1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * Copyright (c) 2003-2017 The DragonFly Project. All rights reserved. 5 * 6 * This code is derived from software contributed to Berkeley by 7 * The Mach Operating System project at Carnegie-Mellon University. 8 * 9 * This code is derived from software contributed to The DragonFly Project 10 * by Matthew Dillon <dillon@backplane.com> 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94 37 * 38 * 39 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 40 * All rights reserved. 41 * 42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 43 * 44 * Permission to use, copy, modify and distribute this software and 45 * its documentation is hereby granted, provided that both the copyright 46 * notice and this permission notice appear in all copies of the 47 * software, derivative works or modified versions, and any portions 48 * thereof, and that both notices appear in supporting documentation. 49 * 50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 53 * 54 * Carnegie Mellon requests users of this software to return to 55 * 56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 57 * School of Computer Science 58 * Carnegie Mellon University 59 * Pittsburgh PA 15213-3890 60 * 61 * any improvements or extensions that they make and grant Carnegie the 62 * rights to redistribute these changes. 63 * 64 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $ 65 */ 66 67 /* 68 * Virtual memory mapping module. 69 */ 70 71 #include <sys/param.h> 72 #include <sys/systm.h> 73 #include <sys/kernel.h> 74 #include <sys/proc.h> 75 #include <sys/serialize.h> 76 #include <sys/lock.h> 77 #include <sys/vmmeter.h> 78 #include <sys/mman.h> 79 #include <sys/vnode.h> 80 #include <sys/resourcevar.h> 81 #include <sys/shm.h> 82 #include <sys/tree.h> 83 #include <sys/malloc.h> 84 #include <sys/objcache.h> 85 #include <sys/kern_syscall.h> 86 87 #include <vm/vm.h> 88 #include <vm/vm_param.h> 89 #include <vm/pmap.h> 90 #include <vm/vm_map.h> 91 #include <vm/vm_page.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_pager.h> 94 #include <vm/vm_kern.h> 95 #include <vm/vm_extern.h> 96 #include <vm/swap_pager.h> 97 #include <vm/vm_zone.h> 98 99 #include <sys/random.h> 100 #include <sys/sysctl.h> 101 #include <sys/spinlock.h> 102 103 #include <sys/thread2.h> 104 #include <sys/spinlock2.h> 105 106 /* 107 * Virtual memory maps provide for the mapping, protection, and sharing 108 * of virtual memory objects. In addition, this module provides for an 109 * efficient virtual copy of memory from one map to another. 110 * 111 * Synchronization is required prior to most operations. 112 * 113 * Maps consist of an ordered doubly-linked list of simple entries. 114 * A hint and a RB tree is used to speed-up lookups. 115 * 116 * Callers looking to modify maps specify start/end addresses which cause 117 * the related map entry to be clipped if necessary, and then later 118 * recombined if the pieces remained compatible. 119 * 120 * Virtual copy operations are performed by copying VM object references 121 * from one map to another, and then marking both regions as copy-on-write. 122 */ 123 static boolean_t vmspace_ctor(void *obj, void *privdata, int ocflags); 124 static void vmspace_dtor(void *obj, void *privdata); 125 static void vmspace_terminate(struct vmspace *vm, int final); 126 127 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore"); 128 static struct objcache *vmspace_cache; 129 130 /* 131 * per-cpu page table cross mappings are initialized in early boot 132 * and might require a considerable number of vm_map_entry structures. 133 */ 134 #define MAPENTRYBSP_CACHE (MAXCPU+1) 135 #define MAPENTRYAP_CACHE 8 136 137 /* 138 * Partioning threaded programs with large anonymous memory areas can 139 * improve concurrent fault performance. 140 */ 141 #define MAP_ENTRY_PARTITION_SIZE ((vm_offset_t)(32 * 1024 * 1024)) 142 #define MAP_ENTRY_PARTITION_MASK (MAP_ENTRY_PARTITION_SIZE - 1) 143 144 #define VM_MAP_ENTRY_WITHIN_PARTITION(entry) \ 145 ((((entry)->start ^ (entry)->end) & ~MAP_ENTRY_PARTITION_MASK) == 0) 146 147 static struct vm_zone mapentzone_store; 148 static vm_zone_t mapentzone; 149 150 static struct vm_map_entry map_entry_init[MAX_MAPENT]; 151 static struct vm_map_entry cpu_map_entry_init_bsp[MAPENTRYBSP_CACHE]; 152 static struct vm_map_entry cpu_map_entry_init_ap[MAXCPU][MAPENTRYAP_CACHE]; 153 154 static int randomize_mmap; 155 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0, 156 "Randomize mmap offsets"); 157 static int vm_map_relock_enable = 1; 158 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW, 159 &vm_map_relock_enable, 0, "insert pop pgtable optimization"); 160 static int vm_map_partition_enable = 1; 161 SYSCTL_INT(_vm, OID_AUTO, map_partition_enable, CTLFLAG_RW, 162 &vm_map_partition_enable, 0, "Break up larger vm_map_entry's"); 163 164 static void vmspace_drop_notoken(struct vmspace *vm); 165 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref); 166 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *); 167 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *); 168 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 169 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 170 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *); 171 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t); 172 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t, 173 vm_map_entry_t); 174 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry, 175 vm_offset_t start, vm_offset_t end, int *countp, int flags); 176 static void vm_map_entry_partition(vm_map_t map, vm_map_entry_t entry, 177 vm_offset_t vaddr, int *countp); 178 179 /* 180 * Initialize the vm_map module. Must be called before any other vm_map 181 * routines. 182 * 183 * Map and entry structures are allocated from the general purpose 184 * memory pool with some exceptions: 185 * 186 * - The kernel map is allocated statically. 187 * - Initial kernel map entries are allocated out of a static pool. 188 * - We must set ZONE_SPECIAL here or the early boot code can get 189 * stuck if there are >63 cores. 190 * 191 * These restrictions are necessary since malloc() uses the 192 * maps and requires map entries. 193 * 194 * Called from the low level boot code only. 195 */ 196 void 197 vm_map_startup(void) 198 { 199 mapentzone = &mapentzone_store; 200 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry), 201 map_entry_init, MAX_MAPENT); 202 mapentzone_store.zflags |= ZONE_SPECIAL; 203 } 204 205 /* 206 * Called prior to any vmspace allocations. 207 * 208 * Called from the low level boot code only. 209 */ 210 void 211 vm_init2(void) 212 { 213 vmspace_cache = objcache_create_mbacked(M_VMSPACE, 214 sizeof(struct vmspace), 215 0, ncpus * 4, 216 vmspace_ctor, vmspace_dtor, 217 NULL); 218 zinitna(mapentzone, NULL, 0, 0, ZONE_USE_RESERVE | ZONE_SPECIAL); 219 pmap_init2(); 220 vm_object_init2(); 221 } 222 223 /* 224 * objcache support. We leave the pmap root cached as long as possible 225 * for performance reasons. 226 */ 227 static 228 boolean_t 229 vmspace_ctor(void *obj, void *privdata, int ocflags) 230 { 231 struct vmspace *vm = obj; 232 233 bzero(vm, sizeof(*vm)); 234 vm->vm_refcnt = VM_REF_DELETED; 235 236 return 1; 237 } 238 239 static 240 void 241 vmspace_dtor(void *obj, void *privdata) 242 { 243 struct vmspace *vm = obj; 244 245 KKASSERT(vm->vm_refcnt == VM_REF_DELETED); 246 pmap_puninit(vmspace_pmap(vm)); 247 } 248 249 /* 250 * Red black tree functions 251 * 252 * The caller must hold the related map lock. 253 */ 254 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b); 255 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare); 256 257 /* a->start is address, and the only field has to be initialized */ 258 static int 259 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b) 260 { 261 if (a->start < b->start) 262 return(-1); 263 else if (a->start > b->start) 264 return(1); 265 return(0); 266 } 267 268 /* 269 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for 270 * every refcnt. 271 */ 272 void 273 vmspace_initrefs(struct vmspace *vm) 274 { 275 vm->vm_refcnt = 1; 276 vm->vm_holdcnt = 1; 277 } 278 279 /* 280 * Allocate a vmspace structure, including a vm_map and pmap. 281 * Initialize numerous fields. While the initial allocation is zerod, 282 * subsequence reuse from the objcache leaves elements of the structure 283 * intact (particularly the pmap), so portions must be zerod. 284 * 285 * Returns a referenced vmspace. 286 * 287 * No requirements. 288 */ 289 struct vmspace * 290 vmspace_alloc(vm_offset_t min, vm_offset_t max) 291 { 292 struct vmspace *vm; 293 294 vm = objcache_get(vmspace_cache, M_WAITOK); 295 296 bzero(&vm->vm_startcopy, 297 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy); 298 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */ 299 300 /* 301 * NOTE: hold to acquires token for safety. 302 * 303 * On return vmspace is referenced (refs=1, hold=1). That is, 304 * each refcnt also has a holdcnt. There can be additional holds 305 * (holdcnt) above and beyond the refcnt. Finalization is handled in 306 * two stages, one on refs 1->0, and the the second on hold 1->0. 307 */ 308 KKASSERT(vm->vm_holdcnt == 0); 309 KKASSERT(vm->vm_refcnt == VM_REF_DELETED); 310 vmspace_initrefs(vm); 311 vmspace_hold(vm); 312 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */ 313 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */ 314 vm->vm_shm = NULL; 315 vm->vm_flags = 0; 316 cpu_vmspace_alloc(vm); 317 vmspace_drop(vm); 318 319 return (vm); 320 } 321 322 /* 323 * NOTE: Can return 0 if the vmspace is exiting. 324 */ 325 int 326 vmspace_getrefs(struct vmspace *vm) 327 { 328 int32_t n; 329 330 n = vm->vm_refcnt; 331 cpu_ccfence(); 332 if (n & VM_REF_DELETED) 333 n = -1; 334 return n; 335 } 336 337 void 338 vmspace_hold(struct vmspace *vm) 339 { 340 atomic_add_int(&vm->vm_holdcnt, 1); 341 lwkt_gettoken(&vm->vm_map.token); 342 } 343 344 /* 345 * Drop with final termination interlock. 346 */ 347 void 348 vmspace_drop(struct vmspace *vm) 349 { 350 lwkt_reltoken(&vm->vm_map.token); 351 vmspace_drop_notoken(vm); 352 } 353 354 static void 355 vmspace_drop_notoken(struct vmspace *vm) 356 { 357 if (atomic_fetchadd_int(&vm->vm_holdcnt, -1) == 1) { 358 if (vm->vm_refcnt & VM_REF_DELETED) 359 vmspace_terminate(vm, 1); 360 } 361 } 362 363 /* 364 * A vmspace object must not be in a terminated state to be able to obtain 365 * additional refs on it. 366 * 367 * These are official references to the vmspace, the count is used to check 368 * for vmspace sharing. Foreign accessors should use 'hold' and not 'ref'. 369 * 370 * XXX we need to combine hold & ref together into one 64-bit field to allow 371 * holds to prevent stage-1 termination. 372 */ 373 void 374 vmspace_ref(struct vmspace *vm) 375 { 376 uint32_t n; 377 378 atomic_add_int(&vm->vm_holdcnt, 1); 379 n = atomic_fetchadd_int(&vm->vm_refcnt, 1); 380 KKASSERT((n & VM_REF_DELETED) == 0); 381 } 382 383 /* 384 * Release a ref on the vmspace. On the 1->0 transition we do stage-1 385 * termination of the vmspace. Then, on the final drop of the hold we 386 * will do stage-2 final termination. 387 */ 388 void 389 vmspace_rel(struct vmspace *vm) 390 { 391 uint32_t n; 392 393 /* 394 * Drop refs. Each ref also has a hold which is also dropped. 395 * 396 * When refs hits 0 compete to get the VM_REF_DELETED flag (hold 397 * prevent finalization) to start termination processing. 398 * Finalization occurs when the last hold count drops to 0. 399 */ 400 n = atomic_fetchadd_int(&vm->vm_refcnt, -1) - 1; 401 while (n == 0) { 402 if (atomic_cmpset_int(&vm->vm_refcnt, 0, VM_REF_DELETED)) { 403 vmspace_terminate(vm, 0); 404 break; 405 } 406 n = vm->vm_refcnt; 407 cpu_ccfence(); 408 } 409 vmspace_drop_notoken(vm); 410 } 411 412 /* 413 * This is called during exit indicating that the vmspace is no 414 * longer in used by an exiting process, but the process has not yet 415 * been reaped. 416 * 417 * We drop refs, allowing for stage-1 termination, but maintain a holdcnt 418 * to prevent stage-2 until the process is reaped. Note hte order of 419 * operation, we must hold first. 420 * 421 * No requirements. 422 */ 423 void 424 vmspace_relexit(struct vmspace *vm) 425 { 426 atomic_add_int(&vm->vm_holdcnt, 1); 427 vmspace_rel(vm); 428 } 429 430 /* 431 * Called during reap to disconnect the remainder of the vmspace from 432 * the process. On the hold drop the vmspace termination is finalized. 433 * 434 * No requirements. 435 */ 436 void 437 vmspace_exitfree(struct proc *p) 438 { 439 struct vmspace *vm; 440 441 vm = p->p_vmspace; 442 p->p_vmspace = NULL; 443 vmspace_drop_notoken(vm); 444 } 445 446 /* 447 * Called in two cases: 448 * 449 * (1) When the last refcnt is dropped and the vmspace becomes inactive, 450 * called with final == 0. refcnt will be (u_int)-1 at this point, 451 * and holdcnt will still be non-zero. 452 * 453 * (2) When holdcnt becomes 0, called with final == 1. There should no 454 * longer be anyone with access to the vmspace. 455 * 456 * VMSPACE_EXIT1 flags the primary deactivation 457 * VMSPACE_EXIT2 flags the last reap 458 */ 459 static void 460 vmspace_terminate(struct vmspace *vm, int final) 461 { 462 int count; 463 464 lwkt_gettoken(&vm->vm_map.token); 465 if (final == 0) { 466 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0); 467 vm->vm_flags |= VMSPACE_EXIT1; 468 469 /* 470 * Get rid of most of the resources. Leave the kernel pmap 471 * intact. 472 * 473 * If the pmap does not contain wired pages we can bulk-delete 474 * the pmap as a performance optimization before removing the 475 * related mappings. 476 * 477 * If the pmap contains wired pages we cannot do this 478 * pre-optimization because currently vm_fault_unwire() 479 * expects the pmap pages to exist and will not decrement 480 * p->wire_count if they do not. 481 */ 482 shmexit(vm); 483 if (vmspace_pmap(vm)->pm_stats.wired_count) { 484 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 485 VM_MAX_USER_ADDRESS); 486 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 487 VM_MAX_USER_ADDRESS); 488 } else { 489 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 490 VM_MAX_USER_ADDRESS); 491 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 492 VM_MAX_USER_ADDRESS); 493 } 494 lwkt_reltoken(&vm->vm_map.token); 495 } else { 496 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0); 497 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0); 498 499 /* 500 * Get rid of remaining basic resources. 501 */ 502 vm->vm_flags |= VMSPACE_EXIT2; 503 shmexit(vm); 504 505 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 506 vm_map_lock(&vm->vm_map); 507 cpu_vmspace_free(vm); 508 509 /* 510 * Lock the map, to wait out all other references to it. 511 * Delete all of the mappings and pages they hold, then call 512 * the pmap module to reclaim anything left. 513 */ 514 vm_map_delete(&vm->vm_map, 515 vm_map_min(&vm->vm_map), 516 vm_map_max(&vm->vm_map), 517 &count); 518 vm_map_unlock(&vm->vm_map); 519 vm_map_entry_release(count); 520 521 pmap_release(vmspace_pmap(vm)); 522 lwkt_reltoken(&vm->vm_map.token); 523 objcache_put(vmspace_cache, vm); 524 } 525 } 526 527 /* 528 * Swap useage is determined by taking the proportional swap used by 529 * VM objects backing the VM map. To make up for fractional losses, 530 * if the VM object has any swap use at all the associated map entries 531 * count for at least 1 swap page. 532 * 533 * No requirements. 534 */ 535 vm_offset_t 536 vmspace_swap_count(struct vmspace *vm) 537 { 538 vm_map_t map = &vm->vm_map; 539 vm_map_entry_t cur; 540 vm_object_t object; 541 vm_offset_t count = 0; 542 vm_offset_t n; 543 544 vmspace_hold(vm); 545 546 RB_FOREACH(cur, vm_map_rb_tree, &map->rb_root) { 547 switch(cur->maptype) { 548 case VM_MAPTYPE_NORMAL: 549 case VM_MAPTYPE_VPAGETABLE: 550 if ((object = cur->object.vm_object) == NULL) 551 break; 552 if (object->swblock_count) { 553 n = (cur->end - cur->start) / PAGE_SIZE; 554 count += object->swblock_count * 555 SWAP_META_PAGES * n / object->size + 1; 556 } 557 break; 558 default: 559 break; 560 } 561 } 562 vmspace_drop(vm); 563 564 return(count); 565 } 566 567 /* 568 * Calculate the approximate number of anonymous pages in use by 569 * this vmspace. To make up for fractional losses, we count each 570 * VM object as having at least 1 anonymous page. 571 * 572 * No requirements. 573 */ 574 vm_offset_t 575 vmspace_anonymous_count(struct vmspace *vm) 576 { 577 vm_map_t map = &vm->vm_map; 578 vm_map_entry_t cur; 579 vm_object_t object; 580 vm_offset_t count = 0; 581 582 vmspace_hold(vm); 583 RB_FOREACH(cur, vm_map_rb_tree, &map->rb_root) { 584 switch(cur->maptype) { 585 case VM_MAPTYPE_NORMAL: 586 case VM_MAPTYPE_VPAGETABLE: 587 if ((object = cur->object.vm_object) == NULL) 588 break; 589 if (object->type != OBJT_DEFAULT && 590 object->type != OBJT_SWAP) { 591 break; 592 } 593 count += object->resident_page_count; 594 break; 595 default: 596 break; 597 } 598 } 599 vmspace_drop(vm); 600 601 return(count); 602 } 603 604 /* 605 * Initialize an existing vm_map structure such as that in the vmspace 606 * structure. The pmap is initialized elsewhere. 607 * 608 * No requirements. 609 */ 610 void 611 vm_map_init(struct vm_map *map, vm_offset_t min_addr, vm_offset_t max_addr, 612 pmap_t pmap) 613 { 614 RB_INIT(&map->rb_root); 615 spin_init(&map->ilock_spin, "ilock"); 616 map->ilock_base = NULL; 617 map->nentries = 0; 618 map->size = 0; 619 map->system_map = 0; 620 vm_map_min(map) = min_addr; 621 vm_map_max(map) = max_addr; 622 map->pmap = pmap; 623 map->timestamp = 0; 624 map->flags = 0; 625 bzero(&map->freehint, sizeof(map->freehint)); 626 lwkt_token_init(&map->token, "vm_map"); 627 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0); 628 } 629 630 /* 631 * Find the first possible free address for the specified request length. 632 * Returns 0 if we don't have one cached. 633 */ 634 static 635 vm_offset_t 636 vm_map_freehint_find(vm_map_t map, vm_size_t length, vm_size_t align) 637 { 638 vm_map_freehint_t *scan; 639 640 scan = &map->freehint[0]; 641 while (scan < &map->freehint[VM_MAP_FFCOUNT]) { 642 if (scan->length == length && scan->align == align) 643 return(scan->start); 644 ++scan; 645 } 646 return 0; 647 } 648 649 /* 650 * Unconditionally set the freehint. Called by vm_map_findspace() after 651 * it finds an address. This will help us iterate optimally on the next 652 * similar findspace. 653 */ 654 static 655 void 656 vm_map_freehint_update(vm_map_t map, vm_offset_t start, 657 vm_size_t length, vm_size_t align) 658 { 659 vm_map_freehint_t *scan; 660 661 scan = &map->freehint[0]; 662 while (scan < &map->freehint[VM_MAP_FFCOUNT]) { 663 if (scan->length == length && scan->align == align) { 664 scan->start = start; 665 return; 666 } 667 ++scan; 668 } 669 scan = &map->freehint[map->freehint_newindex & VM_MAP_FFMASK]; 670 scan->start = start; 671 scan->align = align; 672 scan->length = length; 673 ++map->freehint_newindex; 674 } 675 676 /* 677 * Update any existing freehints (for any alignment), for the hole we just 678 * added. 679 */ 680 static 681 void 682 vm_map_freehint_hole(vm_map_t map, vm_offset_t start, vm_size_t length) 683 { 684 vm_map_freehint_t *scan; 685 686 scan = &map->freehint[0]; 687 while (scan < &map->freehint[VM_MAP_FFCOUNT]) { 688 if (scan->length <= length && scan->start > start) 689 scan->start = start; 690 ++scan; 691 } 692 } 693 694 /* 695 * Shadow the vm_map_entry's object. This typically needs to be done when 696 * a write fault is taken on an entry which had previously been cloned by 697 * fork(). The shared object (which might be NULL) must become private so 698 * we add a shadow layer above it. 699 * 700 * Object allocation for anonymous mappings is defered as long as possible. 701 * When creating a shadow, however, the underlying object must be instantiated 702 * so it can be shared. 703 * 704 * If the map segment is governed by a virtual page table then it is 705 * possible to address offsets beyond the mapped area. Just allocate 706 * a maximally sized object for this case. 707 * 708 * If addref is non-zero an additional reference is added to the returned 709 * entry. This mechanic exists because the additional reference might have 710 * to be added atomically and not after return to prevent a premature 711 * collapse. 712 * 713 * The vm_map must be exclusively locked. 714 * No other requirements. 715 */ 716 static 717 void 718 vm_map_entry_shadow(vm_map_entry_t entry, int addref) 719 { 720 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 721 vm_object_shadow(&entry->object.vm_object, &entry->offset, 722 0x7FFFFFFF, addref); /* XXX */ 723 } else { 724 vm_object_shadow(&entry->object.vm_object, &entry->offset, 725 atop(entry->end - entry->start), addref); 726 } 727 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 728 } 729 730 /* 731 * Allocate an object for a vm_map_entry. 732 * 733 * Object allocation for anonymous mappings is defered as long as possible. 734 * This function is called when we can defer no longer, generally when a map 735 * entry might be split or forked or takes a page fault. 736 * 737 * If the map segment is governed by a virtual page table then it is 738 * possible to address offsets beyond the mapped area. Just allocate 739 * a maximally sized object for this case. 740 * 741 * The vm_map must be exclusively locked. 742 * No other requirements. 743 */ 744 void 745 vm_map_entry_allocate_object(vm_map_entry_t entry) 746 { 747 vm_object_t obj; 748 749 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 750 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */ 751 } else { 752 obj = vm_object_allocate(OBJT_DEFAULT, 753 atop(entry->end - entry->start)); 754 } 755 entry->object.vm_object = obj; 756 entry->offset = 0; 757 } 758 759 /* 760 * Set an initial negative count so the first attempt to reserve 761 * space preloads a bunch of vm_map_entry's for this cpu. Also 762 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 763 * map a new page for vm_map_entry structures. SMP systems are 764 * particularly sensitive. 765 * 766 * This routine is called in early boot so we cannot just call 767 * vm_map_entry_reserve(). 768 * 769 * Called from the low level boot code only (for each cpu) 770 * 771 * WARNING! Take care not to have too-big a static/BSS structure here 772 * as MAXCPU can be 256+, otherwise the loader's 64MB heap 773 * can get blown out by the kernel plus the initrd image. 774 */ 775 void 776 vm_map_entry_reserve_cpu_init(globaldata_t gd) 777 { 778 vm_map_entry_t entry; 779 int count; 780 int i; 781 782 atomic_add_int(&gd->gd_vme_avail, -MAP_RESERVE_COUNT * 2); 783 if (gd->gd_cpuid == 0) { 784 entry = &cpu_map_entry_init_bsp[0]; 785 count = MAPENTRYBSP_CACHE; 786 } else { 787 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0]; 788 count = MAPENTRYAP_CACHE; 789 } 790 for (i = 0; i < count; ++i, ++entry) { 791 MAPENT_FREELIST(entry) = gd->gd_vme_base; 792 gd->gd_vme_base = entry; 793 } 794 } 795 796 /* 797 * Reserves vm_map_entry structures so code later-on can manipulate 798 * map_entry structures within a locked map without blocking trying 799 * to allocate a new vm_map_entry. 800 * 801 * No requirements. 802 * 803 * WARNING! We must not decrement gd_vme_avail until after we have 804 * ensured that sufficient entries exist, otherwise we can 805 * get into an endless call recursion in the zalloc code 806 * itself. 807 */ 808 int 809 vm_map_entry_reserve(int count) 810 { 811 struct globaldata *gd = mycpu; 812 vm_map_entry_t entry; 813 814 /* 815 * Make sure we have enough structures in gd_vme_base to handle 816 * the reservation request. 817 * 818 * Use a critical section to protect against VM faults. It might 819 * not be needed, but we have to be careful here. 820 */ 821 if (gd->gd_vme_avail < count) { 822 crit_enter(); 823 while (gd->gd_vme_avail < count) { 824 entry = zalloc(mapentzone); 825 MAPENT_FREELIST(entry) = gd->gd_vme_base; 826 gd->gd_vme_base = entry; 827 atomic_add_int(&gd->gd_vme_avail, 1); 828 } 829 crit_exit(); 830 } 831 atomic_add_int(&gd->gd_vme_avail, -count); 832 833 return(count); 834 } 835 836 /* 837 * Releases previously reserved vm_map_entry structures that were not 838 * used. If we have too much junk in our per-cpu cache clean some of 839 * it out. 840 * 841 * No requirements. 842 */ 843 void 844 vm_map_entry_release(int count) 845 { 846 struct globaldata *gd = mycpu; 847 vm_map_entry_t entry; 848 vm_map_entry_t efree; 849 850 count = atomic_fetchadd_int(&gd->gd_vme_avail, count) + count; 851 if (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 852 efree = NULL; 853 crit_enter(); 854 while (gd->gd_vme_avail > MAP_RESERVE_HYST) { 855 entry = gd->gd_vme_base; 856 KKASSERT(entry != NULL); 857 gd->gd_vme_base = MAPENT_FREELIST(entry); 858 atomic_add_int(&gd->gd_vme_avail, -1); 859 MAPENT_FREELIST(entry) = efree; 860 efree = entry; 861 } 862 crit_exit(); 863 while ((entry = efree) != NULL) { 864 efree = MAPENT_FREELIST(efree); 865 zfree(mapentzone, entry); 866 } 867 } 868 } 869 870 /* 871 * Reserve map entry structures for use in kernel_map itself. These 872 * entries have *ALREADY* been reserved on a per-cpu basis when the map 873 * was inited. This function is used by zalloc() to avoid a recursion 874 * when zalloc() itself needs to allocate additional kernel memory. 875 * 876 * This function works like the normal reserve but does not load the 877 * vm_map_entry cache (because that would result in an infinite 878 * recursion). Note that gd_vme_avail may go negative. This is expected. 879 * 880 * Any caller of this function must be sure to renormalize after 881 * potentially eating entries to ensure that the reserve supply 882 * remains intact. 883 * 884 * No requirements. 885 */ 886 int 887 vm_map_entry_kreserve(int count) 888 { 889 struct globaldata *gd = mycpu; 890 891 atomic_add_int(&gd->gd_vme_avail, -count); 892 KASSERT(gd->gd_vme_base != NULL, 893 ("no reserved entries left, gd_vme_avail = %d", 894 gd->gd_vme_avail)); 895 return(count); 896 } 897 898 /* 899 * Release previously reserved map entries for kernel_map. We do not 900 * attempt to clean up like the normal release function as this would 901 * cause an unnecessary (but probably not fatal) deep procedure call. 902 * 903 * No requirements. 904 */ 905 void 906 vm_map_entry_krelease(int count) 907 { 908 struct globaldata *gd = mycpu; 909 910 atomic_add_int(&gd->gd_vme_avail, count); 911 } 912 913 /* 914 * Allocates a VM map entry for insertion. No entry fields are filled in. 915 * 916 * The entries should have previously been reserved. The reservation count 917 * is tracked in (*countp). 918 * 919 * No requirements. 920 */ 921 static vm_map_entry_t 922 vm_map_entry_create(vm_map_t map, int *countp) 923 { 924 struct globaldata *gd = mycpu; 925 vm_map_entry_t entry; 926 927 KKASSERT(*countp > 0); 928 --*countp; 929 crit_enter(); 930 entry = gd->gd_vme_base; 931 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 932 gd->gd_vme_base = MAPENT_FREELIST(entry); 933 crit_exit(); 934 935 return(entry); 936 } 937 938 /* 939 * Dispose of a vm_map_entry that is no longer being referenced. 940 * 941 * No requirements. 942 */ 943 static void 944 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 945 { 946 struct globaldata *gd = mycpu; 947 948 ++*countp; 949 crit_enter(); 950 MAPENT_FREELIST(entry) = gd->gd_vme_base; 951 gd->gd_vme_base = entry; 952 crit_exit(); 953 } 954 955 956 /* 957 * Insert/remove entries from maps. 958 * 959 * The related map must be exclusively locked. 960 * The caller must hold map->token 961 * No other requirements. 962 */ 963 static __inline void 964 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) 965 { 966 ASSERT_VM_MAP_LOCKED(map); 967 968 map->nentries++; 969 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry)) 970 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry); 971 } 972 973 static __inline void 974 vm_map_entry_unlink(vm_map_t map, 975 vm_map_entry_t entry) 976 { 977 ASSERT_VM_MAP_LOCKED(map); 978 979 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 980 panic("vm_map_entry_unlink: attempt to mess with " 981 "locked entry! %p", entry); 982 } 983 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry); 984 map->nentries--; 985 } 986 987 /* 988 * Finds the map entry containing (or immediately preceding) the specified 989 * address in the given map. The entry is returned in (*entry). 990 * 991 * The boolean result indicates whether the address is actually contained 992 * in the map. 993 * 994 * The related map must be locked. 995 * No other requirements. 996 */ 997 boolean_t 998 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry) 999 { 1000 vm_map_entry_t tmp; 1001 vm_map_entry_t last; 1002 1003 ASSERT_VM_MAP_LOCKED(map); 1004 1005 /* 1006 * Locate the record from the top of the tree. 'last' tracks the 1007 * closest prior record and is returned if no match is found, which 1008 * in binary tree terms means tracking the most recent right-branch 1009 * taken. If there is no prior record, *entry is set to NULL. 1010 */ 1011 last = NULL; 1012 tmp = RB_ROOT(&map->rb_root); 1013 1014 while (tmp) { 1015 if (address >= tmp->start) { 1016 if (address < tmp->end) { 1017 *entry = tmp; 1018 return(TRUE); 1019 } 1020 last = tmp; 1021 tmp = RB_RIGHT(tmp, rb_entry); 1022 } else { 1023 tmp = RB_LEFT(tmp, rb_entry); 1024 } 1025 } 1026 *entry = last; 1027 return (FALSE); 1028 } 1029 1030 /* 1031 * Inserts the given whole VM object into the target map at the specified 1032 * address range. The object's size should match that of the address range. 1033 * 1034 * The map must be exclusively locked. 1035 * The object must be held. 1036 * The caller must have reserved sufficient vm_map_entry structures. 1037 * 1038 * If object is non-NULL, ref count must be bumped by caller prior to 1039 * making call to account for the new entry. 1040 */ 1041 int 1042 vm_map_insert(vm_map_t map, int *countp, void *map_object, void *map_aux, 1043 vm_ooffset_t offset, vm_offset_t start, vm_offset_t end, 1044 vm_maptype_t maptype, vm_subsys_t id, 1045 vm_prot_t prot, vm_prot_t max, int cow) 1046 { 1047 vm_map_entry_t new_entry; 1048 vm_map_entry_t prev_entry; 1049 vm_map_entry_t next; 1050 vm_map_entry_t temp_entry; 1051 vm_eflags_t protoeflags; 1052 int must_drop = 0; 1053 vm_object_t object; 1054 1055 if (maptype == VM_MAPTYPE_UKSMAP) 1056 object = NULL; 1057 else 1058 object = map_object; 1059 1060 ASSERT_VM_MAP_LOCKED(map); 1061 if (object) 1062 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1063 1064 /* 1065 * Check that the start and end points are not bogus. 1066 */ 1067 if ((start < vm_map_min(map)) || (end > vm_map_max(map)) || 1068 (start >= end)) { 1069 return (KERN_INVALID_ADDRESS); 1070 } 1071 1072 /* 1073 * Find the entry prior to the proposed starting address; if it's part 1074 * of an existing entry, this range is bogus. 1075 */ 1076 if (vm_map_lookup_entry(map, start, &temp_entry)) 1077 return (KERN_NO_SPACE); 1078 prev_entry = temp_entry; 1079 1080 /* 1081 * Assert that the next entry doesn't overlap the end point. 1082 */ 1083 if (prev_entry) 1084 next = vm_map_rb_tree_RB_NEXT(prev_entry); 1085 else 1086 next = RB_MIN(vm_map_rb_tree, &map->rb_root); 1087 if (next && next->start < end) 1088 return (KERN_NO_SPACE); 1089 1090 protoeflags = 0; 1091 1092 if (cow & MAP_COPY_ON_WRITE) 1093 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 1094 1095 if (cow & MAP_NOFAULT) { 1096 protoeflags |= MAP_ENTRY_NOFAULT; 1097 1098 KASSERT(object == NULL, 1099 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 1100 } 1101 if (cow & MAP_DISABLE_SYNCER) 1102 protoeflags |= MAP_ENTRY_NOSYNC; 1103 if (cow & MAP_DISABLE_COREDUMP) 1104 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1105 if (cow & MAP_IS_STACK) 1106 protoeflags |= MAP_ENTRY_STACK; 1107 if (cow & MAP_IS_KSTACK) 1108 protoeflags |= MAP_ENTRY_KSTACK; 1109 1110 lwkt_gettoken(&map->token); 1111 1112 if (object) { 1113 /* 1114 * When object is non-NULL, it could be shared with another 1115 * process. We have to set or clear OBJ_ONEMAPPING 1116 * appropriately. 1117 * 1118 * NOTE: This flag is only applicable to DEFAULT and SWAP 1119 * objects and will already be clear in other types 1120 * of objects, so a shared object lock is ok for 1121 * VNODE objects. 1122 */ 1123 if ((object->ref_count > 1) || (object->shadow_count != 0)) { 1124 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1125 } 1126 } 1127 else if (prev_entry && 1128 (prev_entry->eflags == protoeflags) && 1129 (prev_entry->end == start) && 1130 (prev_entry->wired_count == 0) && 1131 (prev_entry->id == id) && 1132 prev_entry->maptype == maptype && 1133 maptype == VM_MAPTYPE_NORMAL && 1134 ((prev_entry->object.vm_object == NULL) || 1135 vm_object_coalesce(prev_entry->object.vm_object, 1136 OFF_TO_IDX(prev_entry->offset), 1137 (vm_size_t)(prev_entry->end - prev_entry->start), 1138 (vm_size_t)(end - prev_entry->end)))) { 1139 /* 1140 * We were able to extend the object. Determine if we 1141 * can extend the previous map entry to include the 1142 * new range as well. 1143 */ 1144 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 1145 (prev_entry->protection == prot) && 1146 (prev_entry->max_protection == max)) { 1147 map->size += (end - prev_entry->end); 1148 prev_entry->end = end; 1149 vm_map_simplify_entry(map, prev_entry, countp); 1150 lwkt_reltoken(&map->token); 1151 return (KERN_SUCCESS); 1152 } 1153 1154 /* 1155 * If we can extend the object but cannot extend the 1156 * map entry, we have to create a new map entry. We 1157 * must bump the ref count on the extended object to 1158 * account for it. object may be NULL. 1159 * 1160 * XXX if object is NULL should we set offset to 0 here ? 1161 */ 1162 object = prev_entry->object.vm_object; 1163 offset = prev_entry->offset + 1164 (prev_entry->end - prev_entry->start); 1165 if (object) { 1166 vm_object_hold(object); 1167 vm_object_chain_wait(object, 0); 1168 vm_object_reference_locked(object); 1169 must_drop = 1; 1170 map_object = object; 1171 } 1172 } 1173 1174 /* 1175 * NOTE: if conditionals fail, object can be NULL here. This occurs 1176 * in things like the buffer map where we manage kva but do not manage 1177 * backing objects. 1178 */ 1179 1180 /* 1181 * Create a new entry 1182 */ 1183 1184 new_entry = vm_map_entry_create(map, countp); 1185 new_entry->start = start; 1186 new_entry->end = end; 1187 new_entry->id = id; 1188 1189 new_entry->maptype = maptype; 1190 new_entry->eflags = protoeflags; 1191 new_entry->object.map_object = map_object; 1192 new_entry->aux.master_pde = 0; /* in case size is different */ 1193 new_entry->aux.map_aux = map_aux; 1194 new_entry->offset = offset; 1195 1196 new_entry->inheritance = VM_INHERIT_DEFAULT; 1197 new_entry->protection = prot; 1198 new_entry->max_protection = max; 1199 new_entry->wired_count = 0; 1200 1201 /* 1202 * Insert the new entry into the list 1203 */ 1204 1205 vm_map_entry_link(map, new_entry); 1206 map->size += new_entry->end - new_entry->start; 1207 1208 /* 1209 * Don't worry about updating freehint[] when inserting, allow 1210 * addresses to be lower than the actual first free spot. 1211 */ 1212 #if 0 1213 /* 1214 * Temporarily removed to avoid MAP_STACK panic, due to 1215 * MAP_STACK being a huge hack. Will be added back in 1216 * when MAP_STACK (and the user stack mapping) is fixed. 1217 */ 1218 /* 1219 * It may be possible to simplify the entry 1220 */ 1221 vm_map_simplify_entry(map, new_entry, countp); 1222 #endif 1223 1224 /* 1225 * Try to pre-populate the page table. Mappings governed by virtual 1226 * page tables cannot be prepopulated without a lot of work, so 1227 * don't try. 1228 */ 1229 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) && 1230 maptype != VM_MAPTYPE_VPAGETABLE && 1231 maptype != VM_MAPTYPE_UKSMAP) { 1232 int dorelock = 0; 1233 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) { 1234 dorelock = 1; 1235 vm_object_lock_swap(); 1236 vm_object_drop(object); 1237 } 1238 pmap_object_init_pt(map->pmap, start, prot, 1239 object, OFF_TO_IDX(offset), end - start, 1240 cow & MAP_PREFAULT_PARTIAL); 1241 if (dorelock) { 1242 vm_object_hold(object); 1243 vm_object_lock_swap(); 1244 } 1245 } 1246 if (must_drop) 1247 vm_object_drop(object); 1248 1249 lwkt_reltoken(&map->token); 1250 return (KERN_SUCCESS); 1251 } 1252 1253 /* 1254 * Find sufficient space for `length' bytes in the given map, starting at 1255 * `start'. Returns 0 on success, 1 on no space. 1256 * 1257 * This function will returned an arbitrarily aligned pointer. If no 1258 * particular alignment is required you should pass align as 1. Note that 1259 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 1260 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 1261 * argument. 1262 * 1263 * 'align' should be a power of 2 but is not required to be. 1264 * 1265 * The map must be exclusively locked. 1266 * No other requirements. 1267 */ 1268 int 1269 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length, 1270 vm_size_t align, int flags, vm_offset_t *addr) 1271 { 1272 vm_map_entry_t entry; 1273 vm_map_entry_t tmp; 1274 vm_offset_t hole_start; 1275 vm_offset_t end; 1276 vm_offset_t align_mask; 1277 1278 if (start < vm_map_min(map)) 1279 start = vm_map_min(map); 1280 if (start > vm_map_max(map)) 1281 return (1); 1282 1283 /* 1284 * If the alignment is not a power of 2 we will have to use 1285 * a mod/division, set align_mask to a special value. 1286 */ 1287 if ((align | (align - 1)) + 1 != (align << 1)) 1288 align_mask = (vm_offset_t)-1; 1289 else 1290 align_mask = align - 1; 1291 1292 /* 1293 * Use freehint to adjust the start point, hopefully reducing 1294 * the iteration to O(1). 1295 */ 1296 hole_start = vm_map_freehint_find(map, length, align); 1297 if (start < hole_start) 1298 start = hole_start; 1299 if (vm_map_lookup_entry(map, start, &tmp)) 1300 start = tmp->end; 1301 entry = tmp; /* may be NULL */ 1302 1303 /* 1304 * Look through the rest of the map, trying to fit a new region in the 1305 * gap between existing regions, or after the very last region. 1306 */ 1307 for (;;) { 1308 /* 1309 * Adjust the proposed start by the requested alignment, 1310 * be sure that we didn't wrap the address. 1311 */ 1312 if (align_mask == (vm_offset_t)-1) 1313 end = roundup(start, align); 1314 else 1315 end = (start + align_mask) & ~align_mask; 1316 if (end < start) 1317 return (1); 1318 start = end; 1319 1320 /* 1321 * Find the end of the proposed new region. Be sure we didn't 1322 * go beyond the end of the map, or wrap around the address. 1323 * Then check to see if this is the last entry or if the 1324 * proposed end fits in the gap between this and the next 1325 * entry. 1326 */ 1327 end = start + length; 1328 if (end > vm_map_max(map) || end < start) 1329 return (1); 1330 1331 /* 1332 * Locate the next entry, we can stop if this is the 1333 * last entry (we know we are in-bounds so that would 1334 * be a sucess). 1335 */ 1336 if (entry) 1337 entry = vm_map_rb_tree_RB_NEXT(entry); 1338 else 1339 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 1340 if (entry == NULL) 1341 break; 1342 1343 /* 1344 * Determine if the proposed area would overlap the 1345 * next entry. 1346 */ 1347 if (entry->start >= end) { 1348 if ((entry->eflags & MAP_ENTRY_STACK) == 0) 1349 break; 1350 if (flags & MAP_STACK) 1351 break; 1352 if (entry->start - entry->aux.avail_ssize >= end) 1353 break; 1354 } 1355 start = entry->end; 1356 } 1357 1358 /* 1359 * Update the freehint 1360 */ 1361 vm_map_freehint_update(map, start, length, align); 1362 1363 /* 1364 * Grow the kernel_map if necessary. pmap_growkernel() will panic 1365 * if it fails. The kernel_map is locked and nothing can steal 1366 * our address space if pmap_growkernel() blocks. 1367 * 1368 * NOTE: This may be unconditionally called for kldload areas on 1369 * x86_64 because these do not bump kernel_vm_end (which would 1370 * fill 128G worth of page tables!). Therefore we must not 1371 * retry. 1372 */ 1373 if (map == &kernel_map) { 1374 vm_offset_t kstop; 1375 1376 kstop = round_page(start + length); 1377 if (kstop > kernel_vm_end) 1378 pmap_growkernel(start, kstop); 1379 } 1380 *addr = start; 1381 return (0); 1382 } 1383 1384 /* 1385 * vm_map_find finds an unallocated region in the target address map with 1386 * the given length and allocates it. The search is defined to be first-fit 1387 * from the specified address; the region found is returned in the same 1388 * parameter. 1389 * 1390 * If object is non-NULL, ref count must be bumped by caller 1391 * prior to making call to account for the new entry. 1392 * 1393 * No requirements. This function will lock the map temporarily. 1394 */ 1395 int 1396 vm_map_find(vm_map_t map, void *map_object, void *map_aux, 1397 vm_ooffset_t offset, vm_offset_t *addr, 1398 vm_size_t length, vm_size_t align, boolean_t fitit, 1399 vm_maptype_t maptype, vm_subsys_t id, 1400 vm_prot_t prot, vm_prot_t max, int cow) 1401 { 1402 vm_offset_t start; 1403 vm_object_t object; 1404 int result; 1405 int count; 1406 1407 if (maptype == VM_MAPTYPE_UKSMAP) 1408 object = NULL; 1409 else 1410 object = map_object; 1411 1412 start = *addr; 1413 1414 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1415 vm_map_lock(map); 1416 if (object) 1417 vm_object_hold_shared(object); 1418 if (fitit) { 1419 if (vm_map_findspace(map, start, length, align, 0, addr)) { 1420 if (object) 1421 vm_object_drop(object); 1422 vm_map_unlock(map); 1423 vm_map_entry_release(count); 1424 return (KERN_NO_SPACE); 1425 } 1426 start = *addr; 1427 } 1428 result = vm_map_insert(map, &count, map_object, map_aux, 1429 offset, start, start + length, 1430 maptype, id, prot, max, cow); 1431 if (object) 1432 vm_object_drop(object); 1433 vm_map_unlock(map); 1434 vm_map_entry_release(count); 1435 1436 return (result); 1437 } 1438 1439 /* 1440 * Simplify the given map entry by merging with either neighbor. This 1441 * routine also has the ability to merge with both neighbors. 1442 * 1443 * This routine guarentees that the passed entry remains valid (though 1444 * possibly extended). When merging, this routine may delete one or 1445 * both neighbors. No action is taken on entries which have their 1446 * in-transition flag set. 1447 * 1448 * The map must be exclusively locked. 1449 */ 1450 void 1451 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 1452 { 1453 vm_map_entry_t next, prev; 1454 vm_size_t prevsize, esize; 1455 1456 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1457 ++mycpu->gd_cnt.v_intrans_coll; 1458 return; 1459 } 1460 1461 if (entry->maptype == VM_MAPTYPE_SUBMAP) 1462 return; 1463 if (entry->maptype == VM_MAPTYPE_UKSMAP) 1464 return; 1465 1466 prev = vm_map_rb_tree_RB_PREV(entry); 1467 if (prev) { 1468 prevsize = prev->end - prev->start; 1469 if ( (prev->end == entry->start) && 1470 (prev->maptype == entry->maptype) && 1471 (prev->object.vm_object == entry->object.vm_object) && 1472 (!prev->object.vm_object || 1473 (prev->offset + prevsize == entry->offset)) && 1474 (prev->eflags == entry->eflags) && 1475 (prev->protection == entry->protection) && 1476 (prev->max_protection == entry->max_protection) && 1477 (prev->inheritance == entry->inheritance) && 1478 (prev->id == entry->id) && 1479 (prev->wired_count == entry->wired_count)) { 1480 vm_map_entry_unlink(map, prev); 1481 entry->start = prev->start; 1482 entry->offset = prev->offset; 1483 if (prev->object.vm_object) 1484 vm_object_deallocate(prev->object.vm_object); 1485 vm_map_entry_dispose(map, prev, countp); 1486 } 1487 } 1488 1489 next = vm_map_rb_tree_RB_NEXT(entry); 1490 if (next) { 1491 esize = entry->end - entry->start; 1492 if ((entry->end == next->start) && 1493 (next->maptype == entry->maptype) && 1494 (next->object.vm_object == entry->object.vm_object) && 1495 (!entry->object.vm_object || 1496 (entry->offset + esize == next->offset)) && 1497 (next->eflags == entry->eflags) && 1498 (next->protection == entry->protection) && 1499 (next->max_protection == entry->max_protection) && 1500 (next->inheritance == entry->inheritance) && 1501 (next->id == entry->id) && 1502 (next->wired_count == entry->wired_count)) { 1503 vm_map_entry_unlink(map, next); 1504 entry->end = next->end; 1505 if (next->object.vm_object) 1506 vm_object_deallocate(next->object.vm_object); 1507 vm_map_entry_dispose(map, next, countp); 1508 } 1509 } 1510 } 1511 1512 /* 1513 * Asserts that the given entry begins at or after the specified address. 1514 * If necessary, it splits the entry into two. 1515 */ 1516 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1517 { \ 1518 if (startaddr > entry->start) \ 1519 _vm_map_clip_start(map, entry, startaddr, countp); \ 1520 } 1521 1522 /* 1523 * This routine is called only when it is known that the entry must be split. 1524 * 1525 * The map must be exclusively locked. 1526 */ 1527 static void 1528 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, 1529 int *countp) 1530 { 1531 vm_map_entry_t new_entry; 1532 1533 /* 1534 * Split off the front portion -- note that we must insert the new 1535 * entry BEFORE this one, so that this entry has the specified 1536 * starting address. 1537 */ 1538 1539 vm_map_simplify_entry(map, entry, countp); 1540 1541 /* 1542 * If there is no object backing this entry, we might as well create 1543 * one now. If we defer it, an object can get created after the map 1544 * is clipped, and individual objects will be created for the split-up 1545 * map. This is a bit of a hack, but is also about the best place to 1546 * put this improvement. 1547 */ 1548 if (entry->object.vm_object == NULL && !map->system_map && 1549 VM_MAP_ENTRY_WITHIN_PARTITION(entry)) { 1550 vm_map_entry_allocate_object(entry); 1551 } 1552 1553 new_entry = vm_map_entry_create(map, countp); 1554 *new_entry = *entry; 1555 1556 new_entry->end = start; 1557 entry->offset += (start - entry->start); 1558 entry->start = start; 1559 1560 vm_map_entry_link(map, new_entry); 1561 1562 switch(entry->maptype) { 1563 case VM_MAPTYPE_NORMAL: 1564 case VM_MAPTYPE_VPAGETABLE: 1565 if (new_entry->object.vm_object) { 1566 vm_object_hold(new_entry->object.vm_object); 1567 vm_object_chain_wait(new_entry->object.vm_object, 0); 1568 vm_object_reference_locked(new_entry->object.vm_object); 1569 vm_object_drop(new_entry->object.vm_object); 1570 } 1571 break; 1572 default: 1573 break; 1574 } 1575 } 1576 1577 /* 1578 * Asserts that the given entry ends at or before the specified address. 1579 * If necessary, it splits the entry into two. 1580 * 1581 * The map must be exclusively locked. 1582 */ 1583 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1584 { \ 1585 if (endaddr < entry->end) \ 1586 _vm_map_clip_end(map, entry, endaddr, countp); \ 1587 } 1588 1589 /* 1590 * This routine is called only when it is known that the entry must be split. 1591 * 1592 * The map must be exclusively locked. 1593 */ 1594 static void 1595 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, 1596 int *countp) 1597 { 1598 vm_map_entry_t new_entry; 1599 1600 /* 1601 * If there is no object backing this entry, we might as well create 1602 * one now. If we defer it, an object can get created after the map 1603 * is clipped, and individual objects will be created for the split-up 1604 * map. This is a bit of a hack, but is also about the best place to 1605 * put this improvement. 1606 */ 1607 1608 if (entry->object.vm_object == NULL && !map->system_map && 1609 VM_MAP_ENTRY_WITHIN_PARTITION(entry)) { 1610 vm_map_entry_allocate_object(entry); 1611 } 1612 1613 /* 1614 * Create a new entry and insert it AFTER the specified entry 1615 */ 1616 new_entry = vm_map_entry_create(map, countp); 1617 *new_entry = *entry; 1618 1619 new_entry->start = entry->end = end; 1620 new_entry->offset += (end - entry->start); 1621 1622 vm_map_entry_link(map, new_entry); 1623 1624 switch(entry->maptype) { 1625 case VM_MAPTYPE_NORMAL: 1626 case VM_MAPTYPE_VPAGETABLE: 1627 if (new_entry->object.vm_object) { 1628 vm_object_hold(new_entry->object.vm_object); 1629 vm_object_chain_wait(new_entry->object.vm_object, 0); 1630 vm_object_reference_locked(new_entry->object.vm_object); 1631 vm_object_drop(new_entry->object.vm_object); 1632 } 1633 break; 1634 default: 1635 break; 1636 } 1637 } 1638 1639 /* 1640 * Asserts that the starting and ending region addresses fall within the 1641 * valid range for the map. 1642 */ 1643 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1644 { \ 1645 if (start < vm_map_min(map)) \ 1646 start = vm_map_min(map); \ 1647 if (end > vm_map_max(map)) \ 1648 end = vm_map_max(map); \ 1649 if (start > end) \ 1650 start = end; \ 1651 } 1652 1653 /* 1654 * Used to block when an in-transition collison occurs. The map 1655 * is unlocked for the sleep and relocked before the return. 1656 */ 1657 void 1658 vm_map_transition_wait(vm_map_t map, int relock) 1659 { 1660 tsleep_interlock(map, 0); 1661 vm_map_unlock(map); 1662 tsleep(map, PINTERLOCKED, "vment", 0); 1663 if (relock) 1664 vm_map_lock(map); 1665 } 1666 1667 /* 1668 * When we do blocking operations with the map lock held it is 1669 * possible that a clip might have occured on our in-transit entry, 1670 * requiring an adjustment to the entry in our loop. These macros 1671 * help the pageable and clip_range code deal with the case. The 1672 * conditional costs virtually nothing if no clipping has occured. 1673 */ 1674 1675 #define CLIP_CHECK_BACK(entry, save_start) \ 1676 do { \ 1677 while (entry->start != save_start) { \ 1678 entry = vm_map_rb_tree_RB_PREV(entry); \ 1679 KASSERT(entry, ("bad entry clip")); \ 1680 } \ 1681 } while(0) 1682 1683 #define CLIP_CHECK_FWD(entry, save_end) \ 1684 do { \ 1685 while (entry->end != save_end) { \ 1686 entry = vm_map_rb_tree_RB_NEXT(entry); \ 1687 KASSERT(entry, ("bad entry clip")); \ 1688 } \ 1689 } while(0) 1690 1691 1692 /* 1693 * Clip the specified range and return the base entry. The 1694 * range may cover several entries starting at the returned base 1695 * and the first and last entry in the covering sequence will be 1696 * properly clipped to the requested start and end address. 1697 * 1698 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1699 * flag. 1700 * 1701 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1702 * covered by the requested range. 1703 * 1704 * The map must be exclusively locked on entry and will remain locked 1705 * on return. If no range exists or the range contains holes and you 1706 * specified that no holes were allowed, NULL will be returned. This 1707 * routine may temporarily unlock the map in order avoid a deadlock when 1708 * sleeping. 1709 */ 1710 static 1711 vm_map_entry_t 1712 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1713 int *countp, int flags) 1714 { 1715 vm_map_entry_t start_entry; 1716 vm_map_entry_t entry; 1717 vm_map_entry_t next; 1718 1719 /* 1720 * Locate the entry and effect initial clipping. The in-transition 1721 * case does not occur very often so do not try to optimize it. 1722 */ 1723 again: 1724 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1725 return (NULL); 1726 entry = start_entry; 1727 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1728 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1729 ++mycpu->gd_cnt.v_intrans_coll; 1730 ++mycpu->gd_cnt.v_intrans_wait; 1731 vm_map_transition_wait(map, 1); 1732 /* 1733 * entry and/or start_entry may have been clipped while 1734 * we slept, or may have gone away entirely. We have 1735 * to restart from the lookup. 1736 */ 1737 goto again; 1738 } 1739 1740 /* 1741 * Since we hold an exclusive map lock we do not have to restart 1742 * after clipping, even though clipping may block in zalloc. 1743 */ 1744 vm_map_clip_start(map, entry, start, countp); 1745 vm_map_clip_end(map, entry, end, countp); 1746 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1747 1748 /* 1749 * Scan entries covered by the range. When working on the next 1750 * entry a restart need only re-loop on the current entry which 1751 * we have already locked, since 'next' may have changed. Also, 1752 * even though entry is safe, it may have been clipped so we 1753 * have to iterate forwards through the clip after sleeping. 1754 */ 1755 for (;;) { 1756 next = vm_map_rb_tree_RB_NEXT(entry); 1757 if (next == NULL || next->start >= end) 1758 break; 1759 if (flags & MAP_CLIP_NO_HOLES) { 1760 if (next->start > entry->end) { 1761 vm_map_unclip_range(map, start_entry, 1762 start, entry->end, countp, flags); 1763 return(NULL); 1764 } 1765 } 1766 1767 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1768 vm_offset_t save_end = entry->end; 1769 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1770 ++mycpu->gd_cnt.v_intrans_coll; 1771 ++mycpu->gd_cnt.v_intrans_wait; 1772 vm_map_transition_wait(map, 1); 1773 1774 /* 1775 * clips might have occured while we blocked. 1776 */ 1777 CLIP_CHECK_FWD(entry, save_end); 1778 CLIP_CHECK_BACK(start_entry, start); 1779 continue; 1780 } 1781 1782 /* 1783 * No restart necessary even though clip_end may block, we 1784 * are holding the map lock. 1785 */ 1786 vm_map_clip_end(map, next, end, countp); 1787 next->eflags |= MAP_ENTRY_IN_TRANSITION; 1788 entry = next; 1789 } 1790 if (flags & MAP_CLIP_NO_HOLES) { 1791 if (entry->end != end) { 1792 vm_map_unclip_range(map, start_entry, 1793 start, entry->end, countp, flags); 1794 return(NULL); 1795 } 1796 } 1797 return(start_entry); 1798 } 1799 1800 /* 1801 * Undo the effect of vm_map_clip_range(). You should pass the same 1802 * flags and the same range that you passed to vm_map_clip_range(). 1803 * This code will clear the in-transition flag on the entries and 1804 * wake up anyone waiting. This code will also simplify the sequence 1805 * and attempt to merge it with entries before and after the sequence. 1806 * 1807 * The map must be locked on entry and will remain locked on return. 1808 * 1809 * Note that you should also pass the start_entry returned by 1810 * vm_map_clip_range(). However, if you block between the two calls 1811 * with the map unlocked please be aware that the start_entry may 1812 * have been clipped and you may need to scan it backwards to find 1813 * the entry corresponding with the original start address. You are 1814 * responsible for this, vm_map_unclip_range() expects the correct 1815 * start_entry to be passed to it and will KASSERT otherwise. 1816 */ 1817 static 1818 void 1819 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry, 1820 vm_offset_t start, vm_offset_t end, 1821 int *countp, int flags) 1822 { 1823 vm_map_entry_t entry; 1824 1825 entry = start_entry; 1826 1827 KASSERT(entry->start == start, ("unclip_range: illegal base entry")); 1828 while (entry && entry->start < end) { 1829 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, 1830 ("in-transition flag not set during unclip on: %p", 1831 entry)); 1832 KASSERT(entry->end <= end, 1833 ("unclip_range: tail wasn't clipped")); 1834 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 1835 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 1836 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 1837 wakeup(map); 1838 } 1839 entry = vm_map_rb_tree_RB_NEXT(entry); 1840 } 1841 1842 /* 1843 * Simplification does not block so there is no restart case. 1844 */ 1845 entry = start_entry; 1846 while (entry && entry->start < end) { 1847 vm_map_simplify_entry(map, entry, countp); 1848 entry = vm_map_rb_tree_RB_NEXT(entry); 1849 } 1850 } 1851 1852 /* 1853 * Mark the given range as handled by a subordinate map. 1854 * 1855 * This range must have been created with vm_map_find(), and no other 1856 * operations may have been performed on this range prior to calling 1857 * vm_map_submap(). 1858 * 1859 * Submappings cannot be removed. 1860 * 1861 * No requirements. 1862 */ 1863 int 1864 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 1865 { 1866 vm_map_entry_t entry; 1867 int result = KERN_INVALID_ARGUMENT; 1868 int count; 1869 1870 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1871 vm_map_lock(map); 1872 1873 VM_MAP_RANGE_CHECK(map, start, end); 1874 1875 if (vm_map_lookup_entry(map, start, &entry)) { 1876 vm_map_clip_start(map, entry, start, &count); 1877 } else if (entry) { 1878 entry = vm_map_rb_tree_RB_NEXT(entry); 1879 } else { 1880 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 1881 } 1882 1883 vm_map_clip_end(map, entry, end, &count); 1884 1885 if ((entry->start == start) && (entry->end == end) && 1886 ((entry->eflags & MAP_ENTRY_COW) == 0) && 1887 (entry->object.vm_object == NULL)) { 1888 entry->object.sub_map = submap; 1889 entry->maptype = VM_MAPTYPE_SUBMAP; 1890 result = KERN_SUCCESS; 1891 } 1892 vm_map_unlock(map); 1893 vm_map_entry_release(count); 1894 1895 return (result); 1896 } 1897 1898 /* 1899 * Sets the protection of the specified address region in the target map. 1900 * If "set_max" is specified, the maximum protection is to be set; 1901 * otherwise, only the current protection is affected. 1902 * 1903 * The protection is not applicable to submaps, but is applicable to normal 1904 * maps and maps governed by virtual page tables. For example, when operating 1905 * on a virtual page table our protection basically controls how COW occurs 1906 * on the backing object, whereas the virtual page table abstraction itself 1907 * is an abstraction for userland. 1908 * 1909 * No requirements. 1910 */ 1911 int 1912 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 1913 vm_prot_t new_prot, boolean_t set_max) 1914 { 1915 vm_map_entry_t current; 1916 vm_map_entry_t entry; 1917 int count; 1918 1919 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1920 vm_map_lock(map); 1921 1922 VM_MAP_RANGE_CHECK(map, start, end); 1923 1924 if (vm_map_lookup_entry(map, start, &entry)) { 1925 vm_map_clip_start(map, entry, start, &count); 1926 } else if (entry) { 1927 entry = vm_map_rb_tree_RB_NEXT(entry); 1928 } else { 1929 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 1930 } 1931 1932 /* 1933 * Make a first pass to check for protection violations. 1934 */ 1935 current = entry; 1936 while (current && current->start < end) { 1937 if (current->maptype == VM_MAPTYPE_SUBMAP) { 1938 vm_map_unlock(map); 1939 vm_map_entry_release(count); 1940 return (KERN_INVALID_ARGUMENT); 1941 } 1942 if ((new_prot & current->max_protection) != new_prot) { 1943 vm_map_unlock(map); 1944 vm_map_entry_release(count); 1945 return (KERN_PROTECTION_FAILURE); 1946 } 1947 1948 /* 1949 * When making a SHARED+RW file mmap writable, update 1950 * v_lastwrite_ts. 1951 */ 1952 if (new_prot & PROT_WRITE && 1953 (current->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && 1954 (current->maptype == VM_MAPTYPE_NORMAL || 1955 current->maptype == VM_MAPTYPE_VPAGETABLE) && 1956 current->object.vm_object && 1957 current->object.vm_object->type == OBJT_VNODE) { 1958 struct vnode *vp; 1959 1960 vp = current->object.vm_object->handle; 1961 if (vp && vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_NOWAIT) == 0) { 1962 vfs_timestamp(&vp->v_lastwrite_ts); 1963 vsetflags(vp, VLASTWRITETS); 1964 vn_unlock(vp); 1965 } 1966 } 1967 current = vm_map_rb_tree_RB_NEXT(current); 1968 } 1969 1970 /* 1971 * Go back and fix up protections. [Note that clipping is not 1972 * necessary the second time.] 1973 */ 1974 current = entry; 1975 1976 while (current && current->start < end) { 1977 vm_prot_t old_prot; 1978 1979 vm_map_clip_end(map, current, end, &count); 1980 1981 old_prot = current->protection; 1982 if (set_max) { 1983 current->max_protection = new_prot; 1984 current->protection = new_prot & old_prot; 1985 } else { 1986 current->protection = new_prot; 1987 } 1988 1989 /* 1990 * Update physical map if necessary. Worry about copy-on-write 1991 * here -- CHECK THIS XXX 1992 */ 1993 if (current->protection != old_prot) { 1994 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 1995 VM_PROT_ALL) 1996 1997 pmap_protect(map->pmap, current->start, 1998 current->end, 1999 current->protection & MASK(current)); 2000 #undef MASK 2001 } 2002 2003 vm_map_simplify_entry(map, current, &count); 2004 2005 current = vm_map_rb_tree_RB_NEXT(current); 2006 } 2007 vm_map_unlock(map); 2008 vm_map_entry_release(count); 2009 return (KERN_SUCCESS); 2010 } 2011 2012 /* 2013 * This routine traverses a processes map handling the madvise 2014 * system call. Advisories are classified as either those effecting 2015 * the vm_map_entry structure, or those effecting the underlying 2016 * objects. 2017 * 2018 * The <value> argument is used for extended madvise calls. 2019 * 2020 * No requirements. 2021 */ 2022 int 2023 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, 2024 int behav, off_t value) 2025 { 2026 vm_map_entry_t current, entry; 2027 int modify_map = 0; 2028 int error = 0; 2029 int count; 2030 2031 /* 2032 * Some madvise calls directly modify the vm_map_entry, in which case 2033 * we need to use an exclusive lock on the map and we need to perform 2034 * various clipping operations. Otherwise we only need a read-lock 2035 * on the map. 2036 */ 2037 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2038 2039 switch(behav) { 2040 case MADV_NORMAL: 2041 case MADV_SEQUENTIAL: 2042 case MADV_RANDOM: 2043 case MADV_NOSYNC: 2044 case MADV_AUTOSYNC: 2045 case MADV_NOCORE: 2046 case MADV_CORE: 2047 case MADV_SETMAP: 2048 modify_map = 1; 2049 vm_map_lock(map); 2050 break; 2051 case MADV_INVAL: 2052 case MADV_WILLNEED: 2053 case MADV_DONTNEED: 2054 case MADV_FREE: 2055 vm_map_lock_read(map); 2056 break; 2057 default: 2058 vm_map_entry_release(count); 2059 return (EINVAL); 2060 } 2061 2062 /* 2063 * Locate starting entry and clip if necessary. 2064 */ 2065 2066 VM_MAP_RANGE_CHECK(map, start, end); 2067 2068 if (vm_map_lookup_entry(map, start, &entry)) { 2069 if (modify_map) 2070 vm_map_clip_start(map, entry, start, &count); 2071 } else if (entry) { 2072 entry = vm_map_rb_tree_RB_NEXT(entry); 2073 } else { 2074 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 2075 } 2076 2077 if (modify_map) { 2078 /* 2079 * madvise behaviors that are implemented in the vm_map_entry. 2080 * 2081 * We clip the vm_map_entry so that behavioral changes are 2082 * limited to the specified address range. 2083 */ 2084 for (current = entry; 2085 current && current->start < end; 2086 current = vm_map_rb_tree_RB_NEXT(current)) { 2087 /* 2088 * Ignore submaps 2089 */ 2090 if (current->maptype == VM_MAPTYPE_SUBMAP) 2091 continue; 2092 2093 vm_map_clip_end(map, current, end, &count); 2094 2095 switch (behav) { 2096 case MADV_NORMAL: 2097 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 2098 break; 2099 case MADV_SEQUENTIAL: 2100 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 2101 break; 2102 case MADV_RANDOM: 2103 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 2104 break; 2105 case MADV_NOSYNC: 2106 current->eflags |= MAP_ENTRY_NOSYNC; 2107 break; 2108 case MADV_AUTOSYNC: 2109 current->eflags &= ~MAP_ENTRY_NOSYNC; 2110 break; 2111 case MADV_NOCORE: 2112 current->eflags |= MAP_ENTRY_NOCOREDUMP; 2113 break; 2114 case MADV_CORE: 2115 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 2116 break; 2117 case MADV_SETMAP: 2118 /* 2119 * Set the page directory page for a map 2120 * governed by a virtual page table. Mark 2121 * the entry as being governed by a virtual 2122 * page table if it is not. 2123 * 2124 * XXX the page directory page is stored 2125 * in the avail_ssize field if the map_entry. 2126 * 2127 * XXX the map simplification code does not 2128 * compare this field so weird things may 2129 * happen if you do not apply this function 2130 * to the entire mapping governed by the 2131 * virtual page table. 2132 */ 2133 if (current->maptype != VM_MAPTYPE_VPAGETABLE) { 2134 error = EINVAL; 2135 break; 2136 } 2137 current->aux.master_pde = value; 2138 pmap_remove(map->pmap, 2139 current->start, current->end); 2140 break; 2141 case MADV_INVAL: 2142 /* 2143 * Invalidate the related pmap entries, used 2144 * to flush portions of the real kernel's 2145 * pmap when the caller has removed or 2146 * modified existing mappings in a virtual 2147 * page table. 2148 * 2149 * (exclusive locked map version does not 2150 * need the range interlock). 2151 */ 2152 pmap_remove(map->pmap, 2153 current->start, current->end); 2154 break; 2155 default: 2156 error = EINVAL; 2157 break; 2158 } 2159 vm_map_simplify_entry(map, current, &count); 2160 } 2161 vm_map_unlock(map); 2162 } else { 2163 vm_pindex_t pindex; 2164 vm_pindex_t delta; 2165 2166 /* 2167 * madvise behaviors that are implemented in the underlying 2168 * vm_object. 2169 * 2170 * Since we don't clip the vm_map_entry, we have to clip 2171 * the vm_object pindex and count. 2172 * 2173 * NOTE! These functions are only supported on normal maps, 2174 * except MADV_INVAL which is also supported on 2175 * virtual page tables. 2176 */ 2177 for (current = entry; 2178 current && current->start < end; 2179 current = vm_map_rb_tree_RB_NEXT(current)) { 2180 vm_offset_t useStart; 2181 2182 if (current->maptype != VM_MAPTYPE_NORMAL && 2183 (current->maptype != VM_MAPTYPE_VPAGETABLE || 2184 behav != MADV_INVAL)) { 2185 continue; 2186 } 2187 2188 pindex = OFF_TO_IDX(current->offset); 2189 delta = atop(current->end - current->start); 2190 useStart = current->start; 2191 2192 if (current->start < start) { 2193 pindex += atop(start - current->start); 2194 delta -= atop(start - current->start); 2195 useStart = start; 2196 } 2197 if (current->end > end) 2198 delta -= atop(current->end - end); 2199 2200 if ((vm_spindex_t)delta <= 0) 2201 continue; 2202 2203 if (behav == MADV_INVAL) { 2204 /* 2205 * Invalidate the related pmap entries, used 2206 * to flush portions of the real kernel's 2207 * pmap when the caller has removed or 2208 * modified existing mappings in a virtual 2209 * page table. 2210 * 2211 * (shared locked map version needs the 2212 * interlock, see vm_fault()). 2213 */ 2214 struct vm_map_ilock ilock; 2215 2216 KASSERT(useStart >= VM_MIN_USER_ADDRESS && 2217 useStart + ptoa(delta) <= 2218 VM_MAX_USER_ADDRESS, 2219 ("Bad range %016jx-%016jx (%016jx)", 2220 useStart, useStart + ptoa(delta), 2221 delta)); 2222 vm_map_interlock(map, &ilock, 2223 useStart, 2224 useStart + ptoa(delta)); 2225 pmap_remove(map->pmap, 2226 useStart, 2227 useStart + ptoa(delta)); 2228 vm_map_deinterlock(map, &ilock); 2229 } else { 2230 vm_object_madvise(current->object.vm_object, 2231 pindex, delta, behav); 2232 } 2233 2234 /* 2235 * Try to populate the page table. Mappings governed 2236 * by virtual page tables cannot be pre-populated 2237 * without a lot of work so don't try. 2238 */ 2239 if (behav == MADV_WILLNEED && 2240 current->maptype != VM_MAPTYPE_VPAGETABLE) { 2241 pmap_object_init_pt( 2242 map->pmap, 2243 useStart, 2244 current->protection, 2245 current->object.vm_object, 2246 pindex, 2247 (count << PAGE_SHIFT), 2248 MAP_PREFAULT_MADVISE 2249 ); 2250 } 2251 } 2252 vm_map_unlock_read(map); 2253 } 2254 vm_map_entry_release(count); 2255 return(error); 2256 } 2257 2258 2259 /* 2260 * Sets the inheritance of the specified address range in the target map. 2261 * Inheritance affects how the map will be shared with child maps at the 2262 * time of vm_map_fork. 2263 */ 2264 int 2265 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 2266 vm_inherit_t new_inheritance) 2267 { 2268 vm_map_entry_t entry; 2269 vm_map_entry_t temp_entry; 2270 int count; 2271 2272 switch (new_inheritance) { 2273 case VM_INHERIT_NONE: 2274 case VM_INHERIT_COPY: 2275 case VM_INHERIT_SHARE: 2276 break; 2277 default: 2278 return (KERN_INVALID_ARGUMENT); 2279 } 2280 2281 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2282 vm_map_lock(map); 2283 2284 VM_MAP_RANGE_CHECK(map, start, end); 2285 2286 if (vm_map_lookup_entry(map, start, &temp_entry)) { 2287 entry = temp_entry; 2288 vm_map_clip_start(map, entry, start, &count); 2289 } else if (temp_entry) { 2290 entry = vm_map_rb_tree_RB_NEXT(temp_entry); 2291 } else { 2292 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 2293 } 2294 2295 while (entry && entry->start < end) { 2296 vm_map_clip_end(map, entry, end, &count); 2297 2298 entry->inheritance = new_inheritance; 2299 2300 vm_map_simplify_entry(map, entry, &count); 2301 2302 entry = vm_map_rb_tree_RB_NEXT(entry); 2303 } 2304 vm_map_unlock(map); 2305 vm_map_entry_release(count); 2306 return (KERN_SUCCESS); 2307 } 2308 2309 /* 2310 * Implement the semantics of mlock 2311 */ 2312 int 2313 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 2314 boolean_t new_pageable) 2315 { 2316 vm_map_entry_t entry; 2317 vm_map_entry_t start_entry; 2318 vm_offset_t end; 2319 int rv = KERN_SUCCESS; 2320 int count; 2321 2322 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2323 vm_map_lock(map); 2324 VM_MAP_RANGE_CHECK(map, start, real_end); 2325 end = real_end; 2326 2327 start_entry = vm_map_clip_range(map, start, end, &count, 2328 MAP_CLIP_NO_HOLES); 2329 if (start_entry == NULL) { 2330 vm_map_unlock(map); 2331 vm_map_entry_release(count); 2332 return (KERN_INVALID_ADDRESS); 2333 } 2334 2335 if (new_pageable == 0) { 2336 entry = start_entry; 2337 while (entry && entry->start < end) { 2338 vm_offset_t save_start; 2339 vm_offset_t save_end; 2340 2341 /* 2342 * Already user wired or hard wired (trivial cases) 2343 */ 2344 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 2345 entry = vm_map_rb_tree_RB_NEXT(entry); 2346 continue; 2347 } 2348 if (entry->wired_count != 0) { 2349 entry->wired_count++; 2350 entry->eflags |= MAP_ENTRY_USER_WIRED; 2351 entry = vm_map_rb_tree_RB_NEXT(entry); 2352 continue; 2353 } 2354 2355 /* 2356 * A new wiring requires instantiation of appropriate 2357 * management structures and the faulting in of the 2358 * page. 2359 */ 2360 if (entry->maptype == VM_MAPTYPE_NORMAL || 2361 entry->maptype == VM_MAPTYPE_VPAGETABLE) { 2362 int copyflag = entry->eflags & 2363 MAP_ENTRY_NEEDS_COPY; 2364 if (copyflag && ((entry->protection & 2365 VM_PROT_WRITE) != 0)) { 2366 vm_map_entry_shadow(entry, 0); 2367 } else if (entry->object.vm_object == NULL && 2368 !map->system_map) { 2369 vm_map_entry_allocate_object(entry); 2370 } 2371 } 2372 entry->wired_count++; 2373 entry->eflags |= MAP_ENTRY_USER_WIRED; 2374 2375 /* 2376 * Now fault in the area. Note that vm_fault_wire() 2377 * may release the map lock temporarily, it will be 2378 * relocked on return. The in-transition 2379 * flag protects the entries. 2380 */ 2381 save_start = entry->start; 2382 save_end = entry->end; 2383 rv = vm_fault_wire(map, entry, TRUE, 0); 2384 if (rv) { 2385 CLIP_CHECK_BACK(entry, save_start); 2386 for (;;) { 2387 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 2388 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2389 entry->wired_count = 0; 2390 if (entry->end == save_end) 2391 break; 2392 entry = vm_map_rb_tree_RB_NEXT(entry); 2393 KASSERT(entry, 2394 ("bad entry clip during backout")); 2395 } 2396 end = save_start; /* unwire the rest */ 2397 break; 2398 } 2399 /* 2400 * note that even though the entry might have been 2401 * clipped, the USER_WIRED flag we set prevents 2402 * duplication so we do not have to do a 2403 * clip check. 2404 */ 2405 entry = vm_map_rb_tree_RB_NEXT(entry); 2406 } 2407 2408 /* 2409 * If we failed fall through to the unwiring section to 2410 * unwire what we had wired so far. 'end' has already 2411 * been adjusted. 2412 */ 2413 if (rv) 2414 new_pageable = 1; 2415 2416 /* 2417 * start_entry might have been clipped if we unlocked the 2418 * map and blocked. No matter how clipped it has gotten 2419 * there should be a fragment that is on our start boundary. 2420 */ 2421 CLIP_CHECK_BACK(start_entry, start); 2422 } 2423 2424 /* 2425 * Deal with the unwiring case. 2426 */ 2427 if (new_pageable) { 2428 /* 2429 * This is the unwiring case. We must first ensure that the 2430 * range to be unwired is really wired down. We know there 2431 * are no holes. 2432 */ 2433 entry = start_entry; 2434 while (entry && entry->start < end) { 2435 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 2436 rv = KERN_INVALID_ARGUMENT; 2437 goto done; 2438 } 2439 KASSERT(entry->wired_count != 0, 2440 ("wired count was 0 with USER_WIRED set! %p", 2441 entry)); 2442 entry = vm_map_rb_tree_RB_NEXT(entry); 2443 } 2444 2445 /* 2446 * Now decrement the wiring count for each region. If a region 2447 * becomes completely unwired, unwire its physical pages and 2448 * mappings. 2449 */ 2450 /* 2451 * The map entries are processed in a loop, checking to 2452 * make sure the entry is wired and asserting it has a wired 2453 * count. However, another loop was inserted more-or-less in 2454 * the middle of the unwiring path. This loop picks up the 2455 * "entry" loop variable from the first loop without first 2456 * setting it to start_entry. Naturally, the secound loop 2457 * is never entered and the pages backing the entries are 2458 * never unwired. This can lead to a leak of wired pages. 2459 */ 2460 entry = start_entry; 2461 while (entry && entry->start < end) { 2462 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 2463 ("expected USER_WIRED on entry %p", entry)); 2464 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2465 entry->wired_count--; 2466 if (entry->wired_count == 0) 2467 vm_fault_unwire(map, entry); 2468 entry = vm_map_rb_tree_RB_NEXT(entry); 2469 } 2470 } 2471 done: 2472 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2473 MAP_CLIP_NO_HOLES); 2474 vm_map_unlock(map); 2475 vm_map_entry_release(count); 2476 2477 return (rv); 2478 } 2479 2480 /* 2481 * Sets the pageability of the specified address range in the target map. 2482 * Regions specified as not pageable require locked-down physical 2483 * memory and physical page maps. 2484 * 2485 * The map must not be locked, but a reference must remain to the map 2486 * throughout the call. 2487 * 2488 * This function may be called via the zalloc path and must properly 2489 * reserve map entries for kernel_map. 2490 * 2491 * No requirements. 2492 */ 2493 int 2494 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags) 2495 { 2496 vm_map_entry_t entry; 2497 vm_map_entry_t start_entry; 2498 vm_offset_t end; 2499 int rv = KERN_SUCCESS; 2500 int count; 2501 2502 if (kmflags & KM_KRESERVE) 2503 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 2504 else 2505 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2506 vm_map_lock(map); 2507 VM_MAP_RANGE_CHECK(map, start, real_end); 2508 end = real_end; 2509 2510 start_entry = vm_map_clip_range(map, start, end, &count, 2511 MAP_CLIP_NO_HOLES); 2512 if (start_entry == NULL) { 2513 vm_map_unlock(map); 2514 rv = KERN_INVALID_ADDRESS; 2515 goto failure; 2516 } 2517 if ((kmflags & KM_PAGEABLE) == 0) { 2518 /* 2519 * Wiring. 2520 * 2521 * 1. Holding the write lock, we create any shadow or zero-fill 2522 * objects that need to be created. Then we clip each map 2523 * entry to the region to be wired and increment its wiring 2524 * count. We create objects before clipping the map entries 2525 * to avoid object proliferation. 2526 * 2527 * 2. We downgrade to a read lock, and call vm_fault_wire to 2528 * fault in the pages for any newly wired area (wired_count is 2529 * 1). 2530 * 2531 * Downgrading to a read lock for vm_fault_wire avoids a 2532 * possible deadlock with another process that may have faulted 2533 * on one of the pages to be wired (it would mark the page busy, 2534 * blocking us, then in turn block on the map lock that we 2535 * hold). Because of problems in the recursive lock package, 2536 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 2537 * any actions that require the write lock must be done 2538 * beforehand. Because we keep the read lock on the map, the 2539 * copy-on-write status of the entries we modify here cannot 2540 * change. 2541 */ 2542 entry = start_entry; 2543 while (entry && entry->start < end) { 2544 /* 2545 * Trivial case if the entry is already wired 2546 */ 2547 if (entry->wired_count) { 2548 entry->wired_count++; 2549 entry = vm_map_rb_tree_RB_NEXT(entry); 2550 continue; 2551 } 2552 2553 /* 2554 * The entry is being newly wired, we have to setup 2555 * appropriate management structures. A shadow 2556 * object is required for a copy-on-write region, 2557 * or a normal object for a zero-fill region. We 2558 * do not have to do this for entries that point to sub 2559 * maps because we won't hold the lock on the sub map. 2560 */ 2561 if (entry->maptype == VM_MAPTYPE_NORMAL || 2562 entry->maptype == VM_MAPTYPE_VPAGETABLE) { 2563 int copyflag = entry->eflags & 2564 MAP_ENTRY_NEEDS_COPY; 2565 if (copyflag && ((entry->protection & 2566 VM_PROT_WRITE) != 0)) { 2567 vm_map_entry_shadow(entry, 0); 2568 } else if (entry->object.vm_object == NULL && 2569 !map->system_map) { 2570 vm_map_entry_allocate_object(entry); 2571 } 2572 } 2573 entry->wired_count++; 2574 entry = vm_map_rb_tree_RB_NEXT(entry); 2575 } 2576 2577 /* 2578 * Pass 2. 2579 */ 2580 2581 /* 2582 * HACK HACK HACK HACK 2583 * 2584 * vm_fault_wire() temporarily unlocks the map to avoid 2585 * deadlocks. The in-transition flag from vm_map_clip_range 2586 * call should protect us from changes while the map is 2587 * unlocked. T 2588 * 2589 * NOTE: Previously this comment stated that clipping might 2590 * still occur while the entry is unlocked, but from 2591 * what I can tell it actually cannot. 2592 * 2593 * It is unclear whether the CLIP_CHECK_*() calls 2594 * are still needed but we keep them in anyway. 2595 * 2596 * HACK HACK HACK HACK 2597 */ 2598 2599 entry = start_entry; 2600 while (entry && entry->start < end) { 2601 /* 2602 * If vm_fault_wire fails for any page we need to undo 2603 * what has been done. We decrement the wiring count 2604 * for those pages which have not yet been wired (now) 2605 * and unwire those that have (later). 2606 */ 2607 vm_offset_t save_start = entry->start; 2608 vm_offset_t save_end = entry->end; 2609 2610 if (entry->wired_count == 1) 2611 rv = vm_fault_wire(map, entry, FALSE, kmflags); 2612 if (rv) { 2613 CLIP_CHECK_BACK(entry, save_start); 2614 for (;;) { 2615 KASSERT(entry->wired_count == 1, 2616 ("wired_count changed unexpectedly")); 2617 entry->wired_count = 0; 2618 if (entry->end == save_end) 2619 break; 2620 entry = vm_map_rb_tree_RB_NEXT(entry); 2621 KASSERT(entry, 2622 ("bad entry clip during backout")); 2623 } 2624 end = save_start; 2625 break; 2626 } 2627 CLIP_CHECK_FWD(entry, save_end); 2628 entry = vm_map_rb_tree_RB_NEXT(entry); 2629 } 2630 2631 /* 2632 * If a failure occured undo everything by falling through 2633 * to the unwiring code. 'end' has already been adjusted 2634 * appropriately. 2635 */ 2636 if (rv) 2637 kmflags |= KM_PAGEABLE; 2638 2639 /* 2640 * start_entry is still IN_TRANSITION but may have been 2641 * clipped since vm_fault_wire() unlocks and relocks the 2642 * map. No matter how clipped it has gotten there should 2643 * be a fragment that is on our start boundary. 2644 */ 2645 CLIP_CHECK_BACK(start_entry, start); 2646 } 2647 2648 if (kmflags & KM_PAGEABLE) { 2649 /* 2650 * This is the unwiring case. We must first ensure that the 2651 * range to be unwired is really wired down. We know there 2652 * are no holes. 2653 */ 2654 entry = start_entry; 2655 while (entry && entry->start < end) { 2656 if (entry->wired_count == 0) { 2657 rv = KERN_INVALID_ARGUMENT; 2658 goto done; 2659 } 2660 entry = vm_map_rb_tree_RB_NEXT(entry); 2661 } 2662 2663 /* 2664 * Now decrement the wiring count for each region. If a region 2665 * becomes completely unwired, unwire its physical pages and 2666 * mappings. 2667 */ 2668 entry = start_entry; 2669 while (entry && entry->start < end) { 2670 entry->wired_count--; 2671 if (entry->wired_count == 0) 2672 vm_fault_unwire(map, entry); 2673 entry = vm_map_rb_tree_RB_NEXT(entry); 2674 } 2675 } 2676 done: 2677 vm_map_unclip_range(map, start_entry, start, real_end, 2678 &count, MAP_CLIP_NO_HOLES); 2679 vm_map_unlock(map); 2680 failure: 2681 if (kmflags & KM_KRESERVE) 2682 vm_map_entry_krelease(count); 2683 else 2684 vm_map_entry_release(count); 2685 return (rv); 2686 } 2687 2688 /* 2689 * Mark a newly allocated address range as wired but do not fault in 2690 * the pages. The caller is expected to load the pages into the object. 2691 * 2692 * The map must be locked on entry and will remain locked on return. 2693 * No other requirements. 2694 */ 2695 void 2696 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, 2697 int *countp) 2698 { 2699 vm_map_entry_t scan; 2700 vm_map_entry_t entry; 2701 2702 entry = vm_map_clip_range(map, addr, addr + size, 2703 countp, MAP_CLIP_NO_HOLES); 2704 scan = entry; 2705 while (scan && scan->start < addr + size) { 2706 KKASSERT(scan->wired_count == 0); 2707 scan->wired_count = 1; 2708 scan = vm_map_rb_tree_RB_NEXT(scan); 2709 } 2710 vm_map_unclip_range(map, entry, addr, addr + size, 2711 countp, MAP_CLIP_NO_HOLES); 2712 } 2713 2714 /* 2715 * Push any dirty cached pages in the address range to their pager. 2716 * If syncio is TRUE, dirty pages are written synchronously. 2717 * If invalidate is TRUE, any cached pages are freed as well. 2718 * 2719 * This routine is called by sys_msync() 2720 * 2721 * Returns an error if any part of the specified range is not mapped. 2722 * 2723 * No requirements. 2724 */ 2725 int 2726 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, 2727 boolean_t syncio, boolean_t invalidate) 2728 { 2729 vm_map_entry_t current; 2730 vm_map_entry_t next; 2731 vm_map_entry_t entry; 2732 vm_size_t size; 2733 vm_object_t object; 2734 vm_object_t tobj; 2735 vm_ooffset_t offset; 2736 2737 vm_map_lock_read(map); 2738 VM_MAP_RANGE_CHECK(map, start, end); 2739 if (!vm_map_lookup_entry(map, start, &entry)) { 2740 vm_map_unlock_read(map); 2741 return (KERN_INVALID_ADDRESS); 2742 } 2743 lwkt_gettoken(&map->token); 2744 2745 /* 2746 * Make a first pass to check for holes. 2747 */ 2748 current = entry; 2749 while (current && current->start < end) { 2750 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2751 lwkt_reltoken(&map->token); 2752 vm_map_unlock_read(map); 2753 return (KERN_INVALID_ARGUMENT); 2754 } 2755 next = vm_map_rb_tree_RB_NEXT(current); 2756 if (end > current->end && 2757 (next == NULL || 2758 current->end != next->start)) { 2759 lwkt_reltoken(&map->token); 2760 vm_map_unlock_read(map); 2761 return (KERN_INVALID_ADDRESS); 2762 } 2763 current = next; 2764 } 2765 2766 if (invalidate) 2767 pmap_remove(vm_map_pmap(map), start, end); 2768 2769 /* 2770 * Make a second pass, cleaning/uncaching pages from the indicated 2771 * objects as we go. 2772 */ 2773 current = entry; 2774 while (current && current->start < end) { 2775 offset = current->offset + (start - current->start); 2776 size = (end <= current->end ? end : current->end) - start; 2777 2778 switch(current->maptype) { 2779 case VM_MAPTYPE_SUBMAP: 2780 { 2781 vm_map_t smap; 2782 vm_map_entry_t tentry; 2783 vm_size_t tsize; 2784 2785 smap = current->object.sub_map; 2786 vm_map_lock_read(smap); 2787 vm_map_lookup_entry(smap, offset, &tentry); 2788 if (tentry == NULL) { 2789 tsize = vm_map_max(smap) - offset; 2790 object = NULL; 2791 offset = 0 + (offset - vm_map_min(smap)); 2792 } else { 2793 tsize = tentry->end - offset; 2794 object = tentry->object.vm_object; 2795 offset = tentry->offset + 2796 (offset - tentry->start); 2797 } 2798 vm_map_unlock_read(smap); 2799 if (tsize < size) 2800 size = tsize; 2801 break; 2802 } 2803 case VM_MAPTYPE_NORMAL: 2804 case VM_MAPTYPE_VPAGETABLE: 2805 object = current->object.vm_object; 2806 break; 2807 default: 2808 object = NULL; 2809 break; 2810 } 2811 2812 if (object) 2813 vm_object_hold(object); 2814 2815 /* 2816 * Note that there is absolutely no sense in writing out 2817 * anonymous objects, so we track down the vnode object 2818 * to write out. 2819 * We invalidate (remove) all pages from the address space 2820 * anyway, for semantic correctness. 2821 * 2822 * note: certain anonymous maps, such as MAP_NOSYNC maps, 2823 * may start out with a NULL object. 2824 */ 2825 while (object && (tobj = object->backing_object) != NULL) { 2826 vm_object_hold(tobj); 2827 if (tobj == object->backing_object) { 2828 vm_object_lock_swap(); 2829 offset += object->backing_object_offset; 2830 vm_object_drop(object); 2831 object = tobj; 2832 if (object->size < OFF_TO_IDX(offset + size)) 2833 size = IDX_TO_OFF(object->size) - 2834 offset; 2835 break; 2836 } 2837 vm_object_drop(tobj); 2838 } 2839 if (object && (object->type == OBJT_VNODE) && 2840 (current->protection & VM_PROT_WRITE) && 2841 (object->flags & OBJ_NOMSYNC) == 0) { 2842 /* 2843 * Flush pages if writing is allowed, invalidate them 2844 * if invalidation requested. Pages undergoing I/O 2845 * will be ignored by vm_object_page_remove(). 2846 * 2847 * We cannot lock the vnode and then wait for paging 2848 * to complete without deadlocking against vm_fault. 2849 * Instead we simply call vm_object_page_remove() and 2850 * allow it to block internally on a page-by-page 2851 * basis when it encounters pages undergoing async 2852 * I/O. 2853 */ 2854 int flags; 2855 2856 /* no chain wait needed for vnode objects */ 2857 vm_object_reference_locked(object); 2858 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY); 2859 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 2860 flags |= invalidate ? OBJPC_INVAL : 0; 2861 2862 /* 2863 * When operating on a virtual page table just 2864 * flush the whole object. XXX we probably ought 2865 * to 2866 */ 2867 switch(current->maptype) { 2868 case VM_MAPTYPE_NORMAL: 2869 vm_object_page_clean(object, 2870 OFF_TO_IDX(offset), 2871 OFF_TO_IDX(offset + size + PAGE_MASK), 2872 flags); 2873 break; 2874 case VM_MAPTYPE_VPAGETABLE: 2875 vm_object_page_clean(object, 0, 0, flags); 2876 break; 2877 } 2878 vn_unlock(((struct vnode *)object->handle)); 2879 vm_object_deallocate_locked(object); 2880 } 2881 if (object && invalidate && 2882 ((object->type == OBJT_VNODE) || 2883 (object->type == OBJT_DEVICE) || 2884 (object->type == OBJT_MGTDEVICE))) { 2885 int clean_only = 2886 ((object->type == OBJT_DEVICE) || 2887 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE; 2888 /* no chain wait needed for vnode/device objects */ 2889 vm_object_reference_locked(object); 2890 switch(current->maptype) { 2891 case VM_MAPTYPE_NORMAL: 2892 vm_object_page_remove(object, 2893 OFF_TO_IDX(offset), 2894 OFF_TO_IDX(offset + size + PAGE_MASK), 2895 clean_only); 2896 break; 2897 case VM_MAPTYPE_VPAGETABLE: 2898 vm_object_page_remove(object, 0, 0, clean_only); 2899 break; 2900 } 2901 vm_object_deallocate_locked(object); 2902 } 2903 start += size; 2904 if (object) 2905 vm_object_drop(object); 2906 current = vm_map_rb_tree_RB_NEXT(current); 2907 } 2908 2909 lwkt_reltoken(&map->token); 2910 vm_map_unlock_read(map); 2911 2912 return (KERN_SUCCESS); 2913 } 2914 2915 /* 2916 * Make the region specified by this entry pageable. 2917 * 2918 * The vm_map must be exclusively locked. 2919 */ 2920 static void 2921 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 2922 { 2923 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2924 entry->wired_count = 0; 2925 vm_fault_unwire(map, entry); 2926 } 2927 2928 /* 2929 * Deallocate the given entry from the target map. 2930 * 2931 * The vm_map must be exclusively locked. 2932 */ 2933 static void 2934 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 2935 { 2936 vm_map_entry_unlink(map, entry); 2937 map->size -= entry->end - entry->start; 2938 2939 switch(entry->maptype) { 2940 case VM_MAPTYPE_NORMAL: 2941 case VM_MAPTYPE_VPAGETABLE: 2942 case VM_MAPTYPE_SUBMAP: 2943 vm_object_deallocate(entry->object.vm_object); 2944 break; 2945 case VM_MAPTYPE_UKSMAP: 2946 /* XXX TODO */ 2947 break; 2948 default: 2949 break; 2950 } 2951 2952 vm_map_entry_dispose(map, entry, countp); 2953 } 2954 2955 /* 2956 * Deallocates the given address range from the target map. 2957 * 2958 * The vm_map must be exclusively locked. 2959 */ 2960 int 2961 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 2962 { 2963 vm_object_t object; 2964 vm_map_entry_t entry; 2965 vm_map_entry_t first_entry; 2966 vm_offset_t hole_start; 2967 2968 ASSERT_VM_MAP_LOCKED(map); 2969 lwkt_gettoken(&map->token); 2970 again: 2971 /* 2972 * Find the start of the region, and clip it. Set entry to point 2973 * at the first record containing the requested address or, if no 2974 * such record exists, the next record with a greater address. The 2975 * loop will run from this point until a record beyond the termination 2976 * address is encountered. 2977 * 2978 * Adjust freehint[] for either the clip case or the extension case. 2979 * 2980 * GGG see other GGG comment. 2981 */ 2982 if (vm_map_lookup_entry(map, start, &first_entry)) { 2983 entry = first_entry; 2984 vm_map_clip_start(map, entry, start, countp); 2985 hole_start = start; 2986 } else { 2987 if (first_entry) { 2988 entry = vm_map_rb_tree_RB_NEXT(first_entry); 2989 if (entry == NULL) 2990 hole_start = first_entry->start; 2991 else 2992 hole_start = first_entry->end; 2993 } else { 2994 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 2995 if (entry == NULL) 2996 hole_start = vm_map_min(map); 2997 else 2998 hole_start = vm_map_max(map); 2999 } 3000 } 3001 3002 /* 3003 * Step through all entries in this region 3004 */ 3005 while (entry && entry->start < end) { 3006 vm_map_entry_t next; 3007 vm_offset_t s, e; 3008 vm_pindex_t offidxstart, offidxend, count; 3009 3010 /* 3011 * If we hit an in-transition entry we have to sleep and 3012 * retry. It's easier (and not really slower) to just retry 3013 * since this case occurs so rarely and the hint is already 3014 * pointing at the right place. We have to reset the 3015 * start offset so as not to accidently delete an entry 3016 * another process just created in vacated space. 3017 */ 3018 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3019 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3020 start = entry->start; 3021 ++mycpu->gd_cnt.v_intrans_coll; 3022 ++mycpu->gd_cnt.v_intrans_wait; 3023 vm_map_transition_wait(map, 1); 3024 goto again; 3025 } 3026 vm_map_clip_end(map, entry, end, countp); 3027 3028 s = entry->start; 3029 e = entry->end; 3030 next = vm_map_rb_tree_RB_NEXT(entry); 3031 3032 offidxstart = OFF_TO_IDX(entry->offset); 3033 count = OFF_TO_IDX(e - s); 3034 3035 switch(entry->maptype) { 3036 case VM_MAPTYPE_NORMAL: 3037 case VM_MAPTYPE_VPAGETABLE: 3038 case VM_MAPTYPE_SUBMAP: 3039 object = entry->object.vm_object; 3040 break; 3041 default: 3042 object = NULL; 3043 break; 3044 } 3045 3046 /* 3047 * Unwire before removing addresses from the pmap; otherwise, 3048 * unwiring will put the entries back in the pmap. 3049 * 3050 * Generally speaking, doing a bulk pmap_remove() before 3051 * removing the pages from the VM object is better at 3052 * reducing unnecessary IPIs. The pmap code is now optimized 3053 * to not blindly iterate the range when pt and pd pages 3054 * are missing. 3055 */ 3056 if (entry->wired_count != 0) 3057 vm_map_entry_unwire(map, entry); 3058 3059 offidxend = offidxstart + count; 3060 3061 if (object == &kernel_object) { 3062 pmap_remove(map->pmap, s, e); 3063 vm_object_hold(object); 3064 vm_object_page_remove(object, offidxstart, 3065 offidxend, FALSE); 3066 vm_object_drop(object); 3067 } else if (object && object->type != OBJT_DEFAULT && 3068 object->type != OBJT_SWAP) { 3069 /* 3070 * vnode object routines cannot be chain-locked, 3071 * but since we aren't removing pages from the 3072 * object here we can use a shared hold. 3073 */ 3074 vm_object_hold_shared(object); 3075 pmap_remove(map->pmap, s, e); 3076 vm_object_drop(object); 3077 } else if (object) { 3078 vm_object_hold(object); 3079 vm_object_chain_acquire(object, 0); 3080 pmap_remove(map->pmap, s, e); 3081 3082 if (object != NULL && 3083 object->ref_count != 1 && 3084 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == 3085 OBJ_ONEMAPPING && 3086 (object->type == OBJT_DEFAULT || 3087 object->type == OBJT_SWAP)) { 3088 /* 3089 * When ONEMAPPING is set we can destroy the 3090 * pages underlying the entry's range. 3091 */ 3092 vm_object_collapse(object, NULL); 3093 vm_object_page_remove(object, offidxstart, 3094 offidxend, FALSE); 3095 if (object->type == OBJT_SWAP) { 3096 swap_pager_freespace(object, 3097 offidxstart, 3098 count); 3099 } 3100 if (offidxend >= object->size && 3101 offidxstart < object->size) { 3102 object->size = offidxstart; 3103 } 3104 } 3105 vm_object_chain_release(object); 3106 vm_object_drop(object); 3107 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) { 3108 pmap_remove(map->pmap, s, e); 3109 } 3110 3111 /* 3112 * Delete the entry (which may delete the object) only after 3113 * removing all pmap entries pointing to its pages. 3114 * (Otherwise, its page frames may be reallocated, and any 3115 * modify bits will be set in the wrong object!) 3116 */ 3117 vm_map_entry_delete(map, entry, countp); 3118 entry = next; 3119 } 3120 3121 /* 3122 * We either reached the end and use vm_map_max as the end 3123 * address, or we didn't and we use the next entry as the 3124 * end address. 3125 */ 3126 if (entry == NULL) { 3127 vm_map_freehint_hole(map, hole_start, 3128 vm_map_max(map) - hole_start); 3129 } else { 3130 vm_map_freehint_hole(map, hole_start, 3131 entry->start - hole_start); 3132 } 3133 3134 lwkt_reltoken(&map->token); 3135 3136 return (KERN_SUCCESS); 3137 } 3138 3139 /* 3140 * Remove the given address range from the target map. 3141 * This is the exported form of vm_map_delete. 3142 * 3143 * No requirements. 3144 */ 3145 int 3146 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 3147 { 3148 int result; 3149 int count; 3150 3151 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3152 vm_map_lock(map); 3153 VM_MAP_RANGE_CHECK(map, start, end); 3154 result = vm_map_delete(map, start, end, &count); 3155 vm_map_unlock(map); 3156 vm_map_entry_release(count); 3157 3158 return (result); 3159 } 3160 3161 /* 3162 * Assert that the target map allows the specified privilege on the 3163 * entire address region given. The entire region must be allocated. 3164 * 3165 * The caller must specify whether the vm_map is already locked or not. 3166 */ 3167 boolean_t 3168 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 3169 vm_prot_t protection, boolean_t have_lock) 3170 { 3171 vm_map_entry_t entry; 3172 vm_map_entry_t tmp_entry; 3173 boolean_t result; 3174 3175 if (have_lock == FALSE) 3176 vm_map_lock_read(map); 3177 3178 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 3179 if (have_lock == FALSE) 3180 vm_map_unlock_read(map); 3181 return (FALSE); 3182 } 3183 entry = tmp_entry; 3184 3185 result = TRUE; 3186 while (start < end) { 3187 if (entry == NULL) { 3188 result = FALSE; 3189 break; 3190 } 3191 3192 /* 3193 * No holes allowed! 3194 */ 3195 3196 if (start < entry->start) { 3197 result = FALSE; 3198 break; 3199 } 3200 /* 3201 * Check protection associated with entry. 3202 */ 3203 3204 if ((entry->protection & protection) != protection) { 3205 result = FALSE; 3206 break; 3207 } 3208 /* go to next entry */ 3209 start = entry->end; 3210 entry = vm_map_rb_tree_RB_NEXT(entry); 3211 } 3212 if (have_lock == FALSE) 3213 vm_map_unlock_read(map); 3214 return (result); 3215 } 3216 3217 /* 3218 * If appropriate this function shadows the original object with a new object 3219 * and moves the VM pages from the original object to the new object. 3220 * The original object will also be collapsed, if possible. 3221 * 3222 * Caller must supply entry->object.vm_object held and chain_acquired, and 3223 * should chain_release and drop the object upon return. 3224 * 3225 * We can only do this for normal memory objects with a single mapping, and 3226 * it only makes sense to do it if there are 2 or more refs on the original 3227 * object. i.e. typically a memory object that has been extended into 3228 * multiple vm_map_entry's with non-overlapping ranges. 3229 * 3230 * This makes it easier to remove unused pages and keeps object inheritance 3231 * from being a negative impact on memory usage. 3232 * 3233 * On return the (possibly new) entry->object.vm_object will have an 3234 * additional ref on it for the caller to dispose of (usually by cloning 3235 * the vm_map_entry). The additional ref had to be done in this routine 3236 * to avoid racing a collapse. The object's ONEMAPPING flag will also be 3237 * cleared. 3238 * 3239 * The vm_map must be locked and its token held. 3240 */ 3241 static void 3242 vm_map_split(vm_map_entry_t entry, vm_object_t oobject) 3243 { 3244 /* OPTIMIZED */ 3245 vm_object_t nobject, bobject; 3246 vm_offset_t s, e; 3247 vm_page_t m; 3248 vm_pindex_t offidxstart, offidxend, idx; 3249 vm_size_t size; 3250 vm_ooffset_t offset; 3251 int useshadowlist; 3252 3253 /* 3254 * Optimize away object locks for vnode objects. Important exit/exec 3255 * critical path. 3256 * 3257 * OBJ_ONEMAPPING doesn't apply to vnode objects but clear the flag 3258 * anyway. 3259 */ 3260 if (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) { 3261 vm_object_reference_quick(oobject); 3262 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3263 return; 3264 } 3265 3266 #if 0 3267 /* 3268 * Original object cannot be split? 3269 */ 3270 if (oobject->handle == NULL) { 3271 vm_object_reference_locked_chain_held(oobject); 3272 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3273 return; 3274 } 3275 #endif 3276 3277 /* 3278 * Collapse original object with its backing store as an 3279 * optimization to reduce chain lengths when possible. 3280 * 3281 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's 3282 * for oobject, so there's no point collapsing it. 3283 * 3284 * Then re-check whether the object can be split. 3285 */ 3286 vm_object_collapse(oobject, NULL); 3287 3288 if (oobject->ref_count <= 1 || 3289 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) || 3290 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) { 3291 vm_object_reference_locked_chain_held(oobject); 3292 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3293 return; 3294 } 3295 3296 /* 3297 * Acquire the chain lock on the backing object. 3298 * 3299 * Give bobject an additional ref count for when it will be shadowed 3300 * by nobject. 3301 */ 3302 useshadowlist = 0; 3303 if ((bobject = oobject->backing_object) != NULL) { 3304 if (bobject->type != OBJT_VNODE) { 3305 useshadowlist = 1; 3306 vm_object_hold(bobject); 3307 vm_object_chain_wait(bobject, 0); 3308 /* ref for shadowing below */ 3309 vm_object_reference_locked(bobject); 3310 vm_object_chain_acquire(bobject, 0); 3311 KKASSERT(oobject->backing_object == bobject); 3312 KKASSERT((bobject->flags & OBJ_DEAD) == 0); 3313 } else { 3314 /* 3315 * vnodes are not placed on the shadow list but 3316 * they still get another ref for the backing_object 3317 * reference. 3318 */ 3319 vm_object_reference_quick(bobject); 3320 } 3321 } 3322 3323 /* 3324 * Calculate the object page range and allocate the new object. 3325 */ 3326 offset = entry->offset; 3327 s = entry->start; 3328 e = entry->end; 3329 3330 offidxstart = OFF_TO_IDX(offset); 3331 offidxend = offidxstart + OFF_TO_IDX(e - s); 3332 size = offidxend - offidxstart; 3333 3334 switch(oobject->type) { 3335 case OBJT_DEFAULT: 3336 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size), 3337 VM_PROT_ALL, 0); 3338 break; 3339 case OBJT_SWAP: 3340 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size), 3341 VM_PROT_ALL, 0); 3342 break; 3343 default: 3344 /* not reached */ 3345 nobject = NULL; 3346 KKASSERT(0); 3347 } 3348 3349 /* 3350 * If we could not allocate nobject just clear ONEMAPPING on 3351 * oobject and return. 3352 */ 3353 if (nobject == NULL) { 3354 if (bobject) { 3355 if (useshadowlist) { 3356 vm_object_chain_release(bobject); 3357 vm_object_deallocate(bobject); 3358 vm_object_drop(bobject); 3359 } else { 3360 vm_object_deallocate(bobject); 3361 } 3362 } 3363 vm_object_reference_locked_chain_held(oobject); 3364 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3365 return; 3366 } 3367 3368 /* 3369 * The new object will replace entry->object.vm_object so it needs 3370 * a second reference (the caller expects an additional ref). 3371 */ 3372 vm_object_hold(nobject); 3373 vm_object_reference_locked(nobject); 3374 vm_object_chain_acquire(nobject, 0); 3375 3376 /* 3377 * nobject shadows bobject (oobject already shadows bobject). 3378 * 3379 * Adding an object to bobject's shadow list requires refing bobject 3380 * which we did above in the useshadowlist case. 3381 * 3382 * XXX it is unclear if we need to clear ONEMAPPING on bobject here 3383 * or not. 3384 */ 3385 if (bobject) { 3386 nobject->backing_object_offset = 3387 oobject->backing_object_offset + IDX_TO_OFF(offidxstart); 3388 nobject->backing_object = bobject; 3389 if (useshadowlist) { 3390 bobject->shadow_count++; 3391 atomic_add_int(&bobject->generation, 1); 3392 LIST_INSERT_HEAD(&bobject->shadow_head, 3393 nobject, shadow_list); 3394 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /*XXX*/ 3395 vm_object_set_flag(nobject, OBJ_ONSHADOW); 3396 } 3397 } 3398 3399 /* 3400 * Move the VM pages from oobject to nobject 3401 */ 3402 for (idx = 0; idx < size; idx++) { 3403 vm_page_t m; 3404 3405 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx, 3406 TRUE, "vmpg"); 3407 if (m == NULL) 3408 continue; 3409 3410 /* 3411 * We must wait for pending I/O to complete before we can 3412 * rename the page. 3413 * 3414 * We do not have to VM_PROT_NONE the page as mappings should 3415 * not be changed by this operation. 3416 * 3417 * NOTE: The act of renaming a page updates chaingen for both 3418 * objects. 3419 */ 3420 vm_page_rename(m, nobject, idx); 3421 /* page automatically made dirty by rename and cache handled */ 3422 /* page remains busy */ 3423 } 3424 3425 if (oobject->type == OBJT_SWAP) { 3426 vm_object_pip_add(oobject, 1); 3427 /* 3428 * copy oobject pages into nobject and destroy unneeded 3429 * pages in shadow object. 3430 */ 3431 swap_pager_copy(oobject, nobject, offidxstart, 0); 3432 vm_object_pip_wakeup(oobject); 3433 } 3434 3435 /* 3436 * Wakeup the pages we played with. No spl protection is needed 3437 * for a simple wakeup. 3438 */ 3439 for (idx = 0; idx < size; idx++) { 3440 m = vm_page_lookup(nobject, idx); 3441 if (m) { 3442 KKASSERT(m->busy_count & PBUSY_LOCKED); 3443 vm_page_wakeup(m); 3444 } 3445 } 3446 entry->object.vm_object = nobject; 3447 entry->offset = 0LL; 3448 3449 /* 3450 * The map is being split and nobject is going to wind up on both 3451 * vm_map_entry's, so make sure OBJ_ONEMAPPING is cleared on 3452 * nobject. 3453 */ 3454 vm_object_clear_flag(nobject, OBJ_ONEMAPPING); 3455 3456 /* 3457 * Cleanup 3458 * 3459 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the 3460 * related pages were moved and are no longer applicable to the 3461 * original object. 3462 * 3463 * NOTE: Deallocate oobject (due to its entry->object.vm_object being 3464 * replaced by nobject). 3465 */ 3466 vm_object_chain_release(nobject); 3467 vm_object_drop(nobject); 3468 if (bobject && useshadowlist) { 3469 vm_object_chain_release(bobject); 3470 vm_object_drop(bobject); 3471 } 3472 3473 #if 0 3474 if (oobject->resident_page_count) { 3475 kprintf("oobject %p still contains %jd pages!\n", 3476 oobject, (intmax_t)oobject->resident_page_count); 3477 for (idx = 0; idx < size; idx++) { 3478 vm_page_t m; 3479 3480 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx, 3481 TRUE, "vmpg"); 3482 if (m) { 3483 kprintf("oobject %p idx %jd\n", 3484 oobject, 3485 offidxstart + idx); 3486 vm_page_wakeup(m); 3487 } 3488 } 3489 } 3490 #endif 3491 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/ 3492 vm_object_deallocate_locked(oobject); 3493 } 3494 3495 /* 3496 * Copies the contents of the source entry to the destination 3497 * entry. The entries *must* be aligned properly. 3498 * 3499 * The vm_maps must be exclusively locked. 3500 * The vm_map's token must be held. 3501 * 3502 * Because the maps are locked no faults can be in progress during the 3503 * operation. 3504 */ 3505 static void 3506 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 3507 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 3508 { 3509 vm_object_t src_object; 3510 vm_object_t oobject; 3511 3512 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP || 3513 dst_entry->maptype == VM_MAPTYPE_UKSMAP) 3514 return; 3515 if (src_entry->maptype == VM_MAPTYPE_SUBMAP || 3516 src_entry->maptype == VM_MAPTYPE_UKSMAP) 3517 return; 3518 3519 if (src_entry->wired_count == 0) { 3520 /* 3521 * If the source entry is marked needs_copy, it is already 3522 * write-protected. 3523 * 3524 * To avoid interacting with a vm_fault that might have 3525 * released its vm_map, we must acquire the fronting 3526 * object. 3527 */ 3528 oobject = src_entry->object.vm_object; 3529 if (oobject) { 3530 vm_object_hold(oobject); 3531 vm_object_chain_acquire(oobject, 0); 3532 } 3533 3534 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 3535 pmap_protect(src_map->pmap, 3536 src_entry->start, 3537 src_entry->end, 3538 src_entry->protection & ~VM_PROT_WRITE); 3539 } 3540 3541 /* 3542 * Make a copy of the object. 3543 * 3544 * The object must be locked prior to checking the object type 3545 * and for the call to vm_object_collapse() and vm_map_split(). 3546 * We cannot use *_hold() here because the split code will 3547 * probably try to destroy the object. The lock is a pool 3548 * token and doesn't care. 3549 * 3550 * We must bump src_map->timestamp when setting 3551 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault 3552 * to retry, otherwise the concurrent fault might improperly 3553 * install a RW pte when its supposed to be a RO(COW) pte. 3554 * This race can occur because a vnode-backed fault may have 3555 * to temporarily release the map lock. This was handled 3556 * when the caller locked the map exclusively. 3557 */ 3558 if (oobject) { 3559 vm_map_split(src_entry, oobject); 3560 3561 src_object = src_entry->object.vm_object; 3562 dst_entry->object.vm_object = src_object; 3563 src_entry->eflags |= (MAP_ENTRY_COW | 3564 MAP_ENTRY_NEEDS_COPY); 3565 dst_entry->eflags |= (MAP_ENTRY_COW | 3566 MAP_ENTRY_NEEDS_COPY); 3567 dst_entry->offset = src_entry->offset; 3568 } else { 3569 dst_entry->object.vm_object = NULL; 3570 dst_entry->offset = 0; 3571 } 3572 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, 3573 dst_entry->end - dst_entry->start, 3574 src_entry->start); 3575 if (oobject) { 3576 vm_object_chain_release(oobject); 3577 vm_object_drop(oobject); 3578 } 3579 } else { 3580 /* 3581 * Of course, wired down pages can't be set copy-on-write. 3582 * Cause wired pages to be copied into the new map by 3583 * simulating faults (the new pages are pageable) 3584 */ 3585 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 3586 } 3587 } 3588 3589 /* 3590 * vmspace_fork: 3591 * Create a new process vmspace structure and vm_map 3592 * based on those of an existing process. The new map 3593 * is based on the old map, according to the inheritance 3594 * values on the regions in that map. 3595 * 3596 * The source map must not be locked. 3597 * No requirements. 3598 */ 3599 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map, 3600 vm_map_entry_t old_entry, int *countp); 3601 static void vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map, 3602 vm_map_entry_t old_entry, int *countp); 3603 3604 struct vmspace * 3605 vmspace_fork(struct vmspace *vm1) 3606 { 3607 struct vmspace *vm2; 3608 vm_map_t old_map = &vm1->vm_map; 3609 vm_map_t new_map; 3610 vm_map_entry_t old_entry; 3611 int count; 3612 3613 lwkt_gettoken(&vm1->vm_map.token); 3614 vm_map_lock(old_map); 3615 3616 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map)); 3617 lwkt_gettoken(&vm2->vm_map.token); 3618 3619 /* 3620 * We must bump the timestamp to force any concurrent fault 3621 * to retry. 3622 */ 3623 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 3624 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 3625 new_map = &vm2->vm_map; /* XXX */ 3626 new_map->timestamp = 1; 3627 3628 vm_map_lock(new_map); 3629 3630 count = old_map->nentries; 3631 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 3632 3633 RB_FOREACH(old_entry, vm_map_rb_tree, &old_map->rb_root) { 3634 switch(old_entry->maptype) { 3635 case VM_MAPTYPE_SUBMAP: 3636 panic("vm_map_fork: encountered a submap"); 3637 break; 3638 case VM_MAPTYPE_UKSMAP: 3639 vmspace_fork_uksmap_entry(old_map, new_map, 3640 old_entry, &count); 3641 break; 3642 case VM_MAPTYPE_NORMAL: 3643 case VM_MAPTYPE_VPAGETABLE: 3644 vmspace_fork_normal_entry(old_map, new_map, 3645 old_entry, &count); 3646 break; 3647 } 3648 } 3649 3650 new_map->size = old_map->size; 3651 vm_map_unlock(old_map); 3652 vm_map_unlock(new_map); 3653 vm_map_entry_release(count); 3654 3655 lwkt_reltoken(&vm2->vm_map.token); 3656 lwkt_reltoken(&vm1->vm_map.token); 3657 3658 return (vm2); 3659 } 3660 3661 static 3662 void 3663 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map, 3664 vm_map_entry_t old_entry, int *countp) 3665 { 3666 vm_map_entry_t new_entry; 3667 vm_object_t object; 3668 3669 switch (old_entry->inheritance) { 3670 case VM_INHERIT_NONE: 3671 break; 3672 case VM_INHERIT_SHARE: 3673 /* 3674 * Clone the entry, creating the shared object if 3675 * necessary. 3676 */ 3677 if (old_entry->object.vm_object == NULL) 3678 vm_map_entry_allocate_object(old_entry); 3679 3680 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3681 /* 3682 * Shadow a map_entry which needs a copy, 3683 * replacing its object with a new object 3684 * that points to the old one. Ask the 3685 * shadow code to automatically add an 3686 * additional ref. We can't do it afterwords 3687 * because we might race a collapse. The call 3688 * to vm_map_entry_shadow() will also clear 3689 * OBJ_ONEMAPPING. 3690 */ 3691 vm_map_entry_shadow(old_entry, 1); 3692 } else if (old_entry->object.vm_object) { 3693 /* 3694 * We will make a shared copy of the object, 3695 * and must clear OBJ_ONEMAPPING. 3696 * 3697 * Optimize vnode objects. OBJ_ONEMAPPING 3698 * is non-applicable but clear it anyway, 3699 * and its terminal so we don't have to deal 3700 * with chains. Reduces SMP conflicts. 3701 * 3702 * XXX assert that object.vm_object != NULL 3703 * since we allocate it above. 3704 */ 3705 object = old_entry->object.vm_object; 3706 if (object->type == OBJT_VNODE) { 3707 vm_object_reference_quick(object); 3708 vm_object_clear_flag(object, 3709 OBJ_ONEMAPPING); 3710 } else { 3711 vm_object_hold(object); 3712 vm_object_chain_wait(object, 0); 3713 vm_object_reference_locked(object); 3714 vm_object_clear_flag(object, 3715 OBJ_ONEMAPPING); 3716 vm_object_drop(object); 3717 } 3718 } 3719 3720 /* 3721 * Clone the entry. We've already bumped the ref on 3722 * any vm_object. 3723 */ 3724 new_entry = vm_map_entry_create(new_map, countp); 3725 *new_entry = *old_entry; 3726 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3727 new_entry->wired_count = 0; 3728 3729 /* 3730 * Insert the entry into the new map -- we know we're 3731 * inserting at the end of the new map. 3732 */ 3733 vm_map_entry_link(new_map, new_entry); 3734 3735 /* 3736 * Update the physical map 3737 */ 3738 pmap_copy(new_map->pmap, old_map->pmap, 3739 new_entry->start, 3740 (old_entry->end - old_entry->start), 3741 old_entry->start); 3742 break; 3743 case VM_INHERIT_COPY: 3744 /* 3745 * Clone the entry and link into the map. 3746 */ 3747 new_entry = vm_map_entry_create(new_map, countp); 3748 *new_entry = *old_entry; 3749 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3750 new_entry->wired_count = 0; 3751 new_entry->object.vm_object = NULL; 3752 vm_map_entry_link(new_map, new_entry); 3753 vm_map_copy_entry(old_map, new_map, old_entry, 3754 new_entry); 3755 break; 3756 } 3757 } 3758 3759 /* 3760 * When forking user-kernel shared maps, the map might change in the 3761 * child so do not try to copy the underlying pmap entries. 3762 */ 3763 static 3764 void 3765 vmspace_fork_uksmap_entry(vm_map_t old_map, vm_map_t new_map, 3766 vm_map_entry_t old_entry, int *countp) 3767 { 3768 vm_map_entry_t new_entry; 3769 3770 new_entry = vm_map_entry_create(new_map, countp); 3771 *new_entry = *old_entry; 3772 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3773 new_entry->wired_count = 0; 3774 vm_map_entry_link(new_map, new_entry); 3775 } 3776 3777 /* 3778 * Create an auto-grow stack entry 3779 * 3780 * No requirements. 3781 */ 3782 int 3783 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 3784 int flags, vm_prot_t prot, vm_prot_t max, int cow) 3785 { 3786 vm_map_entry_t prev_entry; 3787 vm_map_entry_t next; 3788 vm_size_t init_ssize; 3789 int rv; 3790 int count; 3791 vm_offset_t tmpaddr; 3792 3793 cow |= MAP_IS_STACK; 3794 3795 if (max_ssize < sgrowsiz) 3796 init_ssize = max_ssize; 3797 else 3798 init_ssize = sgrowsiz; 3799 3800 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3801 vm_map_lock(map); 3802 3803 /* 3804 * Find space for the mapping 3805 */ 3806 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) { 3807 if (vm_map_findspace(map, addrbos, max_ssize, 1, 3808 flags, &tmpaddr)) { 3809 vm_map_unlock(map); 3810 vm_map_entry_release(count); 3811 return (KERN_NO_SPACE); 3812 } 3813 addrbos = tmpaddr; 3814 } 3815 3816 /* If addr is already mapped, no go */ 3817 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) { 3818 vm_map_unlock(map); 3819 vm_map_entry_release(count); 3820 return (KERN_NO_SPACE); 3821 } 3822 3823 #if 0 3824 /* XXX already handled by kern_mmap() */ 3825 /* If we would blow our VMEM resource limit, no go */ 3826 if (map->size + init_ssize > 3827 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3828 vm_map_unlock(map); 3829 vm_map_entry_release(count); 3830 return (KERN_NO_SPACE); 3831 } 3832 #endif 3833 3834 /* 3835 * If we can't accomodate max_ssize in the current mapping, 3836 * no go. However, we need to be aware that subsequent user 3837 * mappings might map into the space we have reserved for 3838 * stack, and currently this space is not protected. 3839 * 3840 * Hopefully we will at least detect this condition 3841 * when we try to grow the stack. 3842 */ 3843 if (prev_entry) 3844 next = vm_map_rb_tree_RB_NEXT(prev_entry); 3845 else 3846 next = RB_MIN(vm_map_rb_tree, &map->rb_root); 3847 3848 if (next && next->start < addrbos + max_ssize) { 3849 vm_map_unlock(map); 3850 vm_map_entry_release(count); 3851 return (KERN_NO_SPACE); 3852 } 3853 3854 /* 3855 * We initially map a stack of only init_ssize. We will 3856 * grow as needed later. Since this is to be a grow 3857 * down stack, we map at the top of the range. 3858 * 3859 * Note: we would normally expect prot and max to be 3860 * VM_PROT_ALL, and cow to be 0. Possibly we should 3861 * eliminate these as input parameters, and just 3862 * pass these values here in the insert call. 3863 */ 3864 rv = vm_map_insert(map, &count, NULL, NULL, 3865 0, addrbos + max_ssize - init_ssize, 3866 addrbos + max_ssize, 3867 VM_MAPTYPE_NORMAL, 3868 VM_SUBSYS_STACK, prot, max, cow); 3869 3870 /* Now set the avail_ssize amount */ 3871 if (rv == KERN_SUCCESS) { 3872 if (prev_entry) 3873 next = vm_map_rb_tree_RB_NEXT(prev_entry); 3874 else 3875 next = RB_MIN(vm_map_rb_tree, &map->rb_root); 3876 if (prev_entry != NULL) { 3877 vm_map_clip_end(map, 3878 prev_entry, 3879 addrbos + max_ssize - init_ssize, 3880 &count); 3881 } 3882 if (next->end != addrbos + max_ssize || 3883 next->start != addrbos + max_ssize - init_ssize){ 3884 panic ("Bad entry start/end for new stack entry"); 3885 } else { 3886 next->aux.avail_ssize = max_ssize - init_ssize; 3887 } 3888 } 3889 3890 vm_map_unlock(map); 3891 vm_map_entry_release(count); 3892 return (rv); 3893 } 3894 3895 /* 3896 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 3897 * desired address is already mapped, or if we successfully grow 3898 * the stack. Also returns KERN_SUCCESS if addr is outside the 3899 * stack range (this is strange, but preserves compatibility with 3900 * the grow function in vm_machdep.c). 3901 * 3902 * No requirements. 3903 */ 3904 int 3905 vm_map_growstack (vm_map_t map, vm_offset_t addr) 3906 { 3907 vm_map_entry_t prev_entry; 3908 vm_map_entry_t stack_entry; 3909 vm_map_entry_t next; 3910 struct vmspace *vm; 3911 struct lwp *lp; 3912 struct proc *p; 3913 vm_offset_t end; 3914 int grow_amount; 3915 int rv = KERN_SUCCESS; 3916 int is_procstack; 3917 int use_read_lock = 1; 3918 int count; 3919 3920 /* 3921 * Find the vm 3922 */ 3923 lp = curthread->td_lwp; 3924 p = curthread->td_proc; 3925 KKASSERT(lp != NULL); 3926 vm = lp->lwp_vmspace; 3927 3928 /* 3929 * Growstack is only allowed on the current process. We disallow 3930 * other use cases, e.g. trying to access memory via procfs that 3931 * the stack hasn't grown into. 3932 */ 3933 if (map != &vm->vm_map) { 3934 return KERN_FAILURE; 3935 } 3936 3937 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3938 Retry: 3939 if (use_read_lock) 3940 vm_map_lock_read(map); 3941 else 3942 vm_map_lock(map); 3943 3944 /* 3945 * If addr is already in the entry range, no need to grow. 3946 * prev_entry returns NULL if addr is at the head. 3947 */ 3948 if (vm_map_lookup_entry(map, addr, &prev_entry)) 3949 goto done; 3950 if (prev_entry) 3951 stack_entry = vm_map_rb_tree_RB_NEXT(prev_entry); 3952 else 3953 stack_entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 3954 3955 if (stack_entry == NULL) 3956 goto done; 3957 if (prev_entry == NULL) 3958 end = stack_entry->start - stack_entry->aux.avail_ssize; 3959 else 3960 end = prev_entry->end; 3961 3962 /* 3963 * This next test mimics the old grow function in vm_machdep.c. 3964 * It really doesn't quite make sense, but we do it anyway 3965 * for compatibility. 3966 * 3967 * If not growable stack, return success. This signals the 3968 * caller to proceed as he would normally with normal vm. 3969 */ 3970 if (stack_entry->aux.avail_ssize < 1 || 3971 addr >= stack_entry->start || 3972 addr < stack_entry->start - stack_entry->aux.avail_ssize) { 3973 goto done; 3974 } 3975 3976 /* Find the minimum grow amount */ 3977 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE); 3978 if (grow_amount > stack_entry->aux.avail_ssize) { 3979 rv = KERN_NO_SPACE; 3980 goto done; 3981 } 3982 3983 /* 3984 * If there is no longer enough space between the entries 3985 * nogo, and adjust the available space. Note: this 3986 * should only happen if the user has mapped into the 3987 * stack area after the stack was created, and is 3988 * probably an error. 3989 * 3990 * This also effectively destroys any guard page the user 3991 * might have intended by limiting the stack size. 3992 */ 3993 if (grow_amount > stack_entry->start - end) { 3994 if (use_read_lock && vm_map_lock_upgrade(map)) { 3995 /* lost lock */ 3996 use_read_lock = 0; 3997 goto Retry; 3998 } 3999 use_read_lock = 0; 4000 stack_entry->aux.avail_ssize = stack_entry->start - end; 4001 rv = KERN_NO_SPACE; 4002 goto done; 4003 } 4004 4005 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 4006 4007 /* If this is the main process stack, see if we're over the 4008 * stack limit. 4009 */ 4010 if (is_procstack && (vm->vm_ssize + grow_amount > 4011 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 4012 rv = KERN_NO_SPACE; 4013 goto done; 4014 } 4015 4016 /* Round up the grow amount modulo SGROWSIZ */ 4017 grow_amount = roundup (grow_amount, sgrowsiz); 4018 if (grow_amount > stack_entry->aux.avail_ssize) { 4019 grow_amount = stack_entry->aux.avail_ssize; 4020 } 4021 if (is_procstack && (vm->vm_ssize + grow_amount > 4022 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 4023 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - vm->vm_ssize; 4024 } 4025 4026 /* If we would blow our VMEM resource limit, no go */ 4027 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 4028 rv = KERN_NO_SPACE; 4029 goto done; 4030 } 4031 4032 if (use_read_lock && vm_map_lock_upgrade(map)) { 4033 /* lost lock */ 4034 use_read_lock = 0; 4035 goto Retry; 4036 } 4037 use_read_lock = 0; 4038 4039 /* Get the preliminary new entry start value */ 4040 addr = stack_entry->start - grow_amount; 4041 4042 /* If this puts us into the previous entry, cut back our growth 4043 * to the available space. Also, see the note above. 4044 */ 4045 if (addr < end) { 4046 stack_entry->aux.avail_ssize = stack_entry->start - end; 4047 addr = end; 4048 } 4049 4050 rv = vm_map_insert(map, &count, NULL, NULL, 4051 0, addr, stack_entry->start, 4052 VM_MAPTYPE_NORMAL, 4053 VM_SUBSYS_STACK, VM_PROT_ALL, VM_PROT_ALL, 0); 4054 4055 /* Adjust the available stack space by the amount we grew. */ 4056 if (rv == KERN_SUCCESS) { 4057 if (prev_entry) { 4058 vm_map_clip_end(map, prev_entry, addr, &count); 4059 next = vm_map_rb_tree_RB_NEXT(prev_entry); 4060 } else { 4061 next = RB_MIN(vm_map_rb_tree, &map->rb_root); 4062 } 4063 if (next->end != stack_entry->start || 4064 next->start != addr) { 4065 panic ("Bad stack grow start/end in new stack entry"); 4066 } else { 4067 next->aux.avail_ssize = 4068 stack_entry->aux.avail_ssize - 4069 (next->end - next->start); 4070 if (is_procstack) { 4071 vm->vm_ssize += next->end - 4072 next->start; 4073 } 4074 } 4075 4076 if (map->flags & MAP_WIREFUTURE) 4077 vm_map_unwire(map, next->start, next->end, FALSE); 4078 } 4079 4080 done: 4081 if (use_read_lock) 4082 vm_map_unlock_read(map); 4083 else 4084 vm_map_unlock(map); 4085 vm_map_entry_release(count); 4086 return (rv); 4087 } 4088 4089 /* 4090 * Unshare the specified VM space for exec. If other processes are 4091 * mapped to it, then create a new one. The new vmspace is null. 4092 * 4093 * No requirements. 4094 */ 4095 void 4096 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 4097 { 4098 struct vmspace *oldvmspace = p->p_vmspace; 4099 struct vmspace *newvmspace; 4100 vm_map_t map = &p->p_vmspace->vm_map; 4101 4102 /* 4103 * If we are execing a resident vmspace we fork it, otherwise 4104 * we create a new vmspace. Note that exitingcnt is not 4105 * copied to the new vmspace. 4106 */ 4107 lwkt_gettoken(&oldvmspace->vm_map.token); 4108 if (vmcopy) { 4109 newvmspace = vmspace_fork(vmcopy); 4110 lwkt_gettoken(&newvmspace->vm_map.token); 4111 } else { 4112 newvmspace = vmspace_alloc(vm_map_min(map), vm_map_max(map)); 4113 lwkt_gettoken(&newvmspace->vm_map.token); 4114 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 4115 (caddr_t)&oldvmspace->vm_endcopy - 4116 (caddr_t)&oldvmspace->vm_startcopy); 4117 } 4118 4119 /* 4120 * Finish initializing the vmspace before assigning it 4121 * to the process. The vmspace will become the current vmspace 4122 * if p == curproc. 4123 */ 4124 pmap_pinit2(vmspace_pmap(newvmspace)); 4125 pmap_replacevm(p, newvmspace, 0); 4126 lwkt_reltoken(&newvmspace->vm_map.token); 4127 lwkt_reltoken(&oldvmspace->vm_map.token); 4128 vmspace_rel(oldvmspace); 4129 } 4130 4131 /* 4132 * Unshare the specified VM space for forcing COW. This 4133 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 4134 */ 4135 void 4136 vmspace_unshare(struct proc *p) 4137 { 4138 struct vmspace *oldvmspace = p->p_vmspace; 4139 struct vmspace *newvmspace; 4140 4141 lwkt_gettoken(&oldvmspace->vm_map.token); 4142 if (vmspace_getrefs(oldvmspace) == 1) { 4143 lwkt_reltoken(&oldvmspace->vm_map.token); 4144 return; 4145 } 4146 newvmspace = vmspace_fork(oldvmspace); 4147 lwkt_gettoken(&newvmspace->vm_map.token); 4148 pmap_pinit2(vmspace_pmap(newvmspace)); 4149 pmap_replacevm(p, newvmspace, 0); 4150 lwkt_reltoken(&newvmspace->vm_map.token); 4151 lwkt_reltoken(&oldvmspace->vm_map.token); 4152 vmspace_rel(oldvmspace); 4153 } 4154 4155 /* 4156 * vm_map_hint: return the beginning of the best area suitable for 4157 * creating a new mapping with "prot" protection. 4158 * 4159 * No requirements. 4160 */ 4161 vm_offset_t 4162 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot) 4163 { 4164 struct vmspace *vms = p->p_vmspace; 4165 struct rlimit limit; 4166 rlim_t dsiz; 4167 4168 /* 4169 * Acquire datasize limit for mmap() operation, 4170 * calculate nearest power of 2. 4171 */ 4172 if (kern_getrlimit(RLIMIT_DATA, &limit)) 4173 limit.rlim_cur = maxdsiz; 4174 dsiz = limit.rlim_cur; 4175 4176 if (!randomize_mmap || addr != 0) { 4177 /* 4178 * Set a reasonable start point for the hint if it was 4179 * not specified or if it falls within the heap space. 4180 * Hinted mmap()s do not allocate out of the heap space. 4181 */ 4182 if (addr == 0 || 4183 (addr >= round_page((vm_offset_t)vms->vm_taddr) && 4184 addr < round_page((vm_offset_t)vms->vm_daddr + dsiz))) { 4185 addr = round_page((vm_offset_t)vms->vm_daddr + dsiz); 4186 } 4187 4188 return addr; 4189 } 4190 4191 /* 4192 * randomize_mmap && addr == 0. For now randomize the 4193 * address within a dsiz range beyond the data limit. 4194 */ 4195 addr = (vm_offset_t)vms->vm_daddr + dsiz; 4196 if (dsiz) 4197 addr += (karc4random64() & 0x7FFFFFFFFFFFFFFFLU) % dsiz; 4198 return (round_page(addr)); 4199 } 4200 4201 /* 4202 * Finds the VM object, offset, and protection for a given virtual address 4203 * in the specified map, assuming a page fault of the type specified. 4204 * 4205 * Leaves the map in question locked for read; return values are guaranteed 4206 * until a vm_map_lookup_done call is performed. Note that the map argument 4207 * is in/out; the returned map must be used in the call to vm_map_lookup_done. 4208 * 4209 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make 4210 * that fast. 4211 * 4212 * If a lookup is requested with "write protection" specified, the map may 4213 * be changed to perform virtual copying operations, although the data 4214 * referenced will remain the same. 4215 * 4216 * No requirements. 4217 */ 4218 int 4219 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4220 vm_offset_t vaddr, 4221 vm_prot_t fault_typea, 4222 vm_map_entry_t *out_entry, /* OUT */ 4223 vm_object_t *object, /* OUT */ 4224 vm_pindex_t *pindex, /* OUT */ 4225 vm_prot_t *out_prot, /* OUT */ 4226 int *wflags) /* OUT */ 4227 { 4228 vm_map_entry_t entry; 4229 vm_map_t map = *var_map; 4230 vm_prot_t prot; 4231 vm_prot_t fault_type = fault_typea; 4232 int use_read_lock = 1; 4233 int rv = KERN_SUCCESS; 4234 int count; 4235 thread_t td = curthread; 4236 4237 /* 4238 * vm_map_entry_reserve() implements an important mitigation 4239 * against mmap() span running the kernel out of vm_map_entry 4240 * structures, but it can also cause an infinite call recursion. 4241 * Use td_nest_count to prevent an infinite recursion (allows 4242 * the vm_map code to dig into the pcpu vm_map_entry reserve). 4243 */ 4244 count = 0; 4245 if (td->td_nest_count == 0) { 4246 ++td->td_nest_count; 4247 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 4248 --td->td_nest_count; 4249 } 4250 RetryLookup: 4251 if (use_read_lock) 4252 vm_map_lock_read(map); 4253 else 4254 vm_map_lock(map); 4255 4256 /* 4257 * Always do a full lookup. The hint doesn't get us much anymore 4258 * now that the map is RB'd. 4259 */ 4260 cpu_ccfence(); 4261 *out_entry = NULL; 4262 *object = NULL; 4263 4264 { 4265 vm_map_entry_t tmp_entry; 4266 4267 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 4268 rv = KERN_INVALID_ADDRESS; 4269 goto done; 4270 } 4271 entry = tmp_entry; 4272 *out_entry = entry; 4273 } 4274 4275 /* 4276 * Handle submaps. 4277 */ 4278 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 4279 vm_map_t old_map = map; 4280 4281 *var_map = map = entry->object.sub_map; 4282 if (use_read_lock) 4283 vm_map_unlock_read(old_map); 4284 else 4285 vm_map_unlock(old_map); 4286 use_read_lock = 1; 4287 goto RetryLookup; 4288 } 4289 4290 /* 4291 * Check whether this task is allowed to have this page. 4292 * Note the special case for MAP_ENTRY_COW pages with an override. 4293 * This is to implement a forced COW for debuggers. 4294 */ 4295 if (fault_type & VM_PROT_OVERRIDE_WRITE) 4296 prot = entry->max_protection; 4297 else 4298 prot = entry->protection; 4299 4300 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 4301 if ((fault_type & prot) != fault_type) { 4302 rv = KERN_PROTECTION_FAILURE; 4303 goto done; 4304 } 4305 4306 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 4307 (entry->eflags & MAP_ENTRY_COW) && 4308 (fault_type & VM_PROT_WRITE) && 4309 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 4310 rv = KERN_PROTECTION_FAILURE; 4311 goto done; 4312 } 4313 4314 /* 4315 * If this page is not pageable, we have to get it for all possible 4316 * accesses. 4317 */ 4318 *wflags = 0; 4319 if (entry->wired_count) { 4320 *wflags |= FW_WIRED; 4321 prot = fault_type = entry->protection; 4322 } 4323 4324 /* 4325 * Virtual page tables may need to update the accessed (A) bit 4326 * in a page table entry. Upgrade the fault to a write fault for 4327 * that case if the map will support it. If the map does not support 4328 * it the page table entry simply will not be updated. 4329 */ 4330 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 4331 if (prot & VM_PROT_WRITE) 4332 fault_type |= VM_PROT_WRITE; 4333 } 4334 4335 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 4336 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 4337 if ((prot & VM_PROT_WRITE) == 0) 4338 fault_type |= VM_PROT_WRITE; 4339 } 4340 4341 /* 4342 * Only NORMAL and VPAGETABLE maps are object-based. UKSMAPs are not. 4343 */ 4344 if (entry->maptype != VM_MAPTYPE_NORMAL && 4345 entry->maptype != VM_MAPTYPE_VPAGETABLE) { 4346 *object = NULL; 4347 goto skip; 4348 } 4349 4350 /* 4351 * If the entry was copy-on-write, we either ... 4352 */ 4353 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4354 /* 4355 * If we want to write the page, we may as well handle that 4356 * now since we've got the map locked. 4357 * 4358 * If we don't need to write the page, we just demote the 4359 * permissions allowed. 4360 */ 4361 if (fault_type & VM_PROT_WRITE) { 4362 /* 4363 * Not allowed if TDF_NOFAULT is set as the shadowing 4364 * operation can deadlock against the faulting 4365 * function due to the copy-on-write. 4366 */ 4367 if (curthread->td_flags & TDF_NOFAULT) { 4368 rv = KERN_FAILURE_NOFAULT; 4369 goto done; 4370 } 4371 4372 /* 4373 * Make a new object, and place it in the object 4374 * chain. Note that no new references have appeared 4375 * -- one just moved from the map to the new 4376 * object. 4377 */ 4378 if (use_read_lock && vm_map_lock_upgrade(map)) { 4379 /* lost lock */ 4380 use_read_lock = 0; 4381 goto RetryLookup; 4382 } 4383 use_read_lock = 0; 4384 vm_map_entry_shadow(entry, 0); 4385 *wflags |= FW_DIDCOW; 4386 } else { 4387 /* 4388 * We're attempting to read a copy-on-write page -- 4389 * don't allow writes. 4390 */ 4391 prot &= ~VM_PROT_WRITE; 4392 } 4393 } 4394 4395 /* 4396 * Create an object if necessary. This code also handles 4397 * partitioning large entries to improve vm_fault performance. 4398 */ 4399 if (entry->object.vm_object == NULL && !map->system_map) { 4400 if (use_read_lock && vm_map_lock_upgrade(map)) { 4401 /* lost lock */ 4402 use_read_lock = 0; 4403 goto RetryLookup; 4404 } 4405 use_read_lock = 0; 4406 4407 /* 4408 * Partition large entries, giving each its own VM object, 4409 * to improve concurrent fault performance. This is only 4410 * applicable to userspace. 4411 */ 4412 if (map != &kernel_map && 4413 entry->maptype == VM_MAPTYPE_NORMAL && 4414 ((entry->start ^ entry->end) & ~MAP_ENTRY_PARTITION_MASK) && 4415 vm_map_partition_enable) { 4416 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 4417 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 4418 ++mycpu->gd_cnt.v_intrans_coll; 4419 ++mycpu->gd_cnt.v_intrans_wait; 4420 vm_map_transition_wait(map, 0); 4421 goto RetryLookup; 4422 } 4423 vm_map_entry_partition(map, entry, vaddr, &count); 4424 } 4425 vm_map_entry_allocate_object(entry); 4426 } 4427 4428 /* 4429 * Return the object/offset from this entry. If the entry was 4430 * copy-on-write or empty, it has been fixed up. 4431 */ 4432 *object = entry->object.vm_object; 4433 4434 skip: 4435 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 4436 4437 /* 4438 * Return whether this is the only map sharing this data. On 4439 * success we return with a read lock held on the map. On failure 4440 * we return with the map unlocked. 4441 */ 4442 *out_prot = prot; 4443 done: 4444 if (rv == KERN_SUCCESS) { 4445 if (use_read_lock == 0) 4446 vm_map_lock_downgrade(map); 4447 } else if (use_read_lock) { 4448 vm_map_unlock_read(map); 4449 } else { 4450 vm_map_unlock(map); 4451 } 4452 if (count > 0) 4453 vm_map_entry_release(count); 4454 4455 return (rv); 4456 } 4457 4458 /* 4459 * Releases locks acquired by a vm_map_lookup() 4460 * (according to the handle returned by that lookup). 4461 * 4462 * No other requirements. 4463 */ 4464 void 4465 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 4466 { 4467 /* 4468 * Unlock the main-level map 4469 */ 4470 vm_map_unlock_read(map); 4471 if (count) 4472 vm_map_entry_release(count); 4473 } 4474 4475 static void 4476 vm_map_entry_partition(vm_map_t map, vm_map_entry_t entry, 4477 vm_offset_t vaddr, int *countp) 4478 { 4479 vaddr &= ~MAP_ENTRY_PARTITION_MASK; 4480 vm_map_clip_start(map, entry, vaddr, countp); 4481 vaddr += MAP_ENTRY_PARTITION_SIZE; 4482 vm_map_clip_end(map, entry, vaddr, countp); 4483 } 4484 4485 /* 4486 * Quick hack, needs some help to make it more SMP friendly. 4487 */ 4488 void 4489 vm_map_interlock(vm_map_t map, struct vm_map_ilock *ilock, 4490 vm_offset_t ran_beg, vm_offset_t ran_end) 4491 { 4492 struct vm_map_ilock *scan; 4493 4494 ilock->ran_beg = ran_beg; 4495 ilock->ran_end = ran_end; 4496 ilock->flags = 0; 4497 4498 spin_lock(&map->ilock_spin); 4499 restart: 4500 for (scan = map->ilock_base; scan; scan = scan->next) { 4501 if (ran_end > scan->ran_beg && ran_beg < scan->ran_end) { 4502 scan->flags |= ILOCK_WAITING; 4503 ssleep(scan, &map->ilock_spin, 0, "ilock", 0); 4504 goto restart; 4505 } 4506 } 4507 ilock->next = map->ilock_base; 4508 map->ilock_base = ilock; 4509 spin_unlock(&map->ilock_spin); 4510 } 4511 4512 void 4513 vm_map_deinterlock(vm_map_t map, struct vm_map_ilock *ilock) 4514 { 4515 struct vm_map_ilock *scan; 4516 struct vm_map_ilock **scanp; 4517 4518 spin_lock(&map->ilock_spin); 4519 scanp = &map->ilock_base; 4520 while ((scan = *scanp) != NULL) { 4521 if (scan == ilock) { 4522 *scanp = ilock->next; 4523 spin_unlock(&map->ilock_spin); 4524 if (ilock->flags & ILOCK_WAITING) 4525 wakeup(ilock); 4526 return; 4527 } 4528 scanp = &scan->next; 4529 } 4530 spin_unlock(&map->ilock_spin); 4531 panic("vm_map_deinterlock: missing ilock!"); 4532 } 4533 4534 #include "opt_ddb.h" 4535 #ifdef DDB 4536 #include <ddb/ddb.h> 4537 4538 /* 4539 * Debugging only 4540 */ 4541 DB_SHOW_COMMAND(map, vm_map_print) 4542 { 4543 static int nlines; 4544 /* XXX convert args. */ 4545 vm_map_t map = (vm_map_t)addr; 4546 boolean_t full = have_addr; 4547 4548 vm_map_entry_t entry; 4549 4550 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 4551 (void *)map, 4552 (void *)map->pmap, map->nentries, map->timestamp); 4553 nlines++; 4554 4555 if (!full && db_indent) 4556 return; 4557 4558 db_indent += 2; 4559 RB_FOREACH(entry, vm_map_rb_tree, &map->rb_root) { 4560 db_iprintf("map entry %p: start=%p, end=%p\n", 4561 (void *)entry, (void *)entry->start, (void *)entry->end); 4562 nlines++; 4563 { 4564 static char *inheritance_name[4] = 4565 {"share", "copy", "none", "donate_copy"}; 4566 4567 db_iprintf(" prot=%x/%x/%s", 4568 entry->protection, 4569 entry->max_protection, 4570 inheritance_name[(int)(unsigned char) 4571 entry->inheritance]); 4572 if (entry->wired_count != 0) 4573 db_printf(", wired"); 4574 } 4575 switch(entry->maptype) { 4576 case VM_MAPTYPE_SUBMAP: 4577 /* XXX no %qd in kernel. Truncate entry->offset. */ 4578 db_printf(", share=%p, offset=0x%lx\n", 4579 (void *)entry->object.sub_map, 4580 (long)entry->offset); 4581 nlines++; 4582 4583 db_indent += 2; 4584 vm_map_print((db_expr_t)(intptr_t) 4585 entry->object.sub_map, 4586 full, 0, NULL); 4587 db_indent -= 2; 4588 break; 4589 case VM_MAPTYPE_NORMAL: 4590 case VM_MAPTYPE_VPAGETABLE: 4591 /* XXX no %qd in kernel. Truncate entry->offset. */ 4592 db_printf(", object=%p, offset=0x%lx", 4593 (void *)entry->object.vm_object, 4594 (long)entry->offset); 4595 if (entry->eflags & MAP_ENTRY_COW) 4596 db_printf(", copy (%s)", 4597 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4598 db_printf("\n"); 4599 nlines++; 4600 4601 if (entry->object.vm_object) { 4602 db_indent += 2; 4603 vm_object_print((db_expr_t)(intptr_t) 4604 entry->object.vm_object, 4605 full, 0, NULL); 4606 nlines += 4; 4607 db_indent -= 2; 4608 } 4609 break; 4610 case VM_MAPTYPE_UKSMAP: 4611 db_printf(", uksmap=%p, offset=0x%lx", 4612 (void *)entry->object.uksmap, 4613 (long)entry->offset); 4614 if (entry->eflags & MAP_ENTRY_COW) 4615 db_printf(", copy (%s)", 4616 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4617 db_printf("\n"); 4618 nlines++; 4619 break; 4620 default: 4621 break; 4622 } 4623 } 4624 db_indent -= 2; 4625 if (db_indent == 0) 4626 nlines = 0; 4627 } 4628 4629 /* 4630 * Debugging only 4631 */ 4632 DB_SHOW_COMMAND(procvm, procvm) 4633 { 4634 struct proc *p; 4635 4636 if (have_addr) { 4637 p = (struct proc *) addr; 4638 } else { 4639 p = curproc; 4640 } 4641 4642 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 4643 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 4644 (void *)vmspace_pmap(p->p_vmspace)); 4645 4646 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 4647 } 4648 4649 #endif /* DDB */ 4650