1 /* 2 * Copyright (c) 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * Copyright (c) 2003-2019 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 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 39 * All rights reserved. 40 * 41 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 42 * 43 * Permission to use, copy, modify and distribute this software and 44 * its documentation is hereby granted, provided that both the copyright 45 * notice and this permission notice appear in all copies of the 46 * software, derivative works or modified versions, and any portions 47 * thereof, and that both notices appear in supporting documentation. 48 * 49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 50 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 52 * 53 * Carnegie Mellon requests users of this software to return to 54 * 55 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 56 * School of Computer Science 57 * Carnegie Mellon University 58 * Pittsburgh PA 15213-3890 59 * 60 * any improvements or extensions that they make and grant Carnegie the 61 * rights to redistribute these changes. 62 */ 63 #include <sys/param.h> 64 #include <sys/systm.h> 65 #include <sys/kernel.h> 66 #include <sys/proc.h> 67 #include <sys/serialize.h> 68 #include <sys/lock.h> 69 #include <sys/vmmeter.h> 70 #include <sys/mman.h> 71 #include <sys/vnode.h> 72 #include <sys/resourcevar.h> 73 #include <sys/shm.h> 74 #include <sys/tree.h> 75 #include <sys/malloc.h> 76 #include <sys/objcache.h> 77 #include <sys/kern_syscall.h> 78 79 #include <vm/vm.h> 80 #include <vm/vm_param.h> 81 #include <vm/pmap.h> 82 #include <vm/vm_map.h> 83 #include <vm/vm_page.h> 84 #include <vm/vm_object.h> 85 #include <vm/vm_pager.h> 86 #include <vm/vm_kern.h> 87 #include <vm/vm_extern.h> 88 #include <vm/swap_pager.h> 89 #include <vm/vm_zone.h> 90 91 #include <sys/random.h> 92 #include <sys/sysctl.h> 93 #include <sys/spinlock.h> 94 95 #include <sys/thread2.h> 96 #include <sys/spinlock2.h> 97 98 /* 99 * Virtual memory maps provide for the mapping, protection, and sharing 100 * of virtual memory objects. In addition, this module provides for an 101 * efficient virtual copy of memory from one map to another. 102 * 103 * Synchronization is required prior to most operations. 104 * 105 * Maps consist of an ordered doubly-linked list of simple entries. 106 * A hint and a RB tree is used to speed-up lookups. 107 * 108 * Callers looking to modify maps specify start/end addresses which cause 109 * the related map entry to be clipped if necessary, and then later 110 * recombined if the pieces remained compatible. 111 * 112 * Virtual copy operations are performed by copying VM object references 113 * from one map to another, and then marking both regions as copy-on-write. 114 */ 115 static boolean_t vmspace_ctor(void *obj, void *privdata, int ocflags); 116 static void vmspace_dtor(void *obj, void *privdata); 117 static void vmspace_terminate(struct vmspace *vm, int final); 118 119 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore"); 120 MALLOC_DEFINE(M_MAP_BACKING, "map_backing", "vm_map_backing to entry"); 121 static struct objcache *vmspace_cache; 122 123 /* 124 * per-cpu page table cross mappings are initialized in early boot 125 * and might require a considerable number of vm_map_entry structures. 126 */ 127 #define MAPENTRYBSP_CACHE (MAXCPU+1) 128 #define MAPENTRYAP_CACHE 8 129 130 /* 131 * Partioning threaded programs with large anonymous memory areas can 132 * improve concurrent fault performance. 133 */ 134 #define MAP_ENTRY_PARTITION_SIZE ((vm_offset_t)(32 * 1024 * 1024)) 135 #define MAP_ENTRY_PARTITION_MASK (MAP_ENTRY_PARTITION_SIZE - 1) 136 137 #define VM_MAP_ENTRY_WITHIN_PARTITION(entry) \ 138 ((((entry)->ba.start ^ (entry)->ba.end) & ~MAP_ENTRY_PARTITION_MASK) == 0) 139 140 static struct vm_zone mapentzone_store; 141 __read_mostly static vm_zone_t mapentzone; 142 143 static struct vm_map_entry map_entry_init[MAX_MAPENT]; 144 static struct vm_map_entry cpu_map_entry_init_bsp[MAPENTRYBSP_CACHE]; 145 static struct vm_map_entry cpu_map_entry_init_ap[MAXCPU][MAPENTRYAP_CACHE]; 146 147 __read_mostly static int randomize_mmap; 148 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0, 149 "Randomize mmap offsets"); 150 __read_mostly static int vm_map_relock_enable = 1; 151 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW, 152 &vm_map_relock_enable, 0, "insert pop pgtable optimization"); 153 __read_mostly static int vm_map_partition_enable = 1; 154 SYSCTL_INT(_vm, OID_AUTO, map_partition_enable, CTLFLAG_RW, 155 &vm_map_partition_enable, 0, "Break up larger vm_map_entry's"); 156 __read_mostly static int vm_map_backing_limit = 5; 157 SYSCTL_INT(_vm, OID_AUTO, map_backing_limit, CTLFLAG_RW, 158 &vm_map_backing_limit, 0, "ba.backing_ba link depth"); 159 __read_mostly static int vm_map_backing_shadow_test = 1; 160 SYSCTL_INT(_vm, OID_AUTO, map_backing_shadow_test, CTLFLAG_RW, 161 &vm_map_backing_shadow_test, 0, "ba.object shadow test"); 162 163 static void vmspace_drop_notoken(struct vmspace *vm); 164 static void vm_map_entry_shadow(vm_map_entry_t entry); 165 static vm_map_entry_t vm_map_entry_create(int *); 166 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *); 167 static void vm_map_entry_dispose_ba (vm_map_entry_t entry, vm_map_backing_t ba); 168 static void vm_map_backing_replicated(vm_map_t map, 169 vm_map_entry_t entry, int flags); 170 static void vm_map_backing_adjust_start(vm_map_entry_t entry, 171 vm_ooffset_t start); 172 static void vm_map_backing_adjust_end(vm_map_entry_t entry, 173 vm_ooffset_t end); 174 static void vm_map_backing_attach (vm_map_entry_t entry, vm_map_backing_t ba); 175 static void vm_map_backing_detach (vm_map_entry_t entry, vm_map_backing_t ba); 176 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 177 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 178 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *); 179 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t); 180 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t, 181 vm_map_entry_t); 182 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry, 183 vm_offset_t start, vm_offset_t end, int *countp, int flags); 184 static void vm_map_entry_partition(vm_map_t map, vm_map_entry_t entry, 185 vm_offset_t vaddr, int *countp); 186 187 #define MAP_BACK_CLIPPED 0x0001 188 #define MAP_BACK_BASEOBJREFD 0x0002 189 190 /* 191 * Initialize the vm_map module. Must be called before any other vm_map 192 * routines. 193 * 194 * Map and entry structures are allocated from the general purpose 195 * memory pool with some exceptions: 196 * 197 * - The kernel map is allocated statically. 198 * - Initial kernel map entries are allocated out of a static pool. 199 * - We must set ZONE_SPECIAL here or the early boot code can get 200 * stuck if there are >63 cores. 201 * 202 * These restrictions are necessary since malloc() uses the 203 * maps and requires map entries. 204 * 205 * Called from the low level boot code only. 206 */ 207 void 208 vm_map_startup(void) 209 { 210 mapentzone = &mapentzone_store; 211 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry), 212 map_entry_init, MAX_MAPENT); 213 mapentzone_store.zflags |= ZONE_SPECIAL; 214 } 215 216 /* 217 * Called prior to any vmspace allocations. 218 * 219 * Called from the low level boot code only. 220 */ 221 void 222 vm_init2(void) 223 { 224 vmspace_cache = objcache_create_mbacked(M_VMSPACE, 225 sizeof(struct vmspace), 226 0, ncpus * 4, 227 vmspace_ctor, vmspace_dtor, 228 NULL); 229 zinitna(mapentzone, NULL, 0, 0, ZONE_USE_RESERVE | ZONE_SPECIAL); 230 pmap_init2(); 231 vm_object_init2(); 232 } 233 234 /* 235 * objcache support. We leave the pmap root cached as long as possible 236 * for performance reasons. 237 */ 238 static 239 boolean_t 240 vmspace_ctor(void *obj, void *privdata, int ocflags) 241 { 242 struct vmspace *vm = obj; 243 244 bzero(vm, sizeof(*vm)); 245 vm->vm_refcnt = VM_REF_DELETED; 246 247 return 1; 248 } 249 250 static 251 void 252 vmspace_dtor(void *obj, void *privdata) 253 { 254 struct vmspace *vm = obj; 255 256 KKASSERT(vm->vm_refcnt == VM_REF_DELETED); 257 pmap_puninit(vmspace_pmap(vm)); 258 } 259 260 /* 261 * Red black tree functions 262 * 263 * The caller must hold the related map lock. 264 */ 265 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b); 266 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare); 267 268 /* a->ba.start is address, and the only field which must be initialized */ 269 static int 270 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b) 271 { 272 if (a->ba.start < b->ba.start) 273 return(-1); 274 else if (a->ba.start > b->ba.start) 275 return(1); 276 return(0); 277 } 278 279 /* 280 * Initialize vmspace ref/hold counts vmspace0. There is a holdcnt for 281 * every refcnt. 282 */ 283 void 284 vmspace_initrefs(struct vmspace *vm) 285 { 286 vm->vm_refcnt = 1; 287 vm->vm_holdcnt = 1; 288 } 289 290 /* 291 * Allocate a vmspace structure, including a vm_map and pmap. 292 * Initialize numerous fields. While the initial allocation is zerod, 293 * subsequence reuse from the objcache leaves elements of the structure 294 * intact (particularly the pmap), so portions must be zerod. 295 * 296 * Returns a referenced vmspace. 297 * 298 * No requirements. 299 */ 300 struct vmspace * 301 vmspace_alloc(vm_offset_t min, vm_offset_t max) 302 { 303 struct vmspace *vm; 304 305 vm = objcache_get(vmspace_cache, M_WAITOK); 306 307 bzero(&vm->vm_startcopy, 308 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy); 309 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */ 310 311 /* 312 * NOTE: hold to acquires token for safety. 313 * 314 * On return vmspace is referenced (refs=1, hold=1). That is, 315 * each refcnt also has a holdcnt. There can be additional holds 316 * (holdcnt) above and beyond the refcnt. Finalization is handled in 317 * two stages, one on refs 1->0, and the the second on hold 1->0. 318 */ 319 KKASSERT(vm->vm_holdcnt == 0); 320 KKASSERT(vm->vm_refcnt == VM_REF_DELETED); 321 vmspace_initrefs(vm); 322 vmspace_hold(vm); 323 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */ 324 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */ 325 vm->vm_shm = NULL; 326 vm->vm_flags = 0; 327 cpu_vmspace_alloc(vm); 328 vmspace_drop(vm); 329 330 return (vm); 331 } 332 333 /* 334 * NOTE: Can return 0 if the vmspace is exiting. 335 */ 336 int 337 vmspace_getrefs(struct vmspace *vm) 338 { 339 int32_t n; 340 341 n = vm->vm_refcnt; 342 cpu_ccfence(); 343 if (n & VM_REF_DELETED) 344 n = -1; 345 return n; 346 } 347 348 void 349 vmspace_hold(struct vmspace *vm) 350 { 351 atomic_add_int(&vm->vm_holdcnt, 1); 352 lwkt_gettoken(&vm->vm_map.token); 353 } 354 355 /* 356 * Drop with final termination interlock. 357 */ 358 void 359 vmspace_drop(struct vmspace *vm) 360 { 361 lwkt_reltoken(&vm->vm_map.token); 362 vmspace_drop_notoken(vm); 363 } 364 365 static void 366 vmspace_drop_notoken(struct vmspace *vm) 367 { 368 if (atomic_fetchadd_int(&vm->vm_holdcnt, -1) == 1) { 369 if (vm->vm_refcnt & VM_REF_DELETED) 370 vmspace_terminate(vm, 1); 371 } 372 } 373 374 /* 375 * A vmspace object must not be in a terminated state to be able to obtain 376 * additional refs on it. 377 * 378 * These are official references to the vmspace, the count is used to check 379 * for vmspace sharing. Foreign accessors should use 'hold' and not 'ref'. 380 * 381 * XXX we need to combine hold & ref together into one 64-bit field to allow 382 * holds to prevent stage-1 termination. 383 */ 384 void 385 vmspace_ref(struct vmspace *vm) 386 { 387 uint32_t n; 388 389 atomic_add_int(&vm->vm_holdcnt, 1); 390 n = atomic_fetchadd_int(&vm->vm_refcnt, 1); 391 KKASSERT((n & VM_REF_DELETED) == 0); 392 } 393 394 /* 395 * Release a ref on the vmspace. On the 1->0 transition we do stage-1 396 * termination of the vmspace. Then, on the final drop of the hold we 397 * will do stage-2 final termination. 398 */ 399 void 400 vmspace_rel(struct vmspace *vm) 401 { 402 uint32_t n; 403 404 /* 405 * Drop refs. Each ref also has a hold which is also dropped. 406 * 407 * When refs hits 0 compete to get the VM_REF_DELETED flag (hold 408 * prevent finalization) to start termination processing. 409 * Finalization occurs when the last hold count drops to 0. 410 */ 411 n = atomic_fetchadd_int(&vm->vm_refcnt, -1) - 1; 412 while (n == 0) { 413 if (atomic_cmpset_int(&vm->vm_refcnt, 0, VM_REF_DELETED)) { 414 vmspace_terminate(vm, 0); 415 break; 416 } 417 n = vm->vm_refcnt; 418 cpu_ccfence(); 419 } 420 vmspace_drop_notoken(vm); 421 } 422 423 /* 424 * This is called during exit indicating that the vmspace is no 425 * longer in used by an exiting process, but the process has not yet 426 * been reaped. 427 * 428 * We drop refs, allowing for stage-1 termination, but maintain a holdcnt 429 * to prevent stage-2 until the process is reaped. Note hte order of 430 * operation, we must hold first. 431 * 432 * No requirements. 433 */ 434 void 435 vmspace_relexit(struct vmspace *vm) 436 { 437 atomic_add_int(&vm->vm_holdcnt, 1); 438 vmspace_rel(vm); 439 } 440 441 /* 442 * Called during reap to disconnect the remainder of the vmspace from 443 * the process. On the hold drop the vmspace termination is finalized. 444 * 445 * No requirements. 446 */ 447 void 448 vmspace_exitfree(struct proc *p) 449 { 450 struct vmspace *vm; 451 452 vm = p->p_vmspace; 453 p->p_vmspace = NULL; 454 vmspace_drop_notoken(vm); 455 } 456 457 /* 458 * Called in two cases: 459 * 460 * (1) When the last refcnt is dropped and the vmspace becomes inactive, 461 * called with final == 0. refcnt will be (u_int)-1 at this point, 462 * and holdcnt will still be non-zero. 463 * 464 * (2) When holdcnt becomes 0, called with final == 1. There should no 465 * longer be anyone with access to the vmspace. 466 * 467 * VMSPACE_EXIT1 flags the primary deactivation 468 * VMSPACE_EXIT2 flags the last reap 469 */ 470 static void 471 vmspace_terminate(struct vmspace *vm, int final) 472 { 473 int count; 474 475 lwkt_gettoken(&vm->vm_map.token); 476 if (final == 0) { 477 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) == 0); 478 vm->vm_flags |= VMSPACE_EXIT1; 479 480 /* 481 * Get rid of most of the resources. Leave the kernel pmap 482 * intact. 483 * 484 * If the pmap does not contain wired pages we can bulk-delete 485 * the pmap as a performance optimization before removing the 486 * related mappings. 487 * 488 * If the pmap contains wired pages we cannot do this 489 * pre-optimization because currently vm_fault_unwire() 490 * expects the pmap pages to exist and will not decrement 491 * p->wire_count if they do not. 492 */ 493 shmexit(vm); 494 if (vmspace_pmap(vm)->pm_stats.wired_count) { 495 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 496 VM_MAX_USER_ADDRESS); 497 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 498 VM_MAX_USER_ADDRESS); 499 } else { 500 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 501 VM_MAX_USER_ADDRESS); 502 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 503 VM_MAX_USER_ADDRESS); 504 } 505 lwkt_reltoken(&vm->vm_map.token); 506 } else { 507 KKASSERT((vm->vm_flags & VMSPACE_EXIT1) != 0); 508 KKASSERT((vm->vm_flags & VMSPACE_EXIT2) == 0); 509 510 /* 511 * Get rid of remaining basic resources. 512 */ 513 vm->vm_flags |= VMSPACE_EXIT2; 514 shmexit(vm); 515 516 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 517 vm_map_lock(&vm->vm_map); 518 cpu_vmspace_free(vm); 519 520 /* 521 * Lock the map, to wait out all other references to it. 522 * Delete all of the mappings and pages they hold, then call 523 * the pmap module to reclaim anything left. 524 */ 525 vm_map_delete(&vm->vm_map, 526 vm_map_min(&vm->vm_map), 527 vm_map_max(&vm->vm_map), 528 &count); 529 vm_map_unlock(&vm->vm_map); 530 vm_map_entry_release(count); 531 532 pmap_release(vmspace_pmap(vm)); 533 lwkt_reltoken(&vm->vm_map.token); 534 objcache_put(vmspace_cache, vm); 535 } 536 } 537 538 /* 539 * Swap useage is determined by taking the proportional swap used by 540 * VM objects backing the VM map. To make up for fractional losses, 541 * if the VM object has any swap use at all the associated map entries 542 * count for at least 1 swap page. 543 * 544 * No requirements. 545 */ 546 vm_offset_t 547 vmspace_swap_count(struct vmspace *vm) 548 { 549 vm_map_t map = &vm->vm_map; 550 vm_map_entry_t cur; 551 vm_object_t object; 552 vm_offset_t count = 0; 553 vm_offset_t n; 554 555 vmspace_hold(vm); 556 557 RB_FOREACH(cur, vm_map_rb_tree, &map->rb_root) { 558 switch(cur->maptype) { 559 case VM_MAPTYPE_NORMAL: 560 if ((object = cur->ba.object) == NULL) 561 break; 562 if (object->swblock_count) { 563 n = (cur->ba.end - cur->ba.start) / PAGE_SIZE; 564 count += object->swblock_count * 565 SWAP_META_PAGES * n / object->size + 1; 566 } 567 break; 568 default: 569 break; 570 } 571 } 572 vmspace_drop(vm); 573 574 return(count); 575 } 576 577 /* 578 * Calculate the approximate number of anonymous pages in use by 579 * this vmspace. To make up for fractional losses, we count each 580 * VM object as having at least 1 anonymous page. 581 * 582 * No requirements. 583 */ 584 vm_offset_t 585 vmspace_anonymous_count(struct vmspace *vm) 586 { 587 vm_map_t map = &vm->vm_map; 588 vm_map_entry_t cur; 589 vm_object_t object; 590 vm_offset_t count = 0; 591 592 vmspace_hold(vm); 593 RB_FOREACH(cur, vm_map_rb_tree, &map->rb_root) { 594 switch(cur->maptype) { 595 case VM_MAPTYPE_NORMAL: 596 if ((object = cur->ba.object) == NULL) 597 break; 598 if (object->type != OBJT_DEFAULT && 599 object->type != OBJT_SWAP) { 600 break; 601 } 602 count += object->resident_page_count; 603 break; 604 default: 605 break; 606 } 607 } 608 vmspace_drop(vm); 609 610 return(count); 611 } 612 613 /* 614 * Initialize an existing vm_map structure such as that in the vmspace 615 * structure. The pmap is initialized elsewhere. 616 * 617 * No requirements. 618 */ 619 void 620 vm_map_init(struct vm_map *map, vm_offset_t min_addr, vm_offset_t max_addr, 621 pmap_t pmap) 622 { 623 RB_INIT(&map->rb_root); 624 spin_init(&map->ilock_spin, "ilock"); 625 map->ilock_base = NULL; 626 map->nentries = 0; 627 map->size = 0; 628 map->system_map = 0; 629 vm_map_min(map) = min_addr; 630 vm_map_max(map) = max_addr; 631 map->pmap = pmap; 632 map->timestamp = 0; 633 map->flags = 0; 634 bzero(&map->freehint, sizeof(map->freehint)); 635 lwkt_token_init(&map->token, "vm_map"); 636 lockinit(&map->lock, "vm_maplk", (hz + 9) / 10, 0); 637 } 638 639 /* 640 * Find the first possible free address for the specified request length. 641 * Returns 0 if we don't have one cached. 642 */ 643 static 644 vm_offset_t 645 vm_map_freehint_find(vm_map_t map, vm_size_t length, vm_size_t align) 646 { 647 vm_map_freehint_t *scan; 648 649 scan = &map->freehint[0]; 650 while (scan < &map->freehint[VM_MAP_FFCOUNT]) { 651 if (scan->length == length && scan->align == align) 652 return(scan->start); 653 ++scan; 654 } 655 return 0; 656 } 657 658 /* 659 * Unconditionally set the freehint. Called by vm_map_findspace() after 660 * it finds an address. This will help us iterate optimally on the next 661 * similar findspace. 662 */ 663 static 664 void 665 vm_map_freehint_update(vm_map_t map, vm_offset_t start, 666 vm_size_t length, vm_size_t align) 667 { 668 vm_map_freehint_t *scan; 669 670 scan = &map->freehint[0]; 671 while (scan < &map->freehint[VM_MAP_FFCOUNT]) { 672 if (scan->length == length && scan->align == align) { 673 scan->start = start; 674 return; 675 } 676 ++scan; 677 } 678 scan = &map->freehint[map->freehint_newindex & VM_MAP_FFMASK]; 679 scan->start = start; 680 scan->align = align; 681 scan->length = length; 682 ++map->freehint_newindex; 683 } 684 685 /* 686 * Update any existing freehints (for any alignment), for the hole we just 687 * added. 688 */ 689 static 690 void 691 vm_map_freehint_hole(vm_map_t map, vm_offset_t start, vm_size_t length) 692 { 693 vm_map_freehint_t *scan; 694 695 scan = &map->freehint[0]; 696 while (scan < &map->freehint[VM_MAP_FFCOUNT]) { 697 if (scan->length <= length && scan->start > start) 698 scan->start = start; 699 ++scan; 700 } 701 } 702 703 /* 704 * This function handles MAP_ENTRY_NEEDS_COPY by inserting a fronting 705 * object in the entry for COW faults. 706 * 707 * The entire chain including entry->ba (prior to inserting the fronting 708 * object) essentially becomes set in stone... elements of it can be paged 709 * in or out, but cannot be further modified. 710 * 711 * NOTE: If we do not optimize the backing chain then a unique copy is not 712 * needed. Note, however, that because portions of the chain are 713 * shared across pmaps we cannot make any changes to the vm_map_backing 714 * elements themselves. 715 * 716 * If the map segment is governed by a virtual page table then it is 717 * possible to address offsets beyond the mapped area. Just allocate 718 * a maximally sized object for this case. 719 * 720 * If addref is non-zero an additional reference is added to the returned 721 * entry. This mechanic exists because the additional reference might have 722 * to be added atomically and not after return to prevent a premature 723 * collapse. XXX currently there is no collapse code. 724 * 725 * The vm_map must be exclusively locked. 726 * No other requirements. 727 */ 728 static 729 void 730 vm_map_entry_shadow(vm_map_entry_t entry) 731 { 732 vm_map_backing_t ba; 733 vm_size_t length; 734 vm_object_t source; 735 vm_object_t result; 736 737 /* 738 * Number of bytes we have to shadow 739 */ 740 length = atop(entry->ba.end - entry->ba.start); 741 742 /* 743 * Don't create the new object if the old object isn't shared. 744 * This case occurs quite often when programs fork/exec/wait. 745 * 746 * Caller ensures source exists (all backing_ba's must have objects), 747 * typically indirectly by virtue of the NEEDS_COPY flag being set. 748 * We have a ref on source by virtue of the entry and do not need 749 * to lock it to do this test. 750 */ 751 source = entry->ba.object; 752 KKASSERT(source); 753 754 if (source->type != OBJT_VNODE) { 755 if (source->ref_count == 1 && 756 source->handle == NULL && 757 (source->type == OBJT_DEFAULT || 758 source->type == OBJT_SWAP)) { 759 goto done; 760 } 761 } 762 ba = kmalloc(sizeof(*ba), M_MAP_BACKING, M_INTWAIT); /* copied later */ 763 vm_object_hold_shared(source); 764 765 /* 766 * Once it becomes part of a backing_ba chain it can wind up anywhere, 767 * drop the ONEMAPPING flag now. 768 */ 769 vm_object_clear_flag(source, OBJ_ONEMAPPING); 770 771 /* 772 * Allocate a new object with the given length. The new object 773 * is returned referenced but we may have to add another one. 774 * If we are adding a second reference we must clear OBJ_ONEMAPPING. 775 * (typically because the caller is about to clone a vm_map_entry). 776 * 777 * The source object currently has an extra reference to prevent 778 * collapses into it while we mess with its shadow list, which 779 * we will remove later in this routine. 780 * 781 * The target object may require a second reference if asked for one 782 * by the caller. 783 */ 784 result = vm_object_allocate_hold(OBJT_DEFAULT, length); 785 if (result == NULL) 786 panic("vm_object_shadow: no object for shadowing"); 787 788 /* 789 * The new object shadows the source object. 790 * 791 * Try to optimize the result object's page color when shadowing 792 * in order to maintain page coloring consistency in the combined 793 * shadowed object. 794 * 795 * The source object is moved to ba, retaining its existing ref-count. 796 * No additional ref is needed. 797 * 798 * SHADOWING IS NOT APPLICABLE TO OBJT_VNODE OBJECTS 799 */ 800 vm_map_backing_detach(entry, &entry->ba); 801 *ba = entry->ba; /* previous ba */ 802 entry->ba.object = result; /* new ba (at head of entry) */ 803 entry->ba.backing_ba = ba; 804 entry->ba.backing_count = ba->backing_count + 1; 805 entry->ba.offset = 0; 806 807 /* cpu localization twist */ 808 result->pg_color = vm_quickcolor(); 809 810 vm_map_backing_attach(entry, &entry->ba); 811 vm_map_backing_attach(entry, ba); 812 813 /* 814 * Adjust the return storage. Drop the ref on source before 815 * returning. 816 */ 817 vm_object_drop(result); 818 vm_object_drop(source); 819 done: 820 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 821 } 822 823 /* 824 * Allocate an object for a vm_map_entry. 825 * 826 * Object allocation for anonymous mappings is defered as long as possible. 827 * This function is called when we can defer no longer, generally when a map 828 * entry might be split or forked or takes a page fault. 829 * 830 * If the map segment is governed by a virtual page table then it is 831 * possible to address offsets beyond the mapped area. Just allocate 832 * a maximally sized object for this case. 833 * 834 * The vm_map must be exclusively locked. 835 * No other requirements. 836 */ 837 void 838 vm_map_entry_allocate_object(vm_map_entry_t entry) 839 { 840 vm_object_t obj; 841 842 /* 843 * ba.offset is NOT cumulatively added in the backing_ba scan like 844 * it was in the old object chain, so we can assign whatever offset 845 * we like to the new object. 846 * 847 * For now assign a value of 0 to make debugging object sizes 848 * easier. 849 */ 850 entry->ba.offset = 0; 851 852 obj = vm_object_allocate(OBJT_DEFAULT, 853 atop(entry->ba.end - entry->ba.start) + 854 entry->ba.offset); 855 entry->ba.object = obj; 856 vm_map_backing_attach(entry, &entry->ba); 857 } 858 859 /* 860 * Set an initial negative count so the first attempt to reserve 861 * space preloads a bunch of vm_map_entry's for this cpu. Also 862 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 863 * map a new page for vm_map_entry structures. SMP systems are 864 * particularly sensitive. 865 * 866 * This routine is called in early boot so we cannot just call 867 * vm_map_entry_reserve(). 868 * 869 * Called from the low level boot code only (for each cpu) 870 * 871 * WARNING! Take care not to have too-big a static/BSS structure here 872 * as MAXCPU can be 256+, otherwise the loader's 64MB heap 873 * can get blown out by the kernel plus the initrd image. 874 */ 875 void 876 vm_map_entry_reserve_cpu_init(globaldata_t gd) 877 { 878 vm_map_entry_t entry; 879 int count; 880 int i; 881 882 atomic_add_int(&gd->gd_vme_avail, -MAP_RESERVE_COUNT * 2); 883 if (gd->gd_cpuid == 0) { 884 entry = &cpu_map_entry_init_bsp[0]; 885 count = MAPENTRYBSP_CACHE; 886 } else { 887 entry = &cpu_map_entry_init_ap[gd->gd_cpuid][0]; 888 count = MAPENTRYAP_CACHE; 889 } 890 for (i = 0; i < count; ++i, ++entry) { 891 MAPENT_FREELIST(entry) = gd->gd_vme_base; 892 gd->gd_vme_base = entry; 893 } 894 } 895 896 /* 897 * Reserves vm_map_entry structures so code later-on can manipulate 898 * map_entry structures within a locked map without blocking trying 899 * to allocate a new vm_map_entry. 900 * 901 * No requirements. 902 * 903 * WARNING! We must not decrement gd_vme_avail until after we have 904 * ensured that sufficient entries exist, otherwise we can 905 * get into an endless call recursion in the zalloc code 906 * itself. 907 */ 908 int 909 vm_map_entry_reserve(int count) 910 { 911 struct globaldata *gd = mycpu; 912 vm_map_entry_t entry; 913 914 /* 915 * Make sure we have enough structures in gd_vme_base to handle 916 * the reservation request. 917 * 918 * Use a critical section to protect against VM faults. It might 919 * not be needed, but we have to be careful here. 920 */ 921 if (gd->gd_vme_avail < count) { 922 crit_enter(); 923 while (gd->gd_vme_avail < count) { 924 entry = zalloc(mapentzone); 925 MAPENT_FREELIST(entry) = gd->gd_vme_base; 926 gd->gd_vme_base = entry; 927 atomic_add_int(&gd->gd_vme_avail, 1); 928 } 929 crit_exit(); 930 } 931 atomic_add_int(&gd->gd_vme_avail, -count); 932 933 return(count); 934 } 935 936 /* 937 * Releases previously reserved vm_map_entry structures that were not 938 * used. If we have too much junk in our per-cpu cache clean some of 939 * it out. 940 * 941 * No requirements. 942 */ 943 void 944 vm_map_entry_release(int count) 945 { 946 struct globaldata *gd = mycpu; 947 vm_map_entry_t entry; 948 vm_map_entry_t efree; 949 950 count = atomic_fetchadd_int(&gd->gd_vme_avail, count) + count; 951 if (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 952 efree = NULL; 953 crit_enter(); 954 while (gd->gd_vme_avail > MAP_RESERVE_HYST) { 955 entry = gd->gd_vme_base; 956 KKASSERT(entry != NULL); 957 gd->gd_vme_base = MAPENT_FREELIST(entry); 958 atomic_add_int(&gd->gd_vme_avail, -1); 959 MAPENT_FREELIST(entry) = efree; 960 efree = entry; 961 } 962 crit_exit(); 963 while ((entry = efree) != NULL) { 964 efree = MAPENT_FREELIST(efree); 965 zfree(mapentzone, entry); 966 } 967 } 968 } 969 970 /* 971 * Reserve map entry structures for use in kernel_map itself. These 972 * entries have *ALREADY* been reserved on a per-cpu basis when the map 973 * was inited. This function is used by zalloc() to avoid a recursion 974 * when zalloc() itself needs to allocate additional kernel memory. 975 * 976 * This function works like the normal reserve but does not load the 977 * vm_map_entry cache (because that would result in an infinite 978 * recursion). Note that gd_vme_avail may go negative. This is expected. 979 * 980 * Any caller of this function must be sure to renormalize after 981 * potentially eating entries to ensure that the reserve supply 982 * remains intact. 983 * 984 * No requirements. 985 */ 986 int 987 vm_map_entry_kreserve(int count) 988 { 989 struct globaldata *gd = mycpu; 990 991 atomic_add_int(&gd->gd_vme_avail, -count); 992 KASSERT(gd->gd_vme_base != NULL, 993 ("no reserved entries left, gd_vme_avail = %d", 994 gd->gd_vme_avail)); 995 return(count); 996 } 997 998 /* 999 * Release previously reserved map entries for kernel_map. We do not 1000 * attempt to clean up like the normal release function as this would 1001 * cause an unnecessary (but probably not fatal) deep procedure call. 1002 * 1003 * No requirements. 1004 */ 1005 void 1006 vm_map_entry_krelease(int count) 1007 { 1008 struct globaldata *gd = mycpu; 1009 1010 atomic_add_int(&gd->gd_vme_avail, count); 1011 } 1012 1013 /* 1014 * Allocates a VM map entry for insertion. No entry fields are filled in. 1015 * 1016 * The entries should have previously been reserved. The reservation count 1017 * is tracked in (*countp). 1018 * 1019 * No requirements. 1020 */ 1021 static vm_map_entry_t 1022 vm_map_entry_create(int *countp) 1023 { 1024 struct globaldata *gd = mycpu; 1025 vm_map_entry_t entry; 1026 1027 KKASSERT(*countp > 0); 1028 --*countp; 1029 crit_enter(); 1030 entry = gd->gd_vme_base; 1031 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 1032 gd->gd_vme_base = MAPENT_FREELIST(entry); 1033 crit_exit(); 1034 1035 return(entry); 1036 } 1037 1038 /* 1039 * Attach and detach backing store elements 1040 */ 1041 static void 1042 vm_map_backing_attach(vm_map_entry_t entry, vm_map_backing_t ba) 1043 { 1044 vm_object_t obj; 1045 1046 switch(entry->maptype) { 1047 case VM_MAPTYPE_NORMAL: 1048 obj = ba->object; 1049 lockmgr(&obj->backing_lk, LK_EXCLUSIVE); 1050 TAILQ_INSERT_TAIL(&obj->backing_list, ba, entry); 1051 lockmgr(&obj->backing_lk, LK_RELEASE); 1052 break; 1053 case VM_MAPTYPE_UKSMAP: 1054 ba->uksmap(ba, UKSMAPOP_ADD, entry->aux.dev, NULL); 1055 break; 1056 } 1057 } 1058 1059 static void 1060 vm_map_backing_detach(vm_map_entry_t entry, vm_map_backing_t ba) 1061 { 1062 vm_object_t obj; 1063 1064 switch(entry->maptype) { 1065 case VM_MAPTYPE_NORMAL: 1066 obj = ba->object; 1067 lockmgr(&obj->backing_lk, LK_EXCLUSIVE); 1068 TAILQ_REMOVE(&obj->backing_list, ba, entry); 1069 lockmgr(&obj->backing_lk, LK_RELEASE); 1070 break; 1071 case VM_MAPTYPE_UKSMAP: 1072 ba->uksmap(ba, UKSMAPOP_REM, entry->aux.dev, NULL); 1073 break; 1074 } 1075 } 1076 1077 /* 1078 * Dispose of the dynamically allocated backing_ba chain associated 1079 * with a vm_map_entry. 1080 * 1081 * We decrement the (possibly shared) element and kfree() on the 1082 * 1->0 transition. We only iterate to the next backing_ba when 1083 * the previous one went through a 1->0 transition. 1084 * 1085 * These can only be normal vm_object based backings. 1086 */ 1087 static void 1088 vm_map_entry_dispose_ba(vm_map_entry_t entry, vm_map_backing_t ba) 1089 { 1090 vm_map_backing_t next; 1091 1092 while (ba) { 1093 if (ba->map_object) { 1094 vm_map_backing_detach(entry, ba); 1095 vm_object_deallocate(ba->object); 1096 } 1097 next = ba->backing_ba; 1098 kfree(ba, M_MAP_BACKING); 1099 ba = next; 1100 } 1101 } 1102 1103 /* 1104 * Dispose of a vm_map_entry that is no longer being referenced. 1105 * 1106 * No requirements. 1107 */ 1108 static void 1109 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 1110 { 1111 struct globaldata *gd = mycpu; 1112 1113 /* 1114 * Dispose of the base object and the backing link. 1115 */ 1116 switch(entry->maptype) { 1117 case VM_MAPTYPE_NORMAL: 1118 if (entry->ba.map_object) { 1119 vm_map_backing_detach(entry, &entry->ba); 1120 vm_object_deallocate(entry->ba.object); 1121 } 1122 break; 1123 case VM_MAPTYPE_SUBMAP: 1124 break; 1125 case VM_MAPTYPE_UKSMAP: 1126 vm_map_backing_detach(entry, &entry->ba); 1127 break; 1128 default: 1129 break; 1130 } 1131 vm_map_entry_dispose_ba(entry, entry->ba.backing_ba); 1132 1133 /* 1134 * Cleanup for safety. 1135 */ 1136 entry->ba.backing_ba = NULL; 1137 entry->ba.object = NULL; 1138 entry->ba.offset = 0; 1139 1140 ++*countp; 1141 crit_enter(); 1142 MAPENT_FREELIST(entry) = gd->gd_vme_base; 1143 gd->gd_vme_base = entry; 1144 crit_exit(); 1145 } 1146 1147 1148 /* 1149 * Insert/remove entries from maps. 1150 * 1151 * The related map must be exclusively locked. 1152 * The caller must hold map->token 1153 * No other requirements. 1154 */ 1155 static __inline void 1156 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) 1157 { 1158 ASSERT_VM_MAP_LOCKED(map); 1159 1160 map->nentries++; 1161 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry)) 1162 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry); 1163 } 1164 1165 static __inline void 1166 vm_map_entry_unlink(vm_map_t map, 1167 vm_map_entry_t entry) 1168 { 1169 ASSERT_VM_MAP_LOCKED(map); 1170 1171 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1172 panic("vm_map_entry_unlink: attempt to mess with " 1173 "locked entry! %p", entry); 1174 } 1175 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry); 1176 map->nentries--; 1177 } 1178 1179 /* 1180 * Finds the map entry containing (or immediately preceding) the specified 1181 * address in the given map. The entry is returned in (*entry). 1182 * 1183 * The boolean result indicates whether the address is actually contained 1184 * in the map. 1185 * 1186 * The related map must be locked. 1187 * No other requirements. 1188 */ 1189 boolean_t 1190 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry) 1191 { 1192 vm_map_entry_t tmp; 1193 vm_map_entry_t last; 1194 1195 ASSERT_VM_MAP_LOCKED(map); 1196 1197 /* 1198 * Locate the record from the top of the tree. 'last' tracks the 1199 * closest prior record and is returned if no match is found, which 1200 * in binary tree terms means tracking the most recent right-branch 1201 * taken. If there is no prior record, *entry is set to NULL. 1202 */ 1203 last = NULL; 1204 tmp = RB_ROOT(&map->rb_root); 1205 1206 while (tmp) { 1207 if (address >= tmp->ba.start) { 1208 if (address < tmp->ba.end) { 1209 *entry = tmp; 1210 return(TRUE); 1211 } 1212 last = tmp; 1213 tmp = RB_RIGHT(tmp, rb_entry); 1214 } else { 1215 tmp = RB_LEFT(tmp, rb_entry); 1216 } 1217 } 1218 *entry = last; 1219 return (FALSE); 1220 } 1221 1222 /* 1223 * Inserts the given whole VM object into the target map at the specified 1224 * address range. The object's size should match that of the address range. 1225 * 1226 * The map must be exclusively locked. 1227 * The object must be held. 1228 * The caller must have reserved sufficient vm_map_entry structures. 1229 * 1230 * If object is non-NULL, ref count must be bumped by caller prior to 1231 * making call to account for the new entry. XXX API is a bit messy. 1232 */ 1233 int 1234 vm_map_insert(vm_map_t map, int *countp, 1235 void *map_object, void *map_aux, 1236 vm_ooffset_t offset, void *aux_info, 1237 vm_offset_t start, vm_offset_t end, 1238 vm_maptype_t maptype, vm_subsys_t id, 1239 vm_prot_t prot, vm_prot_t max, int cow) 1240 { 1241 vm_map_entry_t new_entry; 1242 vm_map_entry_t prev_entry; 1243 vm_map_entry_t next; 1244 vm_map_entry_t temp_entry; 1245 vm_eflags_t protoeflags; 1246 vm_object_t object; 1247 int must_drop = 0; 1248 1249 if (maptype == VM_MAPTYPE_UKSMAP) 1250 object = NULL; 1251 else 1252 object = map_object; 1253 1254 ASSERT_VM_MAP_LOCKED(map); 1255 if (object) 1256 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 1257 1258 /* 1259 * Check that the start and end points are not bogus. 1260 */ 1261 if ((start < vm_map_min(map)) || (end > vm_map_max(map)) || 1262 (start >= end)) { 1263 return (KERN_INVALID_ADDRESS); 1264 } 1265 1266 /* 1267 * Find the entry prior to the proposed starting address; if it's part 1268 * of an existing entry, this range is bogus. 1269 */ 1270 if (vm_map_lookup_entry(map, start, &temp_entry)) 1271 return (KERN_NO_SPACE); 1272 prev_entry = temp_entry; 1273 1274 /* 1275 * Assert that the next entry doesn't overlap the end point. 1276 */ 1277 if (prev_entry) 1278 next = vm_map_rb_tree_RB_NEXT(prev_entry); 1279 else 1280 next = RB_MIN(vm_map_rb_tree, &map->rb_root); 1281 if (next && next->ba.start < end) 1282 return (KERN_NO_SPACE); 1283 1284 protoeflags = 0; 1285 1286 if (cow & MAP_COPY_ON_WRITE) 1287 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 1288 1289 if (cow & MAP_NOFAULT) { 1290 protoeflags |= MAP_ENTRY_NOFAULT; 1291 1292 KASSERT(object == NULL, 1293 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 1294 } 1295 if (cow & MAP_DISABLE_SYNCER) 1296 protoeflags |= MAP_ENTRY_NOSYNC; 1297 if (cow & MAP_DISABLE_COREDUMP) 1298 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1299 if (cow & MAP_IS_STACK) 1300 protoeflags |= MAP_ENTRY_STACK; 1301 if (cow & MAP_IS_KSTACK) 1302 protoeflags |= MAP_ENTRY_KSTACK; 1303 1304 lwkt_gettoken(&map->token); 1305 1306 if (object) { 1307 ; 1308 } else if (prev_entry && 1309 (prev_entry->eflags == protoeflags) && 1310 (prev_entry->ba.end == start) && 1311 (prev_entry->wired_count == 0) && 1312 (prev_entry->id == id) && 1313 prev_entry->maptype == maptype && 1314 maptype == VM_MAPTYPE_NORMAL && 1315 prev_entry->ba.backing_ba == NULL && /* not backed */ 1316 ((prev_entry->ba.object == NULL) || 1317 vm_object_coalesce(prev_entry->ba.object, 1318 OFF_TO_IDX(prev_entry->ba.offset), 1319 (vm_size_t)(prev_entry->ba.end - prev_entry->ba.start), 1320 (vm_size_t)(end - prev_entry->ba.end)))) { 1321 /* 1322 * We were able to extend the object. Determine if we 1323 * can extend the previous map entry to include the 1324 * new range as well. 1325 */ 1326 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 1327 (prev_entry->protection == prot) && 1328 (prev_entry->max_protection == max)) { 1329 map->size += (end - prev_entry->ba.end); 1330 vm_map_backing_adjust_end(prev_entry, end); 1331 vm_map_simplify_entry(map, prev_entry, countp); 1332 lwkt_reltoken(&map->token); 1333 return (KERN_SUCCESS); 1334 } 1335 1336 /* 1337 * If we can extend the object but cannot extend the 1338 * map entry, we have to create a new map entry. We 1339 * must bump the ref count on the extended object to 1340 * account for it. object may be NULL. 1341 */ 1342 object = prev_entry->ba.object; 1343 offset = prev_entry->ba.offset + 1344 (prev_entry->ba.end - prev_entry->ba.start); 1345 if (object) { 1346 vm_object_hold(object); 1347 vm_object_lock_swap(); /* map->token order */ 1348 vm_object_reference_locked(object); 1349 map_object = object; 1350 must_drop = 1; 1351 } 1352 } 1353 1354 /* 1355 * NOTE: if conditionals fail, object can be NULL here. This occurs 1356 * in things like the buffer map where we manage kva but do not manage 1357 * backing objects. 1358 */ 1359 1360 /* 1361 * Create a new entry 1362 */ 1363 new_entry = vm_map_entry_create(countp); 1364 new_entry->ba.pmap = map->pmap; 1365 new_entry->ba.start = start; 1366 new_entry->ba.end = end; 1367 new_entry->id = id; 1368 1369 new_entry->maptype = maptype; 1370 new_entry->eflags = protoeflags; 1371 new_entry->aux.master_pde = 0; /* in case size is different */ 1372 new_entry->aux.map_aux = map_aux; 1373 new_entry->ba.map_object = map_object; 1374 new_entry->ba.backing_ba = NULL; 1375 new_entry->ba.backing_count = 0; 1376 new_entry->ba.offset = offset; 1377 new_entry->ba.aux_info = aux_info; 1378 new_entry->ba.flags = 0; 1379 new_entry->ba.pmap = map->pmap; 1380 1381 new_entry->inheritance = VM_INHERIT_DEFAULT; 1382 new_entry->protection = prot; 1383 new_entry->max_protection = max; 1384 new_entry->wired_count = 0; 1385 1386 /* 1387 * Insert the new entry into the list 1388 */ 1389 vm_map_backing_replicated(map, new_entry, MAP_BACK_BASEOBJREFD); 1390 vm_map_entry_link(map, new_entry); 1391 map->size += new_entry->ba.end - new_entry->ba.start; 1392 1393 /* 1394 * Don't worry about updating freehint[] when inserting, allow 1395 * addresses to be lower than the actual first free spot. 1396 */ 1397 #if 0 1398 /* 1399 * Temporarily removed to avoid MAP_STACK panic, due to 1400 * MAP_STACK being a huge hack. Will be added back in 1401 * when MAP_STACK (and the user stack mapping) is fixed. 1402 */ 1403 /* 1404 * It may be possible to simplify the entry 1405 */ 1406 vm_map_simplify_entry(map, new_entry, countp); 1407 #endif 1408 1409 /* 1410 * Try to pre-populate the page table. Mappings governed by virtual 1411 * page tables cannot be prepopulated without a lot of work, so 1412 * don't try. 1413 */ 1414 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) && 1415 maptype != VM_MAPTYPE_UKSMAP) { 1416 int dorelock = 0; 1417 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) { 1418 dorelock = 1; 1419 vm_object_lock_swap(); 1420 vm_object_drop(object); 1421 } 1422 pmap_object_init_pt(map->pmap, new_entry, 1423 new_entry->ba.start, 1424 new_entry->ba.end - new_entry->ba.start, 1425 cow & MAP_PREFAULT_PARTIAL); 1426 if (dorelock) { 1427 vm_object_hold(object); 1428 vm_object_lock_swap(); 1429 } 1430 } 1431 lwkt_reltoken(&map->token); 1432 if (must_drop) 1433 vm_object_drop(object); 1434 1435 return (KERN_SUCCESS); 1436 } 1437 1438 /* 1439 * Find sufficient space for `length' bytes in the given map, starting at 1440 * `start'. Returns 0 on success, 1 on no space. 1441 * 1442 * This function will returned an arbitrarily aligned pointer. If no 1443 * particular alignment is required you should pass align as 1. Note that 1444 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 1445 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 1446 * argument. 1447 * 1448 * 'align' should be a power of 2 but is not required to be. 1449 * 1450 * The map must be exclusively locked. 1451 * No other requirements. 1452 */ 1453 int 1454 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length, 1455 vm_size_t align, int flags, vm_offset_t *addr) 1456 { 1457 vm_map_entry_t entry; 1458 vm_map_entry_t tmp; 1459 vm_offset_t hole_start; 1460 vm_offset_t end; 1461 vm_offset_t align_mask; 1462 1463 if (start < vm_map_min(map)) 1464 start = vm_map_min(map); 1465 if (start > vm_map_max(map)) 1466 return (1); 1467 1468 /* 1469 * If the alignment is not a power of 2 we will have to use 1470 * a mod/division, set align_mask to a special value. 1471 */ 1472 if ((align | (align - 1)) + 1 != (align << 1)) 1473 align_mask = (vm_offset_t)-1; 1474 else 1475 align_mask = align - 1; 1476 1477 /* 1478 * Use freehint to adjust the start point, hopefully reducing 1479 * the iteration to O(1). 1480 */ 1481 hole_start = vm_map_freehint_find(map, length, align); 1482 if (start < hole_start) 1483 start = hole_start; 1484 if (vm_map_lookup_entry(map, start, &tmp)) 1485 start = tmp->ba.end; 1486 entry = tmp; /* may be NULL */ 1487 1488 /* 1489 * Look through the rest of the map, trying to fit a new region in the 1490 * gap between existing regions, or after the very last region. 1491 */ 1492 for (;;) { 1493 /* 1494 * Adjust the proposed start by the requested alignment, 1495 * be sure that we didn't wrap the address. 1496 */ 1497 if (align_mask == (vm_offset_t)-1) 1498 end = roundup(start, align); 1499 else 1500 end = (start + align_mask) & ~align_mask; 1501 if (end < start) 1502 return (1); 1503 start = end; 1504 1505 /* 1506 * Find the end of the proposed new region. Be sure we didn't 1507 * go beyond the end of the map, or wrap around the address. 1508 * Then check to see if this is the last entry or if the 1509 * proposed end fits in the gap between this and the next 1510 * entry. 1511 */ 1512 end = start + length; 1513 if (end > vm_map_max(map) || end < start) 1514 return (1); 1515 1516 /* 1517 * Locate the next entry, we can stop if this is the 1518 * last entry (we know we are in-bounds so that would 1519 * be a sucess). 1520 */ 1521 if (entry) 1522 entry = vm_map_rb_tree_RB_NEXT(entry); 1523 else 1524 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 1525 if (entry == NULL) 1526 break; 1527 1528 /* 1529 * Determine if the proposed area would overlap the 1530 * next entry. 1531 * 1532 * When matching against a STACK entry, only allow the 1533 * memory map to intrude on the ungrown portion of the 1534 * STACK entry when MAP_TRYFIXED is set. 1535 */ 1536 if (entry->ba.start >= end) { 1537 if ((entry->eflags & MAP_ENTRY_STACK) == 0) 1538 break; 1539 if (flags & MAP_TRYFIXED) 1540 break; 1541 if (entry->ba.start - entry->aux.avail_ssize >= end) 1542 break; 1543 } 1544 start = entry->ba.end; 1545 } 1546 1547 /* 1548 * Update the freehint 1549 */ 1550 vm_map_freehint_update(map, start, length, align); 1551 1552 /* 1553 * Grow the kernel_map if necessary. pmap_growkernel() will panic 1554 * if it fails. The kernel_map is locked and nothing can steal 1555 * our address space if pmap_growkernel() blocks. 1556 * 1557 * NOTE: This may be unconditionally called for kldload areas on 1558 * x86_64 because these do not bump kernel_vm_end (which would 1559 * fill 128G worth of page tables!). Therefore we must not 1560 * retry. 1561 */ 1562 if (map == kernel_map) { 1563 vm_offset_t kstop; 1564 1565 kstop = round_page(start + length); 1566 if (kstop > kernel_vm_end) 1567 pmap_growkernel(start, kstop); 1568 } 1569 *addr = start; 1570 return (0); 1571 } 1572 1573 /* 1574 * vm_map_find finds an unallocated region in the target address map with 1575 * the given length and allocates it. The search is defined to be first-fit 1576 * from the specified address; the region found is returned in the same 1577 * parameter. 1578 * 1579 * If object is non-NULL, ref count must be bumped by caller 1580 * prior to making call to account for the new entry. 1581 * 1582 * No requirements. This function will lock the map temporarily. 1583 */ 1584 int 1585 vm_map_find(vm_map_t map, void *map_object, void *map_aux, 1586 vm_ooffset_t offset, vm_offset_t *addr, 1587 vm_size_t length, vm_size_t align, boolean_t fitit, 1588 vm_maptype_t maptype, vm_subsys_t id, 1589 vm_prot_t prot, vm_prot_t max, int cow) 1590 { 1591 vm_offset_t start; 1592 vm_object_t object; 1593 void *aux_info; 1594 int result; 1595 int count; 1596 1597 /* 1598 * Certain UKSMAPs may need aux_info. 1599 * 1600 * (map_object is the callback function, aux_info is the process 1601 * or thread, if necessary). 1602 */ 1603 aux_info = NULL; 1604 if (maptype == VM_MAPTYPE_UKSMAP) { 1605 KKASSERT(map_aux != NULL && map_object != NULL); 1606 1607 switch(minor(((struct cdev *)map_aux))) { 1608 case 5: 1609 /* 1610 * /dev/upmap 1611 */ 1612 aux_info = curproc; 1613 break; 1614 case 6: 1615 /* 1616 * /dev/kpmap 1617 */ 1618 break; 1619 case 7: 1620 /* 1621 * /dev/lpmap 1622 */ 1623 aux_info = curthread->td_lwp; 1624 break; 1625 } 1626 object = NULL; 1627 } else { 1628 object = map_object; 1629 } 1630 1631 start = *addr; 1632 1633 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1634 vm_map_lock(map); 1635 if (object) 1636 vm_object_hold_shared(object); 1637 if (fitit) { 1638 if (vm_map_findspace(map, start, length, align, 0, addr)) { 1639 if (object) 1640 vm_object_drop(object); 1641 vm_map_unlock(map); 1642 vm_map_entry_release(count); 1643 return (KERN_NO_SPACE); 1644 } 1645 start = *addr; 1646 } 1647 result = vm_map_insert(map, &count, 1648 map_object, map_aux, 1649 offset, aux_info, 1650 start, start + length, 1651 maptype, id, prot, max, cow); 1652 if (object) 1653 vm_object_drop(object); 1654 vm_map_unlock(map); 1655 vm_map_entry_release(count); 1656 1657 return (result); 1658 } 1659 1660 /* 1661 * Simplify the given map entry by merging with either neighbor. This 1662 * routine also has the ability to merge with both neighbors. 1663 * 1664 * This routine guarentees that the passed entry remains valid (though 1665 * possibly extended). When merging, this routine may delete one or 1666 * both neighbors. No action is taken on entries which have their 1667 * in-transition flag set. 1668 * 1669 * The map must be exclusively locked. 1670 */ 1671 void 1672 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 1673 { 1674 vm_map_entry_t next, prev; 1675 vm_size_t prevsize, esize; 1676 1677 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1678 ++mycpu->gd_cnt.v_intrans_coll; 1679 return; 1680 } 1681 1682 if (entry->maptype == VM_MAPTYPE_SUBMAP) 1683 return; 1684 if (entry->maptype == VM_MAPTYPE_UKSMAP) 1685 return; 1686 1687 prev = vm_map_rb_tree_RB_PREV(entry); 1688 if (prev) { 1689 prevsize = prev->ba.end - prev->ba.start; 1690 if ( (prev->ba.end == entry->ba.start) && 1691 (prev->maptype == entry->maptype) && 1692 (prev->ba.object == entry->ba.object) && 1693 (prev->ba.backing_ba == entry->ba.backing_ba) && 1694 (!prev->ba.object || 1695 (prev->ba.offset + prevsize == entry->ba.offset)) && 1696 (prev->eflags == entry->eflags) && 1697 (prev->protection == entry->protection) && 1698 (prev->max_protection == entry->max_protection) && 1699 (prev->inheritance == entry->inheritance) && 1700 (prev->id == entry->id) && 1701 (prev->wired_count == entry->wired_count)) { 1702 /* 1703 * NOTE: order important. Unlink before gumming up 1704 * the RBTREE w/adjust, adjust before disposal 1705 * of prior entry, to avoid pmap snafus. 1706 */ 1707 vm_map_entry_unlink(map, prev); 1708 vm_map_backing_adjust_start(entry, prev->ba.start); 1709 if (entry->ba.object == NULL) 1710 entry->ba.offset = 0; 1711 vm_map_entry_dispose(map, prev, countp); 1712 } 1713 } 1714 1715 next = vm_map_rb_tree_RB_NEXT(entry); 1716 if (next) { 1717 esize = entry->ba.end - entry->ba.start; 1718 if ((entry->ba.end == next->ba.start) && 1719 (next->maptype == entry->maptype) && 1720 (next->ba.object == entry->ba.object) && 1721 (prev->ba.backing_ba == entry->ba.backing_ba) && 1722 (!entry->ba.object || 1723 (entry->ba.offset + esize == next->ba.offset)) && 1724 (next->eflags == entry->eflags) && 1725 (next->protection == entry->protection) && 1726 (next->max_protection == entry->max_protection) && 1727 (next->inheritance == entry->inheritance) && 1728 (next->id == entry->id) && 1729 (next->wired_count == entry->wired_count)) { 1730 /* 1731 * NOTE: order important. Unlink before gumming up 1732 * the RBTREE w/adjust, adjust before disposal 1733 * of prior entry, to avoid pmap snafus. 1734 */ 1735 vm_map_entry_unlink(map, next); 1736 vm_map_backing_adjust_end(entry, next->ba.end); 1737 vm_map_entry_dispose(map, next, countp); 1738 } 1739 } 1740 } 1741 1742 /* 1743 * Asserts that the given entry begins at or after the specified address. 1744 * If necessary, it splits the entry into two. 1745 */ 1746 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1747 { \ 1748 if (startaddr > entry->ba.start) \ 1749 _vm_map_clip_start(map, entry, startaddr, countp); \ 1750 } 1751 1752 /* 1753 * This routine is called only when it is known that the entry must be split. 1754 * 1755 * The map must be exclusively locked. 1756 */ 1757 static void 1758 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, 1759 int *countp) 1760 { 1761 vm_map_entry_t new_entry; 1762 1763 /* 1764 * Split off the front portion -- note that we must insert the new 1765 * entry BEFORE this one, so that this entry has the specified 1766 * starting address. 1767 */ 1768 1769 vm_map_simplify_entry(map, entry, countp); 1770 1771 /* 1772 * If there is no object backing this entry, we might as well create 1773 * one now. If we defer it, an object can get created after the map 1774 * is clipped, and individual objects will be created for the split-up 1775 * map. This is a bit of a hack, but is also about the best place to 1776 * put this improvement. 1777 */ 1778 if (entry->ba.object == NULL && !map->system_map && 1779 VM_MAP_ENTRY_WITHIN_PARTITION(entry)) { 1780 vm_map_entry_allocate_object(entry); 1781 } 1782 1783 /* 1784 * NOTE: The replicated function will adjust start, end, and offset 1785 * for the remainder of the backing_ba linkages. We must fixup 1786 * the embedded ba. 1787 */ 1788 new_entry = vm_map_entry_create(countp); 1789 *new_entry = *entry; 1790 new_entry->ba.end = start; 1791 1792 /* 1793 * Ordering is important, make sure the new entry is replicated 1794 * before we cut the exiting entry. 1795 */ 1796 vm_map_backing_replicated(map, new_entry, MAP_BACK_CLIPPED); 1797 vm_map_backing_adjust_start(entry, start); 1798 vm_map_entry_link(map, new_entry); 1799 } 1800 1801 /* 1802 * Asserts that the given entry ends at or before the specified address. 1803 * If necessary, it splits the entry into two. 1804 * 1805 * The map must be exclusively locked. 1806 */ 1807 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1808 { \ 1809 if (endaddr < entry->ba.end) \ 1810 _vm_map_clip_end(map, entry, endaddr, countp); \ 1811 } 1812 1813 /* 1814 * This routine is called only when it is known that the entry must be split. 1815 * 1816 * The map must be exclusively locked. 1817 */ 1818 static void 1819 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, 1820 int *countp) 1821 { 1822 vm_map_entry_t new_entry; 1823 1824 /* 1825 * If there is no object backing this entry, we might as well create 1826 * one now. If we defer it, an object can get created after the map 1827 * is clipped, and individual objects will be created for the split-up 1828 * map. This is a bit of a hack, but is also about the best place to 1829 * put this improvement. 1830 */ 1831 1832 if (entry->ba.object == NULL && !map->system_map && 1833 VM_MAP_ENTRY_WITHIN_PARTITION(entry)) { 1834 vm_map_entry_allocate_object(entry); 1835 } 1836 1837 /* 1838 * Create a new entry and insert it AFTER the specified entry 1839 * 1840 * NOTE: The replicated function will adjust start, end, and offset 1841 * for the remainder of the backing_ba linkages. We must fixup 1842 * the embedded ba. 1843 */ 1844 new_entry = vm_map_entry_create(countp); 1845 *new_entry = *entry; 1846 new_entry->ba.start = end; 1847 new_entry->ba.offset += (new_entry->ba.start - entry->ba.start); 1848 1849 /* 1850 * Ordering is important, make sure the new entry is replicated 1851 * before we cut the exiting entry. 1852 */ 1853 vm_map_backing_replicated(map, new_entry, MAP_BACK_CLIPPED); 1854 vm_map_backing_adjust_end(entry, end); 1855 vm_map_entry_link(map, new_entry); 1856 } 1857 1858 /* 1859 * Asserts that the starting and ending region addresses fall within the 1860 * valid range for the map. 1861 */ 1862 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1863 { \ 1864 if (start < vm_map_min(map)) \ 1865 start = vm_map_min(map); \ 1866 if (end > vm_map_max(map)) \ 1867 end = vm_map_max(map); \ 1868 if (start > end) \ 1869 start = end; \ 1870 } 1871 1872 /* 1873 * Used to block when an in-transition collison occurs. The map 1874 * is unlocked for the sleep and relocked before the return. 1875 */ 1876 void 1877 vm_map_transition_wait(vm_map_t map, int relock) 1878 { 1879 tsleep_interlock(map, 0); 1880 vm_map_unlock(map); 1881 tsleep(map, PINTERLOCKED, "vment", 0); 1882 if (relock) 1883 vm_map_lock(map); 1884 } 1885 1886 /* 1887 * When we do blocking operations with the map lock held it is 1888 * possible that a clip might have occured on our in-transit entry, 1889 * requiring an adjustment to the entry in our loop. These macros 1890 * help the pageable and clip_range code deal with the case. The 1891 * conditional costs virtually nothing if no clipping has occured. 1892 */ 1893 1894 #define CLIP_CHECK_BACK(entry, save_start) \ 1895 do { \ 1896 while (entry->ba.start != save_start) { \ 1897 entry = vm_map_rb_tree_RB_PREV(entry); \ 1898 KASSERT(entry, ("bad entry clip")); \ 1899 } \ 1900 } while(0) 1901 1902 #define CLIP_CHECK_FWD(entry, save_end) \ 1903 do { \ 1904 while (entry->ba.end != save_end) { \ 1905 entry = vm_map_rb_tree_RB_NEXT(entry); \ 1906 KASSERT(entry, ("bad entry clip")); \ 1907 } \ 1908 } while(0) 1909 1910 1911 /* 1912 * Clip the specified range and return the base entry. The 1913 * range may cover several entries starting at the returned base 1914 * and the first and last entry in the covering sequence will be 1915 * properly clipped to the requested start and end address. 1916 * 1917 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1918 * flag. 1919 * 1920 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1921 * covered by the requested range. 1922 * 1923 * The map must be exclusively locked on entry and will remain locked 1924 * on return. If no range exists or the range contains holes and you 1925 * specified that no holes were allowed, NULL will be returned. This 1926 * routine may temporarily unlock the map in order avoid a deadlock when 1927 * sleeping. 1928 */ 1929 static 1930 vm_map_entry_t 1931 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1932 int *countp, int flags) 1933 { 1934 vm_map_entry_t start_entry; 1935 vm_map_entry_t entry; 1936 vm_map_entry_t next; 1937 1938 /* 1939 * Locate the entry and effect initial clipping. The in-transition 1940 * case does not occur very often so do not try to optimize it. 1941 */ 1942 again: 1943 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1944 return (NULL); 1945 entry = start_entry; 1946 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1947 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1948 ++mycpu->gd_cnt.v_intrans_coll; 1949 ++mycpu->gd_cnt.v_intrans_wait; 1950 vm_map_transition_wait(map, 1); 1951 /* 1952 * entry and/or start_entry may have been clipped while 1953 * we slept, or may have gone away entirely. We have 1954 * to restart from the lookup. 1955 */ 1956 goto again; 1957 } 1958 1959 /* 1960 * Since we hold an exclusive map lock we do not have to restart 1961 * after clipping, even though clipping may block in zalloc. 1962 */ 1963 vm_map_clip_start(map, entry, start, countp); 1964 vm_map_clip_end(map, entry, end, countp); 1965 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1966 1967 /* 1968 * Scan entries covered by the range. When working on the next 1969 * entry a restart need only re-loop on the current entry which 1970 * we have already locked, since 'next' may have changed. Also, 1971 * even though entry is safe, it may have been clipped so we 1972 * have to iterate forwards through the clip after sleeping. 1973 */ 1974 for (;;) { 1975 next = vm_map_rb_tree_RB_NEXT(entry); 1976 if (next == NULL || next->ba.start >= end) 1977 break; 1978 if (flags & MAP_CLIP_NO_HOLES) { 1979 if (next->ba.start > entry->ba.end) { 1980 vm_map_unclip_range(map, start_entry, 1981 start, entry->ba.end, countp, flags); 1982 return(NULL); 1983 } 1984 } 1985 1986 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1987 vm_offset_t save_end = entry->ba.end; 1988 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1989 ++mycpu->gd_cnt.v_intrans_coll; 1990 ++mycpu->gd_cnt.v_intrans_wait; 1991 vm_map_transition_wait(map, 1); 1992 1993 /* 1994 * clips might have occured while we blocked. 1995 */ 1996 CLIP_CHECK_FWD(entry, save_end); 1997 CLIP_CHECK_BACK(start_entry, start); 1998 continue; 1999 } 2000 2001 /* 2002 * No restart necessary even though clip_end may block, we 2003 * are holding the map lock. 2004 */ 2005 vm_map_clip_end(map, next, end, countp); 2006 next->eflags |= MAP_ENTRY_IN_TRANSITION; 2007 entry = next; 2008 } 2009 if (flags & MAP_CLIP_NO_HOLES) { 2010 if (entry->ba.end != end) { 2011 vm_map_unclip_range(map, start_entry, 2012 start, entry->ba.end, countp, flags); 2013 return(NULL); 2014 } 2015 } 2016 return(start_entry); 2017 } 2018 2019 /* 2020 * Undo the effect of vm_map_clip_range(). You should pass the same 2021 * flags and the same range that you passed to vm_map_clip_range(). 2022 * This code will clear the in-transition flag on the entries and 2023 * wake up anyone waiting. This code will also simplify the sequence 2024 * and attempt to merge it with entries before and after the sequence. 2025 * 2026 * The map must be locked on entry and will remain locked on return. 2027 * 2028 * Note that you should also pass the start_entry returned by 2029 * vm_map_clip_range(). However, if you block between the two calls 2030 * with the map unlocked please be aware that the start_entry may 2031 * have been clipped and you may need to scan it backwards to find 2032 * the entry corresponding with the original start address. You are 2033 * responsible for this, vm_map_unclip_range() expects the correct 2034 * start_entry to be passed to it and will KASSERT otherwise. 2035 */ 2036 static 2037 void 2038 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry, 2039 vm_offset_t start, vm_offset_t end, 2040 int *countp, int flags) 2041 { 2042 vm_map_entry_t entry; 2043 2044 entry = start_entry; 2045 2046 KASSERT(entry->ba.start == start, ("unclip_range: illegal base entry")); 2047 while (entry && entry->ba.start < end) { 2048 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, 2049 ("in-transition flag not set during unclip on: %p", 2050 entry)); 2051 KASSERT(entry->ba.end <= end, 2052 ("unclip_range: tail wasn't clipped")); 2053 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 2054 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 2055 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 2056 wakeup(map); 2057 } 2058 entry = vm_map_rb_tree_RB_NEXT(entry); 2059 } 2060 2061 /* 2062 * Simplification does not block so there is no restart case. 2063 */ 2064 entry = start_entry; 2065 while (entry && entry->ba.start < end) { 2066 vm_map_simplify_entry(map, entry, countp); 2067 entry = vm_map_rb_tree_RB_NEXT(entry); 2068 } 2069 } 2070 2071 /* 2072 * Mark the given range as handled by a subordinate map. 2073 * 2074 * This range must have been created with vm_map_find(), and no other 2075 * operations may have been performed on this range prior to calling 2076 * vm_map_submap(). 2077 * 2078 * Submappings cannot be removed. 2079 * 2080 * No requirements. 2081 */ 2082 int 2083 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 2084 { 2085 vm_map_entry_t entry; 2086 int result = KERN_INVALID_ARGUMENT; 2087 int count; 2088 2089 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2090 vm_map_lock(map); 2091 2092 VM_MAP_RANGE_CHECK(map, start, end); 2093 2094 if (vm_map_lookup_entry(map, start, &entry)) { 2095 vm_map_clip_start(map, entry, start, &count); 2096 } else if (entry) { 2097 entry = vm_map_rb_tree_RB_NEXT(entry); 2098 } else { 2099 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 2100 } 2101 2102 vm_map_clip_end(map, entry, end, &count); 2103 2104 if ((entry->ba.start == start) && (entry->ba.end == end) && 2105 ((entry->eflags & MAP_ENTRY_COW) == 0) && 2106 (entry->ba.object == NULL)) { 2107 entry->ba.sub_map = submap; 2108 entry->maptype = VM_MAPTYPE_SUBMAP; 2109 result = KERN_SUCCESS; 2110 } 2111 vm_map_unlock(map); 2112 vm_map_entry_release(count); 2113 2114 return (result); 2115 } 2116 2117 /* 2118 * Sets the protection of the specified address region in the target map. 2119 * If "set_max" is specified, the maximum protection is to be set; 2120 * otherwise, only the current protection is affected. 2121 * 2122 * The protection is not applicable to submaps, but is applicable to normal 2123 * maps and maps governed by virtual page tables. For example, when operating 2124 * on a virtual page table our protection basically controls how COW occurs 2125 * on the backing object, whereas the virtual page table abstraction itself 2126 * is an abstraction for userland. 2127 * 2128 * No requirements. 2129 */ 2130 int 2131 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 2132 vm_prot_t new_prot, boolean_t set_max) 2133 { 2134 vm_map_entry_t current; 2135 vm_map_entry_t entry; 2136 int count; 2137 2138 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2139 vm_map_lock(map); 2140 2141 VM_MAP_RANGE_CHECK(map, start, end); 2142 2143 if (vm_map_lookup_entry(map, start, &entry)) { 2144 vm_map_clip_start(map, entry, start, &count); 2145 } else if (entry) { 2146 entry = vm_map_rb_tree_RB_NEXT(entry); 2147 } else { 2148 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 2149 } 2150 2151 /* 2152 * Make a first pass to check for protection violations. 2153 */ 2154 current = entry; 2155 while (current && current->ba.start < end) { 2156 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2157 vm_map_unlock(map); 2158 vm_map_entry_release(count); 2159 return (KERN_INVALID_ARGUMENT); 2160 } 2161 if ((new_prot & current->max_protection) != new_prot) { 2162 vm_map_unlock(map); 2163 vm_map_entry_release(count); 2164 return (KERN_PROTECTION_FAILURE); 2165 } 2166 2167 /* 2168 * When making a SHARED+RW file mmap writable, update 2169 * v_lastwrite_ts. 2170 */ 2171 if (new_prot & PROT_WRITE && 2172 (current->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && 2173 current->maptype == VM_MAPTYPE_NORMAL && 2174 current->ba.object && 2175 current->ba.object->type == OBJT_VNODE) { 2176 struct vnode *vp; 2177 2178 vp = current->ba.object->handle; 2179 if (vp && vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_NOWAIT) == 0) { 2180 vfs_timestamp(&vp->v_lastwrite_ts); 2181 vsetflags(vp, VLASTWRITETS); 2182 vn_unlock(vp); 2183 } 2184 } 2185 current = vm_map_rb_tree_RB_NEXT(current); 2186 } 2187 2188 /* 2189 * Go back and fix up protections. [Note that clipping is not 2190 * necessary the second time.] 2191 */ 2192 current = entry; 2193 2194 while (current && current->ba.start < end) { 2195 vm_prot_t old_prot; 2196 2197 vm_map_clip_end(map, current, end, &count); 2198 2199 old_prot = current->protection; 2200 if (set_max) { 2201 current->max_protection = new_prot; 2202 current->protection = new_prot & old_prot; 2203 } else { 2204 current->protection = new_prot; 2205 } 2206 2207 /* 2208 * Update physical map if necessary. Worry about copy-on-write 2209 * here -- CHECK THIS XXX 2210 */ 2211 if (current->protection != old_prot) { 2212 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 2213 VM_PROT_ALL) 2214 2215 pmap_protect(map->pmap, current->ba.start, 2216 current->ba.end, 2217 current->protection & MASK(current)); 2218 #undef MASK 2219 } 2220 2221 vm_map_simplify_entry(map, current, &count); 2222 2223 current = vm_map_rb_tree_RB_NEXT(current); 2224 } 2225 vm_map_unlock(map); 2226 vm_map_entry_release(count); 2227 return (KERN_SUCCESS); 2228 } 2229 2230 /* 2231 * This routine traverses a processes map handling the madvise 2232 * system call. Advisories are classified as either those effecting 2233 * the vm_map_entry structure, or those effecting the underlying 2234 * objects. 2235 * 2236 * The <value> argument is used for extended madvise calls. 2237 * 2238 * No requirements. 2239 */ 2240 int 2241 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, 2242 int behav, off_t value) 2243 { 2244 vm_map_entry_t current, entry; 2245 int modify_map = 0; 2246 int error = 0; 2247 int count; 2248 2249 /* 2250 * Some madvise calls directly modify the vm_map_entry, in which case 2251 * we need to use an exclusive lock on the map and we need to perform 2252 * various clipping operations. Otherwise we only need a read-lock 2253 * on the map. 2254 */ 2255 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2256 2257 switch(behav) { 2258 case MADV_NORMAL: 2259 case MADV_SEQUENTIAL: 2260 case MADV_RANDOM: 2261 case MADV_NOSYNC: 2262 case MADV_AUTOSYNC: 2263 case MADV_NOCORE: 2264 case MADV_CORE: 2265 case MADV_SETMAP: 2266 modify_map = 1; 2267 vm_map_lock(map); 2268 break; 2269 case MADV_INVAL: 2270 case MADV_WILLNEED: 2271 case MADV_DONTNEED: 2272 case MADV_FREE: 2273 vm_map_lock_read(map); 2274 break; 2275 default: 2276 vm_map_entry_release(count); 2277 return (EINVAL); 2278 } 2279 2280 /* 2281 * Locate starting entry and clip if necessary. 2282 */ 2283 2284 VM_MAP_RANGE_CHECK(map, start, end); 2285 2286 if (vm_map_lookup_entry(map, start, &entry)) { 2287 if (modify_map) 2288 vm_map_clip_start(map, entry, start, &count); 2289 } else if (entry) { 2290 entry = vm_map_rb_tree_RB_NEXT(entry); 2291 } else { 2292 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 2293 } 2294 2295 if (modify_map) { 2296 /* 2297 * madvise behaviors that are implemented in the vm_map_entry. 2298 * 2299 * We clip the vm_map_entry so that behavioral changes are 2300 * limited to the specified address range. 2301 */ 2302 for (current = entry; 2303 current && current->ba.start < end; 2304 current = vm_map_rb_tree_RB_NEXT(current)) { 2305 /* 2306 * Ignore submaps 2307 */ 2308 if (current->maptype == VM_MAPTYPE_SUBMAP) 2309 continue; 2310 2311 vm_map_clip_end(map, current, end, &count); 2312 2313 switch (behav) { 2314 case MADV_NORMAL: 2315 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 2316 break; 2317 case MADV_SEQUENTIAL: 2318 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 2319 break; 2320 case MADV_RANDOM: 2321 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 2322 break; 2323 case MADV_NOSYNC: 2324 current->eflags |= MAP_ENTRY_NOSYNC; 2325 break; 2326 case MADV_AUTOSYNC: 2327 current->eflags &= ~MAP_ENTRY_NOSYNC; 2328 break; 2329 case MADV_NOCORE: 2330 current->eflags |= MAP_ENTRY_NOCOREDUMP; 2331 break; 2332 case MADV_CORE: 2333 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 2334 break; 2335 case MADV_SETMAP: 2336 /* 2337 * Set the page directory page for a map 2338 * governed by a virtual page table. 2339 * 2340 * Software virtual page table support has 2341 * been removed, this MADV is no longer 2342 * supported. 2343 */ 2344 error = EINVAL; 2345 break; 2346 case MADV_INVAL: 2347 /* 2348 * Invalidate the related pmap entries, used 2349 * to flush portions of the real kernel's 2350 * pmap when the caller has removed or 2351 * modified existing mappings in a virtual 2352 * page table. 2353 * 2354 * (exclusive locked map version does not 2355 * need the range interlock). 2356 */ 2357 pmap_remove(map->pmap, 2358 current->ba.start, current->ba.end); 2359 break; 2360 default: 2361 error = EINVAL; 2362 break; 2363 } 2364 vm_map_simplify_entry(map, current, &count); 2365 } 2366 vm_map_unlock(map); 2367 } else { 2368 vm_pindex_t pindex; 2369 vm_pindex_t delta; 2370 2371 /* 2372 * madvise behaviors that are implemented in the underlying 2373 * vm_object. 2374 * 2375 * Since we don't clip the vm_map_entry, we have to clip 2376 * the vm_object pindex and count. 2377 * 2378 * NOTE! These functions are only supported on normal maps. 2379 * 2380 * NOTE! These functions only apply to the top-most object. 2381 * It is not applicable to backing objects. 2382 */ 2383 for (current = entry; 2384 current && current->ba.start < end; 2385 current = vm_map_rb_tree_RB_NEXT(current)) { 2386 vm_offset_t useStart; 2387 2388 if (current->maptype != VM_MAPTYPE_NORMAL) 2389 continue; 2390 2391 pindex = OFF_TO_IDX(current->ba.offset); 2392 delta = atop(current->ba.end - current->ba.start); 2393 useStart = current->ba.start; 2394 2395 if (current->ba.start < start) { 2396 pindex += atop(start - current->ba.start); 2397 delta -= atop(start - current->ba.start); 2398 useStart = start; 2399 } 2400 if (current->ba.end > end) 2401 delta -= atop(current->ba.end - end); 2402 2403 if ((vm_spindex_t)delta <= 0) 2404 continue; 2405 2406 if (behav == MADV_INVAL) { 2407 /* 2408 * Invalidate the related pmap entries, used 2409 * to flush portions of the real kernel's 2410 * pmap when the caller has removed or 2411 * modified existing mappings in a virtual 2412 * page table. 2413 * 2414 * (shared locked map version needs the 2415 * interlock, see vm_fault()). 2416 */ 2417 struct vm_map_ilock ilock; 2418 2419 KASSERT(useStart >= VM_MIN_USER_ADDRESS && 2420 useStart + ptoa(delta) <= 2421 VM_MAX_USER_ADDRESS, 2422 ("Bad range %016jx-%016jx (%016jx)", 2423 useStart, useStart + ptoa(delta), 2424 delta)); 2425 vm_map_interlock(map, &ilock, 2426 useStart, 2427 useStart + ptoa(delta)); 2428 pmap_remove(map->pmap, 2429 useStart, 2430 useStart + ptoa(delta)); 2431 vm_map_deinterlock(map, &ilock); 2432 } else { 2433 vm_object_madvise(current->ba.object, 2434 pindex, delta, behav); 2435 } 2436 2437 /* 2438 * Try to pre-populate the page table. 2439 */ 2440 if (behav == MADV_WILLNEED) { 2441 pmap_object_init_pt( 2442 map->pmap, current, 2443 useStart, 2444 (delta << PAGE_SHIFT), 2445 MAP_PREFAULT_MADVISE 2446 ); 2447 } 2448 } 2449 vm_map_unlock_read(map); 2450 } 2451 vm_map_entry_release(count); 2452 return(error); 2453 } 2454 2455 2456 /* 2457 * Sets the inheritance of the specified address range in the target map. 2458 * Inheritance affects how the map will be shared with child maps at the 2459 * time of vm_map_fork. 2460 */ 2461 int 2462 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 2463 vm_inherit_t new_inheritance) 2464 { 2465 vm_map_entry_t entry; 2466 vm_map_entry_t temp_entry; 2467 int count; 2468 2469 switch (new_inheritance) { 2470 case VM_INHERIT_NONE: 2471 case VM_INHERIT_COPY: 2472 case VM_INHERIT_SHARE: 2473 break; 2474 default: 2475 return (KERN_INVALID_ARGUMENT); 2476 } 2477 2478 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2479 vm_map_lock(map); 2480 2481 VM_MAP_RANGE_CHECK(map, start, end); 2482 2483 if (vm_map_lookup_entry(map, start, &temp_entry)) { 2484 entry = temp_entry; 2485 vm_map_clip_start(map, entry, start, &count); 2486 } else if (temp_entry) { 2487 entry = vm_map_rb_tree_RB_NEXT(temp_entry); 2488 } else { 2489 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 2490 } 2491 2492 while (entry && entry->ba.start < end) { 2493 vm_map_clip_end(map, entry, end, &count); 2494 2495 entry->inheritance = new_inheritance; 2496 2497 vm_map_simplify_entry(map, entry, &count); 2498 2499 entry = vm_map_rb_tree_RB_NEXT(entry); 2500 } 2501 vm_map_unlock(map); 2502 vm_map_entry_release(count); 2503 return (KERN_SUCCESS); 2504 } 2505 2506 /* 2507 * Wiring/Unwiring of memory for user-related operation. 2508 * 2509 * Implement the semantics of mlock 2510 */ 2511 int 2512 vm_map_user_wiring(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 2513 boolean_t new_pageable) 2514 { 2515 vm_map_entry_t entry; 2516 vm_map_entry_t start_entry; 2517 vm_offset_t end; 2518 int rv = KERN_SUCCESS; 2519 int count; 2520 2521 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2522 vm_map_lock(map); 2523 VM_MAP_RANGE_CHECK(map, start, real_end); 2524 end = real_end; 2525 2526 start_entry = vm_map_clip_range(map, start, end, &count, 2527 MAP_CLIP_NO_HOLES); 2528 if (start_entry == NULL) { 2529 vm_map_unlock(map); 2530 vm_map_entry_release(count); 2531 return (KERN_INVALID_ADDRESS); 2532 } 2533 2534 if (new_pageable == 0) { 2535 entry = start_entry; 2536 while (entry && entry->ba.start < end) { 2537 vm_offset_t save_start; 2538 vm_offset_t save_end; 2539 2540 /* 2541 * Already user wired or hard wired (trivial cases) 2542 */ 2543 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 2544 entry = vm_map_rb_tree_RB_NEXT(entry); 2545 continue; 2546 } 2547 if (entry->wired_count != 0) { 2548 entry->wired_count++; 2549 entry->eflags |= MAP_ENTRY_USER_WIRED; 2550 entry = vm_map_rb_tree_RB_NEXT(entry); 2551 continue; 2552 } 2553 2554 /* 2555 * A new wiring requires instantiation of appropriate 2556 * management structures and the faulting in of the 2557 * page. 2558 */ 2559 if (entry->maptype == VM_MAPTYPE_NORMAL) { 2560 int copyflag = entry->eflags & 2561 MAP_ENTRY_NEEDS_COPY; 2562 if (copyflag && ((entry->protection & 2563 VM_PROT_WRITE) != 0)) { 2564 vm_map_entry_shadow(entry); 2565 } else if (entry->ba.object == NULL && 2566 !map->system_map) { 2567 vm_map_entry_allocate_object(entry); 2568 } 2569 } 2570 entry->wired_count++; 2571 entry->eflags |= MAP_ENTRY_USER_WIRED; 2572 2573 /* 2574 * Now fault in the area. Note that vm_fault_wire() 2575 * may release the map lock temporarily, it will be 2576 * relocked on return. The in-transition 2577 * flag protects the entries. 2578 */ 2579 save_start = entry->ba.start; 2580 save_end = entry->ba.end; 2581 rv = vm_fault_wire(map, entry, TRUE, 0); 2582 if (rv) { 2583 CLIP_CHECK_BACK(entry, save_start); 2584 for (;;) { 2585 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 2586 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2587 entry->wired_count = 0; 2588 if (entry->ba.end == save_end) 2589 break; 2590 entry = vm_map_rb_tree_RB_NEXT(entry); 2591 KASSERT(entry, 2592 ("bad entry clip during backout")); 2593 } 2594 end = save_start; /* unwire the rest */ 2595 break; 2596 } 2597 /* 2598 * note that even though the entry might have been 2599 * clipped, the USER_WIRED flag we set prevents 2600 * duplication so we do not have to do a 2601 * clip check. 2602 */ 2603 entry = vm_map_rb_tree_RB_NEXT(entry); 2604 } 2605 2606 /* 2607 * If we failed fall through to the unwiring section to 2608 * unwire what we had wired so far. 'end' has already 2609 * been adjusted. 2610 */ 2611 if (rv) 2612 new_pageable = 1; 2613 2614 /* 2615 * start_entry might have been clipped if we unlocked the 2616 * map and blocked. No matter how clipped it has gotten 2617 * there should be a fragment that is on our start boundary. 2618 */ 2619 CLIP_CHECK_BACK(start_entry, start); 2620 } 2621 2622 /* 2623 * Deal with the unwiring case. 2624 */ 2625 if (new_pageable) { 2626 /* 2627 * This is the unwiring case. We must first ensure that the 2628 * range to be unwired is really wired down. We know there 2629 * are no holes. 2630 */ 2631 entry = start_entry; 2632 while (entry && entry->ba.start < end) { 2633 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 2634 rv = KERN_INVALID_ARGUMENT; 2635 goto done; 2636 } 2637 KASSERT(entry->wired_count != 0, 2638 ("wired count was 0 with USER_WIRED set! %p", 2639 entry)); 2640 entry = vm_map_rb_tree_RB_NEXT(entry); 2641 } 2642 2643 /* 2644 * Now decrement the wiring count for each region. If a region 2645 * becomes completely unwired, unwire its physical pages and 2646 * mappings. 2647 */ 2648 /* 2649 * The map entries are processed in a loop, checking to 2650 * make sure the entry is wired and asserting it has a wired 2651 * count. However, another loop was inserted more-or-less in 2652 * the middle of the unwiring path. This loop picks up the 2653 * "entry" loop variable from the first loop without first 2654 * setting it to start_entry. Naturally, the secound loop 2655 * is never entered and the pages backing the entries are 2656 * never unwired. This can lead to a leak of wired pages. 2657 */ 2658 entry = start_entry; 2659 while (entry && entry->ba.start < end) { 2660 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 2661 ("expected USER_WIRED on entry %p", entry)); 2662 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2663 entry->wired_count--; 2664 if (entry->wired_count == 0) 2665 vm_fault_unwire(map, entry); 2666 entry = vm_map_rb_tree_RB_NEXT(entry); 2667 } 2668 } 2669 done: 2670 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2671 MAP_CLIP_NO_HOLES); 2672 vm_map_unlock(map); 2673 vm_map_entry_release(count); 2674 2675 return (rv); 2676 } 2677 2678 /* 2679 * Wiring/Unwiring of memory for kernel-related operation. 2680 * 2681 * Sets the pageability of the specified address range in the target map. 2682 * Regions specified as not pageable require locked-down physical 2683 * memory and physical page maps. 2684 * 2685 * The map must not be locked, but a reference must remain to the map 2686 * throughout the call. 2687 * 2688 * This function may be called via the zalloc path and must properly 2689 * reserve map entries for kernel_map. 2690 * 2691 * No requirements. 2692 */ 2693 int 2694 vm_map_kernel_wiring(vm_map_t map, vm_offset_t start, 2695 vm_offset_t real_end, int kmflags) 2696 { 2697 vm_map_entry_t entry; 2698 vm_map_entry_t start_entry; 2699 vm_offset_t end; 2700 int rv = KERN_SUCCESS; 2701 int count; 2702 2703 if (kmflags & KM_KRESERVE) 2704 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 2705 else 2706 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2707 vm_map_lock(map); 2708 VM_MAP_RANGE_CHECK(map, start, real_end); 2709 end = real_end; 2710 2711 start_entry = vm_map_clip_range(map, start, end, &count, 2712 MAP_CLIP_NO_HOLES); 2713 if (start_entry == NULL) { 2714 vm_map_unlock(map); 2715 rv = KERN_INVALID_ADDRESS; 2716 goto failure; 2717 } 2718 if ((kmflags & KM_PAGEABLE) == 0) { 2719 /* 2720 * Wiring. 2721 * 2722 * 1. Holding the write lock, we create any shadow or zero-fill 2723 * objects that need to be created. Then we clip each map 2724 * entry to the region to be wired and increment its wiring 2725 * count. We create objects before clipping the map entries 2726 * to avoid object proliferation. 2727 * 2728 * 2. We downgrade to a read lock, and call vm_fault_wire to 2729 * fault in the pages for any newly wired area (wired_count is 2730 * 1). 2731 * 2732 * Downgrading to a read lock for vm_fault_wire avoids a 2733 * possible deadlock with another process that may have faulted 2734 * on one of the pages to be wired (it would mark the page busy, 2735 * blocking us, then in turn block on the map lock that we 2736 * hold). Because of problems in the recursive lock package, 2737 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 2738 * any actions that require the write lock must be done 2739 * beforehand. Because we keep the read lock on the map, the 2740 * copy-on-write status of the entries we modify here cannot 2741 * change. 2742 */ 2743 entry = start_entry; 2744 while (entry && entry->ba.start < end) { 2745 /* 2746 * Trivial case if the entry is already wired 2747 */ 2748 if (entry->wired_count) { 2749 entry->wired_count++; 2750 entry = vm_map_rb_tree_RB_NEXT(entry); 2751 continue; 2752 } 2753 2754 /* 2755 * The entry is being newly wired, we have to setup 2756 * appropriate management structures. A shadow 2757 * object is required for a copy-on-write region, 2758 * or a normal object for a zero-fill region. We 2759 * do not have to do this for entries that point to sub 2760 * maps because we won't hold the lock on the sub map. 2761 */ 2762 if (entry->maptype == VM_MAPTYPE_NORMAL) { 2763 int copyflag = entry->eflags & 2764 MAP_ENTRY_NEEDS_COPY; 2765 if (copyflag && ((entry->protection & 2766 VM_PROT_WRITE) != 0)) { 2767 vm_map_entry_shadow(entry); 2768 } else if (entry->ba.object == NULL && 2769 !map->system_map) { 2770 vm_map_entry_allocate_object(entry); 2771 } 2772 } 2773 entry->wired_count++; 2774 entry = vm_map_rb_tree_RB_NEXT(entry); 2775 } 2776 2777 /* 2778 * Pass 2. 2779 */ 2780 2781 /* 2782 * HACK HACK HACK HACK 2783 * 2784 * vm_fault_wire() temporarily unlocks the map to avoid 2785 * deadlocks. The in-transition flag from vm_map_clip_range 2786 * call should protect us from changes while the map is 2787 * unlocked. T 2788 * 2789 * NOTE: Previously this comment stated that clipping might 2790 * still occur while the entry is unlocked, but from 2791 * what I can tell it actually cannot. 2792 * 2793 * It is unclear whether the CLIP_CHECK_*() calls 2794 * are still needed but we keep them in anyway. 2795 * 2796 * HACK HACK HACK HACK 2797 */ 2798 2799 entry = start_entry; 2800 while (entry && entry->ba.start < end) { 2801 /* 2802 * If vm_fault_wire fails for any page we need to undo 2803 * what has been done. We decrement the wiring count 2804 * for those pages which have not yet been wired (now) 2805 * and unwire those that have (later). 2806 */ 2807 vm_offset_t save_start = entry->ba.start; 2808 vm_offset_t save_end = entry->ba.end; 2809 2810 if (entry->wired_count == 1) 2811 rv = vm_fault_wire(map, entry, FALSE, kmflags); 2812 if (rv) { 2813 CLIP_CHECK_BACK(entry, save_start); 2814 for (;;) { 2815 KASSERT(entry->wired_count == 1, 2816 ("wired_count changed unexpectedly")); 2817 entry->wired_count = 0; 2818 if (entry->ba.end == save_end) 2819 break; 2820 entry = vm_map_rb_tree_RB_NEXT(entry); 2821 KASSERT(entry, 2822 ("bad entry clip during backout")); 2823 } 2824 end = save_start; 2825 break; 2826 } 2827 CLIP_CHECK_FWD(entry, save_end); 2828 entry = vm_map_rb_tree_RB_NEXT(entry); 2829 } 2830 2831 /* 2832 * If a failure occured undo everything by falling through 2833 * to the unwiring code. 'end' has already been adjusted 2834 * appropriately. 2835 */ 2836 if (rv) 2837 kmflags |= KM_PAGEABLE; 2838 2839 /* 2840 * start_entry is still IN_TRANSITION but may have been 2841 * clipped since vm_fault_wire() unlocks and relocks the 2842 * map. No matter how clipped it has gotten there should 2843 * be a fragment that is on our start boundary. 2844 */ 2845 CLIP_CHECK_BACK(start_entry, start); 2846 } 2847 2848 if (kmflags & KM_PAGEABLE) { 2849 /* 2850 * This is the unwiring case. We must first ensure that the 2851 * range to be unwired is really wired down. We know there 2852 * are no holes. 2853 */ 2854 entry = start_entry; 2855 while (entry && entry->ba.start < end) { 2856 if (entry->wired_count == 0) { 2857 rv = KERN_INVALID_ARGUMENT; 2858 goto done; 2859 } 2860 entry = vm_map_rb_tree_RB_NEXT(entry); 2861 } 2862 2863 /* 2864 * Now decrement the wiring count for each region. If a region 2865 * becomes completely unwired, unwire its physical pages and 2866 * mappings. 2867 */ 2868 entry = start_entry; 2869 while (entry && entry->ba.start < end) { 2870 entry->wired_count--; 2871 if (entry->wired_count == 0) 2872 vm_fault_unwire(map, entry); 2873 entry = vm_map_rb_tree_RB_NEXT(entry); 2874 } 2875 } 2876 done: 2877 vm_map_unclip_range(map, start_entry, start, real_end, 2878 &count, MAP_CLIP_NO_HOLES); 2879 vm_map_unlock(map); 2880 failure: 2881 if (kmflags & KM_KRESERVE) 2882 vm_map_entry_krelease(count); 2883 else 2884 vm_map_entry_release(count); 2885 return (rv); 2886 } 2887 2888 /* 2889 * Mark a newly allocated address range as wired but do not fault in 2890 * the pages. The caller is expected to load the pages into the object. 2891 * 2892 * The map must be locked on entry and will remain locked on return. 2893 * No other requirements. 2894 */ 2895 void 2896 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, 2897 int *countp) 2898 { 2899 vm_map_entry_t scan; 2900 vm_map_entry_t entry; 2901 2902 entry = vm_map_clip_range(map, addr, addr + size, 2903 countp, MAP_CLIP_NO_HOLES); 2904 scan = entry; 2905 while (scan && scan->ba.start < addr + size) { 2906 KKASSERT(scan->wired_count == 0); 2907 scan->wired_count = 1; 2908 scan = vm_map_rb_tree_RB_NEXT(scan); 2909 } 2910 vm_map_unclip_range(map, entry, addr, addr + size, 2911 countp, MAP_CLIP_NO_HOLES); 2912 } 2913 2914 /* 2915 * Push any dirty cached pages in the address range to their pager. 2916 * If syncio is TRUE, dirty pages are written synchronously. 2917 * If invalidate is TRUE, any cached pages are freed as well. 2918 * 2919 * This routine is called by sys_msync() 2920 * 2921 * Returns an error if any part of the specified range is not mapped. 2922 * 2923 * No requirements. 2924 */ 2925 int 2926 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, 2927 boolean_t syncio, boolean_t invalidate) 2928 { 2929 vm_map_entry_t current; 2930 vm_map_entry_t next; 2931 vm_map_entry_t entry; 2932 vm_map_backing_t ba; 2933 vm_size_t size; 2934 vm_object_t object; 2935 vm_ooffset_t offset; 2936 2937 vm_map_lock_read(map); 2938 VM_MAP_RANGE_CHECK(map, start, end); 2939 if (!vm_map_lookup_entry(map, start, &entry)) { 2940 vm_map_unlock_read(map); 2941 return (KERN_INVALID_ADDRESS); 2942 } 2943 lwkt_gettoken(&map->token); 2944 2945 /* 2946 * Make a first pass to check for holes. 2947 */ 2948 current = entry; 2949 while (current && current->ba.start < end) { 2950 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2951 lwkt_reltoken(&map->token); 2952 vm_map_unlock_read(map); 2953 return (KERN_INVALID_ARGUMENT); 2954 } 2955 next = vm_map_rb_tree_RB_NEXT(current); 2956 if (end > current->ba.end && 2957 (next == NULL || 2958 current->ba.end != next->ba.start)) { 2959 lwkt_reltoken(&map->token); 2960 vm_map_unlock_read(map); 2961 return (KERN_INVALID_ADDRESS); 2962 } 2963 current = next; 2964 } 2965 2966 if (invalidate) 2967 pmap_remove(vm_map_pmap(map), start, end); 2968 2969 /* 2970 * Make a second pass, cleaning/uncaching pages from the indicated 2971 * objects as we go. 2972 */ 2973 current = entry; 2974 while (current && current->ba.start < end) { 2975 offset = current->ba.offset + (start - current->ba.start); 2976 size = (end <= current->ba.end ? end : current->ba.end) - start; 2977 2978 switch(current->maptype) { 2979 case VM_MAPTYPE_SUBMAP: 2980 { 2981 vm_map_t smap; 2982 vm_map_entry_t tentry; 2983 vm_size_t tsize; 2984 2985 smap = current->ba.sub_map; 2986 vm_map_lock_read(smap); 2987 vm_map_lookup_entry(smap, offset, &tentry); 2988 if (tentry == NULL) { 2989 tsize = vm_map_max(smap) - offset; 2990 ba = NULL; 2991 offset = 0 + (offset - vm_map_min(smap)); 2992 } else { 2993 tsize = tentry->ba.end - offset; 2994 ba = &tentry->ba; 2995 offset = tentry->ba.offset + 2996 (offset - tentry->ba.start); 2997 } 2998 vm_map_unlock_read(smap); 2999 if (tsize < size) 3000 size = tsize; 3001 break; 3002 } 3003 case VM_MAPTYPE_NORMAL: 3004 ba = ¤t->ba; 3005 break; 3006 default: 3007 ba = NULL; 3008 break; 3009 } 3010 if (ba) { 3011 object = ba->object; 3012 if (object) 3013 vm_object_hold(object); 3014 } else { 3015 object = NULL; 3016 } 3017 3018 /* 3019 * Note that there is absolutely no sense in writing out 3020 * anonymous objects, so we track down the vnode object 3021 * to write out. 3022 * We invalidate (remove) all pages from the address space 3023 * anyway, for semantic correctness. 3024 * 3025 * note: certain anonymous maps, such as MAP_NOSYNC maps, 3026 * may start out with a NULL object. 3027 * 3028 * XXX do we really want to stop at the first backing store 3029 * here if there are more? XXX 3030 */ 3031 if (ba) { 3032 vm_object_t tobj; 3033 3034 tobj = object; 3035 while (ba->backing_ba != NULL) { 3036 offset -= ba->offset; 3037 ba = ba->backing_ba; 3038 offset += ba->offset; 3039 tobj = ba->object; 3040 if (tobj->size < OFF_TO_IDX(offset + size)) 3041 size = IDX_TO_OFF(tobj->size) - offset; 3042 break; /* XXX this break is not correct */ 3043 } 3044 if (object != tobj) { 3045 if (object) 3046 vm_object_drop(object); 3047 object = tobj; 3048 vm_object_hold(object); 3049 } 3050 } 3051 3052 if (object && (object->type == OBJT_VNODE) && 3053 (current->protection & VM_PROT_WRITE) && 3054 (object->flags & OBJ_NOMSYNC) == 0) { 3055 /* 3056 * Flush pages if writing is allowed, invalidate them 3057 * if invalidation requested. Pages undergoing I/O 3058 * will be ignored by vm_object_page_remove(). 3059 * 3060 * We cannot lock the vnode and then wait for paging 3061 * to complete without deadlocking against vm_fault. 3062 * Instead we simply call vm_object_page_remove() and 3063 * allow it to block internally on a page-by-page 3064 * basis when it encounters pages undergoing async 3065 * I/O. 3066 */ 3067 int flags; 3068 3069 /* no chain wait needed for vnode objects */ 3070 vm_object_reference_locked(object); 3071 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY); 3072 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 3073 flags |= invalidate ? OBJPC_INVAL : 0; 3074 3075 if (current->maptype == VM_MAPTYPE_NORMAL) { 3076 vm_object_page_clean(object, 3077 OFF_TO_IDX(offset), 3078 OFF_TO_IDX(offset + size + PAGE_MASK), 3079 flags); 3080 } 3081 vn_unlock(((struct vnode *)object->handle)); 3082 vm_object_deallocate_locked(object); 3083 } 3084 if (object && invalidate && 3085 ((object->type == OBJT_VNODE) || 3086 (object->type == OBJT_DEVICE) || 3087 (object->type == OBJT_MGTDEVICE))) { 3088 int clean_only = 3089 ((object->type == OBJT_DEVICE) || 3090 (object->type == OBJT_MGTDEVICE)) ? FALSE : TRUE; 3091 /* no chain wait needed for vnode/device objects */ 3092 vm_object_reference_locked(object); 3093 if (current->maptype == VM_MAPTYPE_NORMAL) { 3094 vm_object_page_remove(object, 3095 OFF_TO_IDX(offset), 3096 OFF_TO_IDX(offset + size + PAGE_MASK), 3097 clean_only); 3098 } 3099 vm_object_deallocate_locked(object); 3100 } 3101 start += size; 3102 if (object) 3103 vm_object_drop(object); 3104 current = vm_map_rb_tree_RB_NEXT(current); 3105 } 3106 3107 lwkt_reltoken(&map->token); 3108 vm_map_unlock_read(map); 3109 3110 return (KERN_SUCCESS); 3111 } 3112 3113 /* 3114 * Make the region specified by this entry pageable. 3115 * 3116 * The vm_map must be exclusively locked. 3117 */ 3118 static void 3119 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 3120 { 3121 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3122 entry->wired_count = 0; 3123 vm_fault_unwire(map, entry); 3124 } 3125 3126 /* 3127 * Deallocate the given entry from the target map. 3128 * 3129 * The vm_map must be exclusively locked. 3130 */ 3131 static void 3132 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 3133 { 3134 vm_map_entry_unlink(map, entry); 3135 map->size -= entry->ba.end - entry->ba.start; 3136 vm_map_entry_dispose(map, entry, countp); 3137 } 3138 3139 /* 3140 * Deallocates the given address range from the target map. 3141 * 3142 * The vm_map must be exclusively locked. 3143 */ 3144 int 3145 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 3146 { 3147 vm_object_t object; 3148 vm_map_entry_t entry; 3149 vm_map_entry_t first_entry; 3150 vm_offset_t hole_start; 3151 3152 ASSERT_VM_MAP_LOCKED(map); 3153 lwkt_gettoken(&map->token); 3154 again: 3155 /* 3156 * Find the start of the region, and clip it. Set entry to point 3157 * at the first record containing the requested address or, if no 3158 * such record exists, the next record with a greater address. The 3159 * loop will run from this point until a record beyond the termination 3160 * address is encountered. 3161 * 3162 * Adjust freehint[] for either the clip case or the extension case. 3163 * 3164 * GGG see other GGG comment. 3165 */ 3166 if (vm_map_lookup_entry(map, start, &first_entry)) { 3167 entry = first_entry; 3168 vm_map_clip_start(map, entry, start, countp); 3169 hole_start = start; 3170 } else { 3171 if (first_entry) { 3172 entry = vm_map_rb_tree_RB_NEXT(first_entry); 3173 if (entry == NULL) 3174 hole_start = first_entry->ba.start; 3175 else 3176 hole_start = first_entry->ba.end; 3177 } else { 3178 entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 3179 if (entry == NULL) 3180 hole_start = vm_map_min(map); 3181 else 3182 hole_start = vm_map_max(map); 3183 } 3184 } 3185 3186 /* 3187 * Step through all entries in this region 3188 */ 3189 while (entry && entry->ba.start < end) { 3190 vm_map_entry_t next; 3191 vm_offset_t s, e; 3192 vm_pindex_t offidxstart, offidxend, count; 3193 3194 /* 3195 * If we hit an in-transition entry we have to sleep and 3196 * retry. It's easier (and not really slower) to just retry 3197 * since this case occurs so rarely and the hint is already 3198 * pointing at the right place. We have to reset the 3199 * start offset so as not to accidently delete an entry 3200 * another process just created in vacated space. 3201 */ 3202 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3203 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3204 start = entry->ba.start; 3205 ++mycpu->gd_cnt.v_intrans_coll; 3206 ++mycpu->gd_cnt.v_intrans_wait; 3207 vm_map_transition_wait(map, 1); 3208 goto again; 3209 } 3210 vm_map_clip_end(map, entry, end, countp); 3211 3212 s = entry->ba.start; 3213 e = entry->ba.end; 3214 next = vm_map_rb_tree_RB_NEXT(entry); 3215 3216 offidxstart = OFF_TO_IDX(entry->ba.offset); 3217 count = OFF_TO_IDX(e - s); 3218 3219 switch(entry->maptype) { 3220 case VM_MAPTYPE_NORMAL: 3221 case VM_MAPTYPE_SUBMAP: 3222 object = entry->ba.object; 3223 break; 3224 default: 3225 object = NULL; 3226 break; 3227 } 3228 3229 /* 3230 * Unwire before removing addresses from the pmap; otherwise, 3231 * unwiring will put the entries back in the pmap. 3232 * 3233 * Generally speaking, doing a bulk pmap_remove() before 3234 * removing the pages from the VM object is better at 3235 * reducing unnecessary IPIs. The pmap code is now optimized 3236 * to not blindly iterate the range when pt and pd pages 3237 * are missing. 3238 */ 3239 if (entry->wired_count != 0) 3240 vm_map_entry_unwire(map, entry); 3241 3242 offidxend = offidxstart + count; 3243 3244 if (object == kernel_object) { 3245 pmap_remove(map->pmap, s, e); 3246 vm_object_hold(object); 3247 vm_object_page_remove(object, offidxstart, 3248 offidxend, FALSE); 3249 vm_object_drop(object); 3250 } else if (object && object->type != OBJT_DEFAULT && 3251 object->type != OBJT_SWAP) { 3252 /* 3253 * vnode object routines cannot be chain-locked, 3254 * but since we aren't removing pages from the 3255 * object here we can use a shared hold. 3256 */ 3257 vm_object_hold_shared(object); 3258 pmap_remove(map->pmap, s, e); 3259 vm_object_drop(object); 3260 } else if (object) { 3261 vm_object_hold(object); 3262 pmap_remove(map->pmap, s, e); 3263 3264 if (object != NULL && 3265 object->ref_count != 1 && 3266 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == 3267 OBJ_ONEMAPPING && 3268 (object->type == OBJT_DEFAULT || 3269 object->type == OBJT_SWAP)) { 3270 /* 3271 * When ONEMAPPING is set we can destroy the 3272 * pages underlying the entry's range. 3273 */ 3274 vm_object_page_remove(object, offidxstart, 3275 offidxend, FALSE); 3276 if (object->type == OBJT_SWAP) { 3277 swap_pager_freespace(object, 3278 offidxstart, 3279 count); 3280 } 3281 if (offidxend >= object->size && 3282 offidxstart < object->size) { 3283 object->size = offidxstart; 3284 } 3285 } 3286 vm_object_drop(object); 3287 } else if (entry->maptype == VM_MAPTYPE_UKSMAP) { 3288 pmap_remove(map->pmap, s, e); 3289 } 3290 3291 /* 3292 * Delete the entry (which may delete the object) only after 3293 * removing all pmap entries pointing to its pages. 3294 * (Otherwise, its page frames may be reallocated, and any 3295 * modify bits will be set in the wrong object!) 3296 */ 3297 vm_map_entry_delete(map, entry, countp); 3298 entry = next; 3299 } 3300 3301 /* 3302 * We either reached the end and use vm_map_max as the end 3303 * address, or we didn't and we use the next entry as the 3304 * end address. 3305 */ 3306 if (entry == NULL) { 3307 vm_map_freehint_hole(map, hole_start, 3308 vm_map_max(map) - hole_start); 3309 } else { 3310 vm_map_freehint_hole(map, hole_start, 3311 entry->ba.start - hole_start); 3312 } 3313 3314 lwkt_reltoken(&map->token); 3315 3316 return (KERN_SUCCESS); 3317 } 3318 3319 /* 3320 * Remove the given address range from the target map. 3321 * This is the exported form of vm_map_delete. 3322 * 3323 * No requirements. 3324 */ 3325 int 3326 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 3327 { 3328 int result; 3329 int count; 3330 3331 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3332 vm_map_lock(map); 3333 VM_MAP_RANGE_CHECK(map, start, end); 3334 result = vm_map_delete(map, start, end, &count); 3335 vm_map_unlock(map); 3336 vm_map_entry_release(count); 3337 3338 return (result); 3339 } 3340 3341 /* 3342 * Assert that the target map allows the specified privilege on the 3343 * entire address region given. The entire region must be allocated. 3344 * 3345 * The caller must specify whether the vm_map is already locked or not. 3346 */ 3347 boolean_t 3348 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 3349 vm_prot_t protection, boolean_t have_lock) 3350 { 3351 vm_map_entry_t entry; 3352 vm_map_entry_t tmp_entry; 3353 boolean_t result; 3354 3355 if (have_lock == FALSE) 3356 vm_map_lock_read(map); 3357 3358 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 3359 if (have_lock == FALSE) 3360 vm_map_unlock_read(map); 3361 return (FALSE); 3362 } 3363 entry = tmp_entry; 3364 3365 result = TRUE; 3366 while (start < end) { 3367 if (entry == NULL) { 3368 result = FALSE; 3369 break; 3370 } 3371 3372 /* 3373 * No holes allowed! 3374 */ 3375 3376 if (start < entry->ba.start) { 3377 result = FALSE; 3378 break; 3379 } 3380 /* 3381 * Check protection associated with entry. 3382 */ 3383 3384 if ((entry->protection & protection) != protection) { 3385 result = FALSE; 3386 break; 3387 } 3388 /* go to next entry */ 3389 start = entry->ba.end; 3390 entry = vm_map_rb_tree_RB_NEXT(entry); 3391 } 3392 if (have_lock == FALSE) 3393 vm_map_unlock_read(map); 3394 return (result); 3395 } 3396 3397 /* 3398 * vm_map_backing structures are not shared across forks and must be 3399 * replicated. 3400 * 3401 * Generally speaking we must reallocate the backing_ba sequence and 3402 * also adjust it for any changes made to the base entry->ba.start and 3403 * entry->ba.end. The first ba in the chain is of course &entry->ba, 3404 * so we only need to adjust subsequent ba's start, end, and offset. 3405 * 3406 * MAP_BACK_CLIPPED - Called as part of a clipping replication. 3407 * Do not clear OBJ_ONEMAPPING. 3408 * 3409 * MAP_BACK_BASEOBJREFD - Called from vm_map_insert(). The base object 3410 * has already been referenced. 3411 */ 3412 static 3413 void 3414 vm_map_backing_replicated(vm_map_t map, vm_map_entry_t entry, int flags) 3415 { 3416 vm_map_backing_t ba; 3417 vm_map_backing_t nba; 3418 vm_object_t object; 3419 3420 ba = &entry->ba; 3421 for (;;) { 3422 ba->pmap = map->pmap; 3423 3424 if (ba->map_object) { 3425 switch(entry->maptype) { 3426 case VM_MAPTYPE_NORMAL: 3427 object = ba->object; 3428 if (ba != &entry->ba || 3429 (flags & MAP_BACK_BASEOBJREFD) == 0) { 3430 vm_object_reference_quick(object); 3431 } 3432 vm_map_backing_attach(entry, ba); 3433 if ((flags & MAP_BACK_CLIPPED) == 0 && 3434 object->ref_count > 1) { 3435 vm_object_clear_flag(object, 3436 OBJ_ONEMAPPING); 3437 } 3438 break; 3439 case VM_MAPTYPE_UKSMAP: 3440 vm_map_backing_attach(entry, ba); 3441 break; 3442 default: 3443 break; 3444 } 3445 } 3446 if (ba->backing_ba == NULL) 3447 break; 3448 3449 /* 3450 * NOTE: The aux_info field is retained. 3451 */ 3452 nba = kmalloc(sizeof(*nba), M_MAP_BACKING, M_INTWAIT); 3453 *nba = *ba->backing_ba; 3454 nba->offset += (ba->start - nba->start); /* += (new - old) */ 3455 nba->start = ba->start; 3456 nba->end = ba->end; 3457 ba->backing_ba = nba; 3458 ba = nba; 3459 /* pmap is replaced at the top of the loop */ 3460 } 3461 } 3462 3463 static 3464 void 3465 vm_map_backing_adjust_start(vm_map_entry_t entry, vm_ooffset_t start) 3466 { 3467 vm_map_backing_t ba; 3468 3469 if (entry->maptype == VM_MAPTYPE_NORMAL) { 3470 for (ba = &entry->ba; ba; ba = ba->backing_ba) { 3471 if (ba->object) { 3472 lockmgr(&ba->object->backing_lk, LK_EXCLUSIVE); 3473 ba->offset += (start - ba->start); 3474 ba->start = start; 3475 lockmgr(&ba->object->backing_lk, LK_RELEASE); 3476 } else { 3477 ba->offset += (start - ba->start); 3478 ba->start = start; 3479 } 3480 } 3481 } else { 3482 /* not an object and can't be shadowed */ 3483 } 3484 } 3485 3486 static 3487 void 3488 vm_map_backing_adjust_end(vm_map_entry_t entry, vm_ooffset_t end) 3489 { 3490 vm_map_backing_t ba; 3491 3492 if (entry->maptype == VM_MAPTYPE_NORMAL) { 3493 for (ba = &entry->ba; ba; ba = ba->backing_ba) { 3494 if (ba->object) { 3495 lockmgr(&ba->object->backing_lk, LK_EXCLUSIVE); 3496 ba->end = end; 3497 lockmgr(&ba->object->backing_lk, LK_RELEASE); 3498 } else { 3499 ba->end = end; 3500 } 3501 } 3502 } /* else not an object and/or can't be shadowed */ 3503 } 3504 3505 /* 3506 * Handles the dirty work of making src_entry and dst_entry copy-on-write 3507 * after src_entry has been cloned to dst_entry. For normal entries only. 3508 * 3509 * The vm_maps must be exclusively locked. 3510 * The vm_map's token must be held. 3511 * 3512 * Because the maps are locked no faults can be in progress during the 3513 * operation. 3514 */ 3515 static void 3516 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 3517 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 3518 { 3519 vm_object_t obj; 3520 3521 KKASSERT(dst_entry->maptype == VM_MAPTYPE_NORMAL); 3522 3523 if (src_entry->wired_count) { 3524 /* 3525 * Of course, wired down pages can't be set copy-on-write. 3526 * Cause wired pages to be copied into the new map by 3527 * simulating faults (the new pages are pageable) 3528 * 3529 * Scrap ba.object (its ref-count has not yet been adjusted 3530 * so we can just NULL out the field). Remove the backing 3531 * store. 3532 * 3533 * Then call vm_fault_copy_entry() to create a new object 3534 * in dst_entry and copy the wired pages from src to dst. 3535 * 3536 * The fault-copy code doesn't work with virtual page 3537 * tables. 3538 * 3539 * NOTE: obj is not actually an object for all MAPTYPEs, 3540 * just test against NULL. 3541 */ 3542 if (dst_entry->ba.map_object != NULL) { 3543 vm_map_backing_detach(dst_entry, &dst_entry->ba); 3544 dst_entry->ba.map_object = NULL; 3545 vm_map_entry_dispose_ba(dst_entry, 3546 dst_entry->ba.backing_ba); 3547 dst_entry->ba.backing_ba = NULL; 3548 dst_entry->ba.backing_count = 0; 3549 } 3550 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 3551 } else { 3552 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 3553 /* 3554 * If the source entry is not already marked NEEDS_COPY 3555 * we need to write-protect the PTEs. 3556 */ 3557 pmap_protect(src_map->pmap, 3558 src_entry->ba.start, 3559 src_entry->ba.end, 3560 src_entry->protection & ~VM_PROT_WRITE); 3561 } 3562 3563 /* 3564 * dst_entry.ba_object might be stale. Update it (its 3565 * ref-count has not yet been updated so just overwrite 3566 * the field). 3567 * 3568 * If there is no object then we are golden. Also, in 3569 * this situation if there are no backing_ba linkages then 3570 * we can set ba.offset to whatever we want. For now we 3571 * set the offset for 0 for make debugging object sizes 3572 * easier. 3573 */ 3574 obj = src_entry->ba.object; 3575 3576 if (obj) { 3577 src_entry->eflags |= (MAP_ENTRY_COW | 3578 MAP_ENTRY_NEEDS_COPY); 3579 dst_entry->eflags |= (MAP_ENTRY_COW | 3580 MAP_ENTRY_NEEDS_COPY); 3581 KKASSERT(dst_entry->ba.offset == src_entry->ba.offset); 3582 } else { 3583 dst_entry->ba.offset = 0; 3584 } 3585 3586 /* 3587 * Normal, allow the backing_ba link depth to 3588 * increase. 3589 */ 3590 pmap_copy(dst_map->pmap, src_map->pmap, 3591 dst_entry->ba.start, 3592 dst_entry->ba.end - dst_entry->ba.start, 3593 src_entry->ba.start); 3594 } 3595 } 3596 3597 /* 3598 * Create a vmspace for a new process and its related vm_map based on an 3599 * existing vmspace. The new map inherits information from the old map 3600 * according to inheritance settings. 3601 * 3602 * The source map must not be locked. 3603 * No requirements. 3604 */ 3605 static void vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map, 3606 vm_map_entry_t old_entry, int *countp); 3607 static void vmspace_fork_uksmap_entry(struct proc *p2, struct lwp *lp2, 3608 vm_map_t old_map, vm_map_t new_map, 3609 vm_map_entry_t old_entry, int *countp); 3610 3611 struct vmspace * 3612 vmspace_fork(struct vmspace *vm1, struct proc *p2, struct lwp *lp2) 3613 { 3614 struct vmspace *vm2; 3615 vm_map_t old_map = &vm1->vm_map; 3616 vm_map_t new_map; 3617 vm_map_entry_t old_entry; 3618 int count; 3619 3620 lwkt_gettoken(&vm1->vm_map.token); 3621 vm_map_lock(old_map); 3622 3623 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map)); 3624 lwkt_gettoken(&vm2->vm_map.token); 3625 3626 /* 3627 * We must bump the timestamp to force any concurrent fault 3628 * to retry. 3629 */ 3630 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 3631 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 3632 new_map = &vm2->vm_map; /* XXX */ 3633 new_map->timestamp = 1; 3634 3635 vm_map_lock(new_map); 3636 3637 count = old_map->nentries; 3638 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 3639 3640 RB_FOREACH(old_entry, vm_map_rb_tree, &old_map->rb_root) { 3641 switch(old_entry->maptype) { 3642 case VM_MAPTYPE_SUBMAP: 3643 panic("vm_map_fork: encountered a submap"); 3644 break; 3645 case VM_MAPTYPE_UKSMAP: 3646 vmspace_fork_uksmap_entry(p2, lp2, 3647 old_map, new_map, 3648 old_entry, &count); 3649 break; 3650 case VM_MAPTYPE_NORMAL: 3651 vmspace_fork_normal_entry(old_map, new_map, 3652 old_entry, &count); 3653 break; 3654 default: 3655 /* nothing to do */ 3656 break; 3657 } 3658 } 3659 3660 new_map->size = old_map->size; 3661 vm_map_unlock(new_map); 3662 vm_map_unlock(old_map); 3663 vm_map_entry_release(count); 3664 3665 lwkt_reltoken(&vm2->vm_map.token); 3666 lwkt_reltoken(&vm1->vm_map.token); 3667 3668 return (vm2); 3669 } 3670 3671 static 3672 void 3673 vmspace_fork_normal_entry(vm_map_t old_map, vm_map_t new_map, 3674 vm_map_entry_t old_entry, int *countp) 3675 { 3676 vm_map_entry_t new_entry; 3677 vm_map_backing_t ba; 3678 vm_object_t object; 3679 3680 /* 3681 * If the backing_ba link list gets too long then fault it 3682 * all into the head object and dispose of the list. We do 3683 * this in old_entry prior to cloning in order to benefit both 3684 * parent and child. 3685 * 3686 * We can test our fronting object's size against its 3687 * resident_page_count for a really cheap (but probably not perfect) 3688 * all-shadowed test, allowing us to disconnect the backing_ba 3689 * link list early. 3690 */ 3691 object = old_entry->ba.object; 3692 if (old_entry->ba.backing_ba && 3693 (old_entry->ba.backing_count >= vm_map_backing_limit || 3694 (vm_map_backing_shadow_test && object && 3695 object->size == object->resident_page_count))) { 3696 /* 3697 * If there are too many backing_ba linkages we 3698 * collapse everything into the head 3699 * 3700 * This will also remove all the pte's. 3701 */ 3702 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) 3703 vm_map_entry_shadow(old_entry); 3704 if (object == NULL) 3705 vm_map_entry_allocate_object(old_entry); 3706 if (vm_fault_collapse(old_map, old_entry) == KERN_SUCCESS) { 3707 ba = old_entry->ba.backing_ba; 3708 old_entry->ba.backing_ba = NULL; 3709 old_entry->ba.backing_count = 0; 3710 vm_map_entry_dispose_ba(old_entry, ba); 3711 } 3712 } 3713 object = NULL; /* object variable is now invalid */ 3714 3715 /* 3716 * Fork the entry 3717 */ 3718 switch (old_entry->inheritance) { 3719 case VM_INHERIT_NONE: 3720 break; 3721 case VM_INHERIT_SHARE: 3722 /* 3723 * Clone the entry as a shared entry. This will look like 3724 * shared memory across the old and the new process. We must 3725 * ensure that the object is allocated. 3726 */ 3727 if (old_entry->ba.object == NULL) 3728 vm_map_entry_allocate_object(old_entry); 3729 3730 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3731 /* 3732 * Create the fronting vm_map_backing for 3733 * an entry which needs a copy, plus an extra 3734 * ref because we are going to duplicate it 3735 * in the fork. 3736 * 3737 * The call to vm_map_entry_shadow() will also clear 3738 * OBJ_ONEMAPPING. 3739 * 3740 * XXX no more collapse. Still need extra ref 3741 * for the fork. 3742 */ 3743 vm_map_entry_shadow(old_entry); 3744 } else if (old_entry->ba.object) { 3745 object = old_entry->ba.object; 3746 } 3747 3748 /* 3749 * Clone the entry. We've already bumped the ref on 3750 * the vm_object for our new entry. 3751 */ 3752 new_entry = vm_map_entry_create(countp); 3753 *new_entry = *old_entry; 3754 3755 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3756 new_entry->wired_count = 0; 3757 3758 /* 3759 * Replicate and index the vm_map_backing. Don't share 3760 * the vm_map_backing across vm_map's (only across clips). 3761 * 3762 * Insert the entry into the new map -- we know we're 3763 * inserting at the end of the new map. 3764 */ 3765 vm_map_backing_replicated(new_map, new_entry, 0); 3766 vm_map_entry_link(new_map, new_entry); 3767 3768 /* 3769 * Update the physical map 3770 */ 3771 pmap_copy(new_map->pmap, old_map->pmap, 3772 new_entry->ba.start, 3773 (old_entry->ba.end - old_entry->ba.start), 3774 old_entry->ba.start); 3775 break; 3776 case VM_INHERIT_COPY: 3777 /* 3778 * Clone the entry and link the copy into the new map. 3779 * 3780 * Note that ref-counting adjustment for old_entry->ba.object 3781 * (if it isn't a special map that is) is handled by 3782 * vm_map_copy_entry(). 3783 */ 3784 new_entry = vm_map_entry_create(countp); 3785 *new_entry = *old_entry; 3786 3787 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3788 new_entry->wired_count = 0; 3789 3790 vm_map_backing_replicated(new_map, new_entry, 0); 3791 vm_map_entry_link(new_map, new_entry); 3792 3793 /* 3794 * This does the actual dirty work of making both entries 3795 * copy-on-write, and will also handle the fronting object. 3796 */ 3797 vm_map_copy_entry(old_map, new_map, old_entry, new_entry); 3798 break; 3799 } 3800 } 3801 3802 /* 3803 * When forking user-kernel shared maps, the map might change in the 3804 * child so do not try to copy the underlying pmap entries. 3805 */ 3806 static 3807 void 3808 vmspace_fork_uksmap_entry(struct proc *p2, struct lwp *lp2, 3809 vm_map_t old_map, vm_map_t new_map, 3810 vm_map_entry_t old_entry, int *countp) 3811 { 3812 vm_map_entry_t new_entry; 3813 3814 /* 3815 * Do not fork lpmap entries whos TIDs do not match lp2's tid. 3816 * 3817 * XXX if p2 is NULL and lp2 is non-NULL, we retain the lpmap entry 3818 * (this is for e.g. resident'ing vmspace's) but set the field 3819 * to NULL. Upon restore it should be restored. XXX NOT IMPL YET 3820 */ 3821 if (old_entry->aux.dev) { 3822 switch(minor(old_entry->aux.dev)) { 3823 case 5: 3824 break; 3825 case 6: 3826 break; 3827 case 7: 3828 if (lp2 == NULL) 3829 return; 3830 if (old_entry->ba.aux_info == NULL) 3831 return; 3832 if (((struct lwp *)old_entry->ba.aux_info)->lwp_tid != 3833 lp2->lwp_tid) 3834 return; 3835 break; 3836 } 3837 } 3838 3839 new_entry = vm_map_entry_create(countp); 3840 *new_entry = *old_entry; 3841 3842 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3843 new_entry->wired_count = 0; 3844 KKASSERT(new_entry->ba.backing_ba == NULL); 3845 3846 if (new_entry->aux.dev) { 3847 switch(minor(new_entry->aux.dev)) { 3848 case 5: 3849 /* 3850 * upmap 3851 */ 3852 new_entry->ba.aux_info = p2; 3853 break; 3854 case 6: 3855 /* 3856 * kpmap 3857 */ 3858 new_entry->ba.aux_info = NULL; 3859 break; 3860 case 7: 3861 /* 3862 * lpmap 3863 */ 3864 new_entry->ba.aux_info = lp2; 3865 break; 3866 } 3867 } else { 3868 new_entry->ba.aux_info = NULL; 3869 } 3870 3871 vm_map_backing_replicated(new_map, new_entry, 0); 3872 3873 vm_map_entry_link(new_map, new_entry); 3874 } 3875 3876 /* 3877 * Create an auto-grow stack entry 3878 * 3879 * No requirements. 3880 */ 3881 int 3882 vm_map_stack (vm_map_t map, vm_offset_t *addrbos, vm_size_t max_ssize, 3883 int flags, vm_prot_t prot, vm_prot_t max, int cow) 3884 { 3885 vm_map_entry_t prev_entry; 3886 vm_map_entry_t next; 3887 vm_size_t init_ssize; 3888 int rv; 3889 int count; 3890 vm_offset_t tmpaddr; 3891 3892 cow |= MAP_IS_STACK; 3893 3894 if (max_ssize < sgrowsiz) 3895 init_ssize = max_ssize; 3896 else 3897 init_ssize = sgrowsiz; 3898 3899 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3900 vm_map_lock(map); 3901 3902 /* 3903 * Find space for the mapping 3904 */ 3905 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) { 3906 if (vm_map_findspace(map, *addrbos, max_ssize, 1, 3907 flags, &tmpaddr)) { 3908 vm_map_unlock(map); 3909 vm_map_entry_release(count); 3910 return (KERN_NO_SPACE); 3911 } 3912 *addrbos = tmpaddr; 3913 } 3914 3915 /* If addr is already mapped, no go */ 3916 if (vm_map_lookup_entry(map, *addrbos, &prev_entry)) { 3917 vm_map_unlock(map); 3918 vm_map_entry_release(count); 3919 return (KERN_NO_SPACE); 3920 } 3921 3922 #if 0 3923 /* XXX already handled by kern_mmap() */ 3924 /* If we would blow our VMEM resource limit, no go */ 3925 if (map->size + init_ssize > 3926 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3927 vm_map_unlock(map); 3928 vm_map_entry_release(count); 3929 return (KERN_NO_SPACE); 3930 } 3931 #endif 3932 3933 /* 3934 * If we can't accomodate max_ssize in the current mapping, 3935 * no go. However, we need to be aware that subsequent user 3936 * mappings might map into the space we have reserved for 3937 * stack, and currently this space is not protected. 3938 * 3939 * Hopefully we will at least detect this condition 3940 * when we try to grow the stack. 3941 */ 3942 if (prev_entry) 3943 next = vm_map_rb_tree_RB_NEXT(prev_entry); 3944 else 3945 next = RB_MIN(vm_map_rb_tree, &map->rb_root); 3946 3947 if (next && next->ba.start < *addrbos + max_ssize) { 3948 vm_map_unlock(map); 3949 vm_map_entry_release(count); 3950 return (KERN_NO_SPACE); 3951 } 3952 3953 /* 3954 * We initially map a stack of only init_ssize. We will 3955 * grow as needed later. Since this is to be a grow 3956 * down stack, we map at the top of the range. 3957 * 3958 * Note: we would normally expect prot and max to be 3959 * VM_PROT_ALL, and cow to be 0. Possibly we should 3960 * eliminate these as input parameters, and just 3961 * pass these values here in the insert call. 3962 */ 3963 rv = vm_map_insert(map, &count, 3964 NULL, NULL, 3965 0, NULL, 3966 *addrbos + max_ssize - init_ssize, 3967 *addrbos + max_ssize, 3968 VM_MAPTYPE_NORMAL, 3969 VM_SUBSYS_STACK, prot, max, cow); 3970 3971 /* Now set the avail_ssize amount */ 3972 if (rv == KERN_SUCCESS) { 3973 if (prev_entry) 3974 next = vm_map_rb_tree_RB_NEXT(prev_entry); 3975 else 3976 next = RB_MIN(vm_map_rb_tree, &map->rb_root); 3977 if (prev_entry != NULL) { 3978 vm_map_clip_end(map, 3979 prev_entry, 3980 *addrbos + max_ssize - init_ssize, 3981 &count); 3982 } 3983 if (next->ba.end != *addrbos + max_ssize || 3984 next->ba.start != *addrbos + max_ssize - init_ssize){ 3985 panic ("Bad entry start/end for new stack entry"); 3986 } else { 3987 next->aux.avail_ssize = max_ssize - init_ssize; 3988 } 3989 } 3990 3991 vm_map_unlock(map); 3992 vm_map_entry_release(count); 3993 return (rv); 3994 } 3995 3996 /* 3997 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 3998 * desired address is already mapped, or if we successfully grow 3999 * the stack. Also returns KERN_SUCCESS if addr is outside the 4000 * stack range (this is strange, but preserves compatibility with 4001 * the grow function in vm_machdep.c). 4002 * 4003 * No requirements. 4004 */ 4005 int 4006 vm_map_growstack (vm_map_t map, vm_offset_t addr) 4007 { 4008 vm_map_entry_t prev_entry; 4009 vm_map_entry_t stack_entry; 4010 vm_map_entry_t next; 4011 struct vmspace *vm; 4012 struct lwp *lp; 4013 struct proc *p; 4014 vm_offset_t end; 4015 int grow_amount; 4016 int rv = KERN_SUCCESS; 4017 int is_procstack; 4018 int use_read_lock = 1; 4019 int count; 4020 4021 /* 4022 * Find the vm 4023 */ 4024 lp = curthread->td_lwp; 4025 p = curthread->td_proc; 4026 KKASSERT(lp != NULL); 4027 vm = lp->lwp_vmspace; 4028 4029 /* 4030 * Growstack is only allowed on the current process. We disallow 4031 * other use cases, e.g. trying to access memory via procfs that 4032 * the stack hasn't grown into. 4033 */ 4034 if (map != &vm->vm_map) { 4035 return KERN_FAILURE; 4036 } 4037 4038 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 4039 Retry: 4040 if (use_read_lock) 4041 vm_map_lock_read(map); 4042 else 4043 vm_map_lock(map); 4044 4045 /* 4046 * If addr is already in the entry range, no need to grow. 4047 * prev_entry returns NULL if addr is at the head. 4048 */ 4049 if (vm_map_lookup_entry(map, addr, &prev_entry)) 4050 goto done; 4051 if (prev_entry) 4052 stack_entry = vm_map_rb_tree_RB_NEXT(prev_entry); 4053 else 4054 stack_entry = RB_MIN(vm_map_rb_tree, &map->rb_root); 4055 4056 if (stack_entry == NULL) 4057 goto done; 4058 if (prev_entry == NULL) 4059 end = stack_entry->ba.start - stack_entry->aux.avail_ssize; 4060 else 4061 end = prev_entry->ba.end; 4062 4063 /* 4064 * This next test mimics the old grow function in vm_machdep.c. 4065 * It really doesn't quite make sense, but we do it anyway 4066 * for compatibility. 4067 * 4068 * If not growable stack, return success. This signals the 4069 * caller to proceed as he would normally with normal vm. 4070 */ 4071 if (stack_entry->aux.avail_ssize < 1 || 4072 addr >= stack_entry->ba.start || 4073 addr < stack_entry->ba.start - stack_entry->aux.avail_ssize) { 4074 goto done; 4075 } 4076 4077 /* Find the minimum grow amount */ 4078 grow_amount = roundup (stack_entry->ba.start - addr, PAGE_SIZE); 4079 if (grow_amount > stack_entry->aux.avail_ssize) { 4080 rv = KERN_NO_SPACE; 4081 goto done; 4082 } 4083 4084 /* 4085 * If there is no longer enough space between the entries 4086 * nogo, and adjust the available space. Note: this 4087 * should only happen if the user has mapped into the 4088 * stack area after the stack was created, and is 4089 * probably an error. 4090 * 4091 * This also effectively destroys any guard page the user 4092 * might have intended by limiting the stack size. 4093 */ 4094 if (grow_amount > stack_entry->ba.start - end) { 4095 if (use_read_lock && vm_map_lock_upgrade(map)) { 4096 /* lost lock */ 4097 use_read_lock = 0; 4098 goto Retry; 4099 } 4100 use_read_lock = 0; 4101 stack_entry->aux.avail_ssize = stack_entry->ba.start - end; 4102 rv = KERN_NO_SPACE; 4103 goto done; 4104 } 4105 4106 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 4107 4108 /* If this is the main process stack, see if we're over the 4109 * stack limit. 4110 */ 4111 if (is_procstack && (vm->vm_ssize + grow_amount > 4112 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 4113 rv = KERN_NO_SPACE; 4114 goto done; 4115 } 4116 4117 /* Round up the grow amount modulo SGROWSIZ */ 4118 grow_amount = roundup (grow_amount, sgrowsiz); 4119 if (grow_amount > stack_entry->aux.avail_ssize) { 4120 grow_amount = stack_entry->aux.avail_ssize; 4121 } 4122 if (is_procstack && (vm->vm_ssize + grow_amount > 4123 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 4124 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - vm->vm_ssize; 4125 } 4126 4127 /* If we would blow our VMEM resource limit, no go */ 4128 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 4129 rv = KERN_NO_SPACE; 4130 goto done; 4131 } 4132 4133 if (use_read_lock && vm_map_lock_upgrade(map)) { 4134 /* lost lock */ 4135 use_read_lock = 0; 4136 goto Retry; 4137 } 4138 use_read_lock = 0; 4139 4140 /* Get the preliminary new entry start value */ 4141 addr = stack_entry->ba.start - grow_amount; 4142 4143 /* If this puts us into the previous entry, cut back our growth 4144 * to the available space. Also, see the note above. 4145 */ 4146 if (addr < end) { 4147 stack_entry->aux.avail_ssize = stack_entry->ba.start - end; 4148 addr = end; 4149 } 4150 4151 rv = vm_map_insert(map, &count, 4152 NULL, NULL, 4153 0, NULL, 4154 addr, stack_entry->ba.start, 4155 VM_MAPTYPE_NORMAL, 4156 VM_SUBSYS_STACK, VM_PROT_ALL, VM_PROT_ALL, 0); 4157 4158 /* Adjust the available stack space by the amount we grew. */ 4159 if (rv == KERN_SUCCESS) { 4160 if (prev_entry) { 4161 vm_map_clip_end(map, prev_entry, addr, &count); 4162 next = vm_map_rb_tree_RB_NEXT(prev_entry); 4163 } else { 4164 next = RB_MIN(vm_map_rb_tree, &map->rb_root); 4165 } 4166 if (next->ba.end != stack_entry->ba.start || 4167 next->ba.start != addr) { 4168 panic ("Bad stack grow start/end in new stack entry"); 4169 } else { 4170 next->aux.avail_ssize = 4171 stack_entry->aux.avail_ssize - 4172 (next->ba.end - next->ba.start); 4173 if (is_procstack) { 4174 vm->vm_ssize += next->ba.end - 4175 next->ba.start; 4176 } 4177 } 4178 4179 if (map->flags & MAP_WIREFUTURE) { 4180 vm_map_user_wiring(map, 4181 next->ba.start, 4182 next->ba.end, 4183 FALSE); 4184 } 4185 } 4186 4187 done: 4188 if (use_read_lock) 4189 vm_map_unlock_read(map); 4190 else 4191 vm_map_unlock(map); 4192 vm_map_entry_release(count); 4193 return (rv); 4194 } 4195 4196 /* 4197 * Unshare the specified VM space for exec. If other processes are 4198 * mapped to it, then create a new one. The new vmspace is null. 4199 * 4200 * No requirements. 4201 */ 4202 void 4203 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 4204 { 4205 struct vmspace *oldvmspace = p->p_vmspace; 4206 struct vmspace *newvmspace; 4207 vm_map_t map = &p->p_vmspace->vm_map; 4208 4209 /* 4210 * If we are execing a resident vmspace we fork it, otherwise 4211 * we create a new vmspace. Note that exitingcnt is not 4212 * copied to the new vmspace. 4213 */ 4214 lwkt_gettoken(&oldvmspace->vm_map.token); 4215 if (vmcopy) { 4216 newvmspace = vmspace_fork(vmcopy, NULL, NULL); 4217 lwkt_gettoken(&newvmspace->vm_map.token); 4218 } else { 4219 newvmspace = vmspace_alloc(vm_map_min(map), vm_map_max(map)); 4220 lwkt_gettoken(&newvmspace->vm_map.token); 4221 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 4222 (caddr_t)&oldvmspace->vm_endcopy - 4223 (caddr_t)&oldvmspace->vm_startcopy); 4224 } 4225 4226 /* 4227 * Finish initializing the vmspace before assigning it 4228 * to the process. The vmspace will become the current vmspace 4229 * if p == curproc. 4230 */ 4231 pmap_pinit2(vmspace_pmap(newvmspace)); 4232 pmap_replacevm(p, newvmspace, 0); 4233 lwkt_reltoken(&newvmspace->vm_map.token); 4234 lwkt_reltoken(&oldvmspace->vm_map.token); 4235 vmspace_rel(oldvmspace); 4236 } 4237 4238 /* 4239 * Unshare the specified VM space for forcing COW. This 4240 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 4241 */ 4242 void 4243 vmspace_unshare(struct proc *p) 4244 { 4245 struct vmspace *oldvmspace = p->p_vmspace; 4246 struct vmspace *newvmspace; 4247 4248 lwkt_gettoken(&oldvmspace->vm_map.token); 4249 if (vmspace_getrefs(oldvmspace) == 1) { 4250 lwkt_reltoken(&oldvmspace->vm_map.token); 4251 return; 4252 } 4253 newvmspace = vmspace_fork(oldvmspace, NULL, NULL); 4254 lwkt_gettoken(&newvmspace->vm_map.token); 4255 pmap_pinit2(vmspace_pmap(newvmspace)); 4256 pmap_replacevm(p, newvmspace, 0); 4257 lwkt_reltoken(&newvmspace->vm_map.token); 4258 lwkt_reltoken(&oldvmspace->vm_map.token); 4259 vmspace_rel(oldvmspace); 4260 } 4261 4262 /* 4263 * vm_map_hint: return the beginning of the best area suitable for 4264 * creating a new mapping with "prot" protection. 4265 * 4266 * No requirements. 4267 */ 4268 vm_offset_t 4269 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot) 4270 { 4271 struct vmspace *vms = p->p_vmspace; 4272 struct rlimit limit; 4273 rlim_t dsiz; 4274 4275 /* 4276 * Acquire datasize limit for mmap() operation, 4277 * calculate nearest power of 2. 4278 */ 4279 if (kern_getrlimit(RLIMIT_DATA, &limit)) 4280 limit.rlim_cur = maxdsiz; 4281 dsiz = limit.rlim_cur; 4282 4283 if (!randomize_mmap || addr != 0) { 4284 /* 4285 * Set a reasonable start point for the hint if it was 4286 * not specified or if it falls within the heap space. 4287 * Hinted mmap()s do not allocate out of the heap space. 4288 */ 4289 if (addr == 0 || 4290 (addr >= round_page((vm_offset_t)vms->vm_taddr) && 4291 addr < round_page((vm_offset_t)vms->vm_daddr + dsiz))) { 4292 addr = round_page((vm_offset_t)vms->vm_daddr + dsiz); 4293 } 4294 4295 return addr; 4296 } 4297 4298 /* 4299 * randomize_mmap && addr == 0. For now randomize the 4300 * address within a dsiz range beyond the data limit. 4301 */ 4302 addr = (vm_offset_t)vms->vm_daddr + dsiz; 4303 if (dsiz) 4304 addr += (karc4random64() & 0x7FFFFFFFFFFFFFFFLU) % dsiz; 4305 return (round_page(addr)); 4306 } 4307 4308 /* 4309 * Finds the VM object, offset, and protection for a given virtual address 4310 * in the specified map, assuming a page fault of the type specified. 4311 * 4312 * Leaves the map in question locked for read; return values are guaranteed 4313 * until a vm_map_lookup_done call is performed. Note that the map argument 4314 * is in/out; the returned map must be used in the call to vm_map_lookup_done. 4315 * 4316 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make 4317 * that fast. 4318 * 4319 * If a lookup is requested with "write protection" specified, the map may 4320 * be changed to perform virtual copying operations, although the data 4321 * referenced will remain the same. 4322 * 4323 * No requirements. 4324 */ 4325 int 4326 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4327 vm_offset_t vaddr, 4328 vm_prot_t fault_typea, 4329 vm_map_entry_t *out_entry, /* OUT */ 4330 struct vm_map_backing **bap, /* OUT */ 4331 vm_pindex_t *pindex, /* OUT */ 4332 vm_pindex_t *pcount, /* OUT */ 4333 vm_prot_t *out_prot, /* OUT */ 4334 int *wflags) /* OUT */ 4335 { 4336 vm_map_entry_t entry; 4337 vm_map_t map = *var_map; 4338 vm_prot_t prot; 4339 vm_prot_t fault_type = fault_typea; 4340 int use_read_lock = 1; 4341 int rv = KERN_SUCCESS; 4342 int count; 4343 thread_t td = curthread; 4344 4345 /* 4346 * vm_map_entry_reserve() implements an important mitigation 4347 * against mmap() span running the kernel out of vm_map_entry 4348 * structures, but it can also cause an infinite call recursion. 4349 * Use td_nest_count to prevent an infinite recursion (allows 4350 * the vm_map code to dig into the pcpu vm_map_entry reserve). 4351 */ 4352 count = 0; 4353 if (td->td_nest_count == 0) { 4354 ++td->td_nest_count; 4355 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 4356 --td->td_nest_count; 4357 } 4358 RetryLookup: 4359 if (use_read_lock) 4360 vm_map_lock_read(map); 4361 else 4362 vm_map_lock(map); 4363 4364 /* 4365 * Always do a full lookup. The hint doesn't get us much anymore 4366 * now that the map is RB'd. 4367 */ 4368 cpu_ccfence(); 4369 *out_entry = NULL; 4370 *bap = NULL; 4371 4372 { 4373 vm_map_entry_t tmp_entry; 4374 4375 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 4376 rv = KERN_INVALID_ADDRESS; 4377 goto done; 4378 } 4379 entry = tmp_entry; 4380 *out_entry = entry; 4381 } 4382 4383 /* 4384 * Handle submaps. 4385 */ 4386 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 4387 vm_map_t old_map = map; 4388 4389 *var_map = map = entry->ba.sub_map; 4390 if (use_read_lock) 4391 vm_map_unlock_read(old_map); 4392 else 4393 vm_map_unlock(old_map); 4394 use_read_lock = 1; 4395 goto RetryLookup; 4396 } 4397 4398 /* 4399 * Check whether this task is allowed to have this page. 4400 * Note the special case for MAP_ENTRY_COW pages with an override. 4401 * This is to implement a forced COW for debuggers. 4402 */ 4403 if (fault_type & VM_PROT_OVERRIDE_WRITE) 4404 prot = entry->max_protection; 4405 else 4406 prot = entry->protection; 4407 4408 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 4409 if ((fault_type & prot) != fault_type) { 4410 rv = KERN_PROTECTION_FAILURE; 4411 goto done; 4412 } 4413 4414 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 4415 (entry->eflags & MAP_ENTRY_COW) && 4416 (fault_type & VM_PROT_WRITE) && 4417 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 4418 rv = KERN_PROTECTION_FAILURE; 4419 goto done; 4420 } 4421 4422 /* 4423 * If this page is not pageable, we have to get it for all possible 4424 * accesses. 4425 */ 4426 *wflags = 0; 4427 if (entry->wired_count) { 4428 *wflags |= FW_WIRED; 4429 prot = fault_type = entry->protection; 4430 } 4431 4432 if (curthread->td_lwp && curthread->td_lwp->lwp_vmspace && 4433 pmap_emulate_ad_bits(&curthread->td_lwp->lwp_vmspace->vm_pmap)) { 4434 if ((prot & VM_PROT_WRITE) == 0) 4435 fault_type |= VM_PROT_WRITE; 4436 } 4437 4438 /* 4439 * Only NORMAL maps are object-based. UKSMAPs are not. 4440 */ 4441 if (entry->maptype != VM_MAPTYPE_NORMAL) { 4442 *bap = NULL; 4443 goto skip; 4444 } 4445 4446 /* 4447 * If the entry was copy-on-write, we either ... 4448 */ 4449 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4450 /* 4451 * If we want to write the page, we may as well handle that 4452 * now since we've got the map locked. 4453 * 4454 * If we don't need to write the page, we just demote the 4455 * permissions allowed. 4456 */ 4457 if (fault_type & VM_PROT_WRITE) { 4458 /* 4459 * Not allowed if TDF_NOFAULT is set as the shadowing 4460 * operation can deadlock against the faulting 4461 * function due to the copy-on-write. 4462 */ 4463 if (curthread->td_flags & TDF_NOFAULT) { 4464 rv = KERN_FAILURE_NOFAULT; 4465 goto done; 4466 } 4467 4468 /* 4469 * Make a new vm_map_backing + object, and place it 4470 * in the object chain. Note that no new references 4471 * have appeared -- one just moved from the map to 4472 * the new object. 4473 */ 4474 if (use_read_lock && vm_map_lock_upgrade(map)) { 4475 /* lost lock */ 4476 use_read_lock = 0; 4477 goto RetryLookup; 4478 } 4479 use_read_lock = 0; 4480 vm_map_entry_shadow(entry); 4481 *wflags |= FW_DIDCOW; 4482 } else { 4483 /* 4484 * We're attempting to read a copy-on-write page -- 4485 * don't allow writes. 4486 */ 4487 prot &= ~VM_PROT_WRITE; 4488 } 4489 } 4490 4491 /* 4492 * Create an object if necessary. This code also handles 4493 * partitioning large entries to improve vm_fault performance. 4494 */ 4495 if (entry->ba.object == NULL && !map->system_map) { 4496 if (use_read_lock && vm_map_lock_upgrade(map)) { 4497 /* lost lock */ 4498 use_read_lock = 0; 4499 goto RetryLookup; 4500 } 4501 use_read_lock = 0; 4502 4503 /* 4504 * Partition large entries, giving each its own VM object, 4505 * to improve concurrent fault performance. This is only 4506 * applicable to userspace. 4507 */ 4508 if (map != kernel_map && 4509 entry->maptype == VM_MAPTYPE_NORMAL && 4510 ((entry->ba.start ^ entry->ba.end) & 4511 ~MAP_ENTRY_PARTITION_MASK) && 4512 vm_map_partition_enable) { 4513 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 4514 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 4515 ++mycpu->gd_cnt.v_intrans_coll; 4516 ++mycpu->gd_cnt.v_intrans_wait; 4517 vm_map_transition_wait(map, 0); 4518 goto RetryLookup; 4519 } 4520 vm_map_entry_partition(map, entry, vaddr, &count); 4521 } 4522 vm_map_entry_allocate_object(entry); 4523 } 4524 4525 /* 4526 * Return the object/offset from this entry. If the entry was 4527 * copy-on-write or empty, it has been fixed up. 4528 */ 4529 *bap = &entry->ba; 4530 4531 skip: 4532 *pindex = OFF_TO_IDX((vaddr - entry->ba.start) + entry->ba.offset); 4533 *pcount = OFF_TO_IDX(entry->ba.end - trunc_page(vaddr)); 4534 4535 /* 4536 * Return whether this is the only map sharing this data. On 4537 * success we return with a read lock held on the map. On failure 4538 * we return with the map unlocked. 4539 */ 4540 *out_prot = prot; 4541 done: 4542 if (rv == KERN_SUCCESS) { 4543 if (use_read_lock == 0) 4544 vm_map_lock_downgrade(map); 4545 } else if (use_read_lock) { 4546 vm_map_unlock_read(map); 4547 } else { 4548 vm_map_unlock(map); 4549 } 4550 if (count > 0) 4551 vm_map_entry_release(count); 4552 4553 return (rv); 4554 } 4555 4556 /* 4557 * Releases locks acquired by a vm_map_lookup() 4558 * (according to the handle returned by that lookup). 4559 * 4560 * No other requirements. 4561 */ 4562 void 4563 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 4564 { 4565 /* 4566 * Unlock the main-level map 4567 */ 4568 vm_map_unlock_read(map); 4569 if (count) 4570 vm_map_entry_release(count); 4571 } 4572 4573 static void 4574 vm_map_entry_partition(vm_map_t map, vm_map_entry_t entry, 4575 vm_offset_t vaddr, int *countp) 4576 { 4577 vaddr &= ~MAP_ENTRY_PARTITION_MASK; 4578 vm_map_clip_start(map, entry, vaddr, countp); 4579 vaddr += MAP_ENTRY_PARTITION_SIZE; 4580 vm_map_clip_end(map, entry, vaddr, countp); 4581 } 4582 4583 /* 4584 * Quick hack, needs some help to make it more SMP friendly. 4585 */ 4586 void 4587 vm_map_interlock(vm_map_t map, struct vm_map_ilock *ilock, 4588 vm_offset_t ran_beg, vm_offset_t ran_end) 4589 { 4590 struct vm_map_ilock *scan; 4591 4592 ilock->ran_beg = ran_beg; 4593 ilock->ran_end = ran_end; 4594 ilock->flags = 0; 4595 4596 spin_lock(&map->ilock_spin); 4597 restart: 4598 for (scan = map->ilock_base; scan; scan = scan->next) { 4599 if (ran_end > scan->ran_beg && ran_beg < scan->ran_end) { 4600 scan->flags |= ILOCK_WAITING; 4601 ssleep(scan, &map->ilock_spin, 0, "ilock", 0); 4602 goto restart; 4603 } 4604 } 4605 ilock->next = map->ilock_base; 4606 map->ilock_base = ilock; 4607 spin_unlock(&map->ilock_spin); 4608 } 4609 4610 void 4611 vm_map_deinterlock(vm_map_t map, struct vm_map_ilock *ilock) 4612 { 4613 struct vm_map_ilock *scan; 4614 struct vm_map_ilock **scanp; 4615 4616 spin_lock(&map->ilock_spin); 4617 scanp = &map->ilock_base; 4618 while ((scan = *scanp) != NULL) { 4619 if (scan == ilock) { 4620 *scanp = ilock->next; 4621 spin_unlock(&map->ilock_spin); 4622 if (ilock->flags & ILOCK_WAITING) 4623 wakeup(ilock); 4624 return; 4625 } 4626 scanp = &scan->next; 4627 } 4628 spin_unlock(&map->ilock_spin); 4629 panic("vm_map_deinterlock: missing ilock!"); 4630 } 4631 4632 #include "opt_ddb.h" 4633 #ifdef DDB 4634 #include <ddb/ddb.h> 4635 4636 /* 4637 * Debugging only 4638 */ 4639 DB_SHOW_COMMAND(map, vm_map_print) 4640 { 4641 static int nlines; 4642 /* XXX convert args. */ 4643 vm_map_t map = (vm_map_t)addr; 4644 boolean_t full = have_addr; 4645 4646 vm_map_entry_t entry; 4647 4648 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 4649 (void *)map, 4650 (void *)map->pmap, map->nentries, map->timestamp); 4651 nlines++; 4652 4653 if (!full && db_indent) 4654 return; 4655 4656 db_indent += 2; 4657 RB_FOREACH(entry, vm_map_rb_tree, &map->rb_root) { 4658 db_iprintf("map entry %p: start=%p, end=%p\n", 4659 (void *)entry, 4660 (void *)entry->ba.start, (void *)entry->ba.end); 4661 nlines++; 4662 { 4663 static char *inheritance_name[4] = 4664 {"share", "copy", "none", "donate_copy"}; 4665 4666 db_iprintf(" prot=%x/%x/%s", 4667 entry->protection, 4668 entry->max_protection, 4669 inheritance_name[(int)(unsigned char) 4670 entry->inheritance]); 4671 if (entry->wired_count != 0) 4672 db_printf(", wired"); 4673 } 4674 switch(entry->maptype) { 4675 case VM_MAPTYPE_SUBMAP: 4676 /* XXX no %qd in kernel. Truncate entry->ba.offset. */ 4677 db_printf(", share=%p, offset=0x%lx\n", 4678 (void *)entry->ba.sub_map, 4679 (long)entry->ba.offset); 4680 nlines++; 4681 4682 db_indent += 2; 4683 vm_map_print((db_expr_t)(intptr_t)entry->ba.sub_map, 4684 full, 0, NULL); 4685 db_indent -= 2; 4686 break; 4687 case VM_MAPTYPE_NORMAL: 4688 /* XXX no %qd in kernel. Truncate entry->ba.offset. */ 4689 db_printf(", object=%p, offset=0x%lx", 4690 (void *)entry->ba.object, 4691 (long)entry->ba.offset); 4692 if (entry->eflags & MAP_ENTRY_COW) 4693 db_printf(", copy (%s)", 4694 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) ? 4695 "needed" : "done")); 4696 db_printf("\n"); 4697 nlines++; 4698 4699 if (entry->ba.object) { 4700 db_indent += 2; 4701 vm_object_print((db_expr_t)(intptr_t) 4702 entry->ba.object, 4703 full, 0, NULL); 4704 nlines += 4; 4705 db_indent -= 2; 4706 } 4707 break; 4708 case VM_MAPTYPE_UKSMAP: 4709 db_printf(", uksmap=%p, offset=0x%lx", 4710 (void *)entry->ba.uksmap, 4711 (long)entry->ba.offset); 4712 if (entry->eflags & MAP_ENTRY_COW) 4713 db_printf(", copy (%s)", 4714 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4715 db_printf("\n"); 4716 nlines++; 4717 break; 4718 default: 4719 break; 4720 } 4721 } 4722 db_indent -= 2; 4723 if (db_indent == 0) 4724 nlines = 0; 4725 } 4726 4727 /* 4728 * Debugging only 4729 */ 4730 DB_SHOW_COMMAND(procvm, procvm) 4731 { 4732 struct proc *p; 4733 4734 if (have_addr) { 4735 p = (struct proc *) addr; 4736 } else { 4737 p = curproc; 4738 } 4739 4740 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 4741 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 4742 (void *)vmspace_pmap(p->p_vmspace)); 4743 4744 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 4745 } 4746 4747 #endif /* DDB */ 4748