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/random.h> 97 #include <sys/sysctl.h> 98 #include <sys/spinlock.h> 99 100 #include <sys/thread2.h> 101 #include <sys/spinlock2.h> 102 103 /* 104 * Virtual memory maps provide for the mapping, protection, and sharing 105 * of virtual memory objects. In addition, this module provides for an 106 * efficient virtual copy of memory from one map to another. 107 * 108 * Synchronization is required prior to most operations. 109 * 110 * Maps consist of an ordered doubly-linked list of simple entries. 111 * A hint and a RB tree is used to speed-up lookups. 112 * 113 * Callers looking to modify maps specify start/end addresses which cause 114 * the related map entry to be clipped if necessary, and then later 115 * recombined if the pieces remained compatible. 116 * 117 * Virtual copy operations are performed by copying VM object references 118 * from one map to another, and then marking both regions as copy-on-write. 119 */ 120 static boolean_t vmspace_ctor(void *obj, void *privdata, int ocflags); 121 static void vmspace_dtor(void *obj, void *privdata); 122 static void vmspace_terminate(struct vmspace *vm, int final); 123 124 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore"); 125 static struct objcache *vmspace_cache; 126 127 /* 128 * per-cpu page table cross mappings are initialized in early boot 129 * and might require a considerable number of vm_map_entry structures. 130 */ 131 #define MAPENTRYBSP_CACHE (MAXCPU+1) 132 #define MAPENTRYAP_CACHE 8 133 134 static struct vm_zone mapentzone_store; 135 static vm_zone_t mapentzone; 136 137 static struct vm_map_entry map_entry_init[MAX_MAPENT]; 138 static struct vm_map_entry cpu_map_entry_init_bsp[MAPENTRYBSP_CACHE]; 139 static struct vm_map_entry cpu_map_entry_init_ap[MAXCPU][MAPENTRYAP_CACHE]; 140 141 static int randomize_mmap; 142 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0, 143 "Randomize mmap offsets"); 144 static int vm_map_relock_enable = 1; 145 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW, 146 &vm_map_relock_enable, 0, "Randomize mmap offsets"); 147 148 static void vmspace_drop_notoken(struct vmspace *vm); 149 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref); 150 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *); 151 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *); 152 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 153 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 154 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *); 155 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t); 156 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t, 157 vm_map_entry_t); 158 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); 159 160 /* 161 * Initialize the vm_map module. Must be called before any other vm_map 162 * routines. 163 * 164 * Map and entry structures are allocated from the general purpose 165 * memory pool with some exceptions: 166 * 167 * - The kernel map is allocated statically. 168 * - Initial kernel map entries are allocated out of a static pool. 169 * - We must set ZONE_SPECIAL here or the early boot code can get 170 * stuck if there are >63 cores. 171 * 172 * These restrictions are necessary since malloc() uses the 173 * maps and requires map entries. 174 * 175 * Called from the low level boot code only. 176 */ 177 void 178 vm_map_startup(void) 179 { 180 mapentzone = &mapentzone_store; 181 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry), 182 map_entry_init, MAX_MAPENT); 183 mapentzone_store.zflags |= ZONE_SPECIAL; 184 } 185 186 /* 187 * Called prior to any vmspace allocations. 188 * 189 * Called from the low level boot code only. 190 */ 191 void 192 vm_init2(void) 193 { 194 vmspace_cache = objcache_create_mbacked(M_VMSPACE, 195 sizeof(struct vmspace), 196 0, ncpus * 4, 197 vmspace_ctor, vmspace_dtor, 198 NULL); 199 zinitna(mapentzone, NULL, 0, 0, ZONE_USE_RESERVE | ZONE_SPECIAL); 200 pmap_init2(); 201 vm_object_init2(); 202 } 203 204 /* 205 * objcache support. We leave the pmap root cached as long as possible 206 * for performance reasons. 207 */ 208 static 209 boolean_t 210 vmspace_ctor(void *obj, void *privdata, int ocflags) 211 { 212 struct vmspace *vm = obj; 213 214 bzero(vm, sizeof(*vm)); 215 vm->vm_refcnt = VM_REF_DELETED; 216 217 return 1; 218 } 219 220 static 221 void 222 vmspace_dtor(void *obj, void *privdata) 223 { 224 struct vmspace *vm = obj; 225 226 KKASSERT(vm->vm_refcnt == VM_REF_DELETED); 227 pmap_puninit(vmspace_pmap(vm)); 228 } 229 230 /* 231 * Red black tree functions 232 * 233 * The caller must hold the related map lock. 234 */ 235 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b); 236 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare); 237 238 /* a->start is address, and the only field has to be initialized */ 239 static int 240 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b) 241 { 242 if (a->start < b->start) 243 return(-1); 244 else if (a->start > b->start) 245 return(1); 246 return(0); 247 } 248 249 /* 250 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for 251 * every refcnt. 252 */ 253 void 254 vmspace_initrefs(struct vmspace *vm) 255 { 256 vm->vm_refcnt = 1; 257 vm->vm_holdcnt = 1; 258 } 259 260 /* 261 * Allocate a vmspace structure, including a vm_map and pmap. 262 * Initialize numerous fields. While the initial allocation is zerod, 263 * subsequence reuse from the objcache leaves elements of the structure 264 * intact (particularly the pmap), so portions must be zerod. 265 * 266 * Returns a referenced vmspace. 267 * 268 * No requirements. 269 */ 270 struct vmspace * 271 vmspace_alloc(vm_offset_t min, vm_offset_t max) 272 { 273 struct vmspace *vm; 274 275 vm = objcache_get(vmspace_cache, M_WAITOK); 276 277 bzero(&vm->vm_startcopy, 278 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy); 279 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */ 280 281 /* 282 * NOTE: hold to acquires token for safety. 283 * 284 * On return vmspace is referenced (refs=1, hold=1). That is, 285 * each refcnt also has a holdcnt. There can be additional holds 286 * (holdcnt) above and beyond the refcnt. Finalization is handled in 287 * two stages, one on refs 1->0, and the the second on hold 1->0. 288 */ 289 KKASSERT(vm->vm_holdcnt == 0); 290 KKASSERT(vm->vm_refcnt == VM_REF_DELETED); 291 vmspace_initrefs(vm); 292 vmspace_hold(vm); 293 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */ 294 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */ 295 vm->vm_shm = NULL; 296 vm->vm_flags = 0; 297 cpu_vmspace_alloc(vm); 298 vmspace_drop(vm); 299 300 return (vm); 301 } 302 303 /* 304 * NOTE: Can return 0 if the vmspace is exiting. 305 */ 306 int 307 vmspace_getrefs(struct vmspace *vm) 308 { 309 int32_t n; 310 311 n = vm->vm_refcnt; 312 cpu_ccfence(); 313 if (n & VM_REF_DELETED) 314 n = -1; 315 return n; 316 } 317 318 void 319 vmspace_hold(struct vmspace *vm) 320 { 321 atomic_add_int(&vm->vm_holdcnt, 1); 322 lwkt_gettoken(&vm->vm_map.token); 323 } 324 325 /* 326 * Drop with final termination interlock. 327 */ 328 void 329 vmspace_drop(struct vmspace *vm) 330 { 331 lwkt_reltoken(&vm->vm_map.token); 332 vmspace_drop_notoken(vm); 333 } 334 335 static void 336 vmspace_drop_notoken(struct vmspace *vm) 337 { 338 if (atomic_fetchadd_int(&vm->vm_holdcnt, -1) == 1) { 339 if (vm->vm_refcnt & VM_REF_DELETED) 340 vmspace_terminate(vm, 1); 341 } 342 } 343 344 /* 345 * A vmspace object must not be in a terminated state to be able to obtain 346 * additional refs on it. 347 * 348 * These are official references to the vmspace, the count is used to check 349 * for vmspace sharing. Foreign accessors should use 'hold' and not 'ref'. 350 * 351 * XXX we need to combine hold & ref together into one 64-bit field to allow 352 * holds to prevent stage-1 termination. 353 */ 354 void 355 vmspace_ref(struct vmspace *vm) 356 { 357 uint32_t n; 358 359 atomic_add_int(&vm->vm_holdcnt, 1); 360 n = atomic_fetchadd_int(&vm->vm_refcnt, 1); 361 KKASSERT((n & VM_REF_DELETED) == 0); 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 uint32_t n; 373 374 /* 375 * Drop refs. Each ref also has a hold which is also dropped. 376 * 377 * When refs hits 0 compete to get the VM_REF_DELETED flag (hold 378 * prevent finalization) to start termination processing. 379 * Finalization occurs when the last hold count drops to 0. 380 */ 381 n = atomic_fetchadd_int(&vm->vm_refcnt, -1) - 1; 382 while (n == 0) { 383 if (atomic_cmpset_int(&vm->vm_refcnt, 0, VM_REF_DELETED)) { 384 vmspace_terminate(vm, 0); 385 break; 386 } 387 n = vm->vm_refcnt; 388 cpu_ccfence(); 389 } 390 vmspace_drop_notoken(vm); 391 } 392 393 /* 394 * This is called during exit indicating that the vmspace is no 395 * longer in used by an exiting process, but the process has not yet 396 * been reaped. 397 * 398 * We drop refs, allowing for stage-1 termination, but maintain a holdcnt 399 * to prevent stage-2 until the process is reaped. Note hte order of 400 * operation, we must hold first. 401 * 402 * No requirements. 403 */ 404 void 405 vmspace_relexit(struct vmspace *vm) 406 { 407 atomic_add_int(&vm->vm_holdcnt, 1); 408 vmspace_rel(vm); 409 } 410 411 /* 412 * Called during reap to disconnect the remainder of the vmspace from 413 * the process. On the hold drop the vmspace termination is finalized. 414 * 415 * No requirements. 416 */ 417 void 418 vmspace_exitfree(struct proc *p) 419 { 420 struct vmspace *vm; 421 422 vm = p->p_vmspace; 423 p->p_vmspace = NULL; 424 vmspace_drop_notoken(vm); 425 } 426 427 /* 428 * Called in two cases: 429 * 430 * (1) When the last refcnt is dropped and the vmspace becomes inactive, 431 * called with final == 0. refcnt will be (u_int)-1 at this point, 432 * and holdcnt will still be non-zero. 433 * 434 * (2) When holdcnt becomes 0, called with final == 1. There should no 435 * longer be anyone with access to the vmspace. 436 * 437 * VMSPACE_EXIT1 flags the primary deactivation 438 * VMSPACE_EXIT2 flags the last reap 439 */ 440 static void 441 vmspace_terminate(struct vmspace *vm, int final) 442 { 443 int count; 444 445 lwkt_gettoken(&vm->vm_map.token); 446 if (final == 0) { 447 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0); 448 vm->vm_flags |= VMSPACE_EXIT1; 449 450 /* 451 * Get rid of most of the resources. Leave the kernel pmap 452 * intact. 453 * 454 * If the pmap does not contain wired pages we can bulk-delete 455 * the pmap as a performance optimization before removing the 456 * related mappings. 457 * 458 * If the pmap contains wired pages we cannot do this 459 * pre-optimization because currently vm_fault_unwire() 460 * expects the pmap pages to exist and will not decrement 461 * p->wire_count if they do not. 462 */ 463 shmexit(vm); 464 if (vmspace_pmap(vm)->pm_stats.wired_count) { 465 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 466 VM_MAX_USER_ADDRESS); 467 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 468 VM_MAX_USER_ADDRESS); 469 } else { 470 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 471 VM_MAX_USER_ADDRESS); 472 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 473 VM_MAX_USER_ADDRESS); 474 } 475 lwkt_reltoken(&vm->vm_map.token); 476 } else { 477 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0); 478 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0); 479 480 /* 481 * Get rid of remaining basic resources. 482 */ 483 vm->vm_flags |= VMSPACE_EXIT2; 484 shmexit(vm); 485 486 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 487 vm_map_lock(&vm->vm_map); 488 cpu_vmspace_free(vm); 489 490 /* 491 * Lock the map, to wait out all other references to it. 492 * Delete all of the mappings and pages they hold, then call 493 * the pmap module to reclaim anything left. 494 */ 495 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset, 496 vm->vm_map.max_offset, &count); 497 vm_map_unlock(&vm->vm_map); 498 vm_map_entry_release(count); 499 500 pmap_release(vmspace_pmap(vm)); 501 lwkt_reltoken(&vm->vm_map.token); 502 objcache_put(vmspace_cache, vm); 503 } 504 } 505 506 /* 507 * Swap useage is determined by taking the proportional swap used by 508 * VM objects backing the VM map. To make up for fractional losses, 509 * if the VM object has any swap use at all the associated map entries 510 * count for at least 1 swap page. 511 * 512 * No requirements. 513 */ 514 vm_offset_t 515 vmspace_swap_count(struct vmspace *vm) 516 { 517 vm_map_t map = &vm->vm_map; 518 vm_map_entry_t cur; 519 vm_object_t object; 520 vm_offset_t count = 0; 521 vm_offset_t n; 522 523 vmspace_hold(vm); 524 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 525 switch(cur->maptype) { 526 case VM_MAPTYPE_NORMAL: 527 case VM_MAPTYPE_VPAGETABLE: 528 if ((object = cur->object.vm_object) == NULL) 529 break; 530 if (object->swblock_count) { 531 n = (cur->end - cur->start) / PAGE_SIZE; 532 count += object->swblock_count * 533 SWAP_META_PAGES * n / object->size + 1; 534 } 535 break; 536 default: 537 break; 538 } 539 } 540 vmspace_drop(vm); 541 542 return(count); 543 } 544 545 /* 546 * Calculate the approximate number of anonymous pages in use by 547 * this vmspace. To make up for fractional losses, we count each 548 * VM object as having at least 1 anonymous page. 549 * 550 * No requirements. 551 */ 552 vm_offset_t 553 vmspace_anonymous_count(struct vmspace *vm) 554 { 555 vm_map_t map = &vm->vm_map; 556 vm_map_entry_t cur; 557 vm_object_t object; 558 vm_offset_t count = 0; 559 560 vmspace_hold(vm); 561 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 562 switch(cur->maptype) { 563 case VM_MAPTYPE_NORMAL: 564 case VM_MAPTYPE_VPAGETABLE: 565 if ((object = cur->object.vm_object) == NULL) 566 break; 567 if (object->type != OBJT_DEFAULT && 568 object->type != OBJT_SWAP) { 569 break; 570 } 571 count += object->resident_page_count; 572 break; 573 default: 574 break; 575 } 576 } 577 vmspace_drop(vm); 578 579 return(count); 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 spin_init(&map->ilock_spin, "ilock"); 594 map->ilock_base = NULL; 595 map->nentries = 0; 596 map->size = 0; 597 map->system_map = 0; 598 map->min_offset = min; 599 map->max_offset = max; 600 map->pmap = pmap; 601 map->first_free = &map->header; 602 map->hint = &map->header; 603 map->timestamp = 0; 604 map->flags = 0; 605 lwkt_token_init(&map->token, "vm_map"); 606 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0); 607 } 608 609 /* 610 * Shadow the vm_map_entry's object. This typically needs to be done when 611 * a write fault is taken on an entry which had previously been cloned by 612 * fork(). The shared object (which might be NULL) must become private so 613 * we add a shadow layer above it. 614 * 615 * Object allocation for anonymous mappings is defered as long as possible. 616 * When creating a shadow, however, the underlying object must be instantiated 617 * so it can be shared. 618 * 619 * If the map segment is governed by a virtual page table then it is 620 * possible to address offsets beyond the mapped area. Just allocate 621 * a maximally sized object for this case. 622 * 623 * If addref is non-zero an additional reference is added to the returned 624 * entry. This mechanic exists because the additional reference might have 625 * to be added atomically and not after return to prevent a premature 626 * collapse. 627 * 628 * The vm_map must be exclusively locked. 629 * No other requirements. 630 */ 631 static 632 void 633 vm_map_entry_shadow(vm_map_entry_t entry, int addref) 634 { 635 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 636 vm_object_shadow(&entry->object.vm_object, &entry->offset, 637 0x7FFFFFFF, addref); /* XXX */ 638 } else { 639 vm_object_shadow(&entry->object.vm_object, &entry->offset, 640 atop(entry->end - entry->start), addref); 641 } 642 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 643 } 644 645 /* 646 * Allocate an object for a vm_map_entry. 647 * 648 * Object allocation for anonymous mappings is defered as long as possible. 649 * This function is called when we can defer no longer, generally when a map 650 * entry might be split or forked or takes a page fault. 651 * 652 * If the map segment is governed by a virtual page table then it is 653 * possible to address offsets beyond the mapped area. Just allocate 654 * a maximally sized object for this case. 655 * 656 * The vm_map must be exclusively locked. 657 * No other requirements. 658 */ 659 void 660 vm_map_entry_allocate_object(vm_map_entry_t entry) 661 { 662 vm_object_t obj; 663 664 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 665 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */ 666 } else { 667 obj = vm_object_allocate(OBJT_DEFAULT, 668 atop(entry->end - entry->start)); 669 } 670 entry->object.vm_object = obj; 671 entry->offset = 0; 672 } 673 674 /* 675 * Set an initial negative count so the first attempt to reserve 676 * space preloads a bunch of vm_map_entry's for this cpu. Also 677 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 678 * map a new page for vm_map_entry structures. SMP systems are 679 * particularly sensitive. 680 * 681 * This routine is called in early boot so we cannot just call 682 * vm_map_entry_reserve(). 683 * 684 * Called from the low level boot code only (for each cpu) 685 * 686 * WARNING! Take care not to have too-big a static/BSS structure here 687 * as MAXCPU can be 256+, otherwise the loader's 64MB heap 688 * can get blown out by the kernel plus the initrd image. 689 */ 690 void 691 vm_map_entry_reserve_cpu_init(globaldata_t gd) 692 { 693 vm_map_entry_t entry; 694 int count; 695 int i; 696 697 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2; 698 if (gd->gd_cpuid == 0) { 699 entry = &cpu_map_entry_init_bsp[0]; 700 count = MAPENTRYBSP_CACHE; 701 } else { 702 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0]; 703 count = MAPENTRYAP_CACHE; 704 } 705 for (i = 0; i < count; ++i, ++entry) { 706 entry->next = gd->gd_vme_base; 707 gd->gd_vme_base = entry; 708 } 709 } 710 711 /* 712 * Reserves vm_map_entry structures so code later on can manipulate 713 * map_entry structures within a locked map without blocking trying 714 * to allocate a new vm_map_entry. 715 * 716 * No requirements. 717 */ 718 int 719 vm_map_entry_reserve(int count) 720 { 721 struct globaldata *gd = mycpu; 722 vm_map_entry_t entry; 723 724 /* 725 * Make sure we have enough structures in gd_vme_base to handle 726 * the reservation request. 727 * 728 * The critical section protects access to the per-cpu gd. 729 */ 730 crit_enter(); 731 while (gd->gd_vme_avail < count) { 732 entry = zalloc(mapentzone); 733 entry->next = gd->gd_vme_base; 734 gd->gd_vme_base = entry; 735 ++gd->gd_vme_avail; 736 } 737 gd->gd_vme_avail -= count; 738 crit_exit(); 739 740 return(count); 741 } 742 743 /* 744 * Releases previously reserved vm_map_entry structures that were not 745 * used. If we have too much junk in our per-cpu cache clean some of 746 * it out. 747 * 748 * No requirements. 749 */ 750 void 751 vm_map_entry_release(int count) 752 { 753 struct globaldata *gd = mycpu; 754 vm_map_entry_t entry; 755 756 crit_enter(); 757 gd->gd_vme_avail += count; 758 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 759 entry = gd->gd_vme_base; 760 KKASSERT(entry != NULL); 761 gd->gd_vme_base = entry->next; 762 --gd->gd_vme_avail; 763 crit_exit(); 764 zfree(mapentzone, entry); 765 crit_enter(); 766 } 767 crit_exit(); 768 } 769 770 /* 771 * Reserve map entry structures for use in kernel_map itself. These 772 * entries have *ALREADY* been reserved on a per-cpu basis when the map 773 * was inited. This function is used by zalloc() to avoid a recursion 774 * when zalloc() itself needs to allocate additional kernel memory. 775 * 776 * This function works like the normal reserve but does not load the 777 * vm_map_entry cache (because that would result in an infinite 778 * recursion). Note that gd_vme_avail may go negative. This is expected. 779 * 780 * Any caller of this function must be sure to renormalize after 781 * potentially eating entries to ensure that the reserve supply 782 * remains intact. 783 * 784 * No requirements. 785 */ 786 int 787 vm_map_entry_kreserve(int count) 788 { 789 struct globaldata *gd = mycpu; 790 791 crit_enter(); 792 gd->gd_vme_avail -= count; 793 crit_exit(); 794 KASSERT(gd->gd_vme_base != NULL, 795 ("no reserved entries left, gd_vme_avail = %d", 796 gd->gd_vme_avail)); 797 return(count); 798 } 799 800 /* 801 * Release previously reserved map entries for kernel_map. We do not 802 * attempt to clean up like the normal release function as this would 803 * cause an unnecessary (but probably not fatal) deep procedure call. 804 * 805 * No requirements. 806 */ 807 void 808 vm_map_entry_krelease(int count) 809 { 810 struct globaldata *gd = mycpu; 811 812 crit_enter(); 813 gd->gd_vme_avail += count; 814 crit_exit(); 815 } 816 817 /* 818 * Allocates a VM map entry for insertion. No entry fields are filled in. 819 * 820 * The entries should have previously been reserved. The reservation count 821 * is tracked in (*countp). 822 * 823 * No requirements. 824 */ 825 static vm_map_entry_t 826 vm_map_entry_create(vm_map_t map, int *countp) 827 { 828 struct globaldata *gd = mycpu; 829 vm_map_entry_t entry; 830 831 KKASSERT(*countp > 0); 832 --*countp; 833 crit_enter(); 834 entry = gd->gd_vme_base; 835 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 836 gd->gd_vme_base = entry->next; 837 crit_exit(); 838 839 return(entry); 840 } 841 842 /* 843 * Dispose of a vm_map_entry that is no longer being referenced. 844 * 845 * No requirements. 846 */ 847 static void 848 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 849 { 850 struct globaldata *gd = mycpu; 851 852 KKASSERT(map->hint != entry); 853 KKASSERT(map->first_free != entry); 854 855 ++*countp; 856 crit_enter(); 857 entry->next = gd->gd_vme_base; 858 gd->gd_vme_base = entry; 859 crit_exit(); 860 } 861 862 863 /* 864 * Insert/remove entries from maps. 865 * 866 * The related map must be exclusively locked. 867 * The caller must hold map->token 868 * No other requirements. 869 */ 870 static __inline void 871 vm_map_entry_link(vm_map_t map, 872 vm_map_entry_t after_where, 873 vm_map_entry_t entry) 874 { 875 ASSERT_VM_MAP_LOCKED(map); 876 877 map->nentries++; 878 entry->prev = after_where; 879 entry->next = after_where->next; 880 entry->next->prev = entry; 881 after_where->next = entry; 882 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry)) 883 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry); 884 } 885 886 static __inline void 887 vm_map_entry_unlink(vm_map_t map, 888 vm_map_entry_t entry) 889 { 890 vm_map_entry_t prev; 891 vm_map_entry_t next; 892 893 ASSERT_VM_MAP_LOCKED(map); 894 895 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 896 panic("vm_map_entry_unlink: attempt to mess with " 897 "locked entry! %p", entry); 898 } 899 prev = entry->prev; 900 next = entry->next; 901 next->prev = prev; 902 prev->next = next; 903 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry); 904 map->nentries--; 905 } 906 907 /* 908 * Finds the map entry containing (or immediately preceding) the specified 909 * address in the given map. The entry is returned in (*entry). 910 * 911 * The boolean result indicates whether the address is actually contained 912 * in the map. 913 * 914 * The related map must be locked. 915 * No other requirements. 916 */ 917 boolean_t 918 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry) 919 { 920 vm_map_entry_t tmp; 921 vm_map_entry_t last; 922 923 ASSERT_VM_MAP_LOCKED(map); 924 #if 0 925 /* 926 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive 927 * the hint code with the red-black lookup meets with system crashes 928 * and lockups. We do not yet know why. 929 * 930 * It is possible that the problem is related to the setting 931 * of the hint during map_entry deletion, in the code specified 932 * at the GGG comment later on in this file. 933 * 934 * YYY More likely it's because this function can be called with 935 * a shared lock on the map, resulting in map->hint updates possibly 936 * racing. Fixed now but untested. 937 */ 938 /* 939 * Quickly check the cached hint, there's a good chance of a match. 940 */ 941 tmp = map->hint; 942 cpu_ccfence(); 943 if (tmp != &map->header) { 944 if (address >= tmp->start && address < tmp->end) { 945 *entry = tmp; 946 return(TRUE); 947 } 948 } 949 #endif 950 951 /* 952 * Locate the record from the top of the tree. 'last' tracks the 953 * closest prior record and is returned if no match is found, which 954 * in binary tree terms means tracking the most recent right-branch 955 * taken. If there is no prior record, &map->header is returned. 956 */ 957 last = &map->header; 958 tmp = RB_ROOT(&map->rb_root); 959 960 while (tmp) { 961 if (address >= tmp->start) { 962 if (address < tmp->end) { 963 *entry = tmp; 964 map->hint = tmp; 965 return(TRUE); 966 } 967 last = tmp; 968 tmp = RB_RIGHT(tmp, rb_entry); 969 } else { 970 tmp = RB_LEFT(tmp, rb_entry); 971 } 972 } 973 *entry = last; 974 return (FALSE); 975 } 976 977 /* 978 * Inserts the given whole VM object into the target map at the specified 979 * address range. The object's size should match that of the address range. 980 * 981 * The map must be exclusively locked. 982 * The object must be held. 983 * The caller must have reserved sufficient vm_map_entry structures. 984 * 985 * If object is non-NULL, ref count must be bumped by caller prior to 986 * making call to account for the new entry. 987 */ 988 int 989 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux, 990 vm_ooffset_t offset, vm_offset_t start, vm_offset_t end, 991 vm_maptype_t maptype, vm_subsys_t id, 992 vm_prot_t prot, vm_prot_t max, int cow) 993 { 994 vm_map_entry_t new_entry; 995 vm_map_entry_t prev_entry; 996 vm_map_entry_t temp_entry; 997 vm_eflags_t protoeflags; 998 int must_drop = 0; 999 vm_object_t object; 1000 1001 if (maptype == VM_MAPTYPE_UKSMAP) 1002 object = NULL; 1003 else 1004 object = map_object; 1005 1006 ASSERT_VM_MAP_LOCKED(map); 1007 if (object) 1008 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1009 1010 /* 1011 * Check that the start and end points are not bogus. 1012 */ 1013 if ((start < map->min_offset) || (end > map->max_offset) || 1014 (start >= end)) 1015 return (KERN_INVALID_ADDRESS); 1016 1017 /* 1018 * Find the entry prior to the proposed starting address; if it's part 1019 * of an existing entry, this range is bogus. 1020 */ 1021 if (vm_map_lookup_entry(map, start, &temp_entry)) 1022 return (KERN_NO_SPACE); 1023 1024 prev_entry = temp_entry; 1025 1026 /* 1027 * Assert that the next entry doesn't overlap the end point. 1028 */ 1029 1030 if ((prev_entry->next != &map->header) && 1031 (prev_entry->next->start < end)) 1032 return (KERN_NO_SPACE); 1033 1034 protoeflags = 0; 1035 1036 if (cow & MAP_COPY_ON_WRITE) 1037 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 1038 1039 if (cow & MAP_NOFAULT) { 1040 protoeflags |= MAP_ENTRY_NOFAULT; 1041 1042 KASSERT(object == NULL, 1043 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 1044 } 1045 if (cow & MAP_DISABLE_SYNCER) 1046 protoeflags |= MAP_ENTRY_NOSYNC; 1047 if (cow & MAP_DISABLE_COREDUMP) 1048 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1049 if (cow & MAP_IS_STACK) 1050 protoeflags |= MAP_ENTRY_STACK; 1051 if (cow & MAP_IS_KSTACK) 1052 protoeflags |= MAP_ENTRY_KSTACK; 1053 1054 lwkt_gettoken(&map->token); 1055 1056 if (object) { 1057 /* 1058 * When object is non-NULL, it could be shared with another 1059 * process. We have to set or clear OBJ_ONEMAPPING 1060 * appropriately. 1061 * 1062 * NOTE: This flag is only applicable to DEFAULT and SWAP 1063 * objects and will already be clear in other types 1064 * of objects, so a shared object lock is ok for 1065 * VNODE objects. 1066 */ 1067 if ((object->ref_count > 1) || (object->shadow_count != 0)) { 1068 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1069 } 1070 } 1071 else if ((prev_entry != &map->header) && 1072 (prev_entry->eflags == protoeflags) && 1073 (prev_entry->end == start) && 1074 (prev_entry->wired_count == 0) && 1075 (prev_entry->id == id) && 1076 prev_entry->maptype == maptype && 1077 maptype == VM_MAPTYPE_NORMAL && 1078 ((prev_entry->object.vm_object == NULL) || 1079 vm_object_coalesce(prev_entry->object.vm_object, 1080 OFF_TO_IDX(prev_entry->offset), 1081 (vm_size_t)(prev_entry->end - prev_entry->start), 1082 (vm_size_t)(end - prev_entry->end)))) { 1083 /* 1084 * We were able to extend the object. Determine if we 1085 * can extend the previous map entry to include the 1086 * new range as well. 1087 */ 1088 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 1089 (prev_entry->protection == prot) && 1090 (prev_entry->max_protection == max)) { 1091 map->size += (end - prev_entry->end); 1092 prev_entry->end = end; 1093 vm_map_simplify_entry(map, prev_entry, countp); 1094 lwkt_reltoken(&map->token); 1095 return (KERN_SUCCESS); 1096 } 1097 1098 /* 1099 * If we can extend the object but cannot extend the 1100 * map entry, we have to create a new map entry. We 1101 * must bump the ref count on the extended object to 1102 * account for it. object may be NULL. 1103 * 1104 * XXX if object is NULL should we set offset to 0 here ? 1105 */ 1106 object = prev_entry->object.vm_object; 1107 offset = prev_entry->offset + 1108 (prev_entry->end - prev_entry->start); 1109 if (object) { 1110 vm_object_hold(object); 1111 vm_object_chain_wait(object, 0); 1112 vm_object_reference_locked(object); 1113 must_drop = 1; 1114 map_object = object; 1115 } 1116 } 1117 1118 /* 1119 * NOTE: if conditionals fail, object can be NULL here. This occurs 1120 * in things like the buffer map where we manage kva but do not manage 1121 * backing objects. 1122 */ 1123 1124 /* 1125 * Create a new entry 1126 */ 1127 1128 new_entry = vm_map_entry_create(map, countp); 1129 new_entry->start = start; 1130 new_entry->end = end; 1131 new_entry->id = id; 1132 1133 new_entry->maptype = maptype; 1134 new_entry->eflags = protoeflags; 1135 new_entry->object.map_object = map_object; 1136 new_entry->aux.master_pde = 0; /* in case size is different */ 1137 new_entry->aux.map_aux = map_aux; 1138 new_entry->offset = offset; 1139 1140 new_entry->inheritance = VM_INHERIT_DEFAULT; 1141 new_entry->protection = prot; 1142 new_entry->max_protection = max; 1143 new_entry->wired_count = 0; 1144 1145 /* 1146 * Insert the new entry into the list 1147 */ 1148 1149 vm_map_entry_link(map, prev_entry, new_entry); 1150 map->size += new_entry->end - new_entry->start; 1151 1152 /* 1153 * Update the free space hint. Entries cannot overlap. 1154 * An exact comparison is needed to avoid matching 1155 * against the map->header. 1156 */ 1157 if ((map->first_free == prev_entry) && 1158 (prev_entry->end == new_entry->start)) { 1159 map->first_free = new_entry; 1160 } 1161 1162 #if 0 1163 /* 1164 * Temporarily removed to avoid MAP_STACK panic, due to 1165 * MAP_STACK being a huge hack. Will be added back in 1166 * when MAP_STACK (and the user stack mapping) is fixed. 1167 */ 1168 /* 1169 * It may be possible to simplify the entry 1170 */ 1171 vm_map_simplify_entry(map, new_entry, countp); 1172 #endif 1173 1174 /* 1175 * Try to pre-populate the page table. Mappings governed by virtual 1176 * page tables cannot be prepopulated without a lot of work, so 1177 * don't try. 1178 */ 1179 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) && 1180 maptype != VM_MAPTYPE_VPAGETABLE && 1181 maptype != VM_MAPTYPE_UKSMAP) { 1182 int dorelock = 0; 1183 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) { 1184 dorelock = 1; 1185 vm_object_lock_swap(); 1186 vm_object_drop(object); 1187 } 1188 pmap_object_init_pt(map->pmap, start, prot, 1189 object, OFF_TO_IDX(offset), end - start, 1190 cow & MAP_PREFAULT_PARTIAL); 1191 if (dorelock) { 1192 vm_object_hold(object); 1193 vm_object_lock_swap(); 1194 } 1195 } 1196 if (must_drop) 1197 vm_object_drop(object); 1198 1199 lwkt_reltoken(&map->token); 1200 return (KERN_SUCCESS); 1201 } 1202 1203 /* 1204 * Find sufficient space for `length' bytes in the given map, starting at 1205 * `start'. Returns 0 on success, 1 on no space. 1206 * 1207 * This function will returned an arbitrarily aligned pointer. If no 1208 * particular alignment is required you should pass align as 1. Note that 1209 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 1210 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 1211 * argument. 1212 * 1213 * 'align' should be a power of 2 but is not required to be. 1214 * 1215 * The map must be exclusively locked. 1216 * No other requirements. 1217 */ 1218 int 1219 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length, 1220 vm_size_t align, int flags, vm_offset_t *addr) 1221 { 1222 vm_map_entry_t entry, next; 1223 vm_offset_t end; 1224 vm_offset_t align_mask; 1225 1226 if (start < map->min_offset) 1227 start = map->min_offset; 1228 if (start > map->max_offset) 1229 return (1); 1230 1231 /* 1232 * If the alignment is not a power of 2 we will have to use 1233 * a mod/division, set align_mask to a special value. 1234 */ 1235 if ((align | (align - 1)) + 1 != (align << 1)) 1236 align_mask = (vm_offset_t)-1; 1237 else 1238 align_mask = align - 1; 1239 1240 /* 1241 * Look for the first possible address; if there's already something 1242 * at this address, we have to start after it. 1243 */ 1244 if (start == map->min_offset) { 1245 if ((entry = map->first_free) != &map->header) 1246 start = entry->end; 1247 } else { 1248 vm_map_entry_t tmp; 1249 1250 if (vm_map_lookup_entry(map, start, &tmp)) 1251 start = tmp->end; 1252 entry = tmp; 1253 } 1254 1255 /* 1256 * Look through the rest of the map, trying to fit a new region in the 1257 * gap between existing regions, or after the very last region. 1258 */ 1259 for (;; start = (entry = next)->end) { 1260 /* 1261 * Adjust the proposed start by the requested alignment, 1262 * be sure that we didn't wrap the address. 1263 */ 1264 if (align_mask == (vm_offset_t)-1) 1265 end = roundup(start, align); 1266 else 1267 end = (start + align_mask) & ~align_mask; 1268 if (end < start) 1269 return (1); 1270 start = end; 1271 /* 1272 * Find the end of the proposed new region. Be sure we didn't 1273 * go beyond the end of the map, or wrap around the address. 1274 * Then check to see if this is the last entry or if the 1275 * proposed end fits in the gap between this and the next 1276 * entry. 1277 */ 1278 end = start + length; 1279 if (end > map->max_offset || end < start) 1280 return (1); 1281 next = entry->next; 1282 1283 /* 1284 * If the next entry's start address is beyond the desired 1285 * end address we may have found a good entry. 1286 * 1287 * If the next entry is a stack mapping we do not map into 1288 * the stack's reserved space. 1289 * 1290 * XXX continue to allow mapping into the stack's reserved 1291 * space if doing a MAP_STACK mapping inside a MAP_STACK 1292 * mapping, for backwards compatibility. But the caller 1293 * really should use MAP_STACK | MAP_TRYFIXED if they 1294 * want to do that. 1295 */ 1296 if (next == &map->header) 1297 break; 1298 if (next->start >= end) { 1299 if ((next->eflags & MAP_ENTRY_STACK) == 0) 1300 break; 1301 if (flags & MAP_STACK) 1302 break; 1303 if (next->start - next->aux.avail_ssize >= end) 1304 break; 1305 } 1306 } 1307 map->hint = entry; 1308 1309 /* 1310 * Grow the kernel_map if necessary. pmap_growkernel() will panic 1311 * if it fails. The kernel_map is locked and nothing can steal 1312 * our address space if pmap_growkernel() blocks. 1313 * 1314 * NOTE: This may be unconditionally called for kldload areas on 1315 * x86_64 because these do not bump kernel_vm_end (which would 1316 * fill 128G worth of page tables!). Therefore we must not 1317 * retry. 1318 */ 1319 if (map == &kernel_map) { 1320 vm_offset_t kstop; 1321 1322 kstop = round_page(start + length); 1323 if (kstop > kernel_vm_end) 1324 pmap_growkernel(start, kstop); 1325 } 1326 *addr = start; 1327 return (0); 1328 } 1329 1330 /* 1331 * vm_map_find finds an unallocated region in the target address map with 1332 * the given length and allocates it. The search is defined to be first-fit 1333 * from the specified address; the region found is returned in the same 1334 * parameter. 1335 * 1336 * If object is non-NULL, ref count must be bumped by caller 1337 * prior to making call to account for the new entry. 1338 * 1339 * No requirements. This function will lock the map temporarily. 1340 */ 1341 int 1342 vm_map_find(vm_map_t map, void *map_object, void *map_aux, 1343 vm_ooffset_t offset, vm_offset_t *addr, 1344 vm_size_t length, vm_size_t align, boolean_t fitit, 1345 vm_maptype_t maptype, vm_subsys_t id, 1346 vm_prot_t prot, vm_prot_t max, int cow) 1347 { 1348 vm_offset_t start; 1349 vm_object_t object; 1350 int result; 1351 int count; 1352 1353 if (maptype == VM_MAPTYPE_UKSMAP) 1354 object = NULL; 1355 else 1356 object = map_object; 1357 1358 start = *addr; 1359 1360 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1361 vm_map_lock(map); 1362 if (object) 1363 vm_object_hold_shared(object); 1364 if (fitit) { 1365 if (vm_map_findspace(map, start, length, align, 0, addr)) { 1366 if (object) 1367 vm_object_drop(object); 1368 vm_map_unlock(map); 1369 vm_map_entry_release(count); 1370 return (KERN_NO_SPACE); 1371 } 1372 start = *addr; 1373 } 1374 result = vm_map_insert(map, &count, map_object, map_aux, 1375 offset, start, start + length, 1376 maptype, id, prot, max, cow); 1377 if (object) 1378 vm_object_drop(object); 1379 vm_map_unlock(map); 1380 vm_map_entry_release(count); 1381 1382 return (result); 1383 } 1384 1385 /* 1386 * Simplify the given map entry by merging with either neighbor. This 1387 * routine also has the ability to merge with both neighbors. 1388 * 1389 * This routine guarentees that the passed entry remains valid (though 1390 * possibly extended). When merging, this routine may delete one or 1391 * both neighbors. No action is taken on entries which have their 1392 * in-transition flag set. 1393 * 1394 * The map must be exclusively locked. 1395 */ 1396 void 1397 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 1398 { 1399 vm_map_entry_t next, prev; 1400 vm_size_t prevsize, esize; 1401 1402 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1403 ++mycpu->gd_cnt.v_intrans_coll; 1404 return; 1405 } 1406 1407 if (entry->maptype == VM_MAPTYPE_SUBMAP) 1408 return; 1409 if (entry->maptype == VM_MAPTYPE_UKSMAP) 1410 return; 1411 1412 prev = entry->prev; 1413 if (prev != &map->header) { 1414 prevsize = prev->end - prev->start; 1415 if ( (prev->end == entry->start) && 1416 (prev->maptype == entry->maptype) && 1417 (prev->object.vm_object == entry->object.vm_object) && 1418 (!prev->object.vm_object || 1419 (prev->offset + prevsize == entry->offset)) && 1420 (prev->eflags == entry->eflags) && 1421 (prev->protection == entry->protection) && 1422 (prev->max_protection == entry->max_protection) && 1423 (prev->inheritance == entry->inheritance) && 1424 (prev->id == entry->id) && 1425 (prev->wired_count == entry->wired_count)) { 1426 if (map->first_free == prev) 1427 map->first_free = entry; 1428 if (map->hint == prev) 1429 map->hint = entry; 1430 vm_map_entry_unlink(map, prev); 1431 entry->start = prev->start; 1432 entry->offset = prev->offset; 1433 if (prev->object.vm_object) 1434 vm_object_deallocate(prev->object.vm_object); 1435 vm_map_entry_dispose(map, prev, countp); 1436 } 1437 } 1438 1439 next = entry->next; 1440 if (next != &map->header) { 1441 esize = entry->end - entry->start; 1442 if ((entry->end == next->start) && 1443 (next->maptype == entry->maptype) && 1444 (next->object.vm_object == entry->object.vm_object) && 1445 (!entry->object.vm_object || 1446 (entry->offset + esize == next->offset)) && 1447 (next->eflags == entry->eflags) && 1448 (next->protection == entry->protection) && 1449 (next->max_protection == entry->max_protection) && 1450 (next->inheritance == entry->inheritance) && 1451 (next->id == entry->id) && 1452 (next->wired_count == entry->wired_count)) { 1453 if (map->first_free == next) 1454 map->first_free = entry; 1455 if (map->hint == next) 1456 map->hint = entry; 1457 vm_map_entry_unlink(map, next); 1458 entry->end = next->end; 1459 if (next->object.vm_object) 1460 vm_object_deallocate(next->object.vm_object); 1461 vm_map_entry_dispose(map, next, countp); 1462 } 1463 } 1464 } 1465 1466 /* 1467 * Asserts that the given entry begins at or after the specified address. 1468 * If necessary, it splits the entry into two. 1469 */ 1470 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1471 { \ 1472 if (startaddr > entry->start) \ 1473 _vm_map_clip_start(map, entry, startaddr, countp); \ 1474 } 1475 1476 /* 1477 * This routine is called only when it is known that the entry must be split. 1478 * 1479 * The map must be exclusively locked. 1480 */ 1481 static void 1482 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, 1483 int *countp) 1484 { 1485 vm_map_entry_t new_entry; 1486 1487 /* 1488 * Split off the front portion -- note that we must insert the new 1489 * entry BEFORE this one, so that this entry has the specified 1490 * starting address. 1491 */ 1492 1493 vm_map_simplify_entry(map, entry, countp); 1494 1495 /* 1496 * If there is no object backing this entry, we might as well create 1497 * one now. If we defer it, an object can get created after the map 1498 * is clipped, and individual objects will be created for the split-up 1499 * map. This is a bit of a hack, but is also about the best place to 1500 * put this improvement. 1501 */ 1502 if (entry->object.vm_object == NULL && !map->system_map) { 1503 vm_map_entry_allocate_object(entry); 1504 } 1505 1506 new_entry = vm_map_entry_create(map, countp); 1507 *new_entry = *entry; 1508 1509 new_entry->end = start; 1510 entry->offset += (start - entry->start); 1511 entry->start = start; 1512 1513 vm_map_entry_link(map, entry->prev, new_entry); 1514 1515 switch(entry->maptype) { 1516 case VM_MAPTYPE_NORMAL: 1517 case VM_MAPTYPE_VPAGETABLE: 1518 if (new_entry->object.vm_object) { 1519 vm_object_hold(new_entry->object.vm_object); 1520 vm_object_chain_wait(new_entry->object.vm_object, 0); 1521 vm_object_reference_locked(new_entry->object.vm_object); 1522 vm_object_drop(new_entry->object.vm_object); 1523 } 1524 break; 1525 default: 1526 break; 1527 } 1528 } 1529 1530 /* 1531 * Asserts that the given entry ends at or before the specified address. 1532 * If necessary, it splits the entry into two. 1533 * 1534 * The map must be exclusively locked. 1535 */ 1536 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1537 { \ 1538 if (endaddr < entry->end) \ 1539 _vm_map_clip_end(map, entry, endaddr, countp); \ 1540 } 1541 1542 /* 1543 * This routine is called only when it is known that the entry must be split. 1544 * 1545 * The map must be exclusively locked. 1546 */ 1547 static void 1548 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, 1549 int *countp) 1550 { 1551 vm_map_entry_t new_entry; 1552 1553 /* 1554 * If there is no object backing this entry, we might as well create 1555 * one now. If we defer it, an object can get created after the map 1556 * is clipped, and individual objects will be created for the split-up 1557 * map. This is a bit of a hack, but is also about the best place to 1558 * put this improvement. 1559 */ 1560 1561 if (entry->object.vm_object == NULL && !map->system_map) { 1562 vm_map_entry_allocate_object(entry); 1563 } 1564 1565 /* 1566 * Create a new entry and insert it AFTER the specified entry 1567 */ 1568 1569 new_entry = vm_map_entry_create(map, countp); 1570 *new_entry = *entry; 1571 1572 new_entry->start = entry->end = end; 1573 new_entry->offset += (end - entry->start); 1574 1575 vm_map_entry_link(map, entry, new_entry); 1576 1577 switch(entry->maptype) { 1578 case VM_MAPTYPE_NORMAL: 1579 case VM_MAPTYPE_VPAGETABLE: 1580 if (new_entry->object.vm_object) { 1581 vm_object_hold(new_entry->object.vm_object); 1582 vm_object_chain_wait(new_entry->object.vm_object, 0); 1583 vm_object_reference_locked(new_entry->object.vm_object); 1584 vm_object_drop(new_entry->object.vm_object); 1585 } 1586 break; 1587 default: 1588 break; 1589 } 1590 } 1591 1592 /* 1593 * Asserts that the starting and ending region addresses fall within the 1594 * valid range for the map. 1595 */ 1596 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1597 { \ 1598 if (start < vm_map_min(map)) \ 1599 start = vm_map_min(map); \ 1600 if (end > vm_map_max(map)) \ 1601 end = vm_map_max(map); \ 1602 if (start > end) \ 1603 start = end; \ 1604 } 1605 1606 /* 1607 * Used to block when an in-transition collison occurs. The map 1608 * is unlocked for the sleep and relocked before the return. 1609 */ 1610 void 1611 vm_map_transition_wait(vm_map_t map) 1612 { 1613 tsleep_interlock(map, 0); 1614 vm_map_unlock(map); 1615 tsleep(map, PINTERLOCKED, "vment", 0); 1616 vm_map_lock(map); 1617 } 1618 1619 /* 1620 * When we do blocking operations with the map lock held it is 1621 * possible that a clip might have occured on our in-transit entry, 1622 * requiring an adjustment to the entry in our loop. These macros 1623 * help the pageable and clip_range code deal with the case. The 1624 * conditional costs virtually nothing if no clipping has occured. 1625 */ 1626 1627 #define CLIP_CHECK_BACK(entry, save_start) \ 1628 do { \ 1629 while (entry->start != save_start) { \ 1630 entry = entry->prev; \ 1631 KASSERT(entry != &map->header, ("bad entry clip")); \ 1632 } \ 1633 } while(0) 1634 1635 #define CLIP_CHECK_FWD(entry, save_end) \ 1636 do { \ 1637 while (entry->end != save_end) { \ 1638 entry = entry->next; \ 1639 KASSERT(entry != &map->header, ("bad entry clip")); \ 1640 } \ 1641 } while(0) 1642 1643 1644 /* 1645 * Clip the specified range and return the base entry. The 1646 * range may cover several entries starting at the returned base 1647 * and the first and last entry in the covering sequence will be 1648 * properly clipped to the requested start and end address. 1649 * 1650 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1651 * flag. 1652 * 1653 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1654 * covered by the requested range. 1655 * 1656 * The map must be exclusively locked on entry and will remain locked 1657 * on return. If no range exists or the range contains holes and you 1658 * specified that no holes were allowed, NULL will be returned. This 1659 * routine may temporarily unlock the map in order avoid a deadlock when 1660 * sleeping. 1661 */ 1662 static 1663 vm_map_entry_t 1664 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1665 int *countp, int flags) 1666 { 1667 vm_map_entry_t start_entry; 1668 vm_map_entry_t entry; 1669 1670 /* 1671 * Locate the entry and effect initial clipping. The in-transition 1672 * case does not occur very often so do not try to optimize it. 1673 */ 1674 again: 1675 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1676 return (NULL); 1677 entry = start_entry; 1678 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1679 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1680 ++mycpu->gd_cnt.v_intrans_coll; 1681 ++mycpu->gd_cnt.v_intrans_wait; 1682 vm_map_transition_wait(map); 1683 /* 1684 * entry and/or start_entry may have been clipped while 1685 * we slept, or may have gone away entirely. We have 1686 * to restart from the lookup. 1687 */ 1688 goto again; 1689 } 1690 1691 /* 1692 * Since we hold an exclusive map lock we do not have to restart 1693 * after clipping, even though clipping may block in zalloc. 1694 */ 1695 vm_map_clip_start(map, entry, start, countp); 1696 vm_map_clip_end(map, entry, end, countp); 1697 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1698 1699 /* 1700 * Scan entries covered by the range. When working on the next 1701 * entry a restart need only re-loop on the current entry which 1702 * we have already locked, since 'next' may have changed. Also, 1703 * even though entry is safe, it may have been clipped so we 1704 * have to iterate forwards through the clip after sleeping. 1705 */ 1706 while (entry->next != &map->header && entry->next->start < end) { 1707 vm_map_entry_t next = entry->next; 1708 1709 if (flags & MAP_CLIP_NO_HOLES) { 1710 if (next->start > entry->end) { 1711 vm_map_unclip_range(map, start_entry, 1712 start, entry->end, countp, flags); 1713 return(NULL); 1714 } 1715 } 1716 1717 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1718 vm_offset_t save_end = entry->end; 1719 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1720 ++mycpu->gd_cnt.v_intrans_coll; 1721 ++mycpu->gd_cnt.v_intrans_wait; 1722 vm_map_transition_wait(map); 1723 1724 /* 1725 * clips might have occured while we blocked. 1726 */ 1727 CLIP_CHECK_FWD(entry, save_end); 1728 CLIP_CHECK_BACK(start_entry, start); 1729 continue; 1730 } 1731 /* 1732 * No restart necessary even though clip_end may block, we 1733 * are holding the map lock. 1734 */ 1735 vm_map_clip_end(map, next, end, countp); 1736 next->eflags |= MAP_ENTRY_IN_TRANSITION; 1737 entry = next; 1738 } 1739 if (flags & MAP_CLIP_NO_HOLES) { 1740 if (entry->end != end) { 1741 vm_map_unclip_range(map, start_entry, 1742 start, entry->end, countp, flags); 1743 return(NULL); 1744 } 1745 } 1746 return(start_entry); 1747 } 1748 1749 /* 1750 * Undo the effect of vm_map_clip_range(). You should pass the same 1751 * flags and the same range that you passed to vm_map_clip_range(). 1752 * This code will clear the in-transition flag on the entries and 1753 * wake up anyone waiting. This code will also simplify the sequence 1754 * and attempt to merge it with entries before and after the sequence. 1755 * 1756 * The map must be locked on entry and will remain locked on return. 1757 * 1758 * Note that you should also pass the start_entry returned by 1759 * vm_map_clip_range(). However, if you block between the two calls 1760 * with the map unlocked please be aware that the start_entry may 1761 * have been clipped and you may need to scan it backwards to find 1762 * the entry corresponding with the original start address. You are 1763 * responsible for this, vm_map_unclip_range() expects the correct 1764 * start_entry to be passed to it and will KASSERT otherwise. 1765 */ 1766 static 1767 void 1768 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry, 1769 vm_offset_t start, vm_offset_t end, 1770 int *countp, int flags) 1771 { 1772 vm_map_entry_t entry; 1773 1774 entry = start_entry; 1775 1776 KASSERT(entry->start == start, ("unclip_range: illegal base entry")); 1777 while (entry != &map->header && entry->start < end) { 1778 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, 1779 ("in-transition flag not set during unclip on: %p", 1780 entry)); 1781 KASSERT(entry->end <= end, 1782 ("unclip_range: tail wasn't clipped")); 1783 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 1784 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 1785 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 1786 wakeup(map); 1787 } 1788 entry = entry->next; 1789 } 1790 1791 /* 1792 * Simplification does not block so there is no restart case. 1793 */ 1794 entry = start_entry; 1795 while (entry != &map->header && entry->start < end) { 1796 vm_map_simplify_entry(map, entry, countp); 1797 entry = entry->next; 1798 } 1799 } 1800 1801 /* 1802 * Mark the given range as handled by a subordinate map. 1803 * 1804 * This range must have been created with vm_map_find(), and no other 1805 * operations may have been performed on this range prior to calling 1806 * vm_map_submap(). 1807 * 1808 * Submappings cannot be removed. 1809 * 1810 * No requirements. 1811 */ 1812 int 1813 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 1814 { 1815 vm_map_entry_t entry; 1816 int result = KERN_INVALID_ARGUMENT; 1817 int count; 1818 1819 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1820 vm_map_lock(map); 1821 1822 VM_MAP_RANGE_CHECK(map, start, end); 1823 1824 if (vm_map_lookup_entry(map, start, &entry)) { 1825 vm_map_clip_start(map, entry, start, &count); 1826 } else { 1827 entry = entry->next; 1828 } 1829 1830 vm_map_clip_end(map, entry, end, &count); 1831 1832 if ((entry->start == start) && (entry->end == end) && 1833 ((entry->eflags & MAP_ENTRY_COW) == 0) && 1834 (entry->object.vm_object == NULL)) { 1835 entry->object.sub_map = submap; 1836 entry->maptype = VM_MAPTYPE_SUBMAP; 1837 result = KERN_SUCCESS; 1838 } 1839 vm_map_unlock(map); 1840 vm_map_entry_release(count); 1841 1842 return (result); 1843 } 1844 1845 /* 1846 * Sets the protection of the specified address region in the target map. 1847 * If "set_max" is specified, the maximum protection is to be set; 1848 * otherwise, only the current protection is affected. 1849 * 1850 * The protection is not applicable to submaps, but is applicable to normal 1851 * maps and maps governed by virtual page tables. For example, when operating 1852 * on a virtual page table our protection basically controls how COW occurs 1853 * on the backing object, whereas the virtual page table abstraction itself 1854 * is an abstraction for userland. 1855 * 1856 * No requirements. 1857 */ 1858 int 1859 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 1860 vm_prot_t new_prot, boolean_t set_max) 1861 { 1862 vm_map_entry_t current; 1863 vm_map_entry_t entry; 1864 int count; 1865 1866 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1867 vm_map_lock(map); 1868 1869 VM_MAP_RANGE_CHECK(map, start, end); 1870 1871 if (vm_map_lookup_entry(map, start, &entry)) { 1872 vm_map_clip_start(map, entry, start, &count); 1873 } else { 1874 entry = entry->next; 1875 } 1876 1877 /* 1878 * Make a first pass to check for protection violations. 1879 */ 1880 current = entry; 1881 while ((current != &map->header) && (current->start < end)) { 1882 if (current->maptype == VM_MAPTYPE_SUBMAP) { 1883 vm_map_unlock(map); 1884 vm_map_entry_release(count); 1885 return (KERN_INVALID_ARGUMENT); 1886 } 1887 if ((new_prot & current->max_protection) != new_prot) { 1888 vm_map_unlock(map); 1889 vm_map_entry_release(count); 1890 return (KERN_PROTECTION_FAILURE); 1891 } 1892 current = current->next; 1893 } 1894 1895 /* 1896 * Go back and fix up protections. [Note that clipping is not 1897 * necessary the second time.] 1898 */ 1899 current = entry; 1900 1901 while ((current != &map->header) && (current->start < end)) { 1902 vm_prot_t old_prot; 1903 1904 vm_map_clip_end(map, current, end, &count); 1905 1906 old_prot = current->protection; 1907 if (set_max) { 1908 current->max_protection = new_prot; 1909 current->protection = new_prot & old_prot; 1910 } else { 1911 current->protection = new_prot; 1912 } 1913 1914 /* 1915 * Update physical map if necessary. Worry about copy-on-write 1916 * here -- CHECK THIS XXX 1917 */ 1918 1919 if (current->protection != old_prot) { 1920 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 1921 VM_PROT_ALL) 1922 1923 pmap_protect(map->pmap, current->start, 1924 current->end, 1925 current->protection & MASK(current)); 1926 #undef MASK 1927 } 1928 1929 vm_map_simplify_entry(map, current, &count); 1930 1931 current = current->next; 1932 } 1933 1934 vm_map_unlock(map); 1935 vm_map_entry_release(count); 1936 return (KERN_SUCCESS); 1937 } 1938 1939 /* 1940 * This routine traverses a processes map handling the madvise 1941 * system call. Advisories are classified as either those effecting 1942 * the vm_map_entry structure, or those effecting the underlying 1943 * objects. 1944 * 1945 * The <value> argument is used for extended madvise calls. 1946 * 1947 * No requirements. 1948 */ 1949 int 1950 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, 1951 int behav, off_t value) 1952 { 1953 vm_map_entry_t current, entry; 1954 int modify_map = 0; 1955 int error = 0; 1956 int count; 1957 1958 /* 1959 * Some madvise calls directly modify the vm_map_entry, in which case 1960 * we need to use an exclusive lock on the map and we need to perform 1961 * various clipping operations. Otherwise we only need a read-lock 1962 * on the map. 1963 */ 1964 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1965 1966 switch(behav) { 1967 case MADV_NORMAL: 1968 case MADV_SEQUENTIAL: 1969 case MADV_RANDOM: 1970 case MADV_NOSYNC: 1971 case MADV_AUTOSYNC: 1972 case MADV_NOCORE: 1973 case MADV_CORE: 1974 case MADV_SETMAP: 1975 modify_map = 1; 1976 vm_map_lock(map); 1977 break; 1978 case MADV_INVAL: 1979 case MADV_WILLNEED: 1980 case MADV_DONTNEED: 1981 case MADV_FREE: 1982 vm_map_lock_read(map); 1983 break; 1984 default: 1985 vm_map_entry_release(count); 1986 return (EINVAL); 1987 } 1988 1989 /* 1990 * Locate starting entry and clip if necessary. 1991 */ 1992 1993 VM_MAP_RANGE_CHECK(map, start, end); 1994 1995 if (vm_map_lookup_entry(map, start, &entry)) { 1996 if (modify_map) 1997 vm_map_clip_start(map, entry, start, &count); 1998 } else { 1999 entry = entry->next; 2000 } 2001 2002 if (modify_map) { 2003 /* 2004 * madvise behaviors that are implemented in the vm_map_entry. 2005 * 2006 * We clip the vm_map_entry so that behavioral changes are 2007 * limited to the specified address range. 2008 */ 2009 for (current = entry; 2010 (current != &map->header) && (current->start < end); 2011 current = current->next 2012 ) { 2013 if (current->maptype == VM_MAPTYPE_SUBMAP) 2014 continue; 2015 2016 vm_map_clip_end(map, current, end, &count); 2017 2018 switch (behav) { 2019 case MADV_NORMAL: 2020 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 2021 break; 2022 case MADV_SEQUENTIAL: 2023 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 2024 break; 2025 case MADV_RANDOM: 2026 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 2027 break; 2028 case MADV_NOSYNC: 2029 current->eflags |= MAP_ENTRY_NOSYNC; 2030 break; 2031 case MADV_AUTOSYNC: 2032 current->eflags &= ~MAP_ENTRY_NOSYNC; 2033 break; 2034 case MADV_NOCORE: 2035 current->eflags |= MAP_ENTRY_NOCOREDUMP; 2036 break; 2037 case MADV_CORE: 2038 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 2039 break; 2040 case MADV_SETMAP: 2041 /* 2042 * Set the page directory page for a map 2043 * governed by a virtual page table. Mark 2044 * the entry as being governed by a virtual 2045 * page table if it is not. 2046 * 2047 * XXX the page directory page is stored 2048 * in the avail_ssize field if the map_entry. 2049 * 2050 * XXX the map simplification code does not 2051 * compare this field so weird things may 2052 * happen if you do not apply this function 2053 * to the entire mapping governed by the 2054 * virtual page table. 2055 */ 2056 if (current->maptype != VM_MAPTYPE_VPAGETABLE) { 2057 error = EINVAL; 2058 break; 2059 } 2060 current->aux.master_pde = value; 2061 pmap_remove(map->pmap, 2062 current->start, current->end); 2063 break; 2064 case MADV_INVAL: 2065 /* 2066 * Invalidate the related pmap entries, used 2067 * to flush portions of the real kernel's 2068 * pmap when the caller has removed or 2069 * modified existing mappings in a virtual 2070 * page table. 2071 * 2072 * (exclusive locked map version does not 2073 * need the range interlock). 2074 */ 2075 pmap_remove(map->pmap, 2076 current->start, current->end); 2077 break; 2078 default: 2079 error = EINVAL; 2080 break; 2081 } 2082 vm_map_simplify_entry(map, current, &count); 2083 } 2084 vm_map_unlock(map); 2085 } else { 2086 vm_pindex_t pindex; 2087 vm_pindex_t delta; 2088 2089 /* 2090 * madvise behaviors that are implemented in the underlying 2091 * vm_object. 2092 * 2093 * Since we don't clip the vm_map_entry, we have to clip 2094 * the vm_object pindex and count. 2095 * 2096 * NOTE! These functions are only supported on normal maps, 2097 * except MADV_INVAL which is also supported on 2098 * virtual page tables. 2099 */ 2100 for (current = entry; 2101 (current != &map->header) && (current->start < end); 2102 current = current->next 2103 ) { 2104 vm_offset_t useStart; 2105 2106 if (current->maptype != VM_MAPTYPE_NORMAL && 2107 (current->maptype != VM_MAPTYPE_VPAGETABLE || 2108 behav != MADV_INVAL)) { 2109 continue; 2110 } 2111 2112 pindex = OFF_TO_IDX(current->offset); 2113 delta = atop(current->end - current->start); 2114 useStart = current->start; 2115 2116 if (current->start < start) { 2117 pindex += atop(start - current->start); 2118 delta -= atop(start - current->start); 2119 useStart = start; 2120 } 2121 if (current->end > end) 2122 delta -= atop(current->end - end); 2123 2124 if ((vm_spindex_t)delta <= 0) 2125 continue; 2126 2127 if (behav == MADV_INVAL) { 2128 /* 2129 * Invalidate the related pmap entries, used 2130 * to flush portions of the real kernel's 2131 * pmap when the caller has removed or 2132 * modified existing mappings in a virtual 2133 * page table. 2134 * 2135 * (shared locked map version needs the 2136 * interlock, see vm_fault()). 2137 */ 2138 struct vm_map_ilock ilock; 2139 2140 KASSERT(useStart >= VM_MIN_USER_ADDRESS && 2141 useStart + ptoa(delta) <= 2142 VM_MAX_USER_ADDRESS, 2143 ("Bad range %016jx-%016jx (%016jx)", 2144 useStart, useStart + ptoa(delta), 2145 delta)); 2146 vm_map_interlock(map, &ilock, 2147 useStart, 2148 useStart + ptoa(delta)); 2149 pmap_remove(map->pmap, 2150 useStart, 2151 useStart + ptoa(delta)); 2152 vm_map_deinterlock(map, &ilock); 2153 } else { 2154 vm_object_madvise(current->object.vm_object, 2155 pindex, delta, behav); 2156 } 2157 2158 /* 2159 * Try to populate the page table. Mappings governed 2160 * by virtual page tables cannot be pre-populated 2161 * without a lot of work so don't try. 2162 */ 2163 if (behav == MADV_WILLNEED && 2164 current->maptype != VM_MAPTYPE_VPAGETABLE) { 2165 pmap_object_init_pt( 2166 map->pmap, 2167 useStart, 2168 current->protection, 2169 current->object.vm_object, 2170 pindex, 2171 (count << PAGE_SHIFT), 2172 MAP_PREFAULT_MADVISE 2173 ); 2174 } 2175 } 2176 vm_map_unlock_read(map); 2177 } 2178 vm_map_entry_release(count); 2179 return(error); 2180 } 2181 2182 2183 /* 2184 * Sets the inheritance of the specified address range in the target map. 2185 * Inheritance affects how the map will be shared with child maps at the 2186 * time of vm_map_fork. 2187 */ 2188 int 2189 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 2190 vm_inherit_t new_inheritance) 2191 { 2192 vm_map_entry_t entry; 2193 vm_map_entry_t temp_entry; 2194 int count; 2195 2196 switch (new_inheritance) { 2197 case VM_INHERIT_NONE: 2198 case VM_INHERIT_COPY: 2199 case VM_INHERIT_SHARE: 2200 break; 2201 default: 2202 return (KERN_INVALID_ARGUMENT); 2203 } 2204 2205 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2206 vm_map_lock(map); 2207 2208 VM_MAP_RANGE_CHECK(map, start, end); 2209 2210 if (vm_map_lookup_entry(map, start, &temp_entry)) { 2211 entry = temp_entry; 2212 vm_map_clip_start(map, entry, start, &count); 2213 } else 2214 entry = temp_entry->next; 2215 2216 while ((entry != &map->header) && (entry->start < end)) { 2217 vm_map_clip_end(map, entry, end, &count); 2218 2219 entry->inheritance = new_inheritance; 2220 2221 vm_map_simplify_entry(map, entry, &count); 2222 2223 entry = entry->next; 2224 } 2225 vm_map_unlock(map); 2226 vm_map_entry_release(count); 2227 return (KERN_SUCCESS); 2228 } 2229 2230 /* 2231 * Implement the semantics of mlock 2232 */ 2233 int 2234 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 2235 boolean_t new_pageable) 2236 { 2237 vm_map_entry_t entry; 2238 vm_map_entry_t start_entry; 2239 vm_offset_t end; 2240 int rv = KERN_SUCCESS; 2241 int count; 2242 2243 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2244 vm_map_lock(map); 2245 VM_MAP_RANGE_CHECK(map, start, real_end); 2246 end = real_end; 2247 2248 start_entry = vm_map_clip_range(map, start, end, &count, 2249 MAP_CLIP_NO_HOLES); 2250 if (start_entry == NULL) { 2251 vm_map_unlock(map); 2252 vm_map_entry_release(count); 2253 return (KERN_INVALID_ADDRESS); 2254 } 2255 2256 if (new_pageable == 0) { 2257 entry = start_entry; 2258 while ((entry != &map->header) && (entry->start < end)) { 2259 vm_offset_t save_start; 2260 vm_offset_t save_end; 2261 2262 /* 2263 * Already user wired or hard wired (trivial cases) 2264 */ 2265 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 2266 entry = entry->next; 2267 continue; 2268 } 2269 if (entry->wired_count != 0) { 2270 entry->wired_count++; 2271 entry->eflags |= MAP_ENTRY_USER_WIRED; 2272 entry = entry->next; 2273 continue; 2274 } 2275 2276 /* 2277 * A new wiring requires instantiation of appropriate 2278 * management structures and the faulting in of the 2279 * page. 2280 */ 2281 if (entry->maptype == VM_MAPTYPE_NORMAL || 2282 entry->maptype == VM_MAPTYPE_VPAGETABLE) { 2283 int copyflag = entry->eflags & 2284 MAP_ENTRY_NEEDS_COPY; 2285 if (copyflag && ((entry->protection & 2286 VM_PROT_WRITE) != 0)) { 2287 vm_map_entry_shadow(entry, 0); 2288 } else if (entry->object.vm_object == NULL && 2289 !map->system_map) { 2290 vm_map_entry_allocate_object(entry); 2291 } 2292 } 2293 entry->wired_count++; 2294 entry->eflags |= MAP_ENTRY_USER_WIRED; 2295 2296 /* 2297 * Now fault in the area. Note that vm_fault_wire() 2298 * may release the map lock temporarily, it will be 2299 * relocked on return. The in-transition 2300 * flag protects the entries. 2301 */ 2302 save_start = entry->start; 2303 save_end = entry->end; 2304 rv = vm_fault_wire(map, entry, TRUE, 0); 2305 if (rv) { 2306 CLIP_CHECK_BACK(entry, save_start); 2307 for (;;) { 2308 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 2309 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2310 entry->wired_count = 0; 2311 if (entry->end == save_end) 2312 break; 2313 entry = entry->next; 2314 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2315 } 2316 end = save_start; /* unwire the rest */ 2317 break; 2318 } 2319 /* 2320 * note that even though the entry might have been 2321 * clipped, the USER_WIRED flag we set prevents 2322 * duplication so we do not have to do a 2323 * clip check. 2324 */ 2325 entry = entry->next; 2326 } 2327 2328 /* 2329 * If we failed fall through to the unwiring section to 2330 * unwire what we had wired so far. 'end' has already 2331 * been adjusted. 2332 */ 2333 if (rv) 2334 new_pageable = 1; 2335 2336 /* 2337 * start_entry might have been clipped if we unlocked the 2338 * map and blocked. No matter how clipped it has gotten 2339 * there should be a fragment that is on our start boundary. 2340 */ 2341 CLIP_CHECK_BACK(start_entry, start); 2342 } 2343 2344 /* 2345 * Deal with the unwiring case. 2346 */ 2347 if (new_pageable) { 2348 /* 2349 * This is the unwiring case. We must first ensure that the 2350 * range to be unwired is really wired down. We know there 2351 * are no holes. 2352 */ 2353 entry = start_entry; 2354 while ((entry != &map->header) && (entry->start < end)) { 2355 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 2356 rv = KERN_INVALID_ARGUMENT; 2357 goto done; 2358 } 2359 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry)); 2360 entry = entry->next; 2361 } 2362 2363 /* 2364 * Now decrement the wiring count for each region. If a region 2365 * becomes completely unwired, unwire its physical pages and 2366 * mappings. 2367 */ 2368 /* 2369 * The map entries are processed in a loop, checking to 2370 * make sure the entry is wired and asserting it has a wired 2371 * count. However, another loop was inserted more-or-less in 2372 * the middle of the unwiring path. This loop picks up the 2373 * "entry" loop variable from the first loop without first 2374 * setting it to start_entry. Naturally, the secound loop 2375 * is never entered and the pages backing the entries are 2376 * never unwired. This can lead to a leak of wired pages. 2377 */ 2378 entry = start_entry; 2379 while ((entry != &map->header) && (entry->start < end)) { 2380 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 2381 ("expected USER_WIRED on entry %p", entry)); 2382 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2383 entry->wired_count--; 2384 if (entry->wired_count == 0) 2385 vm_fault_unwire(map, entry); 2386 entry = entry->next; 2387 } 2388 } 2389 done: 2390 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2391 MAP_CLIP_NO_HOLES); 2392 map->timestamp++; 2393 vm_map_unlock(map); 2394 vm_map_entry_release(count); 2395 return (rv); 2396 } 2397 2398 /* 2399 * Sets the pageability of the specified address range in the target map. 2400 * Regions specified as not pageable require locked-down physical 2401 * memory and physical page maps. 2402 * 2403 * The map must not be locked, but a reference must remain to the map 2404 * throughout the call. 2405 * 2406 * This function may be called via the zalloc path and must properly 2407 * reserve map entries for kernel_map. 2408 * 2409 * No requirements. 2410 */ 2411 int 2412 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags) 2413 { 2414 vm_map_entry_t entry; 2415 vm_map_entry_t start_entry; 2416 vm_offset_t end; 2417 int rv = KERN_SUCCESS; 2418 int count; 2419 2420 if (kmflags & KM_KRESERVE) 2421 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 2422 else 2423 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2424 vm_map_lock(map); 2425 VM_MAP_RANGE_CHECK(map, start, real_end); 2426 end = real_end; 2427 2428 start_entry = vm_map_clip_range(map, start, end, &count, 2429 MAP_CLIP_NO_HOLES); 2430 if (start_entry == NULL) { 2431 vm_map_unlock(map); 2432 rv = KERN_INVALID_ADDRESS; 2433 goto failure; 2434 } 2435 if ((kmflags & KM_PAGEABLE) == 0) { 2436 /* 2437 * Wiring. 2438 * 2439 * 1. Holding the write lock, we create any shadow or zero-fill 2440 * objects that need to be created. Then we clip each map 2441 * entry to the region to be wired and increment its wiring 2442 * count. We create objects before clipping the map entries 2443 * to avoid object proliferation. 2444 * 2445 * 2. We downgrade to a read lock, and call vm_fault_wire to 2446 * fault in the pages for any newly wired area (wired_count is 2447 * 1). 2448 * 2449 * Downgrading to a read lock for vm_fault_wire avoids a 2450 * possible deadlock with another process that may have faulted 2451 * on one of the pages to be wired (it would mark the page busy, 2452 * blocking us, then in turn block on the map lock that we 2453 * hold). Because of problems in the recursive lock package, 2454 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 2455 * any actions that require the write lock must be done 2456 * beforehand. Because we keep the read lock on the map, the 2457 * copy-on-write status of the entries we modify here cannot 2458 * change. 2459 */ 2460 entry = start_entry; 2461 while ((entry != &map->header) && (entry->start < end)) { 2462 /* 2463 * Trivial case if the entry is already wired 2464 */ 2465 if (entry->wired_count) { 2466 entry->wired_count++; 2467 entry = entry->next; 2468 continue; 2469 } 2470 2471 /* 2472 * The entry is being newly wired, we have to setup 2473 * appropriate management structures. A shadow 2474 * object is required for a copy-on-write region, 2475 * or a normal object for a zero-fill region. We 2476 * do not have to do this for entries that point to sub 2477 * maps because we won't hold the lock on the sub map. 2478 */ 2479 if (entry->maptype == VM_MAPTYPE_NORMAL || 2480 entry->maptype == VM_MAPTYPE_VPAGETABLE) { 2481 int copyflag = entry->eflags & 2482 MAP_ENTRY_NEEDS_COPY; 2483 if (copyflag && ((entry->protection & 2484 VM_PROT_WRITE) != 0)) { 2485 vm_map_entry_shadow(entry, 0); 2486 } else if (entry->object.vm_object == NULL && 2487 !map->system_map) { 2488 vm_map_entry_allocate_object(entry); 2489 } 2490 } 2491 2492 entry->wired_count++; 2493 entry = entry->next; 2494 } 2495 2496 /* 2497 * Pass 2. 2498 */ 2499 2500 /* 2501 * HACK HACK HACK HACK 2502 * 2503 * vm_fault_wire() temporarily unlocks the map to avoid 2504 * deadlocks. The in-transition flag from vm_map_clip_range 2505 * call should protect us from changes while the map is 2506 * unlocked. T 2507 * 2508 * NOTE: Previously this comment stated that clipping might 2509 * still occur while the entry is unlocked, but from 2510 * what I can tell it actually cannot. 2511 * 2512 * It is unclear whether the CLIP_CHECK_*() calls 2513 * are still needed but we keep them in anyway. 2514 * 2515 * HACK HACK HACK HACK 2516 */ 2517 2518 entry = start_entry; 2519 while (entry != &map->header && entry->start < end) { 2520 /* 2521 * If vm_fault_wire fails for any page we need to undo 2522 * what has been done. We decrement the wiring count 2523 * for those pages which have not yet been wired (now) 2524 * and unwire those that have (later). 2525 */ 2526 vm_offset_t save_start = entry->start; 2527 vm_offset_t save_end = entry->end; 2528 2529 if (entry->wired_count == 1) 2530 rv = vm_fault_wire(map, entry, FALSE, kmflags); 2531 if (rv) { 2532 CLIP_CHECK_BACK(entry, save_start); 2533 for (;;) { 2534 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly")); 2535 entry->wired_count = 0; 2536 if (entry->end == save_end) 2537 break; 2538 entry = entry->next; 2539 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2540 } 2541 end = save_start; 2542 break; 2543 } 2544 CLIP_CHECK_FWD(entry, save_end); 2545 entry = entry->next; 2546 } 2547 2548 /* 2549 * If a failure occured undo everything by falling through 2550 * to the unwiring code. 'end' has already been adjusted 2551 * appropriately. 2552 */ 2553 if (rv) 2554 kmflags |= KM_PAGEABLE; 2555 2556 /* 2557 * start_entry is still IN_TRANSITION but may have been 2558 * clipped since vm_fault_wire() unlocks and relocks the 2559 * map. No matter how clipped it has gotten there should 2560 * be a fragment that is on our start boundary. 2561 */ 2562 CLIP_CHECK_BACK(start_entry, start); 2563 } 2564 2565 if (kmflags & KM_PAGEABLE) { 2566 /* 2567 * This is the unwiring case. We must first ensure that the 2568 * range to be unwired is really wired down. We know there 2569 * are no holes. 2570 */ 2571 entry = start_entry; 2572 while ((entry != &map->header) && (entry->start < end)) { 2573 if (entry->wired_count == 0) { 2574 rv = KERN_INVALID_ARGUMENT; 2575 goto done; 2576 } 2577 entry = entry->next; 2578 } 2579 2580 /* 2581 * Now decrement the wiring count for each region. If a region 2582 * becomes completely unwired, unwire its physical pages and 2583 * mappings. 2584 */ 2585 entry = start_entry; 2586 while ((entry != &map->header) && (entry->start < end)) { 2587 entry->wired_count--; 2588 if (entry->wired_count == 0) 2589 vm_fault_unwire(map, entry); 2590 entry = entry->next; 2591 } 2592 } 2593 done: 2594 vm_map_unclip_range(map, start_entry, start, real_end, 2595 &count, MAP_CLIP_NO_HOLES); 2596 map->timestamp++; 2597 vm_map_unlock(map); 2598 failure: 2599 if (kmflags & KM_KRESERVE) 2600 vm_map_entry_krelease(count); 2601 else 2602 vm_map_entry_release(count); 2603 return (rv); 2604 } 2605 2606 /* 2607 * Mark a newly allocated address range as wired but do not fault in 2608 * the pages. The caller is expected to load the pages into the object. 2609 * 2610 * The map must be locked on entry and will remain locked on return. 2611 * No other requirements. 2612 */ 2613 void 2614 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, 2615 int *countp) 2616 { 2617 vm_map_entry_t scan; 2618 vm_map_entry_t entry; 2619 2620 entry = vm_map_clip_range(map, addr, addr + size, 2621 countp, MAP_CLIP_NO_HOLES); 2622 for (scan = entry; 2623 scan != &map->header && scan->start < addr + size; 2624 scan = scan->next) { 2625 KKASSERT(scan->wired_count == 0); 2626 scan->wired_count = 1; 2627 } 2628 vm_map_unclip_range(map, entry, addr, addr + size, 2629 countp, MAP_CLIP_NO_HOLES); 2630 } 2631 2632 /* 2633 * Push any dirty cached pages in the address range to their pager. 2634 * If syncio is TRUE, dirty pages are written synchronously. 2635 * If invalidate is TRUE, any cached pages are freed as well. 2636 * 2637 * This routine is called by sys_msync() 2638 * 2639 * Returns an error if any part of the specified range is not mapped. 2640 * 2641 * No requirements. 2642 */ 2643 int 2644 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, 2645 boolean_t syncio, boolean_t invalidate) 2646 { 2647 vm_map_entry_t current; 2648 vm_map_entry_t entry; 2649 vm_size_t size; 2650 vm_object_t object; 2651 vm_object_t tobj; 2652 vm_ooffset_t offset; 2653 2654 vm_map_lock_read(map); 2655 VM_MAP_RANGE_CHECK(map, start, end); 2656 if (!vm_map_lookup_entry(map, start, &entry)) { 2657 vm_map_unlock_read(map); 2658 return (KERN_INVALID_ADDRESS); 2659 } 2660 lwkt_gettoken(&map->token); 2661 2662 /* 2663 * Make a first pass to check for holes. 2664 */ 2665 for (current = entry; current->start < end; current = current->next) { 2666 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2667 lwkt_reltoken(&map->token); 2668 vm_map_unlock_read(map); 2669 return (KERN_INVALID_ARGUMENT); 2670 } 2671 if (end > current->end && 2672 (current->next == &map->header || 2673 current->end != current->next->start)) { 2674 lwkt_reltoken(&map->token); 2675 vm_map_unlock_read(map); 2676 return (KERN_INVALID_ADDRESS); 2677 } 2678 } 2679 2680 if (invalidate) 2681 pmap_remove(vm_map_pmap(map), start, end); 2682 2683 /* 2684 * Make a second pass, cleaning/uncaching pages from the indicated 2685 * objects as we go. 2686 */ 2687 for (current = entry; current->start < end; current = current->next) { 2688 offset = current->offset + (start - current->start); 2689 size = (end <= current->end ? end : current->end) - start; 2690 2691 switch(current->maptype) { 2692 case VM_MAPTYPE_SUBMAP: 2693 { 2694 vm_map_t smap; 2695 vm_map_entry_t tentry; 2696 vm_size_t tsize; 2697 2698 smap = current->object.sub_map; 2699 vm_map_lock_read(smap); 2700 vm_map_lookup_entry(smap, offset, &tentry); 2701 tsize = tentry->end - offset; 2702 if (tsize < size) 2703 size = tsize; 2704 object = tentry->object.vm_object; 2705 offset = tentry->offset + (offset - tentry->start); 2706 vm_map_unlock_read(smap); 2707 break; 2708 } 2709 case VM_MAPTYPE_NORMAL: 2710 case VM_MAPTYPE_VPAGETABLE: 2711 object = current->object.vm_object; 2712 break; 2713 default: 2714 object = NULL; 2715 break; 2716 } 2717 2718 if (object) 2719 vm_object_hold(object); 2720 2721 /* 2722 * Note that there is absolutely no sense in writing out 2723 * anonymous objects, so we track down the vnode object 2724 * to write out. 2725 * We invalidate (remove) all pages from the address space 2726 * anyway, for semantic correctness. 2727 * 2728 * note: certain anonymous maps, such as MAP_NOSYNC maps, 2729 * may start out with a NULL object. 2730 */ 2731 while (object && (tobj = object->backing_object) != NULL) { 2732 vm_object_hold(tobj); 2733 if (tobj == object->backing_object) { 2734 vm_object_lock_swap(); 2735 offset += object->backing_object_offset; 2736 vm_object_drop(object); 2737 object = tobj; 2738 if (object->size < OFF_TO_IDX(offset + size)) 2739 size = IDX_TO_OFF(object->size) - 2740 offset; 2741 break; 2742 } 2743 vm_object_drop(tobj); 2744 } 2745 if (object && (object->type == OBJT_VNODE) && 2746 (current->protection & VM_PROT_WRITE) && 2747 (object->flags & OBJ_NOMSYNC) == 0) { 2748 /* 2749 * Flush pages if writing is allowed, invalidate them 2750 * if invalidation requested. Pages undergoing I/O 2751 * will be ignored by vm_object_page_remove(). 2752 * 2753 * We cannot lock the vnode and then wait for paging 2754 * to complete without deadlocking against vm_fault. 2755 * Instead we simply call vm_object_page_remove() and 2756 * allow it to block internally on a page-by-page 2757 * basis when it encounters pages undergoing async 2758 * I/O. 2759 */ 2760 int flags; 2761 2762 /* no chain wait needed for vnode objects */ 2763 vm_object_reference_locked(object); 2764 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY); 2765 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 2766 flags |= invalidate ? OBJPC_INVAL : 0; 2767 2768 /* 2769 * When operating on a virtual page table just 2770 * flush the whole object. XXX we probably ought 2771 * to 2772 */ 2773 switch(current->maptype) { 2774 case VM_MAPTYPE_NORMAL: 2775 vm_object_page_clean(object, 2776 OFF_TO_IDX(offset), 2777 OFF_TO_IDX(offset + size + PAGE_MASK), 2778 flags); 2779 break; 2780 case VM_MAPTYPE_VPAGETABLE: 2781 vm_object_page_clean(object, 0, 0, flags); 2782 break; 2783 } 2784 vn_unlock(((struct vnode *)object->handle)); 2785 vm_object_deallocate_locked(object); 2786 } 2787 if (object && invalidate && 2788 ((object->type == OBJT_VNODE) || 2789 (object->type == OBJT_DEVICE) || 2790 (object->type == OBJT_MGTDEVICE))) { 2791 int clean_only = 2792 ((object->type == OBJT_DEVICE) || 2793 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE; 2794 /* no chain wait needed for vnode/device objects */ 2795 vm_object_reference_locked(object); 2796 switch(current->maptype) { 2797 case VM_MAPTYPE_NORMAL: 2798 vm_object_page_remove(object, 2799 OFF_TO_IDX(offset), 2800 OFF_TO_IDX(offset + size + PAGE_MASK), 2801 clean_only); 2802 break; 2803 case VM_MAPTYPE_VPAGETABLE: 2804 vm_object_page_remove(object, 0, 0, clean_only); 2805 break; 2806 } 2807 vm_object_deallocate_locked(object); 2808 } 2809 start += size; 2810 if (object) 2811 vm_object_drop(object); 2812 } 2813 2814 lwkt_reltoken(&map->token); 2815 vm_map_unlock_read(map); 2816 2817 return (KERN_SUCCESS); 2818 } 2819 2820 /* 2821 * Make the region specified by this entry pageable. 2822 * 2823 * The vm_map must be exclusively locked. 2824 */ 2825 static void 2826 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 2827 { 2828 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2829 entry->wired_count = 0; 2830 vm_fault_unwire(map, entry); 2831 } 2832 2833 /* 2834 * Deallocate the given entry from the target map. 2835 * 2836 * The vm_map must be exclusively locked. 2837 */ 2838 static void 2839 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 2840 { 2841 vm_map_entry_unlink(map, entry); 2842 map->size -= entry->end - entry->start; 2843 2844 switch(entry->maptype) { 2845 case VM_MAPTYPE_NORMAL: 2846 case VM_MAPTYPE_VPAGETABLE: 2847 case VM_MAPTYPE_SUBMAP: 2848 vm_object_deallocate(entry->object.vm_object); 2849 break; 2850 case VM_MAPTYPE_UKSMAP: 2851 /* XXX TODO */ 2852 break; 2853 default: 2854 break; 2855 } 2856 2857 vm_map_entry_dispose(map, entry, countp); 2858 } 2859 2860 /* 2861 * Deallocates the given address range from the target map. 2862 * 2863 * The vm_map must be exclusively locked. 2864 */ 2865 int 2866 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 2867 { 2868 vm_object_t object; 2869 vm_map_entry_t entry; 2870 vm_map_entry_t first_entry; 2871 2872 ASSERT_VM_MAP_LOCKED(map); 2873 lwkt_gettoken(&map->token); 2874 again: 2875 /* 2876 * Find the start of the region, and clip it. Set entry to point 2877 * at the first record containing the requested address or, if no 2878 * such record exists, the next record with a greater address. The 2879 * loop will run from this point until a record beyond the termination 2880 * address is encountered. 2881 * 2882 * map->hint must be adjusted to not point to anything we delete, 2883 * so set it to the entry prior to the one being deleted. 2884 * 2885 * GGG see other GGG comment. 2886 */ 2887 if (vm_map_lookup_entry(map, start, &first_entry)) { 2888 entry = first_entry; 2889 vm_map_clip_start(map, entry, start, countp); 2890 map->hint = entry->prev; /* possible problem XXX */ 2891 } else { 2892 map->hint = first_entry; /* possible problem XXX */ 2893 entry = first_entry->next; 2894 } 2895 2896 /* 2897 * If a hole opens up prior to the current first_free then 2898 * adjust first_free. As with map->hint, map->first_free 2899 * cannot be left set to anything we might delete. 2900 */ 2901 if (entry == &map->header) { 2902 map->first_free = &map->header; 2903 } else if (map->first_free->start >= start) { 2904 map->first_free = entry->prev; 2905 } 2906 2907 /* 2908 * Step through all entries in this region 2909 */ 2910 while ((entry != &map->header) && (entry->start < end)) { 2911 vm_map_entry_t next; 2912 vm_offset_t s, e; 2913 vm_pindex_t offidxstart, offidxend, count; 2914 2915 /* 2916 * If we hit an in-transition entry we have to sleep and 2917 * retry. It's easier (and not really slower) to just retry 2918 * since this case occurs so rarely and the hint is already 2919 * pointing at the right place. We have to reset the 2920 * start offset so as not to accidently delete an entry 2921 * another process just created in vacated space. 2922 */ 2923 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 2924 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2925 start = entry->start; 2926 ++mycpu->gd_cnt.v_intrans_coll; 2927 ++mycpu->gd_cnt.v_intrans_wait; 2928 vm_map_transition_wait(map); 2929 goto again; 2930 } 2931 vm_map_clip_end(map, entry, end, countp); 2932 2933 s = entry->start; 2934 e = entry->end; 2935 next = entry->next; 2936 2937 offidxstart = OFF_TO_IDX(entry->offset); 2938 count = OFF_TO_IDX(e - s); 2939 2940 switch(entry->maptype) { 2941 case VM_MAPTYPE_NORMAL: 2942 case VM_MAPTYPE_VPAGETABLE: 2943 case VM_MAPTYPE_SUBMAP: 2944 object = entry->object.vm_object; 2945 break; 2946 default: 2947 object = NULL; 2948 break; 2949 } 2950 2951 /* 2952 * Unwire before removing addresses from the pmap; otherwise, 2953 * unwiring will put the entries back in the pmap. 2954 */ 2955 if (entry->wired_count != 0) 2956 vm_map_entry_unwire(map, entry); 2957 2958 offidxend = offidxstart + count; 2959 2960 if (object == &kernel_object) { 2961 vm_object_hold(object); 2962 vm_object_page_remove(object, offidxstart, 2963 offidxend, FALSE); 2964 vm_object_drop(object); 2965 } else if (object && object->type != OBJT_DEFAULT && 2966 object->type != OBJT_SWAP) { 2967 /* 2968 * vnode object routines cannot be chain-locked, 2969 * but since we aren't removing pages from the 2970 * object here we can use a shared hold. 2971 */ 2972 vm_object_hold_shared(object); 2973 pmap_remove(map->pmap, s, e); 2974 vm_object_drop(object); 2975 } else if (object) { 2976 vm_object_hold(object); 2977 vm_object_chain_acquire(object, 0); 2978 pmap_remove(map->pmap, s, e); 2979 2980 if (object != NULL && 2981 object->ref_count != 1 && 2982 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == 2983 OBJ_ONEMAPPING && 2984 (object->type == OBJT_DEFAULT || 2985 object->type == OBJT_SWAP)) { 2986 vm_object_collapse(object, NULL); 2987 vm_object_page_remove(object, offidxstart, 2988 offidxend, FALSE); 2989 if (object->type == OBJT_SWAP) { 2990 swap_pager_freespace(object, 2991 offidxstart, 2992 count); 2993 } 2994 if (offidxend >= object->size && 2995 offidxstart < object->size) { 2996 object->size = offidxstart; 2997 } 2998 } 2999 vm_object_chain_release(object); 3000 vm_object_drop(object); 3001 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) { 3002 pmap_remove(map->pmap, s, e); 3003 } 3004 3005 /* 3006 * Delete the entry (which may delete the object) only after 3007 * removing all pmap entries pointing to its pages. 3008 * (Otherwise, its page frames may be reallocated, and any 3009 * modify bits will be set in the wrong object!) 3010 */ 3011 vm_map_entry_delete(map, entry, countp); 3012 entry = next; 3013 } 3014 lwkt_reltoken(&map->token); 3015 return (KERN_SUCCESS); 3016 } 3017 3018 /* 3019 * Remove the given address range from the target map. 3020 * This is the exported form of vm_map_delete. 3021 * 3022 * No requirements. 3023 */ 3024 int 3025 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 3026 { 3027 int result; 3028 int count; 3029 3030 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3031 vm_map_lock(map); 3032 VM_MAP_RANGE_CHECK(map, start, end); 3033 result = vm_map_delete(map, start, end, &count); 3034 vm_map_unlock(map); 3035 vm_map_entry_release(count); 3036 3037 return (result); 3038 } 3039 3040 /* 3041 * Assert that the target map allows the specified privilege on the 3042 * entire address region given. The entire region must be allocated. 3043 * 3044 * The caller must specify whether the vm_map is already locked or not. 3045 */ 3046 boolean_t 3047 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 3048 vm_prot_t protection, boolean_t have_lock) 3049 { 3050 vm_map_entry_t entry; 3051 vm_map_entry_t tmp_entry; 3052 boolean_t result; 3053 3054 if (have_lock == FALSE) 3055 vm_map_lock_read(map); 3056 3057 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 3058 if (have_lock == FALSE) 3059 vm_map_unlock_read(map); 3060 return (FALSE); 3061 } 3062 entry = tmp_entry; 3063 3064 result = TRUE; 3065 while (start < end) { 3066 if (entry == &map->header) { 3067 result = FALSE; 3068 break; 3069 } 3070 /* 3071 * No holes allowed! 3072 */ 3073 3074 if (start < entry->start) { 3075 result = FALSE; 3076 break; 3077 } 3078 /* 3079 * Check protection associated with entry. 3080 */ 3081 3082 if ((entry->protection & protection) != protection) { 3083 result = FALSE; 3084 break; 3085 } 3086 /* go to next entry */ 3087 3088 start = entry->end; 3089 entry = entry->next; 3090 } 3091 if (have_lock == FALSE) 3092 vm_map_unlock_read(map); 3093 return (result); 3094 } 3095 3096 /* 3097 * If appropriate this function shadows the original object with a new object 3098 * and moves the VM pages from the original object to the new object. 3099 * The original object will also be collapsed, if possible. 3100 * 3101 * We can only do this for normal memory objects with a single mapping, and 3102 * it only makes sense to do it if there are 2 or more refs on the original 3103 * object. i.e. typically a memory object that has been extended into 3104 * multiple vm_map_entry's with non-overlapping ranges. 3105 * 3106 * This makes it easier to remove unused pages and keeps object inheritance 3107 * from being a negative impact on memory usage. 3108 * 3109 * On return the (possibly new) entry->object.vm_object will have an 3110 * additional ref on it for the caller to dispose of (usually by cloning 3111 * the vm_map_entry). The additional ref had to be done in this routine 3112 * to avoid racing a collapse. The object's ONEMAPPING flag will also be 3113 * cleared. 3114 * 3115 * The vm_map must be locked and its token held. 3116 */ 3117 static void 3118 vm_map_split(vm_map_entry_t entry) 3119 { 3120 /* OPTIMIZED */ 3121 vm_object_t oobject, nobject, bobject; 3122 vm_offset_t s, e; 3123 vm_page_t m; 3124 vm_pindex_t offidxstart, offidxend, idx; 3125 vm_size_t size; 3126 vm_ooffset_t offset; 3127 int useshadowlist; 3128 3129 /* 3130 * Optimize away object locks for vnode objects. Important exit/exec 3131 * critical path. 3132 * 3133 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag 3134 * anyway. 3135 */ 3136 oobject = entry->object.vm_object; 3137 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) { 3138 vm_object_reference_quick(oobject); 3139 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3140 return; 3141 } 3142 3143 /* 3144 * Setup. Chain lock the original object throughout the entire 3145 * routine to prevent new page faults from occuring. 3146 * 3147 * XXX can madvise WILLNEED interfere with us too? 3148 */ 3149 vm_object_hold(oobject); 3150 vm_object_chain_acquire(oobject, 0); 3151 3152 /* 3153 * Original object cannot be split? Might have also changed state. 3154 */ 3155 if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT && 3156 oobject->type != OBJT_SWAP)) { 3157 vm_object_chain_release(oobject); 3158 vm_object_reference_locked(oobject); 3159 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3160 vm_object_drop(oobject); 3161 return; 3162 } 3163 3164 /* 3165 * Collapse original object with its backing store as an 3166 * optimization to reduce chain lengths when possible. 3167 * 3168 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's 3169 * for oobject, so there's no point collapsing it. 3170 * 3171 * Then re-check whether the object can be split. 3172 */ 3173 vm_object_collapse(oobject, NULL); 3174 3175 if (oobject->ref_count <= 1 || 3176 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) || 3177 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) { 3178 vm_object_chain_release(oobject); 3179 vm_object_reference_locked(oobject); 3180 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3181 vm_object_drop(oobject); 3182 return; 3183 } 3184 3185 /* 3186 * Acquire the chain lock on the backing object. 3187 * 3188 * Give bobject an additional ref count for when it will be shadowed 3189 * by nobject. 3190 */ 3191 useshadowlist = 0; 3192 if ((bobject = oobject->backing_object) != NULL) { 3193 if (bobject->type != OBJT_VNODE) { 3194 useshadowlist = 1; 3195 vm_object_hold(bobject); 3196 vm_object_chain_wait(bobject, 0); 3197 /* ref for shadowing below */ 3198 vm_object_reference_locked(bobject); 3199 vm_object_chain_acquire(bobject, 0); 3200 KKASSERT(bobject->backing_object == bobject); 3201 KKASSERT((bobject->flags & OBJ_DEAD) == 0); 3202 } else { 3203 /* 3204 * vnodes are not placed on the shadow list but 3205 * they still get another ref for the backing_object 3206 * reference. 3207 */ 3208 vm_object_reference_quick(bobject); 3209 } 3210 } 3211 3212 /* 3213 * Calculate the object page range and allocate the new object. 3214 */ 3215 offset = entry->offset; 3216 s = entry->start; 3217 e = entry->end; 3218 3219 offidxstart = OFF_TO_IDX(offset); 3220 offidxend = offidxstart + OFF_TO_IDX(e - s); 3221 size = offidxend - offidxstart; 3222 3223 switch(oobject->type) { 3224 case OBJT_DEFAULT: 3225 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size), 3226 VM_PROT_ALL, 0); 3227 break; 3228 case OBJT_SWAP: 3229 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size), 3230 VM_PROT_ALL, 0); 3231 break; 3232 default: 3233 /* not reached */ 3234 nobject = NULL; 3235 KKASSERT(0); 3236 } 3237 3238 if (nobject == NULL) { 3239 if (bobject) { 3240 if (useshadowlist) { 3241 vm_object_chain_release(bobject); 3242 vm_object_deallocate(bobject); 3243 vm_object_drop(bobject); 3244 } else { 3245 vm_object_deallocate(bobject); 3246 } 3247 } 3248 vm_object_chain_release(oobject); 3249 vm_object_reference_locked(oobject); 3250 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3251 vm_object_drop(oobject); 3252 return; 3253 } 3254 3255 /* 3256 * The new object will replace entry->object.vm_object so it needs 3257 * a second reference (the caller expects an additional ref). 3258 */ 3259 vm_object_hold(nobject); 3260 vm_object_reference_locked(nobject); 3261 vm_object_chain_acquire(nobject, 0); 3262 3263 /* 3264 * nobject shadows bobject (oobject already shadows bobject). 3265 * 3266 * Adding an object to bobject's shadow list requires refing bobject 3267 * which we did above in the useshadowlist case. 3268 */ 3269 if (bobject) { 3270 nobject->backing_object_offset = 3271 oobject->backing_object_offset + IDX_TO_OFF(offidxstart); 3272 nobject->backing_object = bobject; 3273 if (useshadowlist) { 3274 bobject->shadow_count++; 3275 atomic_add_int(&bobject->generation, 1); 3276 LIST_INSERT_HEAD(&bobject->shadow_head, 3277 nobject, shadow_list); 3278 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/ 3279 vm_object_chain_release(bobject); 3280 vm_object_drop(bobject); 3281 vm_object_set_flag(nobject, OBJ_ONSHADOW); 3282 } 3283 } 3284 3285 /* 3286 * Move the VM pages from oobject to nobject 3287 */ 3288 for (idx = 0; idx < size; idx++) { 3289 vm_page_t m; 3290 3291 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx, 3292 TRUE, "vmpg"); 3293 if (m == NULL) 3294 continue; 3295 3296 /* 3297 * We must wait for pending I/O to complete before we can 3298 * rename the page. 3299 * 3300 * We do not have to VM_PROT_NONE the page as mappings should 3301 * not be changed by this operation. 3302 * 3303 * NOTE: The act of renaming a page updates chaingen for both 3304 * objects. 3305 */ 3306 vm_page_rename(m, nobject, idx); 3307 /* page automatically made dirty by rename and cache handled */ 3308 /* page remains busy */ 3309 } 3310 3311 if (oobject->type == OBJT_SWAP) { 3312 vm_object_pip_add(oobject, 1); 3313 /* 3314 * copy oobject pages into nobject and destroy unneeded 3315 * pages in shadow object. 3316 */ 3317 swap_pager_copy(oobject, nobject, offidxstart, 0); 3318 vm_object_pip_wakeup(oobject); 3319 } 3320 3321 /* 3322 * Wakeup the pages we played with. No spl protection is needed 3323 * for a simple wakeup. 3324 */ 3325 for (idx = 0; idx < size; idx++) { 3326 m = vm_page_lookup(nobject, idx); 3327 if (m) { 3328 KKASSERT(m->flags & PG_BUSY); 3329 vm_page_wakeup(m); 3330 } 3331 } 3332 entry->object.vm_object = nobject; 3333 entry->offset = 0LL; 3334 3335 /* 3336 * Cleanup 3337 * 3338 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the 3339 * related pages were moved and are no longer applicable to the 3340 * original object. 3341 * 3342 * NOTE: Deallocate oobject (due to its entry->object.vm_object being 3343 * replaced by nobject). 3344 */ 3345 vm_object_chain_release(nobject); 3346 vm_object_drop(nobject); 3347 if (bobject && useshadowlist) { 3348 vm_object_chain_release(bobject); 3349 vm_object_drop(bobject); 3350 } 3351 vm_object_chain_release(oobject); 3352 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/ 3353 vm_object_deallocate_locked(oobject); 3354 vm_object_drop(oobject); 3355 } 3356 3357 /* 3358 * Copies the contents of the source entry to the destination 3359 * entry. The entries *must* be aligned properly. 3360 * 3361 * The vm_maps must be exclusively locked. 3362 * The vm_map's token must be held. 3363 * 3364 * Because the maps are locked no faults can be in progress during the 3365 * operation. 3366 */ 3367 static void 3368 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 3369 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 3370 { 3371 vm_object_t src_object; 3372 3373 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP || 3374 dst_entry->maptype == VM_MAPTYPE_UKSMAP) 3375 return; 3376 if (src_entry->maptype == VM_MAPTYPE_SUBMAP || 3377 src_entry->maptype == VM_MAPTYPE_UKSMAP) 3378 return; 3379 3380 if (src_entry->wired_count == 0) { 3381 /* 3382 * If the source entry is marked needs_copy, it is already 3383 * write-protected. 3384 */ 3385 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 3386 pmap_protect(src_map->pmap, 3387 src_entry->start, 3388 src_entry->end, 3389 src_entry->protection & ~VM_PROT_WRITE); 3390 } 3391 3392 /* 3393 * Make a copy of the object. 3394 * 3395 * The object must be locked prior to checking the object type 3396 * and for the call to vm_object_collapse() and vm_map_split(). 3397 * We cannot use *_hold() here because the split code will 3398 * probably try to destroy the object. The lock is a pool 3399 * token and doesn't care. 3400 * 3401 * We must bump src_map->timestamp when setting 3402 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault 3403 * to retry, otherwise the concurrent fault might improperly 3404 * install a RW pte when its supposed to be a RO(COW) pte. 3405 * This race can occur because a vnode-backed fault may have 3406 * to temporarily release the map lock. 3407 */ 3408 if (src_entry->object.vm_object != NULL) { 3409 vm_map_split(src_entry); 3410 src_object = src_entry->object.vm_object; 3411 dst_entry->object.vm_object = src_object; 3412 src_entry->eflags |= (MAP_ENTRY_COW | 3413 MAP_ENTRY_NEEDS_COPY); 3414 dst_entry->eflags |= (MAP_ENTRY_COW | 3415 MAP_ENTRY_NEEDS_COPY); 3416 dst_entry->offset = src_entry->offset; 3417 ++src_map->timestamp; 3418 } else { 3419 dst_entry->object.vm_object = NULL; 3420 dst_entry->offset = 0; 3421 } 3422 3423 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, 3424 dst_entry->end - dst_entry->start, src_entry->start); 3425 } else { 3426 /* 3427 * Of course, wired down pages can't be set copy-on-write. 3428 * Cause wired pages to be copied into the new map by 3429 * simulating faults (the new pages are pageable) 3430 */ 3431 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 3432 } 3433 } 3434 3435 /* 3436 * vmspace_fork: 3437 * Create a new process vmspace structure and vm_map 3438 * based on those of an existing process. The new map 3439 * is based on the old map, according to the inheritance 3440 * values on the regions in that map. 3441 * 3442 * The source map must not be locked. 3443 * No requirements. 3444 */ 3445 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map, 3446 vm_map_entry_t old_entry, int *countp); 3447 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map, 3448 vm_map_entry_t old_entry, int *countp); 3449 3450 struct vmspace * 3451 vmspace_fork(struct vmspace *vm1) 3452 { 3453 struct vmspace *vm2; 3454 vm_map_t old_map = &vm1->vm_map; 3455 vm_map_t new_map; 3456 vm_map_entry_t old_entry; 3457 int count; 3458 3459 lwkt_gettoken(&vm1->vm_map.token); 3460 vm_map_lock(old_map); 3461 3462 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset); 3463 lwkt_gettoken(&vm2->vm_map.token); 3464 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 3465 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 3466 new_map = &vm2->vm_map; /* XXX */ 3467 new_map->timestamp = 1; 3468 3469 vm_map_lock(new_map); 3470 3471 count = 0; 3472 old_entry = old_map->header.next; 3473 while (old_entry != &old_map->header) { 3474 ++count; 3475 old_entry = old_entry->next; 3476 } 3477 3478 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 3479 3480 old_entry = old_map->header.next; 3481 while (old_entry != &old_map->header) { 3482 switch(old_entry->maptype) { 3483 case VM_MAPTYPE_SUBMAP: 3484 panic("vm_map_fork: encountered a submap"); 3485 break; 3486 case VM_MAPTYPE_UKSMAP: 3487 vmspace_fork_uksmap_entry(old_map, new_map, 3488 old_entry, &count); 3489 break; 3490 case VM_MAPTYPE_NORMAL: 3491 case VM_MAPTYPE_VPAGETABLE: 3492 vmspace_fork_normal_entry(old_map, new_map, 3493 old_entry, &count); 3494 break; 3495 } 3496 old_entry = old_entry->next; 3497 } 3498 3499 new_map->size = old_map->size; 3500 vm_map_unlock(old_map); 3501 vm_map_unlock(new_map); 3502 vm_map_entry_release(count); 3503 3504 lwkt_reltoken(&vm2->vm_map.token); 3505 lwkt_reltoken(&vm1->vm_map.token); 3506 3507 return (vm2); 3508 } 3509 3510 static 3511 void 3512 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map, 3513 vm_map_entry_t old_entry, int *countp) 3514 { 3515 vm_map_entry_t new_entry; 3516 vm_object_t object; 3517 3518 switch (old_entry->inheritance) { 3519 case VM_INHERIT_NONE: 3520 break; 3521 case VM_INHERIT_SHARE: 3522 /* 3523 * Clone the entry, creating the shared object if 3524 * necessary. 3525 */ 3526 if (old_entry->object.vm_object == NULL) 3527 vm_map_entry_allocate_object(old_entry); 3528 3529 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3530 /* 3531 * Shadow a map_entry which needs a copy, 3532 * replacing its object with a new object 3533 * that points to the old one. Ask the 3534 * shadow code to automatically add an 3535 * additional ref. We can't do it afterwords 3536 * because we might race a collapse. The call 3537 * to vm_map_entry_shadow() will also clear 3538 * OBJ_ONEMAPPING. 3539 */ 3540 vm_map_entry_shadow(old_entry, 1); 3541 } else if (old_entry->object.vm_object) { 3542 /* 3543 * We will make a shared copy of the object, 3544 * and must clear OBJ_ONEMAPPING. 3545 * 3546 * Optimize vnode objects. OBJ_ONEMAPPING 3547 * is non-applicable but clear it anyway, 3548 * and its terminal so we don'th ave to deal 3549 * with chains. Reduces SMP conflicts. 3550 * 3551 * XXX assert that object.vm_object != NULL 3552 * since we allocate it above. 3553 */ 3554 object = old_entry->object.vm_object; 3555 if (object->type == OBJT_VNODE) { 3556 vm_object_reference_quick(object); 3557 vm_object_clear_flag(object, 3558 OBJ_ONEMAPPING); 3559 } else { 3560 vm_object_hold(object); 3561 vm_object_chain_wait(object, 0); 3562 vm_object_reference_locked(object); 3563 vm_object_clear_flag(object, 3564 OBJ_ONEMAPPING); 3565 vm_object_drop(object); 3566 } 3567 } 3568 3569 /* 3570 * Clone the entry. We've already bumped the ref on 3571 * any vm_object. 3572 */ 3573 new_entry = vm_map_entry_create(new_map, countp); 3574 *new_entry = *old_entry; 3575 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3576 new_entry->wired_count = 0; 3577 3578 /* 3579 * Insert the entry into the new map -- we know we're 3580 * inserting at the end of the new map. 3581 */ 3582 3583 vm_map_entry_link(new_map, new_map->header.prev, 3584 new_entry); 3585 3586 /* 3587 * Update the physical map 3588 */ 3589 pmap_copy(new_map->pmap, old_map->pmap, 3590 new_entry->start, 3591 (old_entry->end - old_entry->start), 3592 old_entry->start); 3593 break; 3594 case VM_INHERIT_COPY: 3595 /* 3596 * Clone the entry and link into the map. 3597 */ 3598 new_entry = vm_map_entry_create(new_map, countp); 3599 *new_entry = *old_entry; 3600 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3601 new_entry->wired_count = 0; 3602 new_entry->object.vm_object = NULL; 3603 vm_map_entry_link(new_map, new_map->header.prev, 3604 new_entry); 3605 vm_map_copy_entry(old_map, new_map, old_entry, 3606 new_entry); 3607 break; 3608 } 3609 } 3610 3611 /* 3612 * When forking user-kernel shared maps, the map might change in the 3613 * child so do not try to copy the underlying pmap entries. 3614 */ 3615 static 3616 void 3617 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map, 3618 vm_map_entry_t old_entry, int *countp) 3619 { 3620 vm_map_entry_t new_entry; 3621 3622 new_entry = vm_map_entry_create(new_map, countp); 3623 *new_entry = *old_entry; 3624 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3625 new_entry->wired_count = 0; 3626 vm_map_entry_link(new_map, new_map->header.prev, 3627 new_entry); 3628 } 3629 3630 /* 3631 * Create an auto-grow stack entry 3632 * 3633 * No requirements. 3634 */ 3635 int 3636 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 3637 int flags, vm_prot_t prot, vm_prot_t max, int cow) 3638 { 3639 vm_map_entry_t prev_entry; 3640 vm_map_entry_t new_stack_entry; 3641 vm_size_t init_ssize; 3642 int rv; 3643 int count; 3644 vm_offset_t tmpaddr; 3645 3646 cow |= MAP_IS_STACK; 3647 3648 if (max_ssize < sgrowsiz) 3649 init_ssize = max_ssize; 3650 else 3651 init_ssize = sgrowsiz; 3652 3653 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3654 vm_map_lock(map); 3655 3656 /* 3657 * Find space for the mapping 3658 */ 3659 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) { 3660 if (vm_map_findspace(map, addrbos, max_ssize, 1, 3661 flags, &tmpaddr)) { 3662 vm_map_unlock(map); 3663 vm_map_entry_release(count); 3664 return (KERN_NO_SPACE); 3665 } 3666 addrbos = tmpaddr; 3667 } 3668 3669 /* If addr is already mapped, no go */ 3670 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) { 3671 vm_map_unlock(map); 3672 vm_map_entry_release(count); 3673 return (KERN_NO_SPACE); 3674 } 3675 3676 #if 0 3677 /* XXX already handled by kern_mmap() */ 3678 /* If we would blow our VMEM resource limit, no go */ 3679 if (map->size + init_ssize > 3680 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3681 vm_map_unlock(map); 3682 vm_map_entry_release(count); 3683 return (KERN_NO_SPACE); 3684 } 3685 #endif 3686 3687 /* 3688 * If we can't accomodate max_ssize in the current mapping, 3689 * no go. However, we need to be aware that subsequent user 3690 * mappings might map into the space we have reserved for 3691 * stack, and currently this space is not protected. 3692 * 3693 * Hopefully we will at least detect this condition 3694 * when we try to grow the stack. 3695 */ 3696 if ((prev_entry->next != &map->header) && 3697 (prev_entry->next->start < addrbos + max_ssize)) { 3698 vm_map_unlock(map); 3699 vm_map_entry_release(count); 3700 return (KERN_NO_SPACE); 3701 } 3702 3703 /* 3704 * We initially map a stack of only init_ssize. We will 3705 * grow as needed later. Since this is to be a grow 3706 * down stack, we map at the top of the range. 3707 * 3708 * Note: we would normally expect prot and max to be 3709 * VM_PROT_ALL, and cow to be 0. Possibly we should 3710 * eliminate these as input parameters, and just 3711 * pass these values here in the insert call. 3712 */ 3713 rv = vm_map_insert(map, &count, NULL, NULL, 3714 0, addrbos + max_ssize - init_ssize, 3715 addrbos + max_ssize, 3716 VM_MAPTYPE_NORMAL, 3717 VM_SUBSYS_STACK, prot, max, cow); 3718 3719 /* Now set the avail_ssize amount */ 3720 if (rv == KERN_SUCCESS) { 3721 if (prev_entry != &map->header) 3722 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count); 3723 new_stack_entry = prev_entry->next; 3724 if (new_stack_entry->end != addrbos + max_ssize || 3725 new_stack_entry->start != addrbos + max_ssize - init_ssize) 3726 panic ("Bad entry start/end for new stack entry"); 3727 else 3728 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize; 3729 } 3730 3731 vm_map_unlock(map); 3732 vm_map_entry_release(count); 3733 return (rv); 3734 } 3735 3736 /* 3737 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 3738 * desired address is already mapped, or if we successfully grow 3739 * the stack. Also returns KERN_SUCCESS if addr is outside the 3740 * stack range (this is strange, but preserves compatibility with 3741 * the grow function in vm_machdep.c). 3742 * 3743 * No requirements. 3744 */ 3745 int 3746 vm_map_growstack (vm_map_t map, vm_offset_t addr) 3747 { 3748 vm_map_entry_t prev_entry; 3749 vm_map_entry_t stack_entry; 3750 vm_map_entry_t new_stack_entry; 3751 struct vmspace *vm; 3752 struct lwp *lp; 3753 struct proc *p; 3754 vm_offset_t end; 3755 int grow_amount; 3756 int rv = KERN_SUCCESS; 3757 int is_procstack; 3758 int use_read_lock = 1; 3759 int count; 3760 3761 /* 3762 * Find the vm 3763 */ 3764 lp = curthread->td_lwp; 3765 p = curthread->td_proc; 3766 KKASSERT(lp != NULL); 3767 vm = lp->lwp_vmspace; 3768 3769 /* 3770 * Growstack is only allowed on the current process. We disallow 3771 * other use cases, e.g. trying to access memory via procfs that 3772 * the stack hasn't grown into. 3773 */ 3774 if (map != &vm->vm_map) { 3775 return KERN_FAILURE; 3776 } 3777 3778 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3779 Retry: 3780 if (use_read_lock) 3781 vm_map_lock_read(map); 3782 else 3783 vm_map_lock(map); 3784 3785 /* If addr is already in the entry range, no need to grow.*/ 3786 if (vm_map_lookup_entry(map, addr, &prev_entry)) 3787 goto done; 3788 3789 if ((stack_entry = prev_entry->next) == &map->header) 3790 goto done; 3791 if (prev_entry == &map->header) 3792 end = stack_entry->start - stack_entry->aux.avail_ssize; 3793 else 3794 end = prev_entry->end; 3795 3796 /* 3797 * This next test mimics the old grow function in vm_machdep.c. 3798 * It really doesn't quite make sense, but we do it anyway 3799 * for compatibility. 3800 * 3801 * If not growable stack, return success. This signals the 3802 * caller to proceed as he would normally with normal vm. 3803 */ 3804 if (stack_entry->aux.avail_ssize < 1 || 3805 addr >= stack_entry->start || 3806 addr < stack_entry->start - stack_entry->aux.avail_ssize) { 3807 goto done; 3808 } 3809 3810 /* Find the minimum grow amount */ 3811 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE); 3812 if (grow_amount > stack_entry->aux.avail_ssize) { 3813 rv = KERN_NO_SPACE; 3814 goto done; 3815 } 3816 3817 /* 3818 * If there is no longer enough space between the entries 3819 * nogo, and adjust the available space. Note: this 3820 * should only happen if the user has mapped into the 3821 * stack area after the stack was created, and is 3822 * probably an error. 3823 * 3824 * This also effectively destroys any guard page the user 3825 * might have intended by limiting the stack size. 3826 */ 3827 if (grow_amount > stack_entry->start - end) { 3828 if (use_read_lock && vm_map_lock_upgrade(map)) { 3829 /* lost lock */ 3830 use_read_lock = 0; 3831 goto Retry; 3832 } 3833 use_read_lock = 0; 3834 stack_entry->aux.avail_ssize = stack_entry->start - end; 3835 rv = KERN_NO_SPACE; 3836 goto done; 3837 } 3838 3839 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 3840 3841 /* If this is the main process stack, see if we're over the 3842 * stack limit. 3843 */ 3844 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3845 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3846 rv = KERN_NO_SPACE; 3847 goto done; 3848 } 3849 3850 /* Round up the grow amount modulo SGROWSIZ */ 3851 grow_amount = roundup (grow_amount, sgrowsiz); 3852 if (grow_amount > stack_entry->aux.avail_ssize) { 3853 grow_amount = stack_entry->aux.avail_ssize; 3854 } 3855 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3856 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3857 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - 3858 ctob(vm->vm_ssize); 3859 } 3860 3861 /* If we would blow our VMEM resource limit, no go */ 3862 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3863 rv = KERN_NO_SPACE; 3864 goto done; 3865 } 3866 3867 if (use_read_lock && vm_map_lock_upgrade(map)) { 3868 /* lost lock */ 3869 use_read_lock = 0; 3870 goto Retry; 3871 } 3872 use_read_lock = 0; 3873 3874 /* Get the preliminary new entry start value */ 3875 addr = stack_entry->start - grow_amount; 3876 3877 /* If this puts us into the previous entry, cut back our growth 3878 * to the available space. Also, see the note above. 3879 */ 3880 if (addr < end) { 3881 stack_entry->aux.avail_ssize = stack_entry->start - end; 3882 addr = end; 3883 } 3884 3885 rv = vm_map_insert(map, &count, NULL, NULL, 3886 0, addr, stack_entry->start, 3887 VM_MAPTYPE_NORMAL, 3888 VM_SUBSYS_STACK, VM_PROT_ALL, VM_PROT_ALL, 0); 3889 3890 /* Adjust the available stack space by the amount we grew. */ 3891 if (rv == KERN_SUCCESS) { 3892 if (prev_entry != &map->header) 3893 vm_map_clip_end(map, prev_entry, addr, &count); 3894 new_stack_entry = prev_entry->next; 3895 if (new_stack_entry->end != stack_entry->start || 3896 new_stack_entry->start != addr) 3897 panic ("Bad stack grow start/end in new stack entry"); 3898 else { 3899 new_stack_entry->aux.avail_ssize = 3900 stack_entry->aux.avail_ssize - 3901 (new_stack_entry->end - new_stack_entry->start); 3902 if (is_procstack) 3903 vm->vm_ssize += btoc(new_stack_entry->end - 3904 new_stack_entry->start); 3905 } 3906 3907 if (map->flags & MAP_WIREFUTURE) 3908 vm_map_unwire(map, new_stack_entry->start, 3909 new_stack_entry->end, FALSE); 3910 } 3911 3912 done: 3913 if (use_read_lock) 3914 vm_map_unlock_read(map); 3915 else 3916 vm_map_unlock(map); 3917 vm_map_entry_release(count); 3918 return (rv); 3919 } 3920 3921 /* 3922 * Unshare the specified VM space for exec. If other processes are 3923 * mapped to it, then create a new one. The new vmspace is null. 3924 * 3925 * No requirements. 3926 */ 3927 void 3928 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 3929 { 3930 struct vmspace *oldvmspace = p->p_vmspace; 3931 struct vmspace *newvmspace; 3932 vm_map_t map = &p->p_vmspace->vm_map; 3933 3934 /* 3935 * If we are execing a resident vmspace we fork it, otherwise 3936 * we create a new vmspace. Note that exitingcnt is not 3937 * copied to the new vmspace. 3938 */ 3939 lwkt_gettoken(&oldvmspace->vm_map.token); 3940 if (vmcopy) { 3941 newvmspace = vmspace_fork(vmcopy); 3942 lwkt_gettoken(&newvmspace->vm_map.token); 3943 } else { 3944 newvmspace = vmspace_alloc(map->min_offset, map->max_offset); 3945 lwkt_gettoken(&newvmspace->vm_map.token); 3946 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 3947 (caddr_t)&oldvmspace->vm_endcopy - 3948 (caddr_t)&oldvmspace->vm_startcopy); 3949 } 3950 3951 /* 3952 * Finish initializing the vmspace before assigning it 3953 * to the process. The vmspace will become the current vmspace 3954 * if p == curproc. 3955 */ 3956 pmap_pinit2(vmspace_pmap(newvmspace)); 3957 pmap_replacevm(p, newvmspace, 0); 3958 lwkt_reltoken(&newvmspace->vm_map.token); 3959 lwkt_reltoken(&oldvmspace->vm_map.token); 3960 vmspace_rel(oldvmspace); 3961 } 3962 3963 /* 3964 * Unshare the specified VM space for forcing COW. This 3965 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 3966 */ 3967 void 3968 vmspace_unshare(struct proc *p) 3969 { 3970 struct vmspace *oldvmspace = p->p_vmspace; 3971 struct vmspace *newvmspace; 3972 3973 lwkt_gettoken(&oldvmspace->vm_map.token); 3974 if (vmspace_getrefs(oldvmspace) == 1) { 3975 lwkt_reltoken(&oldvmspace->vm_map.token); 3976 return; 3977 } 3978 newvmspace = vmspace_fork(oldvmspace); 3979 lwkt_gettoken(&newvmspace->vm_map.token); 3980 pmap_pinit2(vmspace_pmap(newvmspace)); 3981 pmap_replacevm(p, newvmspace, 0); 3982 lwkt_reltoken(&newvmspace->vm_map.token); 3983 lwkt_reltoken(&oldvmspace->vm_map.token); 3984 vmspace_rel(oldvmspace); 3985 } 3986 3987 /* 3988 * vm_map_hint: return the beginning of the best area suitable for 3989 * creating a new mapping with "prot" protection. 3990 * 3991 * No requirements. 3992 */ 3993 vm_offset_t 3994 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot) 3995 { 3996 struct vmspace *vms = p->p_vmspace; 3997 3998 if (!randomize_mmap || addr != 0) { 3999 /* 4000 * Set a reasonable start point for the hint if it was 4001 * not specified or if it falls within the heap space. 4002 * Hinted mmap()s do not allocate out of the heap space. 4003 */ 4004 if (addr == 0 || 4005 (addr >= round_page((vm_offset_t)vms->vm_taddr) && 4006 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) { 4007 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz); 4008 } 4009 4010 return addr; 4011 } 4012 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ; 4013 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1); 4014 4015 return (round_page(addr)); 4016 } 4017 4018 /* 4019 * Finds the VM object, offset, and protection for a given virtual address 4020 * in the specified map, assuming a page fault of the type specified. 4021 * 4022 * Leaves the map in question locked for read; return values are guaranteed 4023 * until a vm_map_lookup_done call is performed. Note that the map argument 4024 * is in/out; the returned map must be used in the call to vm_map_lookup_done. 4025 * 4026 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make 4027 * that fast. 4028 * 4029 * If a lookup is requested with "write protection" specified, the map may 4030 * be changed to perform virtual copying operations, although the data 4031 * referenced will remain the same. 4032 * 4033 * No requirements. 4034 */ 4035 int 4036 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4037 vm_offset_t vaddr, 4038 vm_prot_t fault_typea, 4039 vm_map_entry_t *out_entry, /* OUT */ 4040 vm_object_t *object, /* OUT */ 4041 vm_pindex_t *pindex, /* OUT */ 4042 vm_prot_t *out_prot, /* OUT */ 4043 boolean_t *wired) /* OUT */ 4044 { 4045 vm_map_entry_t entry; 4046 vm_map_t map = *var_map; 4047 vm_prot_t prot; 4048 vm_prot_t fault_type = fault_typea; 4049 int use_read_lock = 1; 4050 int rv = KERN_SUCCESS; 4051 4052 RetryLookup: 4053 if (use_read_lock) 4054 vm_map_lock_read(map); 4055 else 4056 vm_map_lock(map); 4057 4058 /* 4059 * If the map has an interesting hint, try it before calling full 4060 * blown lookup routine. 4061 */ 4062 entry = map->hint; 4063 cpu_ccfence(); 4064 *out_entry = entry; 4065 *object = NULL; 4066 4067 if ((entry == &map->header) || 4068 (vaddr < entry->start) || (vaddr >= entry->end)) { 4069 vm_map_entry_t tmp_entry; 4070 4071 /* 4072 * Entry was either not a valid hint, or the vaddr was not 4073 * contained in the entry, so do a full lookup. 4074 */ 4075 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 4076 rv = KERN_INVALID_ADDRESS; 4077 goto done; 4078 } 4079 4080 entry = tmp_entry; 4081 *out_entry = entry; 4082 } 4083 4084 /* 4085 * Handle submaps. 4086 */ 4087 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 4088 vm_map_t old_map = map; 4089 4090 *var_map = map = entry->object.sub_map; 4091 if (use_read_lock) 4092 vm_map_unlock_read(old_map); 4093 else 4094 vm_map_unlock(old_map); 4095 use_read_lock = 1; 4096 goto RetryLookup; 4097 } 4098 4099 /* 4100 * Check whether this task is allowed to have this page. 4101 * Note the special case for MAP_ENTRY_COW pages with an override. 4102 * This is to implement a forced COW for debuggers. 4103 */ 4104 if (fault_type & VM_PROT_OVERRIDE_WRITE) 4105 prot = entry->max_protection; 4106 else 4107 prot = entry->protection; 4108 4109 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 4110 if ((fault_type & prot) != fault_type) { 4111 rv = KERN_PROTECTION_FAILURE; 4112 goto done; 4113 } 4114 4115 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 4116 (entry->eflags & MAP_ENTRY_COW) && 4117 (fault_type & VM_PROT_WRITE) && 4118 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 4119 rv = KERN_PROTECTION_FAILURE; 4120 goto done; 4121 } 4122 4123 /* 4124 * If this page is not pageable, we have to get it for all possible 4125 * accesses. 4126 */ 4127 *wired = (entry->wired_count != 0); 4128 if (*wired) 4129 prot = fault_type = entry->protection; 4130 4131 /* 4132 * Virtual page tables may need to update the accessed (A) bit 4133 * in a page table entry. Upgrade the fault to a write fault for 4134 * that case if the map will support it. If the map does not support 4135 * it the page table entry simply will not be updated. 4136 */ 4137 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 4138 if (prot & VM_PROT_WRITE) 4139 fault_type |= VM_PROT_WRITE; 4140 } 4141 4142 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 4143 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 4144 if ((prot & VM_PROT_WRITE) == 0) 4145 fault_type |= VM_PROT_WRITE; 4146 } 4147 4148 /* 4149 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not. 4150 */ 4151 if (entry->maptype != VM_MAPTYPE_NORMAL && 4152 entry->maptype != VM_MAPTYPE_VPAGETABLE) { 4153 *object = NULL; 4154 goto skip; 4155 } 4156 4157 /* 4158 * If the entry was copy-on-write, we either ... 4159 */ 4160 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4161 /* 4162 * If we want to write the page, we may as well handle that 4163 * now since we've got the map locked. 4164 * 4165 * If we don't need to write the page, we just demote the 4166 * permissions allowed. 4167 */ 4168 4169 if (fault_type & VM_PROT_WRITE) { 4170 /* 4171 * Not allowed if TDF_NOFAULT is set as the shadowing 4172 * operation can deadlock against the faulting 4173 * function due to the copy-on-write. 4174 */ 4175 if (curthread->td_flags & TDF_NOFAULT) { 4176 rv = KERN_FAILURE_NOFAULT; 4177 goto done; 4178 } 4179 4180 /* 4181 * Make a new object, and place it in the object 4182 * chain. Note that no new references have appeared 4183 * -- one just moved from the map to the new 4184 * object. 4185 */ 4186 4187 if (use_read_lock && vm_map_lock_upgrade(map)) { 4188 /* lost lock */ 4189 use_read_lock = 0; 4190 goto RetryLookup; 4191 } 4192 use_read_lock = 0; 4193 4194 vm_map_entry_shadow(entry, 0); 4195 } else { 4196 /* 4197 * We're attempting to read a copy-on-write page -- 4198 * don't allow writes. 4199 */ 4200 4201 prot &= ~VM_PROT_WRITE; 4202 } 4203 } 4204 4205 /* 4206 * Create an object if necessary. 4207 */ 4208 if (entry->object.vm_object == NULL && !map->system_map) { 4209 if (use_read_lock && vm_map_lock_upgrade(map)) { 4210 /* lost lock */ 4211 use_read_lock = 0; 4212 goto RetryLookup; 4213 } 4214 use_read_lock = 0; 4215 vm_map_entry_allocate_object(entry); 4216 } 4217 4218 /* 4219 * Return the object/offset from this entry. If the entry was 4220 * copy-on-write or empty, it has been fixed up. 4221 */ 4222 *object = entry->object.vm_object; 4223 4224 skip: 4225 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 4226 4227 /* 4228 * Return whether this is the only map sharing this data. On 4229 * success we return with a read lock held on the map. On failure 4230 * we return with the map unlocked. 4231 */ 4232 *out_prot = prot; 4233 done: 4234 if (rv == KERN_SUCCESS) { 4235 if (use_read_lock == 0) 4236 vm_map_lock_downgrade(map); 4237 } else if (use_read_lock) { 4238 vm_map_unlock_read(map); 4239 } else { 4240 vm_map_unlock(map); 4241 } 4242 return (rv); 4243 } 4244 4245 /* 4246 * Releases locks acquired by a vm_map_lookup() 4247 * (according to the handle returned by that lookup). 4248 * 4249 * No other requirements. 4250 */ 4251 void 4252 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 4253 { 4254 /* 4255 * Unlock the main-level map 4256 */ 4257 vm_map_unlock_read(map); 4258 if (count) 4259 vm_map_entry_release(count); 4260 } 4261 4262 /* 4263 * Quick hack, needs some help to make it more SMP friendly. 4264 */ 4265 void 4266 vm_map_interlock(vm_map_t map, struct vm_map_ilock *ilock, 4267 vm_offset_t ran_beg, vm_offset_t ran_end) 4268 { 4269 struct vm_map_ilock *scan; 4270 4271 ilock->ran_beg = ran_beg; 4272 ilock->ran_end = ran_end; 4273 ilock->flags = 0; 4274 4275 spin_lock(&map->ilock_spin); 4276 restart: 4277 for (scan = map->ilock_base; scan; scan = scan->next) { 4278 if (ran_end > scan->ran_beg && ran_beg < scan->ran_end) { 4279 scan->flags |= ILOCK_WAITING; 4280 ssleep(scan, &map->ilock_spin, 0, "ilock", 0); 4281 goto restart; 4282 } 4283 } 4284 ilock->next = map->ilock_base; 4285 map->ilock_base = ilock; 4286 spin_unlock(&map->ilock_spin); 4287 } 4288 4289 void 4290 vm_map_deinterlock(vm_map_t map, struct vm_map_ilock *ilock) 4291 { 4292 struct vm_map_ilock *scan; 4293 struct vm_map_ilock **scanp; 4294 4295 spin_lock(&map->ilock_spin); 4296 scanp = &map->ilock_base; 4297 while ((scan = *scanp) != NULL) { 4298 if (scan == ilock) { 4299 *scanp = ilock->next; 4300 spin_unlock(&map->ilock_spin); 4301 if (ilock->flags & ILOCK_WAITING) 4302 wakeup(ilock); 4303 return; 4304 } 4305 scanp = &scan->next; 4306 } 4307 spin_unlock(&map->ilock_spin); 4308 panic("vm_map_deinterlock: missing ilock!"); 4309 } 4310 4311 #include "opt_ddb.h" 4312 #ifdef DDB 4313 #include <ddb/ddb.h> 4314 4315 /* 4316 * Debugging only 4317 */ 4318 DB_SHOW_COMMAND(map, vm_map_print) 4319 { 4320 static int nlines; 4321 /* XXX convert args. */ 4322 vm_map_t map = (vm_map_t)addr; 4323 boolean_t full = have_addr; 4324 4325 vm_map_entry_t entry; 4326 4327 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 4328 (void *)map, 4329 (void *)map->pmap, map->nentries, map->timestamp); 4330 nlines++; 4331 4332 if (!full && db_indent) 4333 return; 4334 4335 db_indent += 2; 4336 for (entry = map->header.next; entry != &map->header; 4337 entry = entry->next) { 4338 db_iprintf("map entry %p: start=%p, end=%p\n", 4339 (void *)entry, (void *)entry->start, (void *)entry->end); 4340 nlines++; 4341 { 4342 static char *inheritance_name[4] = 4343 {"share", "copy", "none", "donate_copy"}; 4344 4345 db_iprintf(" prot=%x/%x/%s", 4346 entry->protection, 4347 entry->max_protection, 4348 inheritance_name[(int)(unsigned char) 4349 entry->inheritance]); 4350 if (entry->wired_count != 0) 4351 db_printf(", wired"); 4352 } 4353 switch(entry->maptype) { 4354 case VM_MAPTYPE_SUBMAP: 4355 /* XXX no %qd in kernel. Truncate entry->offset. */ 4356 db_printf(", share=%p, offset=0x%lx\n", 4357 (void *)entry->object.sub_map, 4358 (long)entry->offset); 4359 nlines++; 4360 if ((entry->prev == &map->header) || 4361 (entry->prev->object.sub_map != 4362 entry->object.sub_map)) { 4363 db_indent += 2; 4364 vm_map_print((db_expr_t)(intptr_t) 4365 entry->object.sub_map, 4366 full, 0, NULL); 4367 db_indent -= 2; 4368 } 4369 break; 4370 case VM_MAPTYPE_NORMAL: 4371 case VM_MAPTYPE_VPAGETABLE: 4372 /* XXX no %qd in kernel. Truncate entry->offset. */ 4373 db_printf(", object=%p, offset=0x%lx", 4374 (void *)entry->object.vm_object, 4375 (long)entry->offset); 4376 if (entry->eflags & MAP_ENTRY_COW) 4377 db_printf(", copy (%s)", 4378 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4379 db_printf("\n"); 4380 nlines++; 4381 4382 if ((entry->prev == &map->header) || 4383 (entry->prev->object.vm_object != 4384 entry->object.vm_object)) { 4385 db_indent += 2; 4386 vm_object_print((db_expr_t)(intptr_t) 4387 entry->object.vm_object, 4388 full, 0, NULL); 4389 nlines += 4; 4390 db_indent -= 2; 4391 } 4392 break; 4393 case VM_MAPTYPE_UKSMAP: 4394 db_printf(", uksmap=%p, offset=0x%lx", 4395 (void *)entry->object.uksmap, 4396 (long)entry->offset); 4397 if (entry->eflags & MAP_ENTRY_COW) 4398 db_printf(", copy (%s)", 4399 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4400 db_printf("\n"); 4401 nlines++; 4402 break; 4403 default: 4404 break; 4405 } 4406 } 4407 db_indent -= 2; 4408 if (db_indent == 0) 4409 nlines = 0; 4410 } 4411 4412 /* 4413 * Debugging only 4414 */ 4415 DB_SHOW_COMMAND(procvm, procvm) 4416 { 4417 struct proc *p; 4418 4419 if (have_addr) { 4420 p = (struct proc *) addr; 4421 } else { 4422 p = curproc; 4423 } 4424 4425 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 4426 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 4427 (void *)vmspace_pmap(p->p_vmspace)); 4428 4429 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 4430 } 4431 4432 #endif /* DDB */ 4433