or those effecting the underlying 1940 * objects. 1941 * 1942 * The <value> argument is used for extended madvise calls. 1943 * 1944 * No requirements. 1945 */ 1946 int 1947 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, 1948 int behav, off_t value) 1949 { 1950 vm_map_entry_t current, entry; 1951 int modify_map = 0; 1952 int error = 0; 1953 int count; 1954 1955 /* 1956 * Some madvise calls directly modify the vm_map_entry, in which case 1957 * we need to use an exclusive lock on the map and we need to perform 1958 * various clipping operations. Otherwise we only need a read-lock 1959 * on the map. 1960 */ 1961 1962 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1963 1964 switch(behav) { 1965 case MADV_NORMAL: 1966 case MADV_SEQUENTIAL: 1967 case MADV_RANDOM: 1968 case MADV_NOSYNC: 1969 case MADV_AUTOSYNC: 1970 case MADV_NOCORE: 1971 case MADV_CORE: 1972 case MADV_SETMAP: 1973 case MADV_INVAL: 1974 modify_map = 1; 1975 vm_map_lock(map); 1976 break; 1977 case MADV_WILLNEED: 1978 case MADV_DONTNEED: 1979 case MADV_FREE: 1980 vm_map_lock_read(map); 1981 break; 1982 default: 1983 vm_map_entry_release(count); 1984 return (EINVAL); 1985 } 1986 1987 /* 1988 * Locate starting entry and clip if necessary. 1989 */ 1990 1991 VM_MAP_RANGE_CHECK(map, start, end); 1992 1993 if (vm_map_lookup_entry(map, start, &entry)) { 1994 if (modify_map) 1995 vm_map_clip_start(map, entry, start, &count); 1996 } else { 1997 entry = entry->next; 1998 } 1999 2000 if (modify_map) { 2001 /* 2002 * madvise behaviors that are implemented in the vm_map_entry. 2003 * 2004 * We clip the vm_map_entry so that behavioral changes are 2005 * limited to the specified address range. 2006 */ 2007 for (current = entry; 2008 (current != &map->header) && (current->start < end); 2009 current = current->next 2010 ) { 2011 if (current->maptype == VM_MAPTYPE_SUBMAP) 2012 continue; 2013 2014 vm_map_clip_end(map, current, end, &count); 2015 2016 switch (behav) { 2017 case MADV_NORMAL: 2018 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 2019 break; 2020 case MADV_SEQUENTIAL: 2021 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 2022 break; 2023 case MADV_RANDOM: 2024 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 2025 break; 2026 case MADV_NOSYNC: 2027 current->eflags |= MAP_ENTRY_NOSYNC; 2028 break; 2029 case MADV_AUTOSYNC: 2030 current->eflags &= ~MAP_ENTRY_NOSYNC; 2031 break; 2032 case MADV_NOCORE: 2033 current->eflags |= MAP_ENTRY_NOCOREDUMP; 2034 break; 2035 case MADV_CORE: 2036 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 2037 break; 2038 case MADV_INVAL: 2039 /* 2040 * Invalidate the related pmap entries, used 2041 * to flush portions of the real kernel's 2042 * pmap when the caller has removed or 2043 * modified existing mappings in a virtual 2044 * page table. 2045 */ 2046 pmap_remove(map->pmap, 2047 current->start, current->end); 2048 break; 2049 case MADV_SETMAP: 2050 /* 2051 * Set the page directory page for a map 2052 * governed by a virtual page table. Mark 2053 * the entry as being governed by a virtual 2054 * page table if it is not. 2055 * 2056 * XXX the page directory page is stored 2057 * in the avail_ssize field if the map_entry. 2058 * 2059 * XXX the map simplification code does not 2060 * compare this field so weird things may 2061 * happen if you do not apply this function 2062 * to the entire mapping governed by the 2063 * virtual page table. 2064 */ 2065 if (current->maptype != VM_MAPTYPE_VPAGETABLE) { 2066 error = EINVAL; 2067 break; 2068 } 2069 current->aux.master_pde = value; 2070 pmap_remove(map->pmap, 2071 current->start, current->end); 2072 break; 2073 default: 2074 error = EINVAL; 2075 break; 2076 } 2077 vm_map_simplify_entry(map, current, &count); 2078 } 2079 vm_map_unlock(map); 2080 } else { 2081 vm_pindex_t pindex; 2082 int count; 2083 2084 /* 2085 * madvise behaviors that are implemented in the underlying 2086 * vm_object. 2087 * 2088 * Since we don't clip the vm_map_entry, we have to clip 2089 * the vm_object pindex and count. 2090 * 2091 * NOTE! We currently do not support these functions on 2092 * virtual page tables. 2093 */ 2094 for (current = entry; 2095 (current != &map->header) && (current->start < end); 2096 current = current->next 2097 ) { 2098 vm_offset_t useStart; 2099 2100 if (current->maptype != VM_MAPTYPE_NORMAL) 2101 continue; 2102 2103 pindex = OFF_TO_IDX(current->offset); 2104 count = atop(current->end - current->start); 2105 useStart = current->start; 2106 2107 if (current->start < start) { 2108 pindex += atop(start - current->start); 2109 count -= atop(start - current->start); 2110 useStart = start; 2111 } 2112 if (current->end > end) 2113 count -= atop(current->end - end); 2114 2115 if (count <= 0) 2116 continue; 2117 2118 vm_object_madvise(current->object.vm_object, 2119 pindex, count, behav); 2120 2121 /* 2122 * Try to populate the page table. Mappings governed 2123 * by virtual page tables cannot be pre-populated 2124 * without a lot of work so don't try. 2125 */ 2126 if (behav == MADV_WILLNEED && 2127 current->maptype != VM_MAPTYPE_VPAGETABLE) { 2128 pmap_object_init_pt( 2129 map->pmap, 2130 useStart, 2131 current->protection, 2132 current->object.vm_object, 2133 pindex, 2134 (count << PAGE_SHIFT), 2135 MAP_PREFAULT_MADVISE 2136 ); 2137 } 2138 } 2139 vm_map_unlock_read(map); 2140 } 2141 vm_map_entry_release(count); 2142 return(error); 2143 } 2144 2145 2146 /* 2147 * Sets the inheritance of the specified address range in the target map. 2148 * Inheritance affects how the map will be shared with child maps at the 2149 * time of vm_map_fork. 2150 */ 2151 int 2152 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 2153 vm_inherit_t new_inheritance) 2154 { 2155 vm_map_entry_t entry; 2156 vm_map_entry_t temp_entry; 2157 int count; 2158 2159 switch (new_inheritance) { 2160 case VM_INHERIT_NONE: 2161 case VM_INHERIT_COPY: 2162 case VM_INHERIT_SHARE: 2163 break; 2164 default: 2165 return (KERN_INVALID_ARGUMENT); 2166 } 2167 2168 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2169 vm_map_lock(map); 2170 2171 VM_MAP_RANGE_CHECK(map, start, end); 2172 2173 if (vm_map_lookup_entry(map, start, &temp_entry)) { 2174 entry = temp_entry; 2175 vm_map_clip_start(map, entry, start, &count); 2176 } else 2177 entry = temp_entry->next; 2178 2179 while ((entry != &map->header) && (entry->start < end)) { 2180 vm_map_clip_end(map, entry, end, &count); 2181 2182 entry->inheritance = new_inheritance; 2183 2184 vm_map_simplify_entry(map, entry, &count); 2185 2186 entry = entry->next; 2187 } 2188 vm_map_unlock(map); 2189 vm_map_entry_release(count); 2190 return (KERN_SUCCESS); 2191 } 2192 2193 /* 2194 * Implement the semantics of mlock 2195 */ 2196 int 2197 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 2198 boolean_t new_pageable) 2199 { 2200 vm_map_entry_t entry; 2201 vm_map_entry_t start_entry; 2202 vm_offset_t end; 2203 int rv = KERN_SUCCESS; 2204 int count; 2205 2206 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2207 vm_map_lock(map); 2208 VM_MAP_RANGE_CHECK(map, start, real_end); 2209 end = real_end; 2210 2211 start_entry = vm_map_clip_range(map, start, end, &count, 2212 MAP_CLIP_NO_HOLES); 2213 if (start_entry == NULL) { 2214 vm_map_unlock(map); 2215 vm_map_entry_release(count); 2216 return (KERN_INVALID_ADDRESS); 2217 } 2218 2219 if (new_pageable == 0) { 2220 entry = start_entry; 2221 while ((entry != &map->header) && (entry->start < end)) { 2222 vm_offset_t save_start; 2223 vm_offset_t save_end; 2224 2225 /* 2226 * Already user wired or hard wired (trivial cases) 2227 */ 2228 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 2229 entry = entry->next; 2230 continue; 2231 } 2232 if (entry->wired_count != 0) { 2233 entry->wired_count++; 2234 entry->eflags |= MAP_ENTRY_USER_WIRED; 2235 entry = entry->next; 2236 continue; 2237 } 2238 2239 /* 2240 * A new wiring requires instantiation of appropriate 2241 * management structures and the faulting in of the 2242 * page. 2243 */ 2244 if (entry->maptype == VM_MAPTYPE_NORMAL || 2245 entry->maptype == VM_MAPTYPE_VPAGETABLE) { 2246 int copyflag = entry->eflags & 2247 MAP_ENTRY_NEEDS_COPY; 2248 if (copyflag && ((entry->protection & 2249 VM_PROT_WRITE) != 0)) { 2250 vm_map_entry_shadow(entry, 0); 2251 } else if (entry->object.vm_object == NULL && 2252 !map->system_map) { 2253 vm_map_entry_allocate_object(entry); 2254 } 2255 } 2256 entry->wired_count++; 2257 entry->eflags |= MAP_ENTRY_USER_WIRED; 2258 2259 /* 2260 * Now fault in the area. Note that vm_fault_wire() 2261 * may release the map lock temporarily, it will be 2262 * relocked on return. The in-transition 2263 * flag protects the entries. 2264 */ 2265 save_start = entry->start; 2266 save_end = entry->end; 2267 rv = vm_fault_wire(map, entry, TRUE, 0); 2268 if (rv) { 2269 CLIP_CHECK_BACK(entry, save_start); 2270 for (;;) { 2271 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 2272 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2273 entry->wired_count = 0; 2274 if (entry->end == save_end) 2275 break; 2276 entry = entry->next; 2277 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2278 } 2279 end = save_start; /* unwire the rest */ 2280 break; 2281 } 2282 /* 2283 * note that even though the entry might have been 2284 * clipped, the USER_WIRED flag we set prevents 2285 * duplication so we do not have to do a 2286 * clip check. 2287 */ 2288 entry = entry->next; 2289 } 2290 2291 /* 2292 * If we failed fall through to the unwiring section to 2293 * unwire what we had wired so far. 'end' has already 2294 * been adjusted. 2295 */ 2296 if (rv) 2297 new_pageable = 1; 2298 2299 /* 2300 * start_entry might have been clipped if we unlocked the 2301 * map and blocked. No matter how clipped it has gotten 2302 * there should be a fragment that is on our start boundary. 2303 */ 2304 CLIP_CHECK_BACK(start_entry, start); 2305 } 2306 2307 /* 2308 * Deal with the unwiring case. 2309 */ 2310 if (new_pageable) { 2311 /* 2312 * This is the unwiring case. We must first ensure that the 2313 * range to be unwired is really wired down. We know there 2314 * are no holes. 2315 */ 2316 entry = start_entry; 2317 while ((entry != &map->header) && (entry->start < end)) { 2318 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 2319 rv = KERN_INVALID_ARGUMENT; 2320 goto done; 2321 } 2322 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry)); 2323 entry = entry->next; 2324 } 2325 2326 /* 2327 * Now decrement the wiring count for each region. If a region 2328 * becomes completely unwired, unwire its physical pages and 2329 * mappings. 2330 */ 2331 /* 2332 * The map entries are processed in a loop, checking to 2333 * make sure the entry is wired and asserting it has a wired 2334 * count. However, another loop was inserted more-or-less in 2335 * the middle of the unwiring path. This loop picks up the 2336 * "entry" loop variable from the first loop without first 2337 * setting it to start_entry. Naturally, the secound loop 2338 * is never entered and the pages backing the entries are 2339 * never unwired. This can lead to a leak of wired pages. 2340 */ 2341 entry = start_entry; 2342 while ((entry != &map->header) && (entry->start < end)) { 2343 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 2344 ("expected USER_WIRED on entry %p", entry)); 2345 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2346 entry->wired_count--; 2347 if (entry->wired_count == 0) 2348 vm_fault_unwire(map, entry); 2349 entry = entry->next; 2350 } 2351 } 2352 done: 2353 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2354 MAP_CLIP_NO_HOLES); 2355 map->timestamp++; 2356 vm_map_unlock(map); 2357 vm_map_entry_release(count); 2358 return (rv); 2359 } 2360 2361 /* 2362 * Sets the pageability of the specified address range in the target map. 2363 * Regions specified as not pageable require locked-down physical 2364 * memory and physical page maps. 2365 * 2366 * The map must not be locked, but a reference must remain to the map 2367 * throughout the call. 2368 * 2369 * This function may be called via the zalloc path and must properly 2370 * reserve map entries for kernel_map. 2371 * 2372 * No requirements. 2373 */ 2374 int 2375 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags) 2376 { 2377 vm_map_entry_t entry; 2378 vm_map_entry_t start_entry; 2379 vm_offset_t end; 2380 int rv = KERN_SUCCESS; 2381 int count; 2382 2383 if (kmflags & KM_KRESERVE) 2384 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 2385 else 2386 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2387 vm_map_lock(map); 2388 VM_MAP_RANGE_CHECK(map, start, real_end); 2389 end = real_end; 2390 2391 start_entry = vm_map_clip_range(map, start, end, &count, 2392 MAP_CLIP_NO_HOLES); 2393 if (start_entry == NULL) { 2394 vm_map_unlock(map); 2395 rv = KERN_INVALID_ADDRESS; 2396 goto failure; 2397 } 2398 if ((kmflags & KM_PAGEABLE) == 0) { 2399 /* 2400 * Wiring. 2401 * 2402 * 1. Holding the write lock, we create any shadow or zero-fill 2403 * objects that need to be created. Then we clip each map 2404 * entry to the region to be wired and increment its wiring 2405 * count. We create objects before clipping the map entries 2406 * to avoid object proliferation. 2407 * 2408 * 2. We downgrade to a read lock, and call vm_fault_wire to 2409 * fault in the pages for any newly wired area (wired_count is 2410 * 1). 2411 * 2412 * Downgrading to a read lock for vm_fault_wire avoids a 2413 * possible deadlock with another process that may have faulted 2414 * on one of the pages to be wired (it would mark the page busy, 2415 * blocking us, then in turn block on the map lock that we 2416 * hold). Because of problems in the recursive lock package, 2417 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 2418 * any actions that require the write lock must be done 2419 * beforehand. Because we keep the read lock on the map, the 2420 * copy-on-write status of the entries we modify here cannot 2421 * change. 2422 */ 2423 entry = start_entry; 2424 while ((entry != &map->header) && (entry->start < end)) { 2425 /* 2426 * Trivial case if the entry is already wired 2427 */ 2428 if (entry->wired_count) { 2429 entry->wired_count++; 2430 entry = entry->next; 2431 continue; 2432 } 2433 2434 /* 2435 * The entry is being newly wired, we have to setup 2436 * appropriate management structures. A shadow 2437 * object is required for a copy-on-write region, 2438 * or a normal object for a zero-fill region. We 2439 * do not have to do this for entries that point to sub 2440 * maps because we won't hold the lock on the sub map. 2441 */ 2442 if (entry->maptype == VM_MAPTYPE_NORMAL || 2443 entry->maptype == VM_MAPTYPE_VPAGETABLE) { 2444 int copyflag = entry->eflags & 2445 MAP_ENTRY_NEEDS_COPY; 2446 if (copyflag && ((entry->protection & 2447 VM_PROT_WRITE) != 0)) { 2448 vm_map_entry_shadow(entry, 0); 2449 } else if (entry->object.vm_object == NULL && 2450 !map->system_map) { 2451 vm_map_entry_allocate_object(entry); 2452 } 2453 } 2454 2455 entry->wired_count++; 2456 entry = entry->next; 2457 } 2458 2459 /* 2460 * Pass 2. 2461 */ 2462 2463 /* 2464 * HACK HACK HACK HACK 2465 * 2466 * vm_fault_wire() temporarily unlocks the map to avoid 2467 * deadlocks. The in-transition flag from vm_map_clip_range 2468 * call should protect us from changes while the map is 2469 * unlocked. T 2470 * 2471 * NOTE: Previously this comment stated that clipping might 2472 * still occur while the entry is unlocked, but from 2473 * what I can tell it actually cannot. 2474 * 2475 * It is unclear whether the CLIP_CHECK_*() calls 2476 * are still needed but we keep them in anyway. 2477 * 2478 * HACK HACK HACK HACK 2479 */ 2480 2481 entry = start_entry; 2482 while (entry != &map->header && entry->start < end) { 2483 /* 2484 * If vm_fault_wire fails for any page we need to undo 2485 * what has been done. We decrement the wiring count 2486 * for those pages which have not yet been wired (now) 2487 * and unwire those that have (later). 2488 */ 2489 vm_offset_t save_start = entry->start; 2490 vm_offset_t save_end = entry->end; 2491 2492 if (entry->wired_count == 1) 2493 rv = vm_fault_wire(map, entry, FALSE, kmflags); 2494 if (rv) { 2495 CLIP_CHECK_BACK(entry, save_start); 2496 for (;;) { 2497 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly")); 2498 entry->wired_count = 0; 2499 if (entry->end == save_end) 2500 break; 2501 entry = entry->next; 2502 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2503 } 2504 end = save_start; 2505 break; 2506 } 2507 CLIP_CHECK_FWD(entry, save_end); 2508 entry = entry->next; 2509 } 2510 2511 /* 2512 * If a failure occured undo everything by falling through 2513 * to the unwiring code. 'end' has already been adjusted 2514 * appropriately. 2515 */ 2516 if (rv) 2517 kmflags |= KM_PAGEABLE; 2518 2519 /* 2520 * start_entry is still IN_TRANSITION but may have been 2521 * clipped since vm_fault_wire() unlocks and relocks the 2522 * map. No matter how clipped it has gotten there should 2523 * be a fragment that is on our start boundary. 2524 */ 2525 CLIP_CHECK_BACK(start_entry, start); 2526 } 2527 2528 if (kmflags & KM_PAGEABLE) { 2529 /* 2530 * This is the unwiring case. We must first ensure that the 2531 * range to be unwired is really wired down. We know there 2532 * are no holes. 2533 */ 2534 entry = start_entry; 2535 while ((entry != &map->header) && (entry->start < end)) { 2536 if (entry->wired_count == 0) { 2537 rv = KERN_INVALID_ARGUMENT; 2538 goto done; 2539 } 2540 entry = entry->next; 2541 } 2542 2543 /* 2544 * Now decrement the wiring count for each region. If a region 2545 * becomes completely unwired, unwire its physical pages and 2546 * mappings. 2547 */ 2548 entry = start_entry; 2549 while ((entry != &map->header) && (entry->start < end)) { 2550 entry->wired_count--; 2551 if (entry->wired_count == 0) 2552 vm_fault_unwire(map, entry); 2553 entry = entry->next; 2554 } 2555 } 2556 done: 2557 vm_map_unclip_range(map, start_entry, start, real_end, 2558 &count, MAP_CLIP_NO_HOLES); 2559 map->timestamp++; 2560 vm_map_unlock(map); 2561 failure: 2562 if (kmflags & KM_KRESERVE) 2563 vm_map_entry_krelease(count); 2564 else 2565 vm_map_entry_release(count); 2566 return (rv); 2567 } 2568 2569 /* 2570 * Mark a newly allocated address range as wired but do not fault in 2571 * the pages. The caller is expected to load the pages into the object. 2572 * 2573 * The map must be locked on entry and will remain locked on return. 2574 * No other requirements. 2575 */ 2576 void 2577 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, 2578 int *countp) 2579 { 2580 vm_map_entry_t scan; 2581 vm_map_entry_t entry; 2582 2583 entry = vm_map_clip_range(map, addr, addr + size, 2584 countp, MAP_CLIP_NO_HOLES); 2585 for (scan = entry; 2586 scan != &map->header && scan->start < addr + size; 2587 scan = scan->next) { 2588 KKASSERT(scan->wired_count == 0); 2589 scan->wired_count = 1; 2590 } 2591 vm_map_unclip_range(map, entry, addr, addr + size, 2592 countp, MAP_CLIP_NO_HOLES); 2593 } 2594 2595 /* 2596 * Push any dirty cached pages in the address range to their pager. 2597 * If syncio is TRUE, dirty pages are written synchronously. 2598 * If invalidate is TRUE, any cached pages are freed as well. 2599 * 2600 * This routine is called by sys_msync() 2601 * 2602 * Returns an error if any part of the specified range is not mapped. 2603 * 2604 * No requirements. 2605 */ 2606 int 2607 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, 2608 boolean_t syncio, boolean_t invalidate) 2609 { 2610 vm_map_entry_t current; 2611 vm_map_entry_t entry; 2612 vm_size_t size; 2613 vm_object_t object; 2614 vm_object_t tobj; 2615 vm_ooffset_t offset; 2616 2617 vm_map_lock_read(map); 2618 VM_MAP_RANGE_CHECK(map, start, end); 2619 if (!vm_map_lookup_entry(map, start, &entry)) { 2620 vm_map_unlock_read(map); 2621 return (KERN_INVALID_ADDRESS); 2622 } 2623 lwkt_gettoken(&map->token); 2624 2625 /* 2626 * Make a first pass to check for holes. 2627 */ 2628 for (current = entry; current->start < end; current = current->next) { 2629 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2630 lwkt_reltoken(&map->token); 2631 vm_map_unlock_read(map); 2632 return (KERN_INVALID_ARGUMENT); 2633 } 2634 if (end > current->end && 2635 (current->next == &map->header || 2636 current->end != current->next->start)) { 2637 lwkt_reltoken(&map->token); 2638 vm_map_unlock_read(map); 2639 return (KERN_INVALID_ADDRESS); 2640 } 2641 } 2642 2643 if (invalidate) 2644 pmap_remove(vm_map_pmap(map), start, end); 2645 2646 /* 2647 * Make a second pass, cleaning/uncaching pages from the indicated 2648 * objects as we go. 2649 */ 2650 for (current = entry; current->start < end; current = current->next) { 2651 offset = current->offset + (start - current->start); 2652 size = (end <= current->end ? end : current->end) - start; 2653 2654 switch(current->maptype) { 2655 case VM_MAPTYPE_SUBMAP: 2656 { 2657 vm_map_t smap; 2658 vm_map_entry_t tentry; 2659 vm_size_t tsize; 2660 2661 smap = current->object.sub_map; 2662 vm_map_lock_read(smap); 2663 vm_map_lookup_entry(smap, offset, &tentry); 2664 tsize = tentry->end - offset; 2665 if (tsize < size) 2666 size = tsize; 2667 object = tentry->object.vm_object; 2668 offset = tentry->offset + (offset - tentry->start); 2669 vm_map_unlock_read(smap); 2670 break; 2671 } 2672 case VM_MAPTYPE_NORMAL: 2673 case VM_MAPTYPE_VPAGETABLE: 2674 object = current->object.vm_object; 2675 break; 2676 default: 2677 object = NULL; 2678 break; 2679 } 2680 2681 if (object) 2682 vm_object_hold(object); 2683 2684 /* 2685 * Note that there is absolutely no sense in writing out 2686 * anonymous objects, so we track down the vnode object 2687 * to write out. 2688 * We invalidate (remove) all pages from the address space 2689 * anyway, for semantic correctness. 2690 * 2691 * note: certain anonymous maps, such as MAP_NOSYNC maps, 2692 * may start out with a NULL object. 2693 */ 2694 while (object && (tobj = object->backing_object) != NULL) { 2695 vm_object_hold(tobj); 2696 if (tobj == object->backing_object) { 2697 vm_object_lock_swap(); 2698 offset += object->backing_object_offset; 2699 vm_object_drop(object); 2700 object = tobj; 2701 if (object->size < OFF_TO_IDX(offset + size)) 2702 size = IDX_TO_OFF(object->size) - 2703 offset; 2704 break; 2705 } 2706 vm_object_drop(tobj); 2707 } 2708 if (object && (object->type == OBJT_VNODE) && 2709 (current->protection & VM_PROT_WRITE) && 2710 (object->flags & OBJ_NOMSYNC) == 0) { 2711 /* 2712 * Flush pages if writing is allowed, invalidate them 2713 * if invalidation requested. Pages undergoing I/O 2714 * will be ignored by vm_object_page_remove(). 2715 * 2716 * We cannot lock the vnode and then wait for paging 2717 * to complete without deadlocking against vm_fault. 2718 * Instead we simply call vm_object_page_remove() and 2719 * allow it to block internally on a page-by-page 2720 * basis when it encounters pages undergoing async 2721 * I/O. 2722 */ 2723 int flags; 2724 2725 /* no chain wait needed for vnode objects */ 2726 vm_object_reference_locked(object); 2727 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY); 2728 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 2729 flags |= invalidate ? OBJPC_INVAL : 0; 2730 2731 /* 2732 * When operating on a virtual page table just 2733 * flush the whole object. XXX we probably ought 2734 * to 2735 */ 2736 switch(current->maptype) { 2737 case VM_MAPTYPE_NORMAL: 2738 vm_object_page_clean(object, 2739 OFF_TO_IDX(offset), 2740 OFF_TO_IDX(offset + size + PAGE_MASK), 2741 flags); 2742 break; 2743 case VM_MAPTYPE_VPAGETABLE: 2744 vm_object_page_clean(object, 0, 0, flags); 2745 break; 2746 } 2747 vn_unlock(((struct vnode *)object->handle)); 2748 vm_object_deallocate_locked(object); 2749 } 2750 if (object && invalidate && 2751 ((object->type == OBJT_VNODE) || 2752 (object->type == OBJT_DEVICE) || 2753 (object->type == OBJT_MGTDEVICE))) { 2754 int clean_only = 2755 ((object->type == OBJT_DEVICE) || 2756 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE; 2757 /* no chain wait needed for vnode/device objects */ 2758 vm_object_reference_locked(object); 2759 switch(current->maptype) { 2760 case VM_MAPTYPE_NORMAL: 2761 vm_object_page_remove(object, 2762 OFF_TO_IDX(offset), 2763 OFF_TO_IDX(offset + size + PAGE_MASK), 2764 clean_only); 2765 break; 2766 case VM_MAPTYPE_VPAGETABLE: 2767 vm_object_page_remove(object, 0, 0, clean_only); 2768 break; 2769 } 2770 vm_object_deallocate_locked(object); 2771 } 2772 start += size; 2773 if (object) 2774 vm_object_drop(object); 2775 } 2776 2777 lwkt_reltoken(&map->token); 2778 vm_map_unlock_read(map); 2779 2780 return (KERN_SUCCESS); 2781 } 2782 2783 /* 2784 * Make the region specified by this entry pageable. 2785 * 2786 * The vm_map must be exclusively locked. 2787 */ 2788 static void 2789 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 2790 { 2791 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2792 entry->wired_count = 0; 2793 vm_fault_unwire(map, entry); 2794 } 2795 2796 /* 2797 * Deallocate the given entry from the target map. 2798 * 2799 * The vm_map must be exclusively locked. 2800 */ 2801 static void 2802 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 2803 { 2804 vm_map_entry_unlink(map, entry); 2805 map->size -= entry->end - entry->start; 2806 2807 switch(entry->maptype) { 2808 case VM_MAPTYPE_NORMAL: 2809 case VM_MAPTYPE_VPAGETABLE: 2810 case VM_MAPTYPE_SUBMAP: 2811 vm_object_deallocate(entry->object.vm_object); 2812 break; 2813 case VM_MAPTYPE_UKSMAP: 2814 /* XXX TODO */ 2815 break; 2816 default: 2817 break; 2818 } 2819 2820 vm_map_entry_dispose(map, entry, countp); 2821 } 2822 2823 /* 2824 * Deallocates the given address range from the target map. 2825 * 2826 * The vm_map must be exclusively locked. 2827 */ 2828 int 2829 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 2830 { 2831 vm_object_t object; 2832 vm_map_entry_t entry; 2833 vm_map_entry_t first_entry; 2834 2835 ASSERT_VM_MAP_LOCKED(map); 2836 lwkt_gettoken(&map->token); 2837 again: 2838 /* 2839 * Find the start of the region, and clip it. Set entry to point 2840 * at the first record containing the requested address or, if no 2841 * such record exists, the next record with a greater address. The 2842 * loop will run from this point until a record beyond the termination 2843 * address is encountered. 2844 * 2845 * map->hint must be adjusted to not point to anything we delete, 2846 * so set it to the entry prior to the one being deleted. 2847 * 2848 * GGG see other GGG comment. 2849 */ 2850 if (vm_map_lookup_entry(map, start, &first_entry)) { 2851 entry = first_entry; 2852 vm_map_clip_start(map, entry, start, countp); 2853 map->hint = entry->prev; /* possible problem XXX */ 2854 } else { 2855 map->hint = first_entry; /* possible problem XXX */ 2856 entry = first_entry->next; 2857 } 2858 2859 /* 2860 * If a hole opens up prior to the current first_free then 2861 * adjust first_free. As with map->hint, map->first_free 2862 * cannot be left set to anything we might delete. 2863 */ 2864 if (entry == &map->header) { 2865 map->first_free = &map->header; 2866 } else if (map->first_free->start >= start) { 2867 map->first_free = entry->prev; 2868 } 2869 2870 /* 2871 * Step through all entries in this region 2872 */ 2873 while ((entry != &map->header) && (entry->start < end)) { 2874 vm_map_entry_t next; 2875 vm_offset_t s, e; 2876 vm_pindex_t offidxstart, offidxend, count; 2877 2878 /* 2879 * If we hit an in-transition entry we have to sleep and 2880 * retry. It's easier (and not really slower) to just retry 2881 * since this case occurs so rarely and the hint is already 2882 * pointing at the right place. We have to reset the 2883 * start offset so as not to accidently delete an entry 2884 * another process just created in vacated space. 2885 */ 2886 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 2887 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2888 start = entry->start; 2889 ++mycpu->gd_cnt.v_intrans_coll; 2890 ++mycpu->gd_cnt.v_intrans_wait; 2891 vm_map_transition_wait(map); 2892 goto again; 2893 } 2894 vm_map_clip_end(map, entry, end, countp); 2895 2896 s = entry->start; 2897 e = entry->end; 2898 next = entry->next; 2899 2900 offidxstart = OFF_TO_IDX(entry->offset); 2901 count = OFF_TO_IDX(e - s); 2902 2903 switch(entry->maptype) { 2904 case VM_MAPTYPE_NORMAL: 2905 case VM_MAPTYPE_VPAGETABLE: 2906 case VM_MAPTYPE_SUBMAP: 2907 object = entry->object.vm_object; 2908 break; 2909 default: 2910 object = NULL; 2911 break; 2912 } 2913 2914 /* 2915 * Unwire before removing addresses from the pmap; otherwise, 2916 * unwiring will put the entries back in the pmap. 2917 */ 2918 if (entry->wired_count != 0) 2919 vm_map_entry_unwire(map, entry); 2920 2921 offidxend = offidxstart + count; 2922 2923 if (object == &kernel_object) { 2924 vm_object_hold(object); 2925 vm_object_page_remove(object, offidxstart, 2926 offidxend, FALSE); 2927 vm_object_drop(object); 2928 } else if (object && object->type != OBJT_DEFAULT && 2929 object->type != OBJT_SWAP) { 2930 /* 2931 * vnode object routines cannot be chain-locked, 2932 * but since we aren't removing pages from the 2933 * object here we can use a shared hold. 2934 */ 2935 vm_object_hold_shared(object); 2936 pmap_remove(map->pmap, s, e); 2937 vm_object_drop(object); 2938 } else if (object) { 2939 vm_object_hold(object); 2940 vm_object_chain_acquire(object, 0); 2941 pmap_remove(map->pmap, s, e); 2942 2943 if (object != NULL && 2944 object->ref_count != 1 && 2945 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == 2946 OBJ_ONEMAPPING && 2947 (object->type == OBJT_DEFAULT || 2948 object->type == OBJT_SWAP)) { 2949 vm_object_collapse(object, NULL); 2950 vm_object_page_remove(object, offidxstart, 2951 offidxend, FALSE); 2952 if (object->type == OBJT_SWAP) { 2953 swap_pager_freespace(object, 2954 offidxstart, 2955 count); 2956 } 2957 if (offidxend >= object->size && 2958 offidxstart < object->size) { 2959 object->size = offidxstart; 2960 } 2961 } 2962 vm_object_chain_release(object); 2963 vm_object_drop(object); 2964 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) { 2965 pmap_remove(map->pmap, s, e); 2966 } 2967 2968 /* 2969 * Delete the entry (which may delete the object) only after 2970 * removing all pmap entries pointing to its pages. 2971 * (Otherwise, its page frames may be reallocated, and any 2972 * modify bits will be set in the wrong object!) 2973 */ 2974 vm_map_entry_delete(map, entry, countp); 2975 entry = next; 2976 } 2977 lwkt_reltoken(&map->token); 2978 return (KERN_SUCCESS); 2979 } 2980 2981 /* 2982 * Remove the given address range from the target map. 2983 * This is the exported form of vm_map_delete. 2984 * 2985 * No requirements. 2986 */ 2987 int 2988 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 2989 { 2990 int result; 2991 int count; 2992 2993 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2994 vm_map_lock(map); 2995 VM_MAP_RANGE_CHECK(map, start, end); 2996 result = vm_map_delete(map, start, end, &count); 2997 vm_map_unlock(map); 2998 vm_map_entry_release(count); 2999 3000 return (result); 3001 } 3002 3003 /* 3004 * Assert that the target map allows the specified privilege on the 3005 * entire address region given. The entire region must be allocated. 3006 * 3007 * The caller must specify whether the vm_map is already locked or not. 3008 */ 3009 boolean_t 3010 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 3011 vm_prot_t protection, boolean_t have_lock) 3012 { 3013 vm_map_entry_t entry; 3014 vm_map_entry_t tmp_entry; 3015 boolean_t result; 3016 3017 if (have_lock == FALSE) 3018 vm_map_lock_read(map); 3019 3020 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 3021 if (have_lock == FALSE) 3022 vm_map_unlock_read(map); 3023 return (FALSE); 3024 } 3025 entry = tmp_entry; 3026 3027 result = TRUE; 3028 while (start < end) { 3029 if (entry == &map->header) { 3030 result = FALSE; 3031 break; 3032 } 3033 /* 3034 * No holes allowed! 3035 */ 3036 3037 if (start < entry->start) { 3038 result = FALSE; 3039 break; 3040 } 3041 /* 3042 * Check protection associated with entry. 3043 */ 3044 3045 if ((entry->protection & protection) != protection) { 3046 result = FALSE; 3047 break; 3048 } 3049 /* go to next entry */ 3050 3051 start = entry->end; 3052 entry = entry->next; 3053 } 3054 if (have_lock == FALSE) 3055 vm_map_unlock_read(map); 3056 return (result); 3057 } 3058 3059 /* 3060 * If appropriate this function shadows the original object with a new object 3061 * and moves the VM pages from the original object to the new object. 3062 * The original object will also be collapsed, if possible. 3063 * 3064 * We can only do this for normal memory objects with a single mapping, and 3065 * it only makes sense to do it if there are 2 or more refs on the original 3066 * object. i.e. typically a memory object that has been extended into 3067 * multiple vm_map_entry's with non-overlapping ranges. 3068 * 3069 * This makes it easier to remove unused pages and keeps object inheritance 3070 * from being a negative impact on memory usage. 3071 * 3072 * On return the (possibly new) entry->object.vm_object will have an 3073 * additional ref on it for the caller to dispose of (usually by cloning 3074 * the vm_map_entry). The additional ref had to be done in this routine 3075 * to avoid racing a collapse. The object's ONEMAPPING flag will also be 3076 * cleared. 3077 * 3078 * The vm_map must be locked and its token held. 3079 */ 3080 static void 3081 vm_map_split(vm_map_entry_t entry) 3082 { 3083 /* OPTIMIZED */ 3084 vm_object_t oobject, nobject, bobject; 3085 vm_offset_t s, e; 3086 vm_page_t m; 3087 vm_pindex_t offidxstart, offidxend, idx; 3088 vm_size_t size; 3089 vm_ooffset_t offset; 3090 int useshadowlist; 3091 3092 /* 3093 * Optimize away object locks for vnode objects. Important exit/exec 3094 * critical path. 3095 * 3096 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag 3097 * anyway. 3098 */ 3099 oobject = entry->object.vm_object; 3100 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) { 3101 vm_object_reference_quick(oobject); 3102 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3103 return; 3104 } 3105 3106 /* 3107 * Setup. Chain lock the original object throughout the entire 3108 * routine to prevent new page faults from occuring. 3109 * 3110 * XXX can madvise WILLNEED interfere with us too? 3111 */ 3112 vm_object_hold(oobject); 3113 vm_object_chain_acquire(oobject, 0); 3114 3115 /* 3116 * Original object cannot be split? Might have also changed state. 3117 */ 3118 if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT && 3119 oobject->type != OBJT_SWAP)) { 3120 vm_object_chain_release(oobject); 3121 vm_object_reference_locked(oobject); 3122 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3123 vm_object_drop(oobject); 3124 return; 3125 } 3126 3127 /* 3128 * Collapse original object with its backing store as an 3129 * optimization to reduce chain lengths when possible. 3130 * 3131 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's 3132 * for oobject, so there's no point collapsing it. 3133 * 3134 * Then re-check whether the object can be split. 3135 */ 3136 vm_object_collapse(oobject, NULL); 3137 3138 if (oobject->ref_count <= 1 || 3139 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) || 3140 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) { 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 * Acquire the chain lock on the backing object. 3150 * 3151 * Give bobject an additional ref count for when it will be shadowed 3152 * by nobject. 3153 */ 3154 useshadowlist = 0; 3155 if ((bobject = oobject->backing_object) != NULL) { 3156 if (bobject->type != OBJT_VNODE) { 3157 useshadowlist = 1; 3158 vm_object_hold(bobject); 3159 vm_object_chain_wait(bobject, 0); 3160 /* ref for shadowing below */ 3161 vm_object_reference_locked(bobject); 3162 vm_object_chain_acquire(bobject, 0); 3163 KKASSERT(bobject->backing_object == bobject); 3164 KKASSERT((bobject->flags & OBJ_DEAD) == 0); 3165 } else { 3166 /* 3167 * vnodes are not placed on the shadow list but 3168 * they still get another ref for the backing_object 3169 * reference. 3170 */ 3171 vm_object_reference_quick(bobject); 3172 } 3173 } 3174 3175 /* 3176 * Calculate the object page range and allocate the new object. 3177 */ 3178 offset = entry->offset; 3179 s = entry->start; 3180 e = entry->end; 3181 3182 offidxstart = OFF_TO_IDX(offset); 3183 offidxend = offidxstart + OFF_TO_IDX(e - s); 3184 size = offidxend - offidxstart; 3185 3186 switch(oobject->type) { 3187 case OBJT_DEFAULT: 3188 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size), 3189 VM_PROT_ALL, 0); 3190 break; 3191 case OBJT_SWAP: 3192 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size), 3193 VM_PROT_ALL, 0); 3194 break; 3195 default: 3196 /* not reached */ 3197 nobject = NULL; 3198 KKASSERT(0); 3199 } 3200 3201 if (nobject == NULL) { 3202 if (bobject) { 3203 if (useshadowlist) { 3204 vm_object_chain_release(bobject); 3205 vm_object_deallocate(bobject); 3206 vm_object_drop(bobject); 3207 } else { 3208 vm_object_deallocate(bobject); 3209 } 3210 } 3211 vm_object_chain_release(oobject); 3212 vm_object_reference_locked(oobject); 3213 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3214 vm_object_drop(oobject); 3215 return; 3216 } 3217 3218 /* 3219 * The new object will replace entry->object.vm_object so it needs 3220 * a second reference (the caller expects an additional ref). 3221 */ 3222 vm_object_hold(nobject); 3223 vm_object_reference_locked(nobject); 3224 vm_object_chain_acquire(nobject, 0); 3225 3226 /* 3227 * nobject shadows bobject (oobject already shadows bobject). 3228 * 3229 * Adding an object to bobject's shadow list requires refing bobject 3230 * which we did above in the useshadowlist case. 3231 */ 3232 if (bobject) { 3233 nobject->backing_object_offset = 3234 oobject->backing_object_offset + IDX_TO_OFF(offidxstart); 3235 nobject->backing_object = bobject; 3236 if (useshadowlist) { 3237 bobject->shadow_count++; 3238 bobject->generation++; 3239 LIST_INSERT_HEAD(&bobject->shadow_head, 3240 nobject, shadow_list); 3241 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/ 3242 vm_object_chain_release(bobject); 3243 vm_object_drop(bobject); 3244 vm_object_set_flag(nobject, OBJ_ONSHADOW); 3245 } 3246 } 3247 3248 /* 3249 * Move the VM pages from oobject to nobject 3250 */ 3251 for (idx = 0; idx < size; idx++) { 3252 vm_page_t m; 3253 3254 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx, 3255 TRUE, "vmpg"); 3256 if (m == NULL) 3257 continue; 3258 3259 /* 3260 * We must wait for pending I/O to complete before we can 3261 * rename the page. 3262 * 3263 * We do not have to VM_PROT_NONE the page as mappings should 3264 * not be changed by this operation. 3265 * 3266 * NOTE: The act of renaming a page updates chaingen for both 3267 * objects. 3268 */ 3269 vm_page_rename(m, nobject, idx); 3270 /* page automatically made dirty by rename and cache handled */ 3271 /* page remains busy */ 3272 } 3273 3274 if (oobject->type == OBJT_SWAP) { 3275 vm_object_pip_add(oobject, 1); 3276 /* 3277 * copy oobject pages into nobject and destroy unneeded 3278 * pages in shadow object. 3279 */ 3280 swap_pager_copy(oobject, nobject, offidxstart, 0); 3281 vm_object_pip_wakeup(oobject); 3282 } 3283 3284 /* 3285 * Wakeup the pages we played with. No spl protection is needed 3286 * for a simple wakeup. 3287 */ 3288 for (idx = 0; idx < size; idx++) { 3289 m = vm_page_lookup(nobject, idx); 3290 if (m) { 3291 KKASSERT(m->flags & PG_BUSY); 3292 vm_page_wakeup(m); 3293 } 3294 } 3295 entry->object.vm_object = nobject; 3296 entry->offset = 0LL; 3297 3298 /* 3299 * Cleanup 3300 * 3301 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the 3302 * related pages were moved and are no longer applicable to the 3303 * original object. 3304 * 3305 * NOTE: Deallocate oobject (due to its entry->object.vm_object being 3306 * replaced by nobject). 3307 */ 3308 vm_object_chain_release(nobject); 3309 vm_object_drop(nobject); 3310 if (bobject && useshadowlist) { 3311 vm_object_chain_release(bobject); 3312 vm_object_drop(bobject); 3313 } 3314 vm_object_chain_release(oobject); 3315 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/ 3316 vm_object_deallocate_locked(oobject); 3317 vm_object_drop(oobject); 3318 } 3319 3320 /* 3321 * Copies the contents of the source entry to the destination 3322 * entry. The entries *must* be aligned properly. 3323 * 3324 * The vm_maps must be exclusively locked. 3325 * The vm_map's token must be held. 3326 * 3327 * Because the maps are locked no faults can be in progress during the 3328 * operation. 3329 */ 3330 static void 3331 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 3332 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 3333 { 3334 vm_object_t src_object; 3335 3336 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP || 3337 dst_entry->maptype == VM_MAPTYPE_UKSMAP) 3338 return; 3339 if (src_entry->maptype == VM_MAPTYPE_SUBMAP || 3340 src_entry->maptype == VM_MAPTYPE_UKSMAP) 3341 return; 3342 3343 if (src_entry->wired_count == 0) { 3344 /* 3345 * If the source entry is marked needs_copy, it is already 3346 * write-protected. 3347 */ 3348 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 3349 pmap_protect(src_map->pmap, 3350 src_entry->start, 3351 src_entry->end, 3352 src_entry->protection & ~VM_PROT_WRITE); 3353 } 3354 3355 /* 3356 * Make a copy of the object. 3357 * 3358 * The object must be locked prior to checking the object type 3359 * and for the call to vm_object_collapse() and vm_map_split(). 3360 * We cannot use *_hold() here because the split code will 3361 * probably try to destroy the object. The lock is a pool 3362 * token and doesn't care. 3363 * 3364 * We must bump src_map->timestamp when setting 3365 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault 3366 * to retry, otherwise the concurrent fault might improperly 3367 * install a RW pte when its supposed to be a RO(COW) pte. 3368 * This race can occur because a vnode-backed fault may have 3369 * to temporarily release the map lock. 3370 */ 3371 if (src_entry->object.vm_object != NULL) { 3372 vm_map_split(src_entry); 3373 src_object = src_entry->object.vm_object; 3374 dst_entry->object.vm_object = src_object; 3375 src_entry->eflags |= (MAP_ENTRY_COW | 3376 MAP_ENTRY_NEEDS_COPY); 3377 dst_entry->eflags |= (MAP_ENTRY_COW | 3378 MAP_ENTRY_NEEDS_COPY); 3379 dst_entry->offset = src_entry->offset; 3380 ++src_map->timestamp; 3381 } else { 3382 dst_entry->object.vm_object = NULL; 3383 dst_entry->offset = 0; 3384 } 3385 3386 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, 3387 dst_entry->end - dst_entry->start, src_entry->start); 3388 } else { 3389 /* 3390 * Of course, wired down pages can't be set copy-on-write. 3391 * Cause wired pages to be copied into the new map by 3392 * simulating faults (the new pages are pageable) 3393 */ 3394 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 3395 } 3396 } 3397 3398 /* 3399 * vmspace_fork: 3400 * Create a new process vmspace structure and vm_map 3401 * based on those of an existing process. The new map 3402 * is based on the old map, according to the inheritance 3403 * values on the regions in that map. 3404 * 3405 * The source map must not be locked. 3406 * No requirements. 3407 */ 3408 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map, 3409 vm_map_entry_t old_entry, int *countp); 3410 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map, 3411 vm_map_entry_t old_entry, int *countp); 3412 3413 struct vmspace * 3414 vmspace_fork(struct vmspace *vm1) 3415 { 3416 struct vmspace *vm2; 3417 vm_map_t old_map = &vm1->vm_map; 3418 vm_map_t new_map; 3419 vm_map_entry_t old_entry; 3420 int count; 3421 3422 lwkt_gettoken(&vm1->vm_map.token); 3423 vm_map_lock(old_map); 3424 3425 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset); 3426 lwkt_gettoken(&vm2->vm_map.token); 3427 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 3428 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 3429 new_map = &vm2->vm_map; /* XXX */ 3430 new_map->timestamp = 1; 3431 3432 vm_map_lock(new_map); 3433 3434 count = 0; 3435 old_entry = old_map->header.next; 3436 while (old_entry != &old_map->header) { 3437 ++count; 3438 old_entry = old_entry->next; 3439 } 3440 3441 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 3442 3443 old_entry = old_map->header.next; 3444 while (old_entry != &old_map->header) { 3445 switch(old_entry->maptype) { 3446 case VM_MAPTYPE_SUBMAP: 3447 panic("vm_map_fork: encountered a submap"); 3448 break; 3449 case VM_MAPTYPE_UKSMAP: 3450 vmspace_fork_uksmap_entry(old_map, new_map, 3451 old_entry, &count); 3452 break; 3453 case VM_MAPTYPE_NORMAL: 3454 case VM_MAPTYPE_VPAGETABLE: 3455 vmspace_fork_normal_entry(old_map, new_map, 3456 old_entry, &count); 3457 break; 3458 } 3459 old_entry = old_entry->next; 3460 } 3461 3462 new_map->size = old_map->size; 3463 vm_map_unlock(old_map); 3464 vm_map_unlock(new_map); 3465 vm_map_entry_release(count); 3466 3467 lwkt_reltoken(&vm2->vm_map.token); 3468 lwkt_reltoken(&vm1->vm_map.token); 3469 3470 return (vm2); 3471 } 3472 3473 static 3474 void 3475 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map, 3476 vm_map_entry_t old_entry, int *countp) 3477 { 3478 vm_map_entry_t new_entry; 3479 vm_object_t object; 3480 3481 switch (old_entry->inheritance) { 3482 case VM_INHERIT_NONE: 3483 break; 3484 case VM_INHERIT_SHARE: 3485 /* 3486 * Clone the entry, creating the shared object if 3487 * necessary. 3488 */ 3489 if (old_entry->object.vm_object == NULL) 3490 vm_map_entry_allocate_object(old_entry); 3491 3492 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3493 /* 3494 * Shadow a map_entry which needs a copy, 3495 * replacing its object with a new object 3496 * that points to the old one. Ask the 3497 * shadow code to automatically add an 3498 * additional ref. We can't do it afterwords 3499 * because we might race a collapse. The call 3500 * to vm_map_entry_shadow() will also clear 3501 * OBJ_ONEMAPPING. 3502 */ 3503 vm_map_entry_shadow(old_entry, 1); 3504 } else if (old_entry->object.vm_object) { 3505 /* 3506 * We will make a shared copy of the object, 3507 * and must clear OBJ_ONEMAPPING. 3508 * 3509 * Optimize vnode objects. OBJ_ONEMAPPING 3510 * is non-applicable but clear it anyway, 3511 * and its terminal so we don'th ave to deal 3512 * with chains. Reduces SMP conflicts. 3513 * 3514 * XXX assert that object.vm_object != NULL 3515 * since we allocate it above. 3516 */ 3517 object = old_entry->object.vm_object; 3518 if (object->type == OBJT_VNODE) { 3519 vm_object_reference_quick(object); 3520 vm_object_clear_flag(object, 3521 OBJ_ONEMAPPING); 3522 } else { 3523 vm_object_hold(object); 3524 vm_object_chain_wait(object, 0); 3525 vm_object_reference_locked(object); 3526 vm_object_clear_flag(object, 3527 OBJ_ONEMAPPING); 3528 vm_object_drop(object); 3529 } 3530 } 3531 3532 /* 3533 * Clone the entry. We've already bumped the ref on 3534 * any vm_object. 3535 */ 3536 new_entry = vm_map_entry_create(new_map, countp); 3537 *new_entry = *old_entry; 3538 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3539 new_entry->wired_count = 0; 3540 3541 /* 3542 * Insert the entry into the new map -- we know we're 3543 * inserting at the end of the new map. 3544 */ 3545 3546 vm_map_entry_link(new_map, new_map->header.prev, 3547 new_entry); 3548 3549 /* 3550 * Update the physical map 3551 */ 3552 pmap_copy(new_map->pmap, old_map->pmap, 3553 new_entry->start, 3554 (old_entry->end - old_entry->start), 3555 old_entry->start); 3556 break; 3557 case VM_INHERIT_COPY: 3558 /* 3559 * Clone the entry and link into the map. 3560 */ 3561 new_entry = vm_map_entry_create(new_map, countp); 3562 *new_entry = *old_entry; 3563 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3564 new_entry->wired_count = 0; 3565 new_entry->object.vm_object = NULL; 3566 vm_map_entry_link(new_map, new_map->header.prev, 3567 new_entry); 3568 vm_map_copy_entry(old_map, new_map, old_entry, 3569 new_entry); 3570 break; 3571 } 3572 } 3573 3574 /* 3575 * When forking user-kernel shared maps, the map might change in the 3576 * child so do not try to copy the underlying pmap entries. 3577 */ 3578 static 3579 void 3580 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map, 3581 vm_map_entry_t old_entry, int *countp) 3582 { 3583 vm_map_entry_t new_entry; 3584 3585 new_entry = vm_map_entry_create(new_map, countp); 3586 *new_entry = *old_entry; 3587 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3588 new_entry->wired_count = 0; 3589 vm_map_entry_link(new_map, new_map->header.prev, 3590 new_entry); 3591 } 3592 3593 /* 3594 * Create an auto-grow stack entry 3595 * 3596 * No requirements. 3597 */ 3598 int 3599 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 3600 int flags, vm_prot_t prot, vm_prot_t max, int cow) 3601 { 3602 vm_map_entry_t prev_entry; 3603 vm_map_entry_t new_stack_entry; 3604 vm_size_t init_ssize; 3605 int rv; 3606 int count; 3607 vm_offset_t tmpaddr; 3608 3609 cow |= MAP_IS_STACK; 3610 3611 if (max_ssize < sgrowsiz) 3612 init_ssize = max_ssize; 3613 else 3614 init_ssize = sgrowsiz; 3615 3616 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3617 vm_map_lock(map); 3618 3619 /* 3620 * Find space for the mapping 3621 */ 3622 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) { 3623 if (vm_map_findspace(map, addrbos, max_ssize, 1, 3624 flags, &tmpaddr)) { 3625 vm_map_unlock(map); 3626 vm_map_entry_release(count); 3627 return (KERN_NO_SPACE); 3628 } 3629 addrbos = tmpaddr; 3630 } 3631 3632 /* If addr is already mapped, no go */ 3633 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) { 3634 vm_map_unlock(map); 3635 vm_map_entry_release(count); 3636 return (KERN_NO_SPACE); 3637 } 3638 3639 #if 0 3640 /* XXX already handled by kern_mmap() */ 3641 /* If we would blow our VMEM resource limit, no go */ 3642 if (map->size + init_ssize > 3643 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3644 vm_map_unlock(map); 3645 vm_map_entry_release(count); 3646 return (KERN_NO_SPACE); 3647 } 3648 #endif 3649 3650 /* 3651 * If we can't accomodate max_ssize in the current mapping, 3652 * no go. However, we need to be aware that subsequent user 3653 * mappings might map into the space we have reserved for 3654 * stack, and currently this space is not protected. 3655 * 3656 * Hopefully we will at least detect this condition 3657 * when we try to grow the stack. 3658 */ 3659 if ((prev_entry->next != &map->header) && 3660 (prev_entry->next->start < addrbos + max_ssize)) { 3661 vm_map_unlock(map); 3662 vm_map_entry_release(count); 3663 return (KERN_NO_SPACE); 3664 } 3665 3666 /* 3667 * We initially map a stack of only init_ssize. We will 3668 * grow as needed later. Since this is to be a grow 3669 * down stack, we map at the top of the range. 3670 * 3671 * Note: we would normally expect prot and max to be 3672 * VM_PROT_ALL, and cow to be 0. Possibly we should 3673 * eliminate these as input parameters, and just 3674 * pass these values here in the insert call. 3675 */ 3676 rv = vm_map_insert(map, &count, NULL, NULL, 3677 0, addrbos + max_ssize - init_ssize, 3678 addrbos + max_ssize, 3679 VM_MAPTYPE_NORMAL, 3680 prot, max, cow); 3681 3682 /* Now set the avail_ssize amount */ 3683 if (rv == KERN_SUCCESS) { 3684 if (prev_entry != &map->header) 3685 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count); 3686 new_stack_entry = prev_entry->next; 3687 if (new_stack_entry->end != addrbos + max_ssize || 3688 new_stack_entry->start != addrbos + max_ssize - init_ssize) 3689 panic ("Bad entry start/end for new stack entry"); 3690 else 3691 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize; 3692 } 3693 3694 vm_map_unlock(map); 3695 vm_map_entry_release(count); 3696 return (rv); 3697 } 3698 3699 /* 3700 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 3701 * desired address is already mapped, or if we successfully grow 3702 * the stack. Also returns KERN_SUCCESS if addr is outside the 3703 * stack range (this is strange, but preserves compatibility with 3704 * the grow function in vm_machdep.c). 3705 * 3706 * No requirements. 3707 */ 3708 int 3709 vm_map_growstack (struct proc *p, vm_offset_t addr) 3710 { 3711 vm_map_entry_t prev_entry; 3712 vm_map_entry_t stack_entry; 3713 vm_map_entry_t new_stack_entry; 3714 struct vmspace *vm = p->p_vmspace; 3715 vm_map_t map = &vm->vm_map; 3716 vm_offset_t end; 3717 int grow_amount; 3718 int rv = KERN_SUCCESS; 3719 int is_procstack; 3720 int use_read_lock = 1; 3721 int count; 3722 3723 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3724 Retry: 3725 if (use_read_lock) 3726 vm_map_lock_read(map); 3727 else 3728 vm_map_lock(map); 3729 3730 /* If addr is already in the entry range, no need to grow.*/ 3731 if (vm_map_lookup_entry(map, addr, &prev_entry)) 3732 goto done; 3733 3734 if ((stack_entry = prev_entry->next) == &map->header) 3735 goto done; 3736 if (prev_entry == &map->header) 3737 end = stack_entry->start - stack_entry->aux.avail_ssize; 3738 else 3739 end = prev_entry->end; 3740 3741 /* 3742 * This next test mimics the old grow function in vm_machdep.c. 3743 * It really doesn't quite make sense, but we do it anyway 3744 * for compatibility. 3745 * 3746 * If not growable stack, return success. This signals the 3747 * caller to proceed as he would normally with normal vm. 3748 */ 3749 if (stack_entry->aux.avail_ssize < 1 || 3750 addr >= stack_entry->start || 3751 addr < stack_entry->start - stack_entry->aux.avail_ssize) { 3752 goto done; 3753 } 3754 3755 /* Find the minimum grow amount */ 3756 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE); 3757 if (grow_amount > stack_entry->aux.avail_ssize) { 3758 rv = KERN_NO_SPACE; 3759 goto done; 3760 } 3761 3762 /* 3763 * If there is no longer enough space between the entries 3764 * nogo, and adjust the available space. Note: this 3765 * should only happen if the user has mapped into the 3766 * stack area after the stack was created, and is 3767 * probably an error. 3768 * 3769 * This also effectively destroys any guard page the user 3770 * might have intended by limiting the stack size. 3771 */ 3772 if (grow_amount > stack_entry->start - end) { 3773 if (use_read_lock && vm_map_lock_upgrade(map)) { 3774 /* lost lock */ 3775 use_read_lock = 0; 3776 goto Retry; 3777 } 3778 use_read_lock = 0; 3779 stack_entry->aux.avail_ssize = stack_entry->start - end; 3780 rv = KERN_NO_SPACE; 3781 goto done; 3782 } 3783 3784 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 3785 3786 /* If this is the main process stack, see if we're over the 3787 * stack limit. 3788 */ 3789 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3790 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3791 rv = KERN_NO_SPACE; 3792 goto done; 3793 } 3794 3795 /* Round up the grow amount modulo SGROWSIZ */ 3796 grow_amount = roundup (grow_amount, sgrowsiz); 3797 if (grow_amount > stack_entry->aux.avail_ssize) { 3798 grow_amount = stack_entry->aux.avail_ssize; 3799 } 3800 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3801 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3802 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - 3803 ctob(vm->vm_ssize); 3804 } 3805 3806 /* If we would blow our VMEM resource limit, no go */ 3807 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3808 rv = KERN_NO_SPACE; 3809 goto done; 3810 } 3811 3812 if (use_read_lock && vm_map_lock_upgrade(map)) { 3813 /* lost lock */ 3814 use_read_lock = 0; 3815 goto Retry; 3816 } 3817 use_read_lock = 0; 3818 3819 /* Get the preliminary new entry start value */ 3820 addr = stack_entry->start - grow_amount; 3821 3822 /* If this puts us into the previous entry, cut back our growth 3823 * to the available space. Also, see the note above. 3824 */ 3825 if (addr < end) { 3826 stack_entry->aux.avail_ssize = stack_entry->start - end; 3827 addr = end; 3828 } 3829 3830 rv = vm_map_insert(map, &count, NULL, NULL, 3831 0, addr, stack_entry->start, 3832 VM_MAPTYPE_NORMAL, 3833 VM_PROT_ALL, VM_PROT_ALL, 0); 3834 3835 /* Adjust the available stack space by the amount we grew. */ 3836 if (rv == KERN_SUCCESS) { 3837 if (prev_entry != &map->header) 3838 vm_map_clip_end(map, prev_entry, addr, &count); 3839 new_stack_entry = prev_entry->next; 3840 if (new_stack_entry->end != stack_entry->start || 3841 new_stack_entry->start != addr) 3842 panic ("Bad stack grow start/end in new stack entry"); 3843 else { 3844 new_stack_entry->aux.avail_ssize = 3845 stack_entry->aux.avail_ssize - 3846 (new_stack_entry->end - new_stack_entry->start); 3847 if (is_procstack) 3848 vm->vm_ssize += btoc(new_stack_entry->end - 3849 new_stack_entry->start); 3850 } 3851 3852 if (map->flags & MAP_WIREFUTURE) 3853 vm_map_unwire(map, new_stack_entry->start, 3854 new_stack_entry->end, FALSE); 3855 } 3856 3857 done: 3858 if (use_read_lock) 3859 vm_map_unlock_read(map); 3860 else 3861 vm_map_unlock(map); 3862 vm_map_entry_release(count); 3863 return (rv); 3864 } 3865 3866 /* 3867 * Unshare the specified VM space for exec. If other processes are 3868 * mapped to it, then create a new one. The new vmspace is null. 3869 * 3870 * No requirements. 3871 */ 3872 void 3873 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 3874 { 3875 struct vmspace *oldvmspace = p->p_vmspace; 3876 struct vmspace *newvmspace; 3877 vm_map_t map = &p->p_vmspace->vm_map; 3878 3879 /* 3880 * If we are execing a resident vmspace we fork it, otherwise 3881 * we create a new vmspace. Note that exitingcnt is not 3882 * copied to the new vmspace. 3883 */ 3884 lwkt_gettoken(&oldvmspace->vm_map.token); 3885 if (vmcopy) { 3886 newvmspace = vmspace_fork(vmcopy); 3887 lwkt_gettoken(&newvmspace->vm_map.token); 3888 } else { 3889 newvmspace = vmspace_alloc(map->min_offset, map->max_offset); 3890 lwkt_gettoken(&newvmspace->vm_map.token); 3891 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 3892 (caddr_t)&oldvmspace->vm_endcopy - 3893 (caddr_t)&oldvmspace->vm_startcopy); 3894 } 3895 3896 /* 3897 * Finish initializing the vmspace before assigning it 3898 * to the process. The vmspace will become the current vmspace 3899 * if p == curproc. 3900 */ 3901 pmap_pinit2(vmspace_pmap(newvmspace)); 3902 pmap_replacevm(p, newvmspace, 0); 3903 lwkt_reltoken(&newvmspace->vm_map.token); 3904 lwkt_reltoken(&oldvmspace->vm_map.token); 3905 vmspace_rel(oldvmspace); 3906 } 3907 3908 /* 3909 * Unshare the specified VM space for forcing COW. This 3910 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 3911 */ 3912 void 3913 vmspace_unshare(struct proc *p) 3914 { 3915 struct vmspace *oldvmspace = p->p_vmspace; 3916 struct vmspace *newvmspace; 3917 3918 lwkt_gettoken(&oldvmspace->vm_map.token); 3919 if (vmspace_getrefs(oldvmspace) == 1) { 3920 lwkt_reltoken(&oldvmspace->vm_map.token); 3921 return; 3922 } 3923 newvmspace = vmspace_fork(oldvmspace); 3924 lwkt_gettoken(&newvmspace->vm_map.token); 3925 pmap_pinit2(vmspace_pmap(newvmspace)); 3926 pmap_replacevm(p, newvmspace, 0); 3927 lwkt_reltoken(&newvmspace->vm_map.token); 3928 lwkt_reltoken(&oldvmspace->vm_map.token); 3929 vmspace_rel(oldvmspace); 3930 } 3931 3932 /* 3933 * vm_map_hint: return the beginning of the best area suitable for 3934 * creating a new mapping with "prot" protection. 3935 * 3936 * No requirements. 3937 */ 3938 vm_offset_t 3939 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot) 3940 { 3941 struct vmspace *vms = p->p_vmspace; 3942 3943 if (!randomize_mmap || addr != 0) { 3944 /* 3945 * Set a reasonable start point for the hint if it was 3946 * not specified or if it falls within the heap space. 3947 * Hinted mmap()s do not allocate out of the heap space. 3948 */ 3949 if (addr == 0 || 3950 (addr >= round_page((vm_offset_t)vms->vm_taddr) && 3951 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) { 3952 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz); 3953 } 3954 3955 return addr; 3956 } 3957 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ; 3958 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1); 3959 3960 return (round_page(addr)); 3961 } 3962 3963 /* 3964 * Finds the VM object, offset, and protection for a given virtual address 3965 * in the specified map, assuming a page fault of the type specified. 3966 * 3967 * Leaves the map in question locked for read; return values are guaranteed 3968 * until a vm_map_lookup_done call is performed. Note that the map argument 3969 * is in/out; the returned map must be used in the call to vm_map_lookup_done. 3970 * 3971 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make 3972 * that fast. 3973 * 3974 * If a lookup is requested with "write protection" specified, the map may 3975 * be changed to perform virtual copying operations, although the data 3976 * referenced will remain the same. 3977 * 3978 * No requirements. 3979 */ 3980 int 3981 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 3982 vm_offset_t vaddr, 3983 vm_prot_t fault_typea, 3984 vm_map_entry_t *out_entry, /* OUT */ 3985 vm_object_t *object, /* OUT */ 3986 vm_pindex_t *pindex, /* OUT */ 3987 vm_prot_t *out_prot, /* OUT */ 3988 boolean_t *wired) /* OUT */ 3989 { 3990 vm_map_entry_t entry; 3991 vm_map_t map = *var_map; 3992 vm_prot_t prot; 3993 vm_prot_t fault_type = fault_typea; 3994 int use_read_lock = 1; 3995 int rv = KERN_SUCCESS; 3996 3997 RetryLookup: 3998 if (use_read_lock) 3999 vm_map_lock_read(map); 4000 else 4001 vm_map_lock(map); 4002 4003 /* 4004 * If the map has an interesting hint, try it before calling full 4005 * blown lookup routine. 4006 */ 4007 entry = map->hint; 4008 cpu_ccfence(); 4009 *out_entry = entry; 4010 *object = NULL; 4011 4012 if ((entry == &map->header) || 4013 (vaddr < entry->start) || (vaddr >= entry->end)) { 4014 vm_map_entry_t tmp_entry; 4015 4016 /* 4017 * Entry was either not a valid hint, or the vaddr was not 4018 * contained in the entry, so do a full lookup. 4019 */ 4020 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 4021 rv = KERN_INVALID_ADDRESS; 4022 goto done; 4023 } 4024 4025 entry = tmp_entry; 4026 *out_entry = entry; 4027 } 4028 4029 /* 4030 * Handle submaps. 4031 */ 4032 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 4033 vm_map_t old_map = map; 4034 4035 *var_map = map = entry->object.sub_map; 4036 if (use_read_lock) 4037 vm_map_unlock_read(old_map); 4038 else 4039 vm_map_unlock(old_map); 4040 use_read_lock = 1; 4041 goto RetryLookup; 4042 } 4043 4044 /* 4045 * Check whether this task is allowed to have this page. 4046 * Note the special case for MAP_ENTRY_COW 4047 * pages with an override. This is to implement a forced 4048 * COW for debuggers. 4049 */ 4050 4051 if (fault_type & VM_PROT_OVERRIDE_WRITE) 4052 prot = entry->max_protection; 4053 else 4054 prot = entry->protection; 4055 4056 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 4057 if ((fault_type & prot) != fault_type) { 4058 rv = KERN_PROTECTION_FAILURE; 4059 goto done; 4060 } 4061 4062 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 4063 (entry->eflags & MAP_ENTRY_COW) && 4064 (fault_type & VM_PROT_WRITE) && 4065 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 4066 rv = KERN_PROTECTION_FAILURE; 4067 goto done; 4068 } 4069 4070 /* 4071 * If this page is not pageable, we have to get it for all possible 4072 * accesses. 4073 */ 4074 *wired = (entry->wired_count != 0); 4075 if (*wired) 4076 prot = fault_type = entry->protection; 4077 4078 /* 4079 * Virtual page tables may need to update the accessed (A) bit 4080 * in a page table entry. Upgrade the fault to a write fault for 4081 * that case if the map will support it. If the map does not support 4082 * it the page table entry simply will not be updated. 4083 */ 4084 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 4085 if (prot & VM_PROT_WRITE) 4086 fault_type |= VM_PROT_WRITE; 4087 } 4088 4089 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 4090 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 4091 if ((prot & VM_PROT_WRITE) == 0) 4092 fault_type |= VM_PROT_WRITE; 4093 } 4094 4095 /* 4096 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not. 4097 */ 4098 if (entry->maptype != VM_MAPTYPE_NORMAL && 4099 entry->maptype != VM_MAPTYPE_VPAGETABLE) { 4100 *object = NULL; 4101 goto skip; 4102 } 4103 4104 /* 4105 * If the entry was copy-on-write, we either ... 4106 */ 4107 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4108 /* 4109 * If we want to write the page, we may as well handle that 4110 * now since we've got the map locked. 4111 * 4112 * If we don't need to write the page, we just demote the 4113 * permissions allowed. 4114 */ 4115 4116 if (fault_type & VM_PROT_WRITE) { 4117 /* 4118 * Not allowed if TDF_NOFAULT is set as the shadowing 4119 * operation can deadlock against the faulting 4120 * function due to the copy-on-write. 4121 */ 4122 if (curthread->td_flags & TDF_NOFAULT) { 4123 rv = KERN_FAILURE_NOFAULT; 4124 goto done; 4125 } 4126 4127 /* 4128 * Make a new object, and place it in the object 4129 * chain. Note that no new references have appeared 4130 * -- one just moved from the map to the new 4131 * object. 4132 */ 4133 4134 if (use_read_lock && vm_map_lock_upgrade(map)) { 4135 /* lost lock */ 4136 use_read_lock = 0; 4137 goto RetryLookup; 4138 } 4139 use_read_lock = 0; 4140 4141 vm_map_entry_shadow(entry, 0); 4142 } else { 4143 /* 4144 * We're attempting to read a copy-on-write page -- 4145 * don't allow writes. 4146 */ 4147 4148 prot &= ~VM_PROT_WRITE; 4149 } 4150 } 4151 4152 /* 4153 * Create an object if necessary. 4154 */ 4155 if (entry->object.vm_object == NULL && !map->system_map) { 4156 if (use_read_lock && vm_map_lock_upgrade(map)) { 4157 /* lost lock */ 4158 use_read_lock = 0; 4159 goto RetryLookup; 4160 } 4161 use_read_lock = 0; 4162 vm_map_entry_allocate_object(entry); 4163 } 4164 4165 /* 4166 * Return the object/offset from this entry. If the entry was 4167 * copy-on-write or empty, it has been fixed up. 4168 */ 4169 *object = entry->object.vm_object; 4170 4171 skip: 4172 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 4173 4174 /* 4175 * Return whether this is the only map sharing this data. On 4176 * success we return with a read lock held on the map. On failure 4177 * we return with the map unlocked. 4178 */ 4179 *out_prot = prot; 4180 done: 4181 if (rv == KERN_SUCCESS) { 4182 if (use_read_lock == 0) 4183 vm_map_lock_downgrade(map); 4184 } else if (use_read_lock) { 4185 vm_map_unlock_read(map); 4186 } else { 4187 vm_map_unlock(map); 4188 } 4189 return (rv); 4190 } 4191 4192 /* 4193 * Releases locks acquired by a vm_map_lookup() 4194 * (according to the handle returned by that lookup). 4195 * 4196 * No other requirements. 4197 */ 4198 void 4199 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 4200 { 4201 /* 4202 * Unlock the main-level map 4203 */ 4204 vm_map_unlock_read(map); 4205 if (count) 4206 vm_map_entry_release(count); 4207 } 4208 4209 #include "opt_ddb.h" 4210 #ifdef DDB 4211 #include <sys/kernel.h> 4212 4213 #include <ddb/ddb.h> 4214 4215 /* 4216 * Debugging only 4217 */ 4218 DB_SHOW_COMMAND(map, vm_map_print) 4219 { 4220 static int nlines; 4221 /* XXX convert args. */ 4222 vm_map_t map = (vm_map_t)addr; 4223 boolean_t full = have_addr; 4224 4225 vm_map_entry_t entry; 4226 4227 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 4228 (void *)map, 4229 (void *)map->pmap, map->nentries, map->timestamp); 4230 nlines++; 4231 4232 if (!full && db_indent) 4233 return; 4234 4235 db_indent += 2; 4236 for (entry = map->header.next; entry != &map->header; 4237 entry = entry->next) { 4238 db_iprintf("map entry %p: start=%p, end=%p\n", 4239 (void *)entry, (void *)entry->start, (void *)entry->end); 4240 nlines++; 4241 { 4242 static char *inheritance_name[4] = 4243 {"share", "copy", "none", "donate_copy"}; 4244 4245 db_iprintf(" prot=%x/%x/%s", 4246 entry->protection, 4247 entry->max_protection, 4248 inheritance_name[(int)(unsigned char)entry->inheritance]); 4249 if (entry->wired_count != 0) 4250 db_printf(", wired"); 4251 } 4252 switch(entry->maptype) { 4253 case VM_MAPTYPE_SUBMAP: 4254 /* XXX no %qd in kernel. Truncate entry->offset. */ 4255 db_printf(", share=%p, offset=0x%lx\n", 4256 (void *)entry->object.sub_map, 4257 (long)entry->offset); 4258 nlines++; 4259 if ((entry->prev == &map->header) || 4260 (entry->prev->object.sub_map != 4261 entry->object.sub_map)) { 4262 db_indent += 2; 4263 vm_map_print((db_expr_t)(intptr_t) 4264 entry->object.sub_map, 4265 full, 0, NULL); 4266 db_indent -= 2; 4267 } 4268 break; 4269 case VM_MAPTYPE_NORMAL: 4270 case VM_MAPTYPE_VPAGETABLE: 4271 /* XXX no %qd in kernel. Truncate entry->offset. */ 4272 db_printf(", object=%p, offset=0x%lx", 4273 (void *)entry->object.vm_object, 4274 (long)entry->offset); 4275 if (entry->eflags & MAP_ENTRY_COW) 4276 db_printf(", copy (%s)", 4277 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4278 db_printf("\n"); 4279 nlines++; 4280 4281 if ((entry->prev == &map->header) || 4282 (entry->prev->object.vm_object != 4283 entry->object.vm_object)) { 4284 db_indent += 2; 4285 vm_object_print((db_expr_t)(intptr_t) 4286 entry->object.vm_object, 4287 full, 0, NULL); 4288 nlines += 4; 4289 db_indent -= 2; 4290 } 4291 break; 4292 case VM_MAPTYPE_UKSMAP: 4293 db_printf(", uksmap=%p, offset=0x%lx", 4294 (void *)entry->object.uksmap, 4295 (long)entry->offset); 4296 if (entry->eflags & MAP_ENTRY_COW) 4297 db_printf(", copy (%s)", 4298 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4299 db_printf("\n"); 4300 nlines++; 4301 break; 4302 default: 4303 break; 4304 } 4305 } 4306 db_indent -= 2; 4307 if (db_indent == 0) 4308 nlines = 0; 4309 } 4310 4311 /* 4312 * Debugging only 4313 */ 4314 DB_SHOW_COMMAND(procvm, procvm) 4315 { 4316 struct proc *p; 4317 4318 if (have_addr) { 4319 p = (struct proc *) addr; 4320 } else { 4321 p = curproc; 4322 } 4323 4324 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 4325 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 4326 (void *)vmspace_pmap(p->p_vmspace)); 4327 4328 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 4329 } 4330 4331 #endif /* DDB */ 4332