1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94 35 * 36 * 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Virtual memory mapping module. 65 */ 66 67 #include <sys/cdefs.h> 68 __FBSDID("$FreeBSD$"); 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/kernel.h> 73 #include <sys/ktr.h> 74 #include <sys/lock.h> 75 #include <sys/mutex.h> 76 #include <sys/proc.h> 77 #include <sys/vmmeter.h> 78 #include <sys/mman.h> 79 #include <sys/vnode.h> 80 #include <sys/racct.h> 81 #include <sys/resourcevar.h> 82 #include <sys/rwlock.h> 83 #include <sys/file.h> 84 #include <sys/sysctl.h> 85 #include <sys/sysent.h> 86 #include <sys/shm.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_page.h> 93 #include <vm/vm_pageout.h> 94 #include <vm/vm_object.h> 95 #include <vm/vm_pager.h> 96 #include <vm/vm_kern.h> 97 #include <vm/vm_extern.h> 98 #include <vm/vnode_pager.h> 99 #include <vm/swap_pager.h> 100 #include <vm/uma.h> 101 102 /* 103 * Virtual memory maps provide for the mapping, protection, 104 * and sharing of virtual memory objects. In addition, 105 * this module provides for an efficient virtual copy of 106 * memory from one map to another. 107 * 108 * Synchronization is required prior to most operations. 109 * 110 * Maps consist of an ordered doubly-linked list of simple 111 * entries; a self-adjusting binary search tree of these 112 * entries is used to speed up lookups. 113 * 114 * Since portions of maps are specified by start/end addresses, 115 * which may not align with existing map entries, all 116 * routines merely "clip" entries to these start/end values. 117 * [That is, an entry is split into two, bordering at a 118 * start or end value.] Note that these clippings may not 119 * always be necessary (as the two resulting entries are then 120 * not changed); however, the clipping is done for convenience. 121 * 122 * As mentioned above, virtual copy operations are performed 123 * by copying VM object references from one map to 124 * another, and then marking both regions as copy-on-write. 125 */ 126 127 static struct mtx map_sleep_mtx; 128 static uma_zone_t mapentzone; 129 static uma_zone_t kmapentzone; 130 static uma_zone_t mapzone; 131 static uma_zone_t vmspace_zone; 132 static int vmspace_zinit(void *mem, int size, int flags); 133 static int vm_map_zinit(void *mem, int ize, int flags); 134 static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, 135 vm_offset_t max); 136 static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map); 137 static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry); 138 static void vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry); 139 static int vm_map_growstack(vm_map_t map, vm_offset_t addr, 140 vm_map_entry_t gap_entry); 141 static void vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 142 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags); 143 #ifdef INVARIANTS 144 static void vm_map_zdtor(void *mem, int size, void *arg); 145 static void vmspace_zdtor(void *mem, int size, void *arg); 146 #endif 147 static int vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, 148 vm_size_t max_ssize, vm_size_t growsize, vm_prot_t prot, vm_prot_t max, 149 int cow); 150 static void vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 151 vm_offset_t failed_addr); 152 153 #define ENTRY_CHARGED(e) ((e)->cred != NULL || \ 154 ((e)->object.vm_object != NULL && (e)->object.vm_object->cred != NULL && \ 155 !((e)->eflags & MAP_ENTRY_NEEDS_COPY))) 156 157 /* 158 * PROC_VMSPACE_{UN,}LOCK() can be a noop as long as vmspaces are type 159 * stable. 160 */ 161 #define PROC_VMSPACE_LOCK(p) do { } while (0) 162 #define PROC_VMSPACE_UNLOCK(p) do { } while (0) 163 164 /* 165 * VM_MAP_RANGE_CHECK: [ internal use only ] 166 * 167 * Asserts that the starting and ending region 168 * addresses fall within the valid range of the map. 169 */ 170 #define VM_MAP_RANGE_CHECK(map, start, end) \ 171 { \ 172 if (start < vm_map_min(map)) \ 173 start = vm_map_min(map); \ 174 if (end > vm_map_max(map)) \ 175 end = vm_map_max(map); \ 176 if (start > end) \ 177 start = end; \ 178 } 179 180 /* 181 * vm_map_startup: 182 * 183 * Initialize the vm_map module. Must be called before 184 * any other vm_map routines. 185 * 186 * Map and entry structures are allocated from the general 187 * purpose memory pool with some exceptions: 188 * 189 * - The kernel map and kmem submap are allocated statically. 190 * - Kernel map entries are allocated out of a static pool. 191 * 192 * These restrictions are necessary since malloc() uses the 193 * maps and requires map entries. 194 */ 195 196 void 197 vm_map_startup(void) 198 { 199 mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF); 200 mapzone = uma_zcreate("MAP", sizeof(struct vm_map), NULL, 201 #ifdef INVARIANTS 202 vm_map_zdtor, 203 #else 204 NULL, 205 #endif 206 vm_map_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 207 uma_prealloc(mapzone, MAX_KMAP); 208 kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry), 209 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 210 UMA_ZONE_MTXCLASS | UMA_ZONE_VM); 211 mapentzone = uma_zcreate("MAP ENTRY", sizeof(struct vm_map_entry), 212 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 213 vmspace_zone = uma_zcreate("VMSPACE", sizeof(struct vmspace), NULL, 214 #ifdef INVARIANTS 215 vmspace_zdtor, 216 #else 217 NULL, 218 #endif 219 vmspace_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 220 } 221 222 static int 223 vmspace_zinit(void *mem, int size, int flags) 224 { 225 struct vmspace *vm; 226 227 vm = (struct vmspace *)mem; 228 229 vm->vm_map.pmap = NULL; 230 (void)vm_map_zinit(&vm->vm_map, sizeof(vm->vm_map), flags); 231 PMAP_LOCK_INIT(vmspace_pmap(vm)); 232 return (0); 233 } 234 235 static int 236 vm_map_zinit(void *mem, int size, int flags) 237 { 238 vm_map_t map; 239 240 map = (vm_map_t)mem; 241 memset(map, 0, sizeof(*map)); 242 mtx_init(&map->system_mtx, "vm map (system)", NULL, MTX_DEF | MTX_DUPOK); 243 sx_init(&map->lock, "vm map (user)"); 244 return (0); 245 } 246 247 #ifdef INVARIANTS 248 static void 249 vmspace_zdtor(void *mem, int size, void *arg) 250 { 251 struct vmspace *vm; 252 253 vm = (struct vmspace *)mem; 254 255 vm_map_zdtor(&vm->vm_map, sizeof(vm->vm_map), arg); 256 } 257 static void 258 vm_map_zdtor(void *mem, int size, void *arg) 259 { 260 vm_map_t map; 261 262 map = (vm_map_t)mem; 263 KASSERT(map->nentries == 0, 264 ("map %p nentries == %d on free.", 265 map, map->nentries)); 266 KASSERT(map->size == 0, 267 ("map %p size == %lu on free.", 268 map, (unsigned long)map->size)); 269 } 270 #endif /* INVARIANTS */ 271 272 /* 273 * Allocate a vmspace structure, including a vm_map and pmap, 274 * and initialize those structures. The refcnt is set to 1. 275 * 276 * If 'pinit' is NULL then the embedded pmap is initialized via pmap_pinit(). 277 */ 278 struct vmspace * 279 vmspace_alloc(vm_offset_t min, vm_offset_t max, pmap_pinit_t pinit) 280 { 281 struct vmspace *vm; 282 283 vm = uma_zalloc(vmspace_zone, M_WAITOK); 284 KASSERT(vm->vm_map.pmap == NULL, ("vm_map.pmap must be NULL")); 285 if (!pinit(vmspace_pmap(vm))) { 286 uma_zfree(vmspace_zone, vm); 287 return (NULL); 288 } 289 CTR1(KTR_VM, "vmspace_alloc: %p", vm); 290 _vm_map_init(&vm->vm_map, vmspace_pmap(vm), min, max); 291 vm->vm_refcnt = 1; 292 vm->vm_shm = NULL; 293 vm->vm_swrss = 0; 294 vm->vm_tsize = 0; 295 vm->vm_dsize = 0; 296 vm->vm_ssize = 0; 297 vm->vm_taddr = 0; 298 vm->vm_daddr = 0; 299 vm->vm_maxsaddr = 0; 300 return (vm); 301 } 302 303 #ifdef RACCT 304 static void 305 vmspace_container_reset(struct proc *p) 306 { 307 308 PROC_LOCK(p); 309 racct_set(p, RACCT_DATA, 0); 310 racct_set(p, RACCT_STACK, 0); 311 racct_set(p, RACCT_RSS, 0); 312 racct_set(p, RACCT_MEMLOCK, 0); 313 racct_set(p, RACCT_VMEM, 0); 314 PROC_UNLOCK(p); 315 } 316 #endif 317 318 static inline void 319 vmspace_dofree(struct vmspace *vm) 320 { 321 322 CTR1(KTR_VM, "vmspace_free: %p", vm); 323 324 /* 325 * Make sure any SysV shm is freed, it might not have been in 326 * exit1(). 327 */ 328 shmexit(vm); 329 330 /* 331 * Lock the map, to wait out all other references to it. 332 * Delete all of the mappings and pages they hold, then call 333 * the pmap module to reclaim anything left. 334 */ 335 (void)vm_map_remove(&vm->vm_map, vm_map_min(&vm->vm_map), 336 vm_map_max(&vm->vm_map)); 337 338 pmap_release(vmspace_pmap(vm)); 339 vm->vm_map.pmap = NULL; 340 uma_zfree(vmspace_zone, vm); 341 } 342 343 void 344 vmspace_free(struct vmspace *vm) 345 { 346 347 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 348 "vmspace_free() called"); 349 350 if (vm->vm_refcnt == 0) 351 panic("vmspace_free: attempt to free already freed vmspace"); 352 353 if (atomic_fetchadd_int(&vm->vm_refcnt, -1) == 1) 354 vmspace_dofree(vm); 355 } 356 357 void 358 vmspace_exitfree(struct proc *p) 359 { 360 struct vmspace *vm; 361 362 PROC_VMSPACE_LOCK(p); 363 vm = p->p_vmspace; 364 p->p_vmspace = NULL; 365 PROC_VMSPACE_UNLOCK(p); 366 KASSERT(vm == &vmspace0, ("vmspace_exitfree: wrong vmspace")); 367 vmspace_free(vm); 368 } 369 370 void 371 vmspace_exit(struct thread *td) 372 { 373 int refcnt; 374 struct vmspace *vm; 375 struct proc *p; 376 377 /* 378 * Release user portion of address space. 379 * This releases references to vnodes, 380 * which could cause I/O if the file has been unlinked. 381 * Need to do this early enough that we can still sleep. 382 * 383 * The last exiting process to reach this point releases as 384 * much of the environment as it can. vmspace_dofree() is the 385 * slower fallback in case another process had a temporary 386 * reference to the vmspace. 387 */ 388 389 p = td->td_proc; 390 vm = p->p_vmspace; 391 atomic_add_int(&vmspace0.vm_refcnt, 1); 392 refcnt = vm->vm_refcnt; 393 do { 394 if (refcnt > 1 && p->p_vmspace != &vmspace0) { 395 /* Switch now since other proc might free vmspace */ 396 PROC_VMSPACE_LOCK(p); 397 p->p_vmspace = &vmspace0; 398 PROC_VMSPACE_UNLOCK(p); 399 pmap_activate(td); 400 } 401 } while (!atomic_fcmpset_int(&vm->vm_refcnt, &refcnt, refcnt - 1)); 402 if (refcnt == 1) { 403 if (p->p_vmspace != vm) { 404 /* vmspace not yet freed, switch back */ 405 PROC_VMSPACE_LOCK(p); 406 p->p_vmspace = vm; 407 PROC_VMSPACE_UNLOCK(p); 408 pmap_activate(td); 409 } 410 pmap_remove_pages(vmspace_pmap(vm)); 411 /* Switch now since this proc will free vmspace */ 412 PROC_VMSPACE_LOCK(p); 413 p->p_vmspace = &vmspace0; 414 PROC_VMSPACE_UNLOCK(p); 415 pmap_activate(td); 416 vmspace_dofree(vm); 417 } 418 #ifdef RACCT 419 if (racct_enable) 420 vmspace_container_reset(p); 421 #endif 422 } 423 424 /* Acquire reference to vmspace owned by another process. */ 425 426 struct vmspace * 427 vmspace_acquire_ref(struct proc *p) 428 { 429 struct vmspace *vm; 430 int refcnt; 431 432 PROC_VMSPACE_LOCK(p); 433 vm = p->p_vmspace; 434 if (vm == NULL) { 435 PROC_VMSPACE_UNLOCK(p); 436 return (NULL); 437 } 438 refcnt = vm->vm_refcnt; 439 do { 440 if (refcnt <= 0) { /* Avoid 0->1 transition */ 441 PROC_VMSPACE_UNLOCK(p); 442 return (NULL); 443 } 444 } while (!atomic_fcmpset_int(&vm->vm_refcnt, &refcnt, refcnt + 1)); 445 if (vm != p->p_vmspace) { 446 PROC_VMSPACE_UNLOCK(p); 447 vmspace_free(vm); 448 return (NULL); 449 } 450 PROC_VMSPACE_UNLOCK(p); 451 return (vm); 452 } 453 454 /* 455 * Switch between vmspaces in an AIO kernel process. 456 * 457 * The new vmspace is either the vmspace of a user process obtained 458 * from an active AIO request or the initial vmspace of the AIO kernel 459 * process (when it is idling). Because user processes will block to 460 * drain any active AIO requests before proceeding in exit() or 461 * execve(), the reference count for vmspaces from AIO requests can 462 * never be 0. Similarly, AIO kernel processes hold an extra 463 * reference on their initial vmspace for the life of the process. As 464 * a result, the 'newvm' vmspace always has a non-zero reference 465 * count. This permits an additional reference on 'newvm' to be 466 * acquired via a simple atomic increment rather than the loop in 467 * vmspace_acquire_ref() above. 468 */ 469 void 470 vmspace_switch_aio(struct vmspace *newvm) 471 { 472 struct vmspace *oldvm; 473 474 /* XXX: Need some way to assert that this is an aio daemon. */ 475 476 KASSERT(newvm->vm_refcnt > 0, 477 ("vmspace_switch_aio: newvm unreferenced")); 478 479 oldvm = curproc->p_vmspace; 480 if (oldvm == newvm) 481 return; 482 483 /* 484 * Point to the new address space and refer to it. 485 */ 486 curproc->p_vmspace = newvm; 487 atomic_add_int(&newvm->vm_refcnt, 1); 488 489 /* Activate the new mapping. */ 490 pmap_activate(curthread); 491 492 vmspace_free(oldvm); 493 } 494 495 void 496 _vm_map_lock(vm_map_t map, const char *file, int line) 497 { 498 499 if (map->system_map) 500 mtx_lock_flags_(&map->system_mtx, 0, file, line); 501 else 502 sx_xlock_(&map->lock, file, line); 503 map->timestamp++; 504 } 505 506 void 507 vm_map_entry_set_vnode_text(vm_map_entry_t entry, bool add) 508 { 509 vm_object_t object, object1; 510 struct vnode *vp; 511 512 if ((entry->eflags & MAP_ENTRY_VN_EXEC) == 0) 513 return; 514 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 515 ("Submap with execs")); 516 object = entry->object.vm_object; 517 KASSERT(object != NULL, ("No object for text, entry %p", entry)); 518 VM_OBJECT_RLOCK(object); 519 while ((object1 = object->backing_object) != NULL) { 520 VM_OBJECT_RLOCK(object1); 521 VM_OBJECT_RUNLOCK(object); 522 object = object1; 523 } 524 525 vp = NULL; 526 if (object->type == OBJT_DEAD) { 527 /* 528 * For OBJT_DEAD objects, v_writecount was handled in 529 * vnode_pager_dealloc(). 530 */ 531 } else if (object->type == OBJT_VNODE) { 532 vp = object->handle; 533 } else if (object->type == OBJT_SWAP) { 534 KASSERT((object->flags & OBJ_TMPFS_NODE) != 0, 535 ("vm_map_entry_set_vnode_text: swap and !TMPFS " 536 "entry %p, object %p, add %d", entry, object, add)); 537 /* 538 * Tmpfs VREG node, which was reclaimed, has 539 * OBJ_TMPFS_NODE flag set, but not OBJ_TMPFS. In 540 * this case there is no v_writecount to adjust. 541 */ 542 if ((object->flags & OBJ_TMPFS) != 0) 543 vp = object->un_pager.swp.swp_tmpfs; 544 } else { 545 KASSERT(0, 546 ("vm_map_entry_set_vnode_text: wrong object type, " 547 "entry %p, object %p, add %d", entry, object, add)); 548 } 549 if (vp != NULL) { 550 if (add) { 551 VOP_SET_TEXT_CHECKED(vp); 552 VM_OBJECT_RUNLOCK(object); 553 } else { 554 vhold(vp); 555 VM_OBJECT_RUNLOCK(object); 556 vn_lock(vp, LK_SHARED | LK_RETRY); 557 VOP_UNSET_TEXT_CHECKED(vp); 558 VOP_UNLOCK(vp, 0); 559 vdrop(vp); 560 } 561 } else { 562 VM_OBJECT_RUNLOCK(object); 563 } 564 } 565 566 static void 567 vm_map_process_deferred(void) 568 { 569 struct thread *td; 570 vm_map_entry_t entry, next; 571 vm_object_t object; 572 573 td = curthread; 574 entry = td->td_map_def_user; 575 td->td_map_def_user = NULL; 576 while (entry != NULL) { 577 next = entry->next; 578 MPASS((entry->eflags & (MAP_ENTRY_WRITECNT | 579 MAP_ENTRY_VN_EXEC)) != (MAP_ENTRY_WRITECNT | 580 MAP_ENTRY_VN_EXEC)); 581 if ((entry->eflags & MAP_ENTRY_WRITECNT) != 0) { 582 /* 583 * Decrement the object's writemappings and 584 * possibly the vnode's v_writecount. 585 */ 586 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 587 ("Submap with writecount")); 588 object = entry->object.vm_object; 589 KASSERT(object != NULL, ("No object for writecount")); 590 vm_pager_release_writecount(object, entry->start, 591 entry->end); 592 } 593 vm_map_entry_set_vnode_text(entry, false); 594 vm_map_entry_deallocate(entry, FALSE); 595 entry = next; 596 } 597 } 598 599 void 600 _vm_map_unlock(vm_map_t map, const char *file, int line) 601 { 602 603 if (map->system_map) 604 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 605 else { 606 sx_xunlock_(&map->lock, file, line); 607 vm_map_process_deferred(); 608 } 609 } 610 611 void 612 _vm_map_lock_read(vm_map_t map, const char *file, int line) 613 { 614 615 if (map->system_map) 616 mtx_lock_flags_(&map->system_mtx, 0, file, line); 617 else 618 sx_slock_(&map->lock, file, line); 619 } 620 621 void 622 _vm_map_unlock_read(vm_map_t map, const char *file, int line) 623 { 624 625 if (map->system_map) 626 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 627 else { 628 sx_sunlock_(&map->lock, file, line); 629 vm_map_process_deferred(); 630 } 631 } 632 633 int 634 _vm_map_trylock(vm_map_t map, const char *file, int line) 635 { 636 int error; 637 638 error = map->system_map ? 639 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 640 !sx_try_xlock_(&map->lock, file, line); 641 if (error == 0) 642 map->timestamp++; 643 return (error == 0); 644 } 645 646 int 647 _vm_map_trylock_read(vm_map_t map, const char *file, int line) 648 { 649 int error; 650 651 error = map->system_map ? 652 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 653 !sx_try_slock_(&map->lock, file, line); 654 return (error == 0); 655 } 656 657 /* 658 * _vm_map_lock_upgrade: [ internal use only ] 659 * 660 * Tries to upgrade a read (shared) lock on the specified map to a write 661 * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a 662 * non-zero value if the upgrade fails. If the upgrade fails, the map is 663 * returned without a read or write lock held. 664 * 665 * Requires that the map be read locked. 666 */ 667 int 668 _vm_map_lock_upgrade(vm_map_t map, const char *file, int line) 669 { 670 unsigned int last_timestamp; 671 672 if (map->system_map) { 673 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 674 } else { 675 if (!sx_try_upgrade_(&map->lock, file, line)) { 676 last_timestamp = map->timestamp; 677 sx_sunlock_(&map->lock, file, line); 678 vm_map_process_deferred(); 679 /* 680 * If the map's timestamp does not change while the 681 * map is unlocked, then the upgrade succeeds. 682 */ 683 sx_xlock_(&map->lock, file, line); 684 if (last_timestamp != map->timestamp) { 685 sx_xunlock_(&map->lock, file, line); 686 return (1); 687 } 688 } 689 } 690 map->timestamp++; 691 return (0); 692 } 693 694 void 695 _vm_map_lock_downgrade(vm_map_t map, const char *file, int line) 696 { 697 698 if (map->system_map) { 699 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 700 } else 701 sx_downgrade_(&map->lock, file, line); 702 } 703 704 /* 705 * vm_map_locked: 706 * 707 * Returns a non-zero value if the caller holds a write (exclusive) lock 708 * on the specified map and the value "0" otherwise. 709 */ 710 int 711 vm_map_locked(vm_map_t map) 712 { 713 714 if (map->system_map) 715 return (mtx_owned(&map->system_mtx)); 716 else 717 return (sx_xlocked(&map->lock)); 718 } 719 720 #ifdef INVARIANTS 721 static void 722 _vm_map_assert_locked(vm_map_t map, const char *file, int line) 723 { 724 725 if (map->system_map) 726 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 727 else 728 sx_assert_(&map->lock, SA_XLOCKED, file, line); 729 } 730 731 #define VM_MAP_ASSERT_LOCKED(map) \ 732 _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE) 733 734 #ifdef DIAGNOSTIC 735 static int enable_vmmap_check = 1; 736 #else 737 static int enable_vmmap_check = 0; 738 #endif 739 SYSCTL_INT(_debug, OID_AUTO, vmmap_check, CTLFLAG_RWTUN, 740 &enable_vmmap_check, 0, "Enable vm map consistency checking"); 741 742 static void 743 _vm_map_assert_consistent(vm_map_t map) 744 { 745 vm_map_entry_t child, entry, prev; 746 vm_size_t max_left, max_right; 747 748 if (!enable_vmmap_check) 749 return; 750 751 for (prev = &map->header; (entry = prev->next) != &map->header; 752 prev = entry) { 753 KASSERT(prev->end <= entry->start, 754 ("map %p prev->end = %jx, start = %jx", map, 755 (uintmax_t)prev->end, (uintmax_t)entry->start)); 756 KASSERT(entry->start < entry->end, 757 ("map %p start = %jx, end = %jx", map, 758 (uintmax_t)entry->start, (uintmax_t)entry->end)); 759 KASSERT(entry->end <= entry->next->start, 760 ("map %p end = %jx, next->start = %jx", map, 761 (uintmax_t)entry->end, (uintmax_t)entry->next->start)); 762 KASSERT(entry->left == NULL || 763 entry->left->start < entry->start, 764 ("map %p left->start = %jx, start = %jx", map, 765 (uintmax_t)entry->left->start, (uintmax_t)entry->start)); 766 KASSERT(entry->right == NULL || 767 entry->start < entry->right->start, 768 ("map %p start = %jx, right->start = %jx", map, 769 (uintmax_t)entry->start, (uintmax_t)entry->right->start)); 770 child = entry->left; 771 max_left = (child != NULL) ? child->max_free : 772 entry->start - prev->end; 773 child = entry->right; 774 max_right = (child != NULL) ? child->max_free : 775 entry->next->start - entry->end; 776 KASSERT(entry->max_free == MAX(max_left, max_right), 777 ("map %p max = %jx, max_left = %jx, max_right = %jx", map, 778 (uintmax_t)entry->max_free, 779 (uintmax_t)max_left, (uintmax_t)max_right)); 780 } 781 } 782 783 #define VM_MAP_ASSERT_CONSISTENT(map) \ 784 _vm_map_assert_consistent(map) 785 #else 786 #define VM_MAP_ASSERT_LOCKED(map) 787 #define VM_MAP_ASSERT_CONSISTENT(map) 788 #endif /* INVARIANTS */ 789 790 /* 791 * _vm_map_unlock_and_wait: 792 * 793 * Atomically releases the lock on the specified map and puts the calling 794 * thread to sleep. The calling thread will remain asleep until either 795 * vm_map_wakeup() is performed on the map or the specified timeout is 796 * exceeded. 797 * 798 * WARNING! This function does not perform deferred deallocations of 799 * objects and map entries. Therefore, the calling thread is expected to 800 * reacquire the map lock after reawakening and later perform an ordinary 801 * unlock operation, such as vm_map_unlock(), before completing its 802 * operation on the map. 803 */ 804 int 805 _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line) 806 { 807 808 mtx_lock(&map_sleep_mtx); 809 if (map->system_map) 810 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 811 else 812 sx_xunlock_(&map->lock, file, line); 813 return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps", 814 timo)); 815 } 816 817 /* 818 * vm_map_wakeup: 819 * 820 * Awaken any threads that have slept on the map using 821 * vm_map_unlock_and_wait(). 822 */ 823 void 824 vm_map_wakeup(vm_map_t map) 825 { 826 827 /* 828 * Acquire and release map_sleep_mtx to prevent a wakeup() 829 * from being performed (and lost) between the map unlock 830 * and the msleep() in _vm_map_unlock_and_wait(). 831 */ 832 mtx_lock(&map_sleep_mtx); 833 mtx_unlock(&map_sleep_mtx); 834 wakeup(&map->root); 835 } 836 837 void 838 vm_map_busy(vm_map_t map) 839 { 840 841 VM_MAP_ASSERT_LOCKED(map); 842 map->busy++; 843 } 844 845 void 846 vm_map_unbusy(vm_map_t map) 847 { 848 849 VM_MAP_ASSERT_LOCKED(map); 850 KASSERT(map->busy, ("vm_map_unbusy: not busy")); 851 if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) { 852 vm_map_modflags(map, 0, MAP_BUSY_WAKEUP); 853 wakeup(&map->busy); 854 } 855 } 856 857 void 858 vm_map_wait_busy(vm_map_t map) 859 { 860 861 VM_MAP_ASSERT_LOCKED(map); 862 while (map->busy) { 863 vm_map_modflags(map, MAP_BUSY_WAKEUP, 0); 864 if (map->system_map) 865 msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0); 866 else 867 sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0); 868 } 869 map->timestamp++; 870 } 871 872 long 873 vmspace_resident_count(struct vmspace *vmspace) 874 { 875 return pmap_resident_count(vmspace_pmap(vmspace)); 876 } 877 878 /* 879 * vm_map_create: 880 * 881 * Creates and returns a new empty VM map with 882 * the given physical map structure, and having 883 * the given lower and upper address bounds. 884 */ 885 vm_map_t 886 vm_map_create(pmap_t pmap, vm_offset_t min, vm_offset_t max) 887 { 888 vm_map_t result; 889 890 result = uma_zalloc(mapzone, M_WAITOK); 891 CTR1(KTR_VM, "vm_map_create: %p", result); 892 _vm_map_init(result, pmap, min, max); 893 return (result); 894 } 895 896 /* 897 * Initialize an existing vm_map structure 898 * such as that in the vmspace structure. 899 */ 900 static void 901 _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 902 { 903 904 map->header.next = map->header.prev = &map->header; 905 map->header.eflags = MAP_ENTRY_HEADER; 906 map->needs_wakeup = FALSE; 907 map->system_map = 0; 908 map->pmap = pmap; 909 map->header.end = min; 910 map->header.start = max; 911 map->flags = 0; 912 map->root = NULL; 913 map->timestamp = 0; 914 map->busy = 0; 915 map->anon_loc = 0; 916 } 917 918 void 919 vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 920 { 921 922 _vm_map_init(map, pmap, min, max); 923 mtx_init(&map->system_mtx, "system map", NULL, MTX_DEF | MTX_DUPOK); 924 sx_init(&map->lock, "user map"); 925 } 926 927 /* 928 * vm_map_entry_dispose: [ internal use only ] 929 * 930 * Inverse of vm_map_entry_create. 931 */ 932 static void 933 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry) 934 { 935 uma_zfree(map->system_map ? kmapentzone : mapentzone, entry); 936 } 937 938 /* 939 * vm_map_entry_create: [ internal use only ] 940 * 941 * Allocates a VM map entry for insertion. 942 * No entry fields are filled in. 943 */ 944 static vm_map_entry_t 945 vm_map_entry_create(vm_map_t map) 946 { 947 vm_map_entry_t new_entry; 948 949 if (map->system_map) 950 new_entry = uma_zalloc(kmapentzone, M_NOWAIT); 951 else 952 new_entry = uma_zalloc(mapentzone, M_WAITOK); 953 if (new_entry == NULL) 954 panic("vm_map_entry_create: kernel resources exhausted"); 955 return (new_entry); 956 } 957 958 /* 959 * vm_map_entry_set_behavior: 960 * 961 * Set the expected access behavior, either normal, random, or 962 * sequential. 963 */ 964 static inline void 965 vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior) 966 { 967 entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) | 968 (behavior & MAP_ENTRY_BEHAV_MASK); 969 } 970 971 /* 972 * vm_map_entry_max_free_{left,right}: 973 * 974 * Compute the size of the largest free gap between two entries, 975 * one the root of a tree and the other the ancestor of that root 976 * that is the least or greatest ancestor found on the search path. 977 */ 978 static inline vm_size_t 979 vm_map_entry_max_free_left(vm_map_entry_t root, vm_map_entry_t left_ancestor) 980 { 981 982 return (root->left != NULL ? 983 root->left->max_free : root->start - left_ancestor->end); 984 } 985 986 static inline vm_size_t 987 vm_map_entry_max_free_right(vm_map_entry_t root, vm_map_entry_t right_ancestor) 988 { 989 990 return (root->right != NULL ? 991 root->right->max_free : right_ancestor->start - root->end); 992 } 993 994 #define SPLAY_LEFT_STEP(root, y, rlist, test) do { \ 995 vm_size_t max_free; \ 996 \ 997 /* \ 998 * Infer root->right->max_free == root->max_free when \ 999 * y->max_free < root->max_free || root->max_free == 0. \ 1000 * Otherwise, look right to find it. \ 1001 */ \ 1002 y = root->left; \ 1003 max_free = root->max_free; \ 1004 KASSERT(max_free >= vm_map_entry_max_free_right(root, rlist), \ 1005 ("%s: max_free invariant fails", __func__)); \ 1006 if (y == NULL ? max_free > 0 : max_free - 1 < y->max_free) \ 1007 max_free = vm_map_entry_max_free_right(root, rlist); \ 1008 if (y != NULL && (test)) { \ 1009 /* Rotate right and make y root. */ \ 1010 root->left = y->right; \ 1011 y->right = root; \ 1012 if (max_free < y->max_free) \ 1013 root->max_free = max_free = MAX(max_free, \ 1014 vm_map_entry_max_free_left(root, y)); \ 1015 root = y; \ 1016 y = root->left; \ 1017 } \ 1018 /* Copy right->max_free. Put root on rlist. */ \ 1019 root->max_free = max_free; \ 1020 KASSERT(max_free == vm_map_entry_max_free_right(root, rlist), \ 1021 ("%s: max_free not copied from right", __func__)); \ 1022 root->left = rlist; \ 1023 rlist = root; \ 1024 root = y; \ 1025 } while (0) 1026 1027 #define SPLAY_RIGHT_STEP(root, y, llist, test) do { \ 1028 vm_size_t max_free; \ 1029 \ 1030 /* \ 1031 * Infer root->left->max_free == root->max_free when \ 1032 * y->max_free < root->max_free || root->max_free == 0. \ 1033 * Otherwise, look left to find it. \ 1034 */ \ 1035 y = root->right; \ 1036 max_free = root->max_free; \ 1037 KASSERT(max_free >= vm_map_entry_max_free_left(root, llist), \ 1038 ("%s: max_free invariant fails", __func__)); \ 1039 if (y == NULL ? max_free > 0 : max_free - 1 < y->max_free) \ 1040 max_free = vm_map_entry_max_free_left(root, llist); \ 1041 if (y != NULL && (test)) { \ 1042 /* Rotate left and make y root. */ \ 1043 root->right = y->left; \ 1044 y->left = root; \ 1045 if (max_free < y->max_free) \ 1046 root->max_free = max_free = MAX(max_free, \ 1047 vm_map_entry_max_free_right(root, y)); \ 1048 root = y; \ 1049 y = root->right; \ 1050 } \ 1051 /* Copy left->max_free. Put root on llist. */ \ 1052 root->max_free = max_free; \ 1053 KASSERT(max_free == vm_map_entry_max_free_left(root, llist), \ 1054 ("%s: max_free not copied from left", __func__)); \ 1055 root->right = llist; \ 1056 llist = root; \ 1057 root = y; \ 1058 } while (0) 1059 1060 /* 1061 * Walk down the tree until we find addr or a NULL pointer where addr would go, 1062 * breaking off left and right subtrees of nodes less than, or greater than 1063 * addr. Treat pointers to nodes with max_free < length as NULL pointers. 1064 * llist and rlist are the two sides in reverse order (bottom-up), with llist 1065 * linked by the right pointer and rlist linked by the left pointer in the 1066 * vm_map_entry, and both lists terminated by &map->header. This function, and 1067 * the subsequent call to vm_map_splay_merge, rely on the start and end address 1068 * values in &map->header. 1069 */ 1070 static vm_map_entry_t 1071 vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length, 1072 vm_map_entry_t *out_llist, vm_map_entry_t *out_rlist) 1073 { 1074 vm_map_entry_t llist, rlist, root, y; 1075 1076 llist = rlist = &map->header; 1077 root = map->root; 1078 while (root != NULL && root->max_free >= length) { 1079 KASSERT(llist->end <= root->start && root->end <= rlist->start, 1080 ("%s: root not within tree bounds", __func__)); 1081 if (addr < root->start) { 1082 SPLAY_LEFT_STEP(root, y, rlist, 1083 y->max_free >= length && addr < y->start); 1084 } else if (addr >= root->end) { 1085 SPLAY_RIGHT_STEP(root, y, llist, 1086 y->max_free >= length && addr >= y->end); 1087 } else 1088 break; 1089 } 1090 *out_llist = llist; 1091 *out_rlist = rlist; 1092 return (root); 1093 } 1094 1095 static void 1096 vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *iolist) 1097 { 1098 vm_map_entry_t rlist, y; 1099 1100 root = root->right; 1101 rlist = *iolist; 1102 while (root != NULL) 1103 SPLAY_LEFT_STEP(root, y, rlist, true); 1104 *iolist = rlist; 1105 } 1106 1107 static void 1108 vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *iolist) 1109 { 1110 vm_map_entry_t llist, y; 1111 1112 root = root->left; 1113 llist = *iolist; 1114 while (root != NULL) 1115 SPLAY_RIGHT_STEP(root, y, llist, true); 1116 *iolist = llist; 1117 } 1118 1119 static inline void 1120 vm_map_entry_swap(vm_map_entry_t *a, vm_map_entry_t *b) 1121 { 1122 vm_map_entry_t tmp; 1123 1124 tmp = *b; 1125 *b = *a; 1126 *a = tmp; 1127 } 1128 1129 /* 1130 * Walk back up the two spines, flip the pointers and set max_free. The 1131 * subtrees of the root go at the bottom of llist and rlist. 1132 */ 1133 static void 1134 vm_map_splay_merge(vm_map_t map, vm_map_entry_t root, 1135 vm_map_entry_t llist, vm_map_entry_t rlist) 1136 { 1137 vm_map_entry_t prev; 1138 vm_size_t max_free_left, max_free_right; 1139 1140 max_free_left = vm_map_entry_max_free_left(root, llist); 1141 if (llist != &map->header) { 1142 prev = root->left; 1143 do { 1144 /* 1145 * The max_free values of the children of llist are in 1146 * llist->max_free and max_free_left. Update with the 1147 * max value. 1148 */ 1149 llist->max_free = max_free_left = 1150 MAX(llist->max_free, max_free_left); 1151 vm_map_entry_swap(&llist->right, &prev); 1152 vm_map_entry_swap(&prev, &llist); 1153 } while (llist != &map->header); 1154 root->left = prev; 1155 } 1156 max_free_right = vm_map_entry_max_free_right(root, rlist); 1157 if (rlist != &map->header) { 1158 prev = root->right; 1159 do { 1160 /* 1161 * The max_free values of the children of rlist are in 1162 * rlist->max_free and max_free_right. Update with the 1163 * max value. 1164 */ 1165 rlist->max_free = max_free_right = 1166 MAX(rlist->max_free, max_free_right); 1167 vm_map_entry_swap(&rlist->left, &prev); 1168 vm_map_entry_swap(&prev, &rlist); 1169 } while (rlist != &map->header); 1170 root->right = prev; 1171 } 1172 root->max_free = MAX(max_free_left, max_free_right); 1173 map->root = root; 1174 } 1175 1176 /* 1177 * vm_map_splay: 1178 * 1179 * The Sleator and Tarjan top-down splay algorithm with the 1180 * following variation. Max_free must be computed bottom-up, so 1181 * on the downward pass, maintain the left and right spines in 1182 * reverse order. Then, make a second pass up each side to fix 1183 * the pointers and compute max_free. The time bound is O(log n) 1184 * amortized. 1185 * 1186 * The new root is the vm_map_entry containing "addr", or else an 1187 * adjacent entry (lower if possible) if addr is not in the tree. 1188 * 1189 * The map must be locked, and leaves it so. 1190 * 1191 * Returns: the new root. 1192 */ 1193 static vm_map_entry_t 1194 vm_map_splay(vm_map_t map, vm_offset_t addr) 1195 { 1196 vm_map_entry_t llist, rlist, root; 1197 1198 root = vm_map_splay_split(map, addr, 0, &llist, &rlist); 1199 if (root != NULL) { 1200 /* do nothing */ 1201 } else if (llist != &map->header) { 1202 /* 1203 * Recover the greatest node in the left 1204 * subtree and make it the root. 1205 */ 1206 root = llist; 1207 llist = root->right; 1208 root->right = NULL; 1209 } else if (rlist != &map->header) { 1210 /* 1211 * Recover the least node in the right 1212 * subtree and make it the root. 1213 */ 1214 root = rlist; 1215 rlist = root->left; 1216 root->left = NULL; 1217 } else { 1218 /* There is no root. */ 1219 return (NULL); 1220 } 1221 vm_map_splay_merge(map, root, llist, rlist); 1222 VM_MAP_ASSERT_CONSISTENT(map); 1223 return (root); 1224 } 1225 1226 /* 1227 * vm_map_entry_{un,}link: 1228 * 1229 * Insert/remove entries from maps. 1230 */ 1231 static void 1232 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) 1233 { 1234 vm_map_entry_t llist, rlist, root; 1235 1236 CTR3(KTR_VM, 1237 "vm_map_entry_link: map %p, nentries %d, entry %p", map, 1238 map->nentries, entry); 1239 VM_MAP_ASSERT_LOCKED(map); 1240 map->nentries++; 1241 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1242 KASSERT(root == NULL, 1243 ("vm_map_entry_link: link object already mapped")); 1244 entry->prev = llist; 1245 entry->next = rlist; 1246 llist->next = rlist->prev = entry; 1247 entry->left = entry->right = NULL; 1248 vm_map_splay_merge(map, entry, llist, rlist); 1249 VM_MAP_ASSERT_CONSISTENT(map); 1250 } 1251 1252 enum unlink_merge_type { 1253 UNLINK_MERGE_PREV, 1254 UNLINK_MERGE_NONE, 1255 UNLINK_MERGE_NEXT 1256 }; 1257 1258 static void 1259 vm_map_entry_unlink(vm_map_t map, vm_map_entry_t entry, 1260 enum unlink_merge_type op) 1261 { 1262 vm_map_entry_t llist, rlist, root, y; 1263 1264 VM_MAP_ASSERT_LOCKED(map); 1265 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1266 KASSERT(root != NULL, 1267 ("vm_map_entry_unlink: unlink object not mapped")); 1268 1269 switch (op) { 1270 case UNLINK_MERGE_PREV: 1271 vm_map_splay_findprev(root, &llist); 1272 llist->end = root->end; 1273 y = root->right; 1274 root = llist; 1275 llist = root->right; 1276 root->right = y; 1277 break; 1278 case UNLINK_MERGE_NEXT: 1279 vm_map_splay_findnext(root, &rlist); 1280 rlist->start = root->start; 1281 rlist->offset = root->offset; 1282 y = root->left; 1283 root = rlist; 1284 rlist = root->left; 1285 root->left = y; 1286 break; 1287 case UNLINK_MERGE_NONE: 1288 vm_map_splay_findprev(root, &llist); 1289 vm_map_splay_findnext(root, &rlist); 1290 if (llist != &map->header) { 1291 root = llist; 1292 llist = root->right; 1293 root->right = NULL; 1294 } else if (rlist != &map->header) { 1295 root = rlist; 1296 rlist = root->left; 1297 root->left = NULL; 1298 } else 1299 root = NULL; 1300 break; 1301 } 1302 y = entry->next; 1303 y->prev = entry->prev; 1304 y->prev->next = y; 1305 if (root != NULL) 1306 vm_map_splay_merge(map, root, llist, rlist); 1307 else 1308 map->root = NULL; 1309 VM_MAP_ASSERT_CONSISTENT(map); 1310 map->nentries--; 1311 CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map, 1312 map->nentries, entry); 1313 } 1314 1315 /* 1316 * vm_map_entry_resize: 1317 * 1318 * Resize a vm_map_entry, recompute the amount of free space that 1319 * follows it and propagate that value up the tree. 1320 * 1321 * The map must be locked, and leaves it so. 1322 */ 1323 static void 1324 vm_map_entry_resize(vm_map_t map, vm_map_entry_t entry, vm_size_t grow_amount) 1325 { 1326 vm_map_entry_t llist, rlist, root; 1327 1328 VM_MAP_ASSERT_LOCKED(map); 1329 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1330 KASSERT(root != NULL, 1331 ("%s: resize object not mapped", __func__)); 1332 vm_map_splay_findnext(root, &rlist); 1333 root->right = NULL; 1334 entry->end += grow_amount; 1335 vm_map_splay_merge(map, root, llist, rlist); 1336 VM_MAP_ASSERT_CONSISTENT(map); 1337 CTR4(KTR_VM, "%s: map %p, nentries %d, entry %p", 1338 __func__, map, map->nentries, entry); 1339 } 1340 1341 /* 1342 * vm_map_lookup_entry: [ internal use only ] 1343 * 1344 * Finds the map entry containing (or 1345 * immediately preceding) the specified address 1346 * in the given map; the entry is returned 1347 * in the "entry" parameter. The boolean 1348 * result indicates whether the address is 1349 * actually contained in the map. 1350 */ 1351 boolean_t 1352 vm_map_lookup_entry( 1353 vm_map_t map, 1354 vm_offset_t address, 1355 vm_map_entry_t *entry) /* OUT */ 1356 { 1357 vm_map_entry_t cur, lbound; 1358 boolean_t locked; 1359 1360 /* 1361 * If the map is empty, then the map entry immediately preceding 1362 * "address" is the map's header. 1363 */ 1364 cur = map->root; 1365 if (cur == NULL) { 1366 *entry = &map->header; 1367 return (FALSE); 1368 } 1369 if (address >= cur->start && cur->end > address) { 1370 *entry = cur; 1371 return (TRUE); 1372 } 1373 if ((locked = vm_map_locked(map)) || 1374 sx_try_upgrade(&map->lock)) { 1375 /* 1376 * Splay requires a write lock on the map. However, it only 1377 * restructures the binary search tree; it does not otherwise 1378 * change the map. Thus, the map's timestamp need not change 1379 * on a temporary upgrade. 1380 */ 1381 cur = vm_map_splay(map, address); 1382 if (!locked) 1383 sx_downgrade(&map->lock); 1384 1385 /* 1386 * If "address" is contained within a map entry, the new root 1387 * is that map entry. Otherwise, the new root is a map entry 1388 * immediately before or after "address". 1389 */ 1390 if (address < cur->start) { 1391 *entry = &map->header; 1392 return (FALSE); 1393 } 1394 *entry = cur; 1395 return (address < cur->end); 1396 } 1397 /* 1398 * Since the map is only locked for read access, perform a 1399 * standard binary search tree lookup for "address". 1400 */ 1401 lbound = &map->header; 1402 do { 1403 if (address < cur->start) { 1404 cur = cur->left; 1405 } else if (cur->end <= address) { 1406 lbound = cur; 1407 cur = cur->right; 1408 } else { 1409 *entry = cur; 1410 return (TRUE); 1411 } 1412 } while (cur != NULL); 1413 *entry = lbound; 1414 return (FALSE); 1415 } 1416 1417 /* 1418 * vm_map_insert: 1419 * 1420 * Inserts the given whole VM object into the target 1421 * map at the specified address range. The object's 1422 * size should match that of the address range. 1423 * 1424 * Requires that the map be locked, and leaves it so. 1425 * 1426 * If object is non-NULL, ref count must be bumped by caller 1427 * prior to making call to account for the new entry. 1428 */ 1429 int 1430 vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1431 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow) 1432 { 1433 vm_map_entry_t new_entry, prev_entry; 1434 struct ucred *cred; 1435 vm_eflags_t protoeflags; 1436 vm_inherit_t inheritance; 1437 1438 VM_MAP_ASSERT_LOCKED(map); 1439 KASSERT(object != kernel_object || 1440 (cow & MAP_COPY_ON_WRITE) == 0, 1441 ("vm_map_insert: kernel object and COW")); 1442 KASSERT(object == NULL || (cow & MAP_NOFAULT) == 0, 1443 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 1444 KASSERT((prot & ~max) == 0, 1445 ("prot %#x is not subset of max_prot %#x", prot, max)); 1446 1447 /* 1448 * Check that the start and end points are not bogus. 1449 */ 1450 if (start < vm_map_min(map) || end > vm_map_max(map) || 1451 start >= end) 1452 return (KERN_INVALID_ADDRESS); 1453 1454 /* 1455 * Find the entry prior to the proposed starting address; if it's part 1456 * of an existing entry, this range is bogus. 1457 */ 1458 if (vm_map_lookup_entry(map, start, &prev_entry)) 1459 return (KERN_NO_SPACE); 1460 1461 /* 1462 * Assert that the next entry doesn't overlap the end point. 1463 */ 1464 if (prev_entry->next->start < end) 1465 return (KERN_NO_SPACE); 1466 1467 if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL || 1468 max != VM_PROT_NONE)) 1469 return (KERN_INVALID_ARGUMENT); 1470 1471 protoeflags = 0; 1472 if (cow & MAP_COPY_ON_WRITE) 1473 protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; 1474 if (cow & MAP_NOFAULT) 1475 protoeflags |= MAP_ENTRY_NOFAULT; 1476 if (cow & MAP_DISABLE_SYNCER) 1477 protoeflags |= MAP_ENTRY_NOSYNC; 1478 if (cow & MAP_DISABLE_COREDUMP) 1479 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1480 if (cow & MAP_STACK_GROWS_DOWN) 1481 protoeflags |= MAP_ENTRY_GROWS_DOWN; 1482 if (cow & MAP_STACK_GROWS_UP) 1483 protoeflags |= MAP_ENTRY_GROWS_UP; 1484 if (cow & MAP_WRITECOUNT) 1485 protoeflags |= MAP_ENTRY_WRITECNT; 1486 if (cow & MAP_VN_EXEC) 1487 protoeflags |= MAP_ENTRY_VN_EXEC; 1488 if ((cow & MAP_CREATE_GUARD) != 0) 1489 protoeflags |= MAP_ENTRY_GUARD; 1490 if ((cow & MAP_CREATE_STACK_GAP_DN) != 0) 1491 protoeflags |= MAP_ENTRY_STACK_GAP_DN; 1492 if ((cow & MAP_CREATE_STACK_GAP_UP) != 0) 1493 protoeflags |= MAP_ENTRY_STACK_GAP_UP; 1494 if (cow & MAP_INHERIT_SHARE) 1495 inheritance = VM_INHERIT_SHARE; 1496 else 1497 inheritance = VM_INHERIT_DEFAULT; 1498 1499 cred = NULL; 1500 if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0) 1501 goto charged; 1502 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) && 1503 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) { 1504 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start)) 1505 return (KERN_RESOURCE_SHORTAGE); 1506 KASSERT(object == NULL || 1507 (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 || 1508 object->cred == NULL, 1509 ("overcommit: vm_map_insert o %p", object)); 1510 cred = curthread->td_ucred; 1511 } 1512 1513 charged: 1514 /* Expand the kernel pmap, if necessary. */ 1515 if (map == kernel_map && end > kernel_vm_end) 1516 pmap_growkernel(end); 1517 if (object != NULL) { 1518 /* 1519 * OBJ_ONEMAPPING must be cleared unless this mapping 1520 * is trivially proven to be the only mapping for any 1521 * of the object's pages. (Object granularity 1522 * reference counting is insufficient to recognize 1523 * aliases with precision.) 1524 */ 1525 VM_OBJECT_WLOCK(object); 1526 if (object->ref_count > 1 || object->shadow_count != 0) 1527 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1528 VM_OBJECT_WUNLOCK(object); 1529 } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) == 1530 protoeflags && 1531 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP | 1532 MAP_VN_EXEC)) == 0 && 1533 prev_entry->end == start && (prev_entry->cred == cred || 1534 (prev_entry->object.vm_object != NULL && 1535 prev_entry->object.vm_object->cred == cred)) && 1536 vm_object_coalesce(prev_entry->object.vm_object, 1537 prev_entry->offset, 1538 (vm_size_t)(prev_entry->end - prev_entry->start), 1539 (vm_size_t)(end - prev_entry->end), cred != NULL && 1540 (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) { 1541 /* 1542 * We were able to extend the object. Determine if we 1543 * can extend the previous map entry to include the 1544 * new range as well. 1545 */ 1546 if (prev_entry->inheritance == inheritance && 1547 prev_entry->protection == prot && 1548 prev_entry->max_protection == max && 1549 prev_entry->wired_count == 0) { 1550 KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) == 1551 0, ("prev_entry %p has incoherent wiring", 1552 prev_entry)); 1553 if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0) 1554 map->size += end - prev_entry->end; 1555 vm_map_entry_resize(map, prev_entry, 1556 end - prev_entry->end); 1557 vm_map_try_merge_entries(map, prev_entry, prev_entry->next); 1558 return (KERN_SUCCESS); 1559 } 1560 1561 /* 1562 * If we can extend the object but cannot extend the 1563 * map entry, we have to create a new map entry. We 1564 * must bump the ref count on the extended object to 1565 * account for it. object may be NULL. 1566 */ 1567 object = prev_entry->object.vm_object; 1568 offset = prev_entry->offset + 1569 (prev_entry->end - prev_entry->start); 1570 vm_object_reference(object); 1571 if (cred != NULL && object != NULL && object->cred != NULL && 1572 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 1573 /* Object already accounts for this uid. */ 1574 cred = NULL; 1575 } 1576 } 1577 if (cred != NULL) 1578 crhold(cred); 1579 1580 /* 1581 * Create a new entry 1582 */ 1583 new_entry = vm_map_entry_create(map); 1584 new_entry->start = start; 1585 new_entry->end = end; 1586 new_entry->cred = NULL; 1587 1588 new_entry->eflags = protoeflags; 1589 new_entry->object.vm_object = object; 1590 new_entry->offset = offset; 1591 1592 new_entry->inheritance = inheritance; 1593 new_entry->protection = prot; 1594 new_entry->max_protection = max; 1595 new_entry->wired_count = 0; 1596 new_entry->wiring_thread = NULL; 1597 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT; 1598 new_entry->next_read = start; 1599 1600 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry), 1601 ("overcommit: vm_map_insert leaks vm_map %p", new_entry)); 1602 new_entry->cred = cred; 1603 1604 /* 1605 * Insert the new entry into the list 1606 */ 1607 vm_map_entry_link(map, new_entry); 1608 if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0) 1609 map->size += new_entry->end - new_entry->start; 1610 1611 /* 1612 * Try to coalesce the new entry with both the previous and next 1613 * entries in the list. Previously, we only attempted to coalesce 1614 * with the previous entry when object is NULL. Here, we handle the 1615 * other cases, which are less common. 1616 */ 1617 vm_map_try_merge_entries(map, prev_entry, new_entry); 1618 vm_map_try_merge_entries(map, new_entry, new_entry->next); 1619 1620 if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) { 1621 vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset), 1622 end - start, cow & MAP_PREFAULT_PARTIAL); 1623 } 1624 1625 return (KERN_SUCCESS); 1626 } 1627 1628 /* 1629 * vm_map_findspace: 1630 * 1631 * Find the first fit (lowest VM address) for "length" free bytes 1632 * beginning at address >= start in the given map. 1633 * 1634 * In a vm_map_entry, "max_free" is the maximum amount of 1635 * contiguous free space between an entry in its subtree and a 1636 * neighbor of that entry. This allows finding a free region in 1637 * one path down the tree, so O(log n) amortized with splay 1638 * trees. 1639 * 1640 * The map must be locked, and leaves it so. 1641 * 1642 * Returns: starting address if sufficient space, 1643 * vm_map_max(map)-length+1 if insufficient space. 1644 */ 1645 vm_offset_t 1646 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length) 1647 { 1648 vm_map_entry_t llist, rlist, root, y; 1649 vm_size_t left_length; 1650 vm_offset_t gap_end; 1651 1652 /* 1653 * Request must fit within min/max VM address and must avoid 1654 * address wrap. 1655 */ 1656 start = MAX(start, vm_map_min(map)); 1657 if (start >= vm_map_max(map) || length > vm_map_max(map) - start) 1658 return (vm_map_max(map) - length + 1); 1659 1660 /* Empty tree means wide open address space. */ 1661 if (map->root == NULL) 1662 return (start); 1663 1664 /* 1665 * After splay_split, if start is within an entry, push it to the start 1666 * of the following gap. If rlist is at the end of the gap containing 1667 * start, save the end of that gap in gap_end to see if the gap is big 1668 * enough; otherwise set gap_end to start skip gap-checking and move 1669 * directly to a search of the right subtree. 1670 */ 1671 root = vm_map_splay_split(map, start, length, &llist, &rlist); 1672 gap_end = rlist->start; 1673 if (root != NULL) { 1674 start = root->end; 1675 if (root->right != NULL) 1676 gap_end = start; 1677 } else if (rlist != &map->header) { 1678 root = rlist; 1679 rlist = root->left; 1680 root->left = NULL; 1681 } else { 1682 root = llist; 1683 llist = root->right; 1684 root->right = NULL; 1685 } 1686 vm_map_splay_merge(map, root, llist, rlist); 1687 VM_MAP_ASSERT_CONSISTENT(map); 1688 if (length <= gap_end - start) 1689 return (start); 1690 1691 /* With max_free, can immediately tell if no solution. */ 1692 if (root->right == NULL || length > root->right->max_free) 1693 return (vm_map_max(map) - length + 1); 1694 1695 /* 1696 * Splay for the least large-enough gap in the right subtree. 1697 */ 1698 llist = rlist = &map->header; 1699 for (left_length = 0;; 1700 left_length = vm_map_entry_max_free_left(root, llist)) { 1701 if (length <= left_length) 1702 SPLAY_LEFT_STEP(root, y, rlist, 1703 length <= vm_map_entry_max_free_left(y, llist)); 1704 else 1705 SPLAY_RIGHT_STEP(root, y, llist, 1706 length > vm_map_entry_max_free_left(y, root)); 1707 if (root == NULL) 1708 break; 1709 } 1710 root = llist; 1711 llist = root->right; 1712 root->right = NULL; 1713 if (rlist != &map->header) { 1714 y = rlist; 1715 rlist = y->left; 1716 y->left = NULL; 1717 vm_map_splay_merge(map, y, &map->header, rlist); 1718 y->max_free = MAX( 1719 vm_map_entry_max_free_left(y, root), 1720 vm_map_entry_max_free_right(y, &map->header)); 1721 root->right = y; 1722 } 1723 vm_map_splay_merge(map, root, llist, &map->header); 1724 VM_MAP_ASSERT_CONSISTENT(map); 1725 return (root->end); 1726 } 1727 1728 int 1729 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1730 vm_offset_t start, vm_size_t length, vm_prot_t prot, 1731 vm_prot_t max, int cow) 1732 { 1733 vm_offset_t end; 1734 int result; 1735 1736 end = start + length; 1737 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 1738 object == NULL, 1739 ("vm_map_fixed: non-NULL backing object for stack")); 1740 vm_map_lock(map); 1741 VM_MAP_RANGE_CHECK(map, start, end); 1742 if ((cow & MAP_CHECK_EXCL) == 0) 1743 vm_map_delete(map, start, end); 1744 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 1745 result = vm_map_stack_locked(map, start, length, sgrowsiz, 1746 prot, max, cow); 1747 } else { 1748 result = vm_map_insert(map, object, offset, start, end, 1749 prot, max, cow); 1750 } 1751 vm_map_unlock(map); 1752 return (result); 1753 } 1754 1755 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10}; 1756 static const int aslr_pages_rnd_32[2] = {0x100, 0x4}; 1757 1758 static int cluster_anon = 1; 1759 SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW, 1760 &cluster_anon, 0, 1761 "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always"); 1762 1763 static bool 1764 clustering_anon_allowed(vm_offset_t addr) 1765 { 1766 1767 switch (cluster_anon) { 1768 case 0: 1769 return (false); 1770 case 1: 1771 return (addr == 0); 1772 case 2: 1773 default: 1774 return (true); 1775 } 1776 } 1777 1778 static long aslr_restarts; 1779 SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD, 1780 &aslr_restarts, 0, 1781 "Number of aslr failures"); 1782 1783 #define MAP_32BIT_MAX_ADDR ((vm_offset_t)1 << 31) 1784 1785 /* 1786 * Searches for the specified amount of free space in the given map with the 1787 * specified alignment. Performs an address-ordered, first-fit search from 1788 * the given address "*addr", with an optional upper bound "max_addr". If the 1789 * parameter "alignment" is zero, then the alignment is computed from the 1790 * given (object, offset) pair so as to enable the greatest possible use of 1791 * superpage mappings. Returns KERN_SUCCESS and the address of the free space 1792 * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE. 1793 * 1794 * The map must be locked. Initially, there must be at least "length" bytes 1795 * of free space at the given address. 1796 */ 1797 static int 1798 vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1799 vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, 1800 vm_offset_t alignment) 1801 { 1802 vm_offset_t aligned_addr, free_addr; 1803 1804 VM_MAP_ASSERT_LOCKED(map); 1805 free_addr = *addr; 1806 KASSERT(free_addr == vm_map_findspace(map, free_addr, length), 1807 ("caller failed to provide space %#jx at address %p", 1808 (uintmax_t)length, (void *)free_addr)); 1809 for (;;) { 1810 /* 1811 * At the start of every iteration, the free space at address 1812 * "*addr" is at least "length" bytes. 1813 */ 1814 if (alignment == 0) 1815 pmap_align_superpage(object, offset, addr, length); 1816 else if ((*addr & (alignment - 1)) != 0) { 1817 *addr &= ~(alignment - 1); 1818 *addr += alignment; 1819 } 1820 aligned_addr = *addr; 1821 if (aligned_addr == free_addr) { 1822 /* 1823 * Alignment did not change "*addr", so "*addr" must 1824 * still provide sufficient free space. 1825 */ 1826 return (KERN_SUCCESS); 1827 } 1828 1829 /* 1830 * Test for address wrap on "*addr". A wrapped "*addr" could 1831 * be a valid address, in which case vm_map_findspace() cannot 1832 * be relied upon to fail. 1833 */ 1834 if (aligned_addr < free_addr) 1835 return (KERN_NO_SPACE); 1836 *addr = vm_map_findspace(map, aligned_addr, length); 1837 if (*addr + length > vm_map_max(map) || 1838 (max_addr != 0 && *addr + length > max_addr)) 1839 return (KERN_NO_SPACE); 1840 free_addr = *addr; 1841 if (free_addr == aligned_addr) { 1842 /* 1843 * If a successful call to vm_map_findspace() did not 1844 * change "*addr", then "*addr" must still be aligned 1845 * and provide sufficient free space. 1846 */ 1847 return (KERN_SUCCESS); 1848 } 1849 } 1850 } 1851 1852 /* 1853 * vm_map_find finds an unallocated region in the target address 1854 * map with the given length. The search is defined to be 1855 * first-fit from the specified address; the region found is 1856 * returned in the same parameter. 1857 * 1858 * If object is non-NULL, ref count must be bumped by caller 1859 * prior to making call to account for the new entry. 1860 */ 1861 int 1862 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1863 vm_offset_t *addr, /* IN/OUT */ 1864 vm_size_t length, vm_offset_t max_addr, int find_space, 1865 vm_prot_t prot, vm_prot_t max, int cow) 1866 { 1867 vm_offset_t alignment, curr_min_addr, min_addr; 1868 int gap, pidx, rv, try; 1869 bool cluster, en_aslr, update_anon; 1870 1871 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 1872 object == NULL, 1873 ("vm_map_find: non-NULL backing object for stack")); 1874 MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE && 1875 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0)); 1876 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL || 1877 (object->flags & OBJ_COLORED) == 0)) 1878 find_space = VMFS_ANY_SPACE; 1879 if (find_space >> 8 != 0) { 1880 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags")); 1881 alignment = (vm_offset_t)1 << (find_space >> 8); 1882 } else 1883 alignment = 0; 1884 en_aslr = (map->flags & MAP_ASLR) != 0; 1885 update_anon = cluster = clustering_anon_allowed(*addr) && 1886 (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 && 1887 find_space != VMFS_NO_SPACE && object == NULL && 1888 (cow & (MAP_INHERIT_SHARE | MAP_STACK_GROWS_UP | 1889 MAP_STACK_GROWS_DOWN)) == 0 && prot != PROT_NONE; 1890 curr_min_addr = min_addr = *addr; 1891 if (en_aslr && min_addr == 0 && !cluster && 1892 find_space != VMFS_NO_SPACE && 1893 (map->flags & MAP_ASLR_IGNSTART) != 0) 1894 curr_min_addr = min_addr = vm_map_min(map); 1895 try = 0; 1896 vm_map_lock(map); 1897 if (cluster) { 1898 curr_min_addr = map->anon_loc; 1899 if (curr_min_addr == 0) 1900 cluster = false; 1901 } 1902 if (find_space != VMFS_NO_SPACE) { 1903 KASSERT(find_space == VMFS_ANY_SPACE || 1904 find_space == VMFS_OPTIMAL_SPACE || 1905 find_space == VMFS_SUPER_SPACE || 1906 alignment != 0, ("unexpected VMFS flag")); 1907 again: 1908 /* 1909 * When creating an anonymous mapping, try clustering 1910 * with an existing anonymous mapping first. 1911 * 1912 * We make up to two attempts to find address space 1913 * for a given find_space value. The first attempt may 1914 * apply randomization or may cluster with an existing 1915 * anonymous mapping. If this first attempt fails, 1916 * perform a first-fit search of the available address 1917 * space. 1918 * 1919 * If all tries failed, and find_space is 1920 * VMFS_OPTIMAL_SPACE, fallback to VMFS_ANY_SPACE. 1921 * Again enable clustering and randomization. 1922 */ 1923 try++; 1924 MPASS(try <= 2); 1925 1926 if (try == 2) { 1927 /* 1928 * Second try: we failed either to find a 1929 * suitable region for randomizing the 1930 * allocation, or to cluster with an existing 1931 * mapping. Retry with free run. 1932 */ 1933 curr_min_addr = (map->flags & MAP_ASLR_IGNSTART) != 0 ? 1934 vm_map_min(map) : min_addr; 1935 atomic_add_long(&aslr_restarts, 1); 1936 } 1937 1938 if (try == 1 && en_aslr && !cluster) { 1939 /* 1940 * Find space for allocation, including 1941 * gap needed for later randomization. 1942 */ 1943 pidx = MAXPAGESIZES > 1 && pagesizes[1] != 0 && 1944 (find_space == VMFS_SUPER_SPACE || find_space == 1945 VMFS_OPTIMAL_SPACE) ? 1 : 0; 1946 gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR && 1947 (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ? 1948 aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx]; 1949 *addr = vm_map_findspace(map, curr_min_addr, 1950 length + gap * pagesizes[pidx]); 1951 if (*addr + length + gap * pagesizes[pidx] > 1952 vm_map_max(map)) 1953 goto again; 1954 /* And randomize the start address. */ 1955 *addr += (arc4random() % gap) * pagesizes[pidx]; 1956 if (max_addr != 0 && *addr + length > max_addr) 1957 goto again; 1958 } else { 1959 *addr = vm_map_findspace(map, curr_min_addr, length); 1960 if (*addr + length > vm_map_max(map) || 1961 (max_addr != 0 && *addr + length > max_addr)) { 1962 if (cluster) { 1963 cluster = false; 1964 MPASS(try == 1); 1965 goto again; 1966 } 1967 rv = KERN_NO_SPACE; 1968 goto done; 1969 } 1970 } 1971 1972 if (find_space != VMFS_ANY_SPACE && 1973 (rv = vm_map_alignspace(map, object, offset, addr, length, 1974 max_addr, alignment)) != KERN_SUCCESS) { 1975 if (find_space == VMFS_OPTIMAL_SPACE) { 1976 find_space = VMFS_ANY_SPACE; 1977 curr_min_addr = min_addr; 1978 cluster = update_anon; 1979 try = 0; 1980 goto again; 1981 } 1982 goto done; 1983 } 1984 } else if ((cow & MAP_REMAP) != 0) { 1985 if (*addr < vm_map_min(map) || 1986 *addr + length > vm_map_max(map) || 1987 *addr + length <= length) { 1988 rv = KERN_INVALID_ADDRESS; 1989 goto done; 1990 } 1991 vm_map_delete(map, *addr, *addr + length); 1992 } 1993 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 1994 rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot, 1995 max, cow); 1996 } else { 1997 rv = vm_map_insert(map, object, offset, *addr, *addr + length, 1998 prot, max, cow); 1999 } 2000 if (rv == KERN_SUCCESS && update_anon) 2001 map->anon_loc = *addr + length; 2002 done: 2003 vm_map_unlock(map); 2004 return (rv); 2005 } 2006 2007 /* 2008 * vm_map_find_min() is a variant of vm_map_find() that takes an 2009 * additional parameter (min_addr) and treats the given address 2010 * (*addr) differently. Specifically, it treats *addr as a hint 2011 * and not as the minimum address where the mapping is created. 2012 * 2013 * This function works in two phases. First, it tries to 2014 * allocate above the hint. If that fails and the hint is 2015 * greater than min_addr, it performs a second pass, replacing 2016 * the hint with min_addr as the minimum address for the 2017 * allocation. 2018 */ 2019 int 2020 vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2021 vm_offset_t *addr, vm_size_t length, vm_offset_t min_addr, 2022 vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max, 2023 int cow) 2024 { 2025 vm_offset_t hint; 2026 int rv; 2027 2028 hint = *addr; 2029 for (;;) { 2030 rv = vm_map_find(map, object, offset, addr, length, max_addr, 2031 find_space, prot, max, cow); 2032 if (rv == KERN_SUCCESS || min_addr >= hint) 2033 return (rv); 2034 *addr = hint = min_addr; 2035 } 2036 } 2037 2038 /* 2039 * A map entry with any of the following flags set must not be merged with 2040 * another entry. 2041 */ 2042 #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP | \ 2043 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC) 2044 2045 static bool 2046 vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry) 2047 { 2048 2049 KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 || 2050 (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0, 2051 ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable", 2052 prev, entry)); 2053 return (prev->end == entry->start && 2054 prev->object.vm_object == entry->object.vm_object && 2055 (prev->object.vm_object == NULL || 2056 prev->offset + (prev->end - prev->start) == entry->offset) && 2057 prev->eflags == entry->eflags && 2058 prev->protection == entry->protection && 2059 prev->max_protection == entry->max_protection && 2060 prev->inheritance == entry->inheritance && 2061 prev->wired_count == entry->wired_count && 2062 prev->cred == entry->cred); 2063 } 2064 2065 static void 2066 vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry) 2067 { 2068 2069 /* 2070 * If the backing object is a vnode object, vm_object_deallocate() 2071 * calls vrele(). However, vrele() does not lock the vnode because 2072 * the vnode has additional references. Thus, the map lock can be 2073 * kept without causing a lock-order reversal with the vnode lock. 2074 * 2075 * Since we count the number of virtual page mappings in 2076 * object->un_pager.vnp.writemappings, the writemappings value 2077 * should not be adjusted when the entry is disposed of. 2078 */ 2079 if (entry->object.vm_object != NULL) 2080 vm_object_deallocate(entry->object.vm_object); 2081 if (entry->cred != NULL) 2082 crfree(entry->cred); 2083 vm_map_entry_dispose(map, entry); 2084 } 2085 2086 /* 2087 * vm_map_try_merge_entries: 2088 * 2089 * Compare the given map entry to its predecessor, and merge its precessor 2090 * into it if possible. The entry remains valid, and may be extended. 2091 * The predecessor may be deleted. 2092 * 2093 * The map must be locked. 2094 */ 2095 void 2096 vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev, vm_map_entry_t entry) 2097 { 2098 2099 VM_MAP_ASSERT_LOCKED(map); 2100 if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 && 2101 vm_map_mergeable_neighbors(prev, entry)) { 2102 vm_map_entry_unlink(map, prev, UNLINK_MERGE_NEXT); 2103 vm_map_merged_neighbor_dispose(map, prev); 2104 } 2105 } 2106 2107 /* 2108 * vm_map_entry_back: 2109 * 2110 * Allocate an object to back a map entry. 2111 */ 2112 static inline void 2113 vm_map_entry_back(vm_map_entry_t entry) 2114 { 2115 vm_object_t object; 2116 2117 KASSERT(entry->object.vm_object == NULL, 2118 ("map entry %p has backing object", entry)); 2119 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2120 ("map entry %p is a submap", entry)); 2121 object = vm_object_allocate(OBJT_DEFAULT, 2122 atop(entry->end - entry->start)); 2123 entry->object.vm_object = object; 2124 entry->offset = 0; 2125 if (entry->cred != NULL) { 2126 object->cred = entry->cred; 2127 object->charge = entry->end - entry->start; 2128 entry->cred = NULL; 2129 } 2130 } 2131 2132 /* 2133 * vm_map_entry_charge_object 2134 * 2135 * If there is no object backing this entry, create one. Otherwise, if 2136 * the entry has cred, give it to the backing object. 2137 */ 2138 static inline void 2139 vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry) 2140 { 2141 2142 VM_MAP_ASSERT_LOCKED(map); 2143 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2144 ("map entry %p is a submap", entry)); 2145 if (entry->object.vm_object == NULL && !map->system_map && 2146 (entry->eflags & MAP_ENTRY_GUARD) == 0) 2147 vm_map_entry_back(entry); 2148 else if (entry->object.vm_object != NULL && 2149 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) && 2150 entry->cred != NULL) { 2151 VM_OBJECT_WLOCK(entry->object.vm_object); 2152 KASSERT(entry->object.vm_object->cred == NULL, 2153 ("OVERCOMMIT: %s: both cred e %p", __func__, entry)); 2154 entry->object.vm_object->cred = entry->cred; 2155 entry->object.vm_object->charge = entry->end - entry->start; 2156 VM_OBJECT_WUNLOCK(entry->object.vm_object); 2157 entry->cred = NULL; 2158 } 2159 } 2160 2161 /* 2162 * vm_map_clip_start: [ internal use only ] 2163 * 2164 * Asserts that the given entry begins at or after 2165 * the specified address; if necessary, 2166 * it splits the entry into two. 2167 */ 2168 #define vm_map_clip_start(map, entry, startaddr) \ 2169 { \ 2170 if (startaddr > entry->start) \ 2171 _vm_map_clip_start(map, entry, startaddr); \ 2172 } 2173 2174 /* 2175 * This routine is called only when it is known that 2176 * the entry must be split. 2177 */ 2178 static void 2179 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start) 2180 { 2181 vm_map_entry_t new_entry; 2182 2183 VM_MAP_ASSERT_LOCKED(map); 2184 KASSERT(entry->end > start && entry->start < start, 2185 ("_vm_map_clip_start: invalid clip of entry %p", entry)); 2186 2187 /* 2188 * Create a backing object now, if none exists, so that more individual 2189 * objects won't be created after the map entry is split. 2190 */ 2191 vm_map_entry_charge_object(map, entry); 2192 2193 /* Clone the entry. */ 2194 new_entry = vm_map_entry_create(map); 2195 *new_entry = *entry; 2196 2197 /* 2198 * Split off the front portion. Insert the new entry BEFORE this one, 2199 * so that this entry has the specified starting address. 2200 */ 2201 new_entry->end = start; 2202 entry->offset += (start - entry->start); 2203 entry->start = start; 2204 if (new_entry->cred != NULL) 2205 crhold(entry->cred); 2206 2207 vm_map_entry_link(map, new_entry); 2208 2209 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2210 vm_object_reference(new_entry->object.vm_object); 2211 vm_map_entry_set_vnode_text(new_entry, true); 2212 /* 2213 * The object->un_pager.vnp.writemappings for the 2214 * object of MAP_ENTRY_WRITECNT type entry shall be 2215 * kept as is here. The virtual pages are 2216 * re-distributed among the clipped entries, so the sum is 2217 * left the same. 2218 */ 2219 } 2220 } 2221 2222 /* 2223 * vm_map_clip_end: [ internal use only ] 2224 * 2225 * Asserts that the given entry ends at or before 2226 * the specified address; if necessary, 2227 * it splits the entry into two. 2228 */ 2229 #define vm_map_clip_end(map, entry, endaddr) \ 2230 { \ 2231 if ((endaddr) < (entry->end)) \ 2232 _vm_map_clip_end((map), (entry), (endaddr)); \ 2233 } 2234 2235 /* 2236 * This routine is called only when it is known that 2237 * the entry must be split. 2238 */ 2239 static void 2240 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end) 2241 { 2242 vm_map_entry_t new_entry; 2243 2244 VM_MAP_ASSERT_LOCKED(map); 2245 KASSERT(entry->start < end && entry->end > end, 2246 ("_vm_map_clip_end: invalid clip of entry %p", entry)); 2247 2248 /* 2249 * Create a backing object now, if none exists, so that more individual 2250 * objects won't be created after the map entry is split. 2251 */ 2252 vm_map_entry_charge_object(map, entry); 2253 2254 /* Clone the entry. */ 2255 new_entry = vm_map_entry_create(map); 2256 *new_entry = *entry; 2257 2258 /* 2259 * Split off the back portion. Insert the new entry AFTER this one, 2260 * so that this entry has the specified ending address. 2261 */ 2262 new_entry->start = entry->end = end; 2263 new_entry->offset += (end - entry->start); 2264 if (new_entry->cred != NULL) 2265 crhold(entry->cred); 2266 2267 vm_map_entry_link(map, new_entry); 2268 2269 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2270 vm_object_reference(new_entry->object.vm_object); 2271 vm_map_entry_set_vnode_text(new_entry, true); 2272 } 2273 } 2274 2275 /* 2276 * vm_map_submap: [ kernel use only ] 2277 * 2278 * Mark the given range as handled by a subordinate map. 2279 * 2280 * This range must have been created with vm_map_find, 2281 * and no other operations may have been performed on this 2282 * range prior to calling vm_map_submap. 2283 * 2284 * Only a limited number of operations can be performed 2285 * within this rage after calling vm_map_submap: 2286 * vm_fault 2287 * [Don't try vm_map_copy!] 2288 * 2289 * To remove a submapping, one must first remove the 2290 * range from the superior map, and then destroy the 2291 * submap (if desired). [Better yet, don't try it.] 2292 */ 2293 int 2294 vm_map_submap( 2295 vm_map_t map, 2296 vm_offset_t start, 2297 vm_offset_t end, 2298 vm_map_t submap) 2299 { 2300 vm_map_entry_t entry; 2301 int result; 2302 2303 result = KERN_INVALID_ARGUMENT; 2304 2305 vm_map_lock(submap); 2306 submap->flags |= MAP_IS_SUB_MAP; 2307 vm_map_unlock(submap); 2308 2309 vm_map_lock(map); 2310 2311 VM_MAP_RANGE_CHECK(map, start, end); 2312 2313 if (vm_map_lookup_entry(map, start, &entry)) { 2314 vm_map_clip_start(map, entry, start); 2315 } else 2316 entry = entry->next; 2317 2318 vm_map_clip_end(map, entry, end); 2319 2320 if ((entry->start == start) && (entry->end == end) && 2321 ((entry->eflags & MAP_ENTRY_COW) == 0) && 2322 (entry->object.vm_object == NULL)) { 2323 entry->object.sub_map = submap; 2324 entry->eflags |= MAP_ENTRY_IS_SUB_MAP; 2325 result = KERN_SUCCESS; 2326 } 2327 vm_map_unlock(map); 2328 2329 if (result != KERN_SUCCESS) { 2330 vm_map_lock(submap); 2331 submap->flags &= ~MAP_IS_SUB_MAP; 2332 vm_map_unlock(submap); 2333 } 2334 return (result); 2335 } 2336 2337 /* 2338 * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified 2339 */ 2340 #define MAX_INIT_PT 96 2341 2342 /* 2343 * vm_map_pmap_enter: 2344 * 2345 * Preload the specified map's pmap with mappings to the specified 2346 * object's memory-resident pages. No further physical pages are 2347 * allocated, and no further virtual pages are retrieved from secondary 2348 * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a 2349 * limited number of page mappings are created at the low-end of the 2350 * specified address range. (For this purpose, a superpage mapping 2351 * counts as one page mapping.) Otherwise, all resident pages within 2352 * the specified address range are mapped. 2353 */ 2354 static void 2355 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 2356 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags) 2357 { 2358 vm_offset_t start; 2359 vm_page_t p, p_start; 2360 vm_pindex_t mask, psize, threshold, tmpidx; 2361 2362 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL) 2363 return; 2364 VM_OBJECT_RLOCK(object); 2365 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2366 VM_OBJECT_RUNLOCK(object); 2367 VM_OBJECT_WLOCK(object); 2368 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2369 pmap_object_init_pt(map->pmap, addr, object, pindex, 2370 size); 2371 VM_OBJECT_WUNLOCK(object); 2372 return; 2373 } 2374 VM_OBJECT_LOCK_DOWNGRADE(object); 2375 } 2376 2377 psize = atop(size); 2378 if (psize + pindex > object->size) { 2379 if (object->size < pindex) { 2380 VM_OBJECT_RUNLOCK(object); 2381 return; 2382 } 2383 psize = object->size - pindex; 2384 } 2385 2386 start = 0; 2387 p_start = NULL; 2388 threshold = MAX_INIT_PT; 2389 2390 p = vm_page_find_least(object, pindex); 2391 /* 2392 * Assert: the variable p is either (1) the page with the 2393 * least pindex greater than or equal to the parameter pindex 2394 * or (2) NULL. 2395 */ 2396 for (; 2397 p != NULL && (tmpidx = p->pindex - pindex) < psize; 2398 p = TAILQ_NEXT(p, listq)) { 2399 /* 2400 * don't allow an madvise to blow away our really 2401 * free pages allocating pv entries. 2402 */ 2403 if (((flags & MAP_PREFAULT_MADVISE) != 0 && 2404 vm_page_count_severe()) || 2405 ((flags & MAP_PREFAULT_PARTIAL) != 0 && 2406 tmpidx >= threshold)) { 2407 psize = tmpidx; 2408 break; 2409 } 2410 if (p->valid == VM_PAGE_BITS_ALL) { 2411 if (p_start == NULL) { 2412 start = addr + ptoa(tmpidx); 2413 p_start = p; 2414 } 2415 /* Jump ahead if a superpage mapping is possible. */ 2416 if (p->psind > 0 && ((addr + ptoa(tmpidx)) & 2417 (pagesizes[p->psind] - 1)) == 0) { 2418 mask = atop(pagesizes[p->psind]) - 1; 2419 if (tmpidx + mask < psize && 2420 vm_page_ps_test(p, PS_ALL_VALID, NULL)) { 2421 p += mask; 2422 threshold += mask; 2423 } 2424 } 2425 } else if (p_start != NULL) { 2426 pmap_enter_object(map->pmap, start, addr + 2427 ptoa(tmpidx), p_start, prot); 2428 p_start = NULL; 2429 } 2430 } 2431 if (p_start != NULL) 2432 pmap_enter_object(map->pmap, start, addr + ptoa(psize), 2433 p_start, prot); 2434 VM_OBJECT_RUNLOCK(object); 2435 } 2436 2437 /* 2438 * vm_map_protect: 2439 * 2440 * Sets the protection of the specified address 2441 * region in the target map. If "set_max" is 2442 * specified, the maximum protection is to be set; 2443 * otherwise, only the current protection is affected. 2444 */ 2445 int 2446 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 2447 vm_prot_t new_prot, boolean_t set_max) 2448 { 2449 vm_map_entry_t current, entry, in_tran; 2450 vm_object_t obj; 2451 struct ucred *cred; 2452 vm_prot_t old_prot; 2453 int rv; 2454 2455 if (start == end) 2456 return (KERN_SUCCESS); 2457 2458 again: 2459 in_tran = NULL; 2460 vm_map_lock(map); 2461 2462 /* 2463 * Ensure that we are not concurrently wiring pages. vm_map_wire() may 2464 * need to fault pages into the map and will drop the map lock while 2465 * doing so, and the VM object may end up in an inconsistent state if we 2466 * update the protection on the map entry in between faults. 2467 */ 2468 vm_map_wait_busy(map); 2469 2470 VM_MAP_RANGE_CHECK(map, start, end); 2471 2472 if (!vm_map_lookup_entry(map, start, &entry)) 2473 entry = entry->next; 2474 2475 /* 2476 * Make a first pass to check for protection violations. 2477 */ 2478 for (current = entry; current->start < end; current = current->next) { 2479 if ((current->eflags & MAP_ENTRY_GUARD) != 0) 2480 continue; 2481 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { 2482 vm_map_unlock(map); 2483 return (KERN_INVALID_ARGUMENT); 2484 } 2485 if ((new_prot & current->max_protection) != new_prot) { 2486 vm_map_unlock(map); 2487 return (KERN_PROTECTION_FAILURE); 2488 } 2489 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0) 2490 in_tran = entry; 2491 } 2492 2493 /* 2494 * Postpone the operation until all in transition map entries 2495 * are stabilized. In-transition entry might already have its 2496 * pages wired and wired_count incremented, but 2497 * MAP_ENTRY_USER_WIRED flag not yet set, and visible to other 2498 * threads because the map lock is dropped. In this case we 2499 * would miss our call to vm_fault_copy_entry(). 2500 */ 2501 if (in_tran != NULL) { 2502 in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2503 vm_map_unlock_and_wait(map, 0); 2504 goto again; 2505 } 2506 2507 /* 2508 * Before changing the protections, try to reserve swap space for any 2509 * private (i.e., copy-on-write) mappings that are transitioning from 2510 * read-only to read/write access. If a reservation fails, break out 2511 * of this loop early and let the next loop simplify the entries, since 2512 * some may now be mergeable. 2513 */ 2514 rv = KERN_SUCCESS; 2515 vm_map_clip_start(map, entry, start); 2516 for (current = entry; current->start < end; current = current->next) { 2517 2518 vm_map_clip_end(map, current, end); 2519 2520 if (set_max || 2521 ((new_prot & ~(current->protection)) & VM_PROT_WRITE) == 0 || 2522 ENTRY_CHARGED(current) || 2523 (current->eflags & MAP_ENTRY_GUARD) != 0) { 2524 continue; 2525 } 2526 2527 cred = curthread->td_ucred; 2528 obj = current->object.vm_object; 2529 2530 if (obj == NULL || (current->eflags & MAP_ENTRY_NEEDS_COPY)) { 2531 if (!swap_reserve(current->end - current->start)) { 2532 rv = KERN_RESOURCE_SHORTAGE; 2533 end = current->end; 2534 break; 2535 } 2536 crhold(cred); 2537 current->cred = cred; 2538 continue; 2539 } 2540 2541 VM_OBJECT_WLOCK(obj); 2542 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) { 2543 VM_OBJECT_WUNLOCK(obj); 2544 continue; 2545 } 2546 2547 /* 2548 * Charge for the whole object allocation now, since 2549 * we cannot distinguish between non-charged and 2550 * charged clipped mapping of the same object later. 2551 */ 2552 KASSERT(obj->charge == 0, 2553 ("vm_map_protect: object %p overcharged (entry %p)", 2554 obj, current)); 2555 if (!swap_reserve(ptoa(obj->size))) { 2556 VM_OBJECT_WUNLOCK(obj); 2557 rv = KERN_RESOURCE_SHORTAGE; 2558 end = current->end; 2559 break; 2560 } 2561 2562 crhold(cred); 2563 obj->cred = cred; 2564 obj->charge = ptoa(obj->size); 2565 VM_OBJECT_WUNLOCK(obj); 2566 } 2567 2568 /* 2569 * If enough swap space was available, go back and fix up protections. 2570 * Otherwise, just simplify entries, since some may have been modified. 2571 * [Note that clipping is not necessary the second time.] 2572 */ 2573 for (current = entry; current->start < end; 2574 vm_map_try_merge_entries(map, current->prev, current), 2575 current = current->next) { 2576 if (rv != KERN_SUCCESS || 2577 (current->eflags & MAP_ENTRY_GUARD) != 0) 2578 continue; 2579 2580 old_prot = current->protection; 2581 2582 if (set_max) 2583 current->protection = 2584 (current->max_protection = new_prot) & 2585 old_prot; 2586 else 2587 current->protection = new_prot; 2588 2589 /* 2590 * For user wired map entries, the normal lazy evaluation of 2591 * write access upgrades through soft page faults is 2592 * undesirable. Instead, immediately copy any pages that are 2593 * copy-on-write and enable write access in the physical map. 2594 */ 2595 if ((current->eflags & MAP_ENTRY_USER_WIRED) != 0 && 2596 (current->protection & VM_PROT_WRITE) != 0 && 2597 (old_prot & VM_PROT_WRITE) == 0) 2598 vm_fault_copy_entry(map, map, current, current, NULL); 2599 2600 /* 2601 * When restricting access, update the physical map. Worry 2602 * about copy-on-write here. 2603 */ 2604 if ((old_prot & ~current->protection) != 0) { 2605 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 2606 VM_PROT_ALL) 2607 pmap_protect(map->pmap, current->start, 2608 current->end, 2609 current->protection & MASK(current)); 2610 #undef MASK 2611 } 2612 } 2613 vm_map_try_merge_entries(map, current->prev, current); 2614 vm_map_unlock(map); 2615 return (rv); 2616 } 2617 2618 /* 2619 * vm_map_madvise: 2620 * 2621 * This routine traverses a processes map handling the madvise 2622 * system call. Advisories are classified as either those effecting 2623 * the vm_map_entry structure, or those effecting the underlying 2624 * objects. 2625 */ 2626 int 2627 vm_map_madvise( 2628 vm_map_t map, 2629 vm_offset_t start, 2630 vm_offset_t end, 2631 int behav) 2632 { 2633 vm_map_entry_t current, entry; 2634 bool modify_map; 2635 2636 /* 2637 * Some madvise calls directly modify the vm_map_entry, in which case 2638 * we need to use an exclusive lock on the map and we need to perform 2639 * various clipping operations. Otherwise we only need a read-lock 2640 * on the map. 2641 */ 2642 switch(behav) { 2643 case MADV_NORMAL: 2644 case MADV_SEQUENTIAL: 2645 case MADV_RANDOM: 2646 case MADV_NOSYNC: 2647 case MADV_AUTOSYNC: 2648 case MADV_NOCORE: 2649 case MADV_CORE: 2650 if (start == end) 2651 return (0); 2652 modify_map = true; 2653 vm_map_lock(map); 2654 break; 2655 case MADV_WILLNEED: 2656 case MADV_DONTNEED: 2657 case MADV_FREE: 2658 if (start == end) 2659 return (0); 2660 modify_map = false; 2661 vm_map_lock_read(map); 2662 break; 2663 default: 2664 return (EINVAL); 2665 } 2666 2667 /* 2668 * Locate starting entry and clip if necessary. 2669 */ 2670 VM_MAP_RANGE_CHECK(map, start, end); 2671 2672 if (vm_map_lookup_entry(map, start, &entry)) { 2673 if (modify_map) 2674 vm_map_clip_start(map, entry, start); 2675 } else { 2676 entry = entry->next; 2677 } 2678 2679 if (modify_map) { 2680 /* 2681 * madvise behaviors that are implemented in the vm_map_entry. 2682 * 2683 * We clip the vm_map_entry so that behavioral changes are 2684 * limited to the specified address range. 2685 */ 2686 for (current = entry; current->start < end; 2687 current = current->next) { 2688 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) 2689 continue; 2690 2691 vm_map_clip_end(map, current, end); 2692 2693 switch (behav) { 2694 case MADV_NORMAL: 2695 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 2696 break; 2697 case MADV_SEQUENTIAL: 2698 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 2699 break; 2700 case MADV_RANDOM: 2701 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 2702 break; 2703 case MADV_NOSYNC: 2704 current->eflags |= MAP_ENTRY_NOSYNC; 2705 break; 2706 case MADV_AUTOSYNC: 2707 current->eflags &= ~MAP_ENTRY_NOSYNC; 2708 break; 2709 case MADV_NOCORE: 2710 current->eflags |= MAP_ENTRY_NOCOREDUMP; 2711 break; 2712 case MADV_CORE: 2713 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 2714 break; 2715 default: 2716 break; 2717 } 2718 vm_map_try_merge_entries(map, current->prev, current); 2719 } 2720 vm_map_try_merge_entries(map, current->prev, current); 2721 vm_map_unlock(map); 2722 } else { 2723 vm_pindex_t pstart, pend; 2724 2725 /* 2726 * madvise behaviors that are implemented in the underlying 2727 * vm_object. 2728 * 2729 * Since we don't clip the vm_map_entry, we have to clip 2730 * the vm_object pindex and count. 2731 */ 2732 for (current = entry; current->start < end; 2733 current = current->next) { 2734 vm_offset_t useEnd, useStart; 2735 2736 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) 2737 continue; 2738 2739 /* 2740 * MADV_FREE would otherwise rewind time to 2741 * the creation of the shadow object. Because 2742 * we hold the VM map read-locked, neither the 2743 * entry's object nor the presence of a 2744 * backing object can change. 2745 */ 2746 if (behav == MADV_FREE && 2747 current->object.vm_object != NULL && 2748 current->object.vm_object->backing_object != NULL) 2749 continue; 2750 2751 pstart = OFF_TO_IDX(current->offset); 2752 pend = pstart + atop(current->end - current->start); 2753 useStart = current->start; 2754 useEnd = current->end; 2755 2756 if (current->start < start) { 2757 pstart += atop(start - current->start); 2758 useStart = start; 2759 } 2760 if (current->end > end) { 2761 pend -= atop(current->end - end); 2762 useEnd = end; 2763 } 2764 2765 if (pstart >= pend) 2766 continue; 2767 2768 /* 2769 * Perform the pmap_advise() before clearing 2770 * PGA_REFERENCED in vm_page_advise(). Otherwise, a 2771 * concurrent pmap operation, such as pmap_remove(), 2772 * could clear a reference in the pmap and set 2773 * PGA_REFERENCED on the page before the pmap_advise() 2774 * had completed. Consequently, the page would appear 2775 * referenced based upon an old reference that 2776 * occurred before this pmap_advise() ran. 2777 */ 2778 if (behav == MADV_DONTNEED || behav == MADV_FREE) 2779 pmap_advise(map->pmap, useStart, useEnd, 2780 behav); 2781 2782 vm_object_madvise(current->object.vm_object, pstart, 2783 pend, behav); 2784 2785 /* 2786 * Pre-populate paging structures in the 2787 * WILLNEED case. For wired entries, the 2788 * paging structures are already populated. 2789 */ 2790 if (behav == MADV_WILLNEED && 2791 current->wired_count == 0) { 2792 vm_map_pmap_enter(map, 2793 useStart, 2794 current->protection, 2795 current->object.vm_object, 2796 pstart, 2797 ptoa(pend - pstart), 2798 MAP_PREFAULT_MADVISE 2799 ); 2800 } 2801 } 2802 vm_map_unlock_read(map); 2803 } 2804 return (0); 2805 } 2806 2807 2808 /* 2809 * vm_map_inherit: 2810 * 2811 * Sets the inheritance of the specified address 2812 * range in the target map. Inheritance 2813 * affects how the map will be shared with 2814 * child maps at the time of vmspace_fork. 2815 */ 2816 int 2817 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 2818 vm_inherit_t new_inheritance) 2819 { 2820 vm_map_entry_t entry; 2821 vm_map_entry_t temp_entry; 2822 2823 switch (new_inheritance) { 2824 case VM_INHERIT_NONE: 2825 case VM_INHERIT_COPY: 2826 case VM_INHERIT_SHARE: 2827 case VM_INHERIT_ZERO: 2828 break; 2829 default: 2830 return (KERN_INVALID_ARGUMENT); 2831 } 2832 if (start == end) 2833 return (KERN_SUCCESS); 2834 vm_map_lock(map); 2835 VM_MAP_RANGE_CHECK(map, start, end); 2836 if (vm_map_lookup_entry(map, start, &temp_entry)) { 2837 entry = temp_entry; 2838 vm_map_clip_start(map, entry, start); 2839 } else 2840 entry = temp_entry->next; 2841 while (entry->start < end) { 2842 vm_map_clip_end(map, entry, end); 2843 if ((entry->eflags & MAP_ENTRY_GUARD) == 0 || 2844 new_inheritance != VM_INHERIT_ZERO) 2845 entry->inheritance = new_inheritance; 2846 vm_map_try_merge_entries(map, entry->prev, entry); 2847 entry = entry->next; 2848 } 2849 vm_map_try_merge_entries(map, entry->prev, entry); 2850 vm_map_unlock(map); 2851 return (KERN_SUCCESS); 2852 } 2853 2854 /* 2855 * vm_map_entry_in_transition: 2856 * 2857 * Release the map lock, and sleep until the entry is no longer in 2858 * transition. Awake and acquire the map lock. If the map changed while 2859 * another held the lock, lookup a possibly-changed entry at or after the 2860 * 'start' position of the old entry. 2861 */ 2862 static vm_map_entry_t 2863 vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start, 2864 vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry) 2865 { 2866 vm_map_entry_t entry; 2867 vm_offset_t start; 2868 u_int last_timestamp; 2869 2870 VM_MAP_ASSERT_LOCKED(map); 2871 KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 2872 ("not in-tranition map entry %p", in_entry)); 2873 /* 2874 * We have not yet clipped the entry. 2875 */ 2876 start = MAX(in_start, in_entry->start); 2877 in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2878 last_timestamp = map->timestamp; 2879 if (vm_map_unlock_and_wait(map, 0)) { 2880 /* 2881 * Allow interruption of user wiring/unwiring? 2882 */ 2883 } 2884 vm_map_lock(map); 2885 if (last_timestamp + 1 == map->timestamp) 2886 return (in_entry); 2887 2888 /* 2889 * Look again for the entry because the map was modified while it was 2890 * unlocked. Specifically, the entry may have been clipped, merged, or 2891 * deleted. 2892 */ 2893 if (!vm_map_lookup_entry(map, start, &entry)) { 2894 if (!holes_ok) { 2895 *io_end = start; 2896 return (NULL); 2897 } 2898 entry = entry->next; 2899 } 2900 return (entry); 2901 } 2902 2903 /* 2904 * vm_map_unwire: 2905 * 2906 * Implements both kernel and user unwiring. 2907 */ 2908 int 2909 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 2910 int flags) 2911 { 2912 vm_map_entry_t entry, first_entry; 2913 int rv; 2914 bool first_iteration, holes_ok, need_wakeup, user_unwire; 2915 2916 if (start == end) 2917 return (KERN_SUCCESS); 2918 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 2919 user_unwire = (flags & VM_MAP_WIRE_USER) != 0; 2920 vm_map_lock(map); 2921 VM_MAP_RANGE_CHECK(map, start, end); 2922 if (!vm_map_lookup_entry(map, start, &first_entry)) { 2923 if (holes_ok) 2924 first_entry = first_entry->next; 2925 else { 2926 vm_map_unlock(map); 2927 return (KERN_INVALID_ADDRESS); 2928 } 2929 } 2930 first_iteration = true; 2931 entry = first_entry; 2932 rv = KERN_SUCCESS; 2933 while (entry->start < end) { 2934 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 2935 /* 2936 * We have not yet clipped the entry. 2937 */ 2938 entry = vm_map_entry_in_transition(map, start, &end, 2939 holes_ok, entry); 2940 if (entry == NULL) { 2941 if (first_iteration) { 2942 vm_map_unlock(map); 2943 return (KERN_INVALID_ADDRESS); 2944 } 2945 rv = KERN_INVALID_ADDRESS; 2946 break; 2947 } 2948 first_entry = first_iteration ? entry : NULL; 2949 continue; 2950 } 2951 first_iteration = false; 2952 vm_map_clip_start(map, entry, start); 2953 vm_map_clip_end(map, entry, end); 2954 /* 2955 * Mark the entry in case the map lock is released. (See 2956 * above.) 2957 */ 2958 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 2959 entry->wiring_thread == NULL, 2960 ("owned map entry %p", entry)); 2961 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 2962 entry->wiring_thread = curthread; 2963 /* 2964 * Check the map for holes in the specified region. 2965 * If holes_ok, skip this check. 2966 */ 2967 if (!holes_ok && 2968 (entry->end < end && entry->next->start > entry->end)) { 2969 end = entry->end; 2970 rv = KERN_INVALID_ADDRESS; 2971 break; 2972 } 2973 /* 2974 * If system unwiring, require that the entry is system wired. 2975 */ 2976 if (!user_unwire && 2977 vm_map_entry_system_wired_count(entry) == 0) { 2978 end = entry->end; 2979 rv = KERN_INVALID_ARGUMENT; 2980 break; 2981 } 2982 entry = entry->next; 2983 } 2984 need_wakeup = false; 2985 if (first_entry == NULL && 2986 !vm_map_lookup_entry(map, start, &first_entry)) { 2987 KASSERT(holes_ok, ("vm_map_unwire: lookup failed")); 2988 first_entry = first_entry->next; 2989 } 2990 for (entry = first_entry; entry->start < end; entry = entry->next) { 2991 /* 2992 * If holes_ok was specified, an empty 2993 * space in the unwired region could have been mapped 2994 * while the map lock was dropped for draining 2995 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread 2996 * could be simultaneously wiring this new mapping 2997 * entry. Detect these cases and skip any entries 2998 * marked as in transition by us. 2999 */ 3000 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3001 entry->wiring_thread != curthread) { 3002 KASSERT(holes_ok, 3003 ("vm_map_unwire: !HOLESOK and new/changed entry")); 3004 continue; 3005 } 3006 3007 if (rv == KERN_SUCCESS && (!user_unwire || 3008 (entry->eflags & MAP_ENTRY_USER_WIRED))) { 3009 if (entry->wired_count == 1) 3010 vm_map_entry_unwire(map, entry); 3011 else 3012 entry->wired_count--; 3013 if (user_unwire) 3014 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3015 } 3016 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3017 ("vm_map_unwire: in-transition flag missing %p", entry)); 3018 KASSERT(entry->wiring_thread == curthread, 3019 ("vm_map_unwire: alien wire %p", entry)); 3020 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 3021 entry->wiring_thread = NULL; 3022 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3023 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3024 need_wakeup = true; 3025 } 3026 vm_map_try_merge_entries(map, entry->prev, entry); 3027 } 3028 vm_map_try_merge_entries(map, entry->prev, entry); 3029 vm_map_unlock(map); 3030 if (need_wakeup) 3031 vm_map_wakeup(map); 3032 return (rv); 3033 } 3034 3035 static void 3036 vm_map_wire_user_count_sub(u_long npages) 3037 { 3038 3039 atomic_subtract_long(&vm_user_wire_count, npages); 3040 } 3041 3042 static bool 3043 vm_map_wire_user_count_add(u_long npages) 3044 { 3045 u_long wired; 3046 3047 wired = vm_user_wire_count; 3048 do { 3049 if (npages + wired > vm_page_max_user_wired) 3050 return (false); 3051 } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired, 3052 npages + wired)); 3053 3054 return (true); 3055 } 3056 3057 /* 3058 * vm_map_wire_entry_failure: 3059 * 3060 * Handle a wiring failure on the given entry. 3061 * 3062 * The map should be locked. 3063 */ 3064 static void 3065 vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 3066 vm_offset_t failed_addr) 3067 { 3068 3069 VM_MAP_ASSERT_LOCKED(map); 3070 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 && 3071 entry->wired_count == 1, 3072 ("vm_map_wire_entry_failure: entry %p isn't being wired", entry)); 3073 KASSERT(failed_addr < entry->end, 3074 ("vm_map_wire_entry_failure: entry %p was fully wired", entry)); 3075 3076 /* 3077 * If any pages at the start of this entry were successfully wired, 3078 * then unwire them. 3079 */ 3080 if (failed_addr > entry->start) { 3081 pmap_unwire(map->pmap, entry->start, failed_addr); 3082 vm_object_unwire(entry->object.vm_object, entry->offset, 3083 failed_addr - entry->start, PQ_ACTIVE); 3084 } 3085 3086 /* 3087 * Assign an out-of-range value to represent the failure to wire this 3088 * entry. 3089 */ 3090 entry->wired_count = -1; 3091 } 3092 3093 int 3094 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3095 { 3096 int rv; 3097 3098 vm_map_lock(map); 3099 rv = vm_map_wire_locked(map, start, end, flags); 3100 vm_map_unlock(map); 3101 return (rv); 3102 } 3103 3104 3105 /* 3106 * vm_map_wire_locked: 3107 * 3108 * Implements both kernel and user wiring. Returns with the map locked, 3109 * the map lock may be dropped. 3110 */ 3111 int 3112 vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3113 { 3114 vm_map_entry_t entry, first_entry, tmp_entry; 3115 vm_offset_t faddr, saved_end, saved_start; 3116 u_long npages; 3117 u_int last_timestamp; 3118 int rv; 3119 bool first_iteration, holes_ok, need_wakeup, user_wire; 3120 vm_prot_t prot; 3121 3122 VM_MAP_ASSERT_LOCKED(map); 3123 3124 if (start == end) 3125 return (KERN_SUCCESS); 3126 prot = 0; 3127 if (flags & VM_MAP_WIRE_WRITE) 3128 prot |= VM_PROT_WRITE; 3129 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3130 user_wire = (flags & VM_MAP_WIRE_USER) != 0; 3131 VM_MAP_RANGE_CHECK(map, start, end); 3132 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3133 if (holes_ok) 3134 first_entry = first_entry->next; 3135 else 3136 return (KERN_INVALID_ADDRESS); 3137 } 3138 first_iteration = true; 3139 entry = first_entry; 3140 while (entry->start < end) { 3141 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3142 /* 3143 * We have not yet clipped the entry. 3144 */ 3145 entry = vm_map_entry_in_transition(map, start, &end, 3146 holes_ok, entry); 3147 if (entry == NULL) { 3148 if (first_iteration) 3149 return (KERN_INVALID_ADDRESS); 3150 rv = KERN_INVALID_ADDRESS; 3151 goto done; 3152 } 3153 first_entry = first_iteration ? entry : NULL; 3154 continue; 3155 } 3156 first_iteration = false; 3157 vm_map_clip_start(map, entry, start); 3158 vm_map_clip_end(map, entry, end); 3159 /* 3160 * Mark the entry in case the map lock is released. (See 3161 * above.) 3162 */ 3163 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3164 entry->wiring_thread == NULL, 3165 ("owned map entry %p", entry)); 3166 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3167 entry->wiring_thread = curthread; 3168 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 3169 || (entry->protection & prot) != prot) { 3170 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED; 3171 if (!holes_ok) { 3172 end = entry->end; 3173 rv = KERN_INVALID_ADDRESS; 3174 goto done; 3175 } 3176 } else if (entry->wired_count == 0) { 3177 entry->wired_count++; 3178 3179 npages = atop(entry->end - entry->start); 3180 if (user_wire && !vm_map_wire_user_count_add(npages)) { 3181 vm_map_wire_entry_failure(map, entry, 3182 entry->start); 3183 end = entry->end; 3184 rv = KERN_RESOURCE_SHORTAGE; 3185 goto done; 3186 } 3187 3188 /* 3189 * Release the map lock, relying on the in-transition 3190 * mark. Mark the map busy for fork. 3191 */ 3192 saved_start = entry->start; 3193 saved_end = entry->end; 3194 last_timestamp = map->timestamp; 3195 vm_map_busy(map); 3196 vm_map_unlock(map); 3197 3198 faddr = saved_start; 3199 do { 3200 /* 3201 * Simulate a fault to get the page and enter 3202 * it into the physical map. 3203 */ 3204 if ((rv = vm_fault(map, faddr, VM_PROT_NONE, 3205 VM_FAULT_WIRE)) != KERN_SUCCESS) 3206 break; 3207 } while ((faddr += PAGE_SIZE) < saved_end); 3208 vm_map_lock(map); 3209 vm_map_unbusy(map); 3210 if (last_timestamp + 1 != map->timestamp) { 3211 /* 3212 * Look again for the entry because the map was 3213 * modified while it was unlocked. The entry 3214 * may have been clipped, but NOT merged or 3215 * deleted. 3216 */ 3217 if (!vm_map_lookup_entry(map, saved_start, 3218 &tmp_entry)) 3219 KASSERT(false, 3220 ("vm_map_wire: lookup failed")); 3221 if (entry == first_entry) 3222 first_entry = tmp_entry; 3223 else 3224 first_entry = NULL; 3225 entry = tmp_entry; 3226 while (entry->end < saved_end) { 3227 /* 3228 * In case of failure, handle entries 3229 * that were not fully wired here; 3230 * fully wired entries are handled 3231 * later. 3232 */ 3233 if (rv != KERN_SUCCESS && 3234 faddr < entry->end) 3235 vm_map_wire_entry_failure(map, 3236 entry, faddr); 3237 entry = entry->next; 3238 } 3239 } 3240 if (rv != KERN_SUCCESS) { 3241 vm_map_wire_entry_failure(map, entry, faddr); 3242 if (user_wire) 3243 vm_map_wire_user_count_sub(npages); 3244 end = entry->end; 3245 goto done; 3246 } 3247 } else if (!user_wire || 3248 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3249 entry->wired_count++; 3250 } 3251 /* 3252 * Check the map for holes in the specified region. 3253 * If holes_ok was specified, skip this check. 3254 */ 3255 if (!holes_ok && 3256 entry->end < end && entry->next->start > entry->end) { 3257 end = entry->end; 3258 rv = KERN_INVALID_ADDRESS; 3259 goto done; 3260 } 3261 entry = entry->next; 3262 } 3263 rv = KERN_SUCCESS; 3264 done: 3265 need_wakeup = false; 3266 if (first_entry == NULL && 3267 !vm_map_lookup_entry(map, start, &first_entry)) { 3268 KASSERT(holes_ok, ("vm_map_wire: lookup failed")); 3269 first_entry = first_entry->next; 3270 } 3271 for (entry = first_entry; entry->start < end; entry = entry->next) { 3272 /* 3273 * If holes_ok was specified, an empty 3274 * space in the unwired region could have been mapped 3275 * while the map lock was dropped for faulting in the 3276 * pages or draining MAP_ENTRY_IN_TRANSITION. 3277 * Moreover, another thread could be simultaneously 3278 * wiring this new mapping entry. Detect these cases 3279 * and skip any entries marked as in transition not by us. 3280 */ 3281 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3282 entry->wiring_thread != curthread) { 3283 KASSERT(holes_ok, 3284 ("vm_map_wire: !HOLESOK and new/changed entry")); 3285 continue; 3286 } 3287 3288 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) { 3289 /* do nothing */ 3290 } else if (rv == KERN_SUCCESS) { 3291 if (user_wire) 3292 entry->eflags |= MAP_ENTRY_USER_WIRED; 3293 } else if (entry->wired_count == -1) { 3294 /* 3295 * Wiring failed on this entry. Thus, unwiring is 3296 * unnecessary. 3297 */ 3298 entry->wired_count = 0; 3299 } else if (!user_wire || 3300 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3301 /* 3302 * Undo the wiring. Wiring succeeded on this entry 3303 * but failed on a later entry. 3304 */ 3305 if (entry->wired_count == 1) { 3306 vm_map_entry_unwire(map, entry); 3307 if (user_wire) 3308 vm_map_wire_user_count_sub( 3309 atop(entry->end - entry->start)); 3310 } else 3311 entry->wired_count--; 3312 } 3313 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3314 ("vm_map_wire: in-transition flag missing %p", entry)); 3315 KASSERT(entry->wiring_thread == curthread, 3316 ("vm_map_wire: alien wire %p", entry)); 3317 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION | 3318 MAP_ENTRY_WIRE_SKIPPED); 3319 entry->wiring_thread = NULL; 3320 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3321 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3322 need_wakeup = true; 3323 } 3324 vm_map_try_merge_entries(map, entry->prev, entry); 3325 } 3326 vm_map_try_merge_entries(map, entry->prev, entry); 3327 if (need_wakeup) 3328 vm_map_wakeup(map); 3329 return (rv); 3330 } 3331 3332 /* 3333 * vm_map_sync 3334 * 3335 * Push any dirty cached pages in the address range to their pager. 3336 * If syncio is TRUE, dirty pages are written synchronously. 3337 * If invalidate is TRUE, any cached pages are freed as well. 3338 * 3339 * If the size of the region from start to end is zero, we are 3340 * supposed to flush all modified pages within the region containing 3341 * start. Unfortunately, a region can be split or coalesced with 3342 * neighboring regions, making it difficult to determine what the 3343 * original region was. Therefore, we approximate this requirement by 3344 * flushing the current region containing start. 3345 * 3346 * Returns an error if any part of the specified range is not mapped. 3347 */ 3348 int 3349 vm_map_sync( 3350 vm_map_t map, 3351 vm_offset_t start, 3352 vm_offset_t end, 3353 boolean_t syncio, 3354 boolean_t invalidate) 3355 { 3356 vm_map_entry_t current; 3357 vm_map_entry_t entry; 3358 vm_size_t size; 3359 vm_object_t object; 3360 vm_ooffset_t offset; 3361 unsigned int last_timestamp; 3362 boolean_t failed; 3363 3364 vm_map_lock_read(map); 3365 VM_MAP_RANGE_CHECK(map, start, end); 3366 if (!vm_map_lookup_entry(map, start, &entry)) { 3367 vm_map_unlock_read(map); 3368 return (KERN_INVALID_ADDRESS); 3369 } else if (start == end) { 3370 start = entry->start; 3371 end = entry->end; 3372 } 3373 /* 3374 * Make a first pass to check for user-wired memory and holes. 3375 */ 3376 for (current = entry; current->start < end; current = current->next) { 3377 if (invalidate && (current->eflags & MAP_ENTRY_USER_WIRED)) { 3378 vm_map_unlock_read(map); 3379 return (KERN_INVALID_ARGUMENT); 3380 } 3381 if (end > current->end && 3382 current->end != current->next->start) { 3383 vm_map_unlock_read(map); 3384 return (KERN_INVALID_ADDRESS); 3385 } 3386 } 3387 3388 if (invalidate) 3389 pmap_remove(map->pmap, start, end); 3390 failed = FALSE; 3391 3392 /* 3393 * Make a second pass, cleaning/uncaching pages from the indicated 3394 * objects as we go. 3395 */ 3396 for (current = entry; current->start < end;) { 3397 offset = current->offset + (start - current->start); 3398 size = (end <= current->end ? end : current->end) - start; 3399 if (current->eflags & MAP_ENTRY_IS_SUB_MAP) { 3400 vm_map_t smap; 3401 vm_map_entry_t tentry; 3402 vm_size_t tsize; 3403 3404 smap = current->object.sub_map; 3405 vm_map_lock_read(smap); 3406 (void) vm_map_lookup_entry(smap, offset, &tentry); 3407 tsize = tentry->end - offset; 3408 if (tsize < size) 3409 size = tsize; 3410 object = tentry->object.vm_object; 3411 offset = tentry->offset + (offset - tentry->start); 3412 vm_map_unlock_read(smap); 3413 } else { 3414 object = current->object.vm_object; 3415 } 3416 vm_object_reference(object); 3417 last_timestamp = map->timestamp; 3418 vm_map_unlock_read(map); 3419 if (!vm_object_sync(object, offset, size, syncio, invalidate)) 3420 failed = TRUE; 3421 start += size; 3422 vm_object_deallocate(object); 3423 vm_map_lock_read(map); 3424 if (last_timestamp == map->timestamp || 3425 !vm_map_lookup_entry(map, start, ¤t)) 3426 current = current->next; 3427 } 3428 3429 vm_map_unlock_read(map); 3430 return (failed ? KERN_FAILURE : KERN_SUCCESS); 3431 } 3432 3433 /* 3434 * vm_map_entry_unwire: [ internal use only ] 3435 * 3436 * Make the region specified by this entry pageable. 3437 * 3438 * The map in question should be locked. 3439 * [This is the reason for this routine's existence.] 3440 */ 3441 static void 3442 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 3443 { 3444 vm_size_t size; 3445 3446 VM_MAP_ASSERT_LOCKED(map); 3447 KASSERT(entry->wired_count > 0, 3448 ("vm_map_entry_unwire: entry %p isn't wired", entry)); 3449 3450 size = entry->end - entry->start; 3451 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) 3452 vm_map_wire_user_count_sub(atop(size)); 3453 pmap_unwire(map->pmap, entry->start, entry->end); 3454 vm_object_unwire(entry->object.vm_object, entry->offset, size, 3455 PQ_ACTIVE); 3456 entry->wired_count = 0; 3457 } 3458 3459 static void 3460 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map) 3461 { 3462 3463 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) 3464 vm_object_deallocate(entry->object.vm_object); 3465 uma_zfree(system_map ? kmapentzone : mapentzone, entry); 3466 } 3467 3468 /* 3469 * vm_map_entry_delete: [ internal use only ] 3470 * 3471 * Deallocate the given entry from the target map. 3472 */ 3473 static void 3474 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry) 3475 { 3476 vm_object_t object; 3477 vm_pindex_t offidxstart, offidxend, count, size1; 3478 vm_size_t size; 3479 3480 vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE); 3481 object = entry->object.vm_object; 3482 3483 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { 3484 MPASS(entry->cred == NULL); 3485 MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0); 3486 MPASS(object == NULL); 3487 vm_map_entry_deallocate(entry, map->system_map); 3488 return; 3489 } 3490 3491 size = entry->end - entry->start; 3492 map->size -= size; 3493 3494 if (entry->cred != NULL) { 3495 swap_release_by_cred(size, entry->cred); 3496 crfree(entry->cred); 3497 } 3498 3499 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0 && 3500 (object != NULL)) { 3501 KASSERT(entry->cred == NULL || object->cred == NULL || 3502 (entry->eflags & MAP_ENTRY_NEEDS_COPY), 3503 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry)); 3504 count = atop(size); 3505 offidxstart = OFF_TO_IDX(entry->offset); 3506 offidxend = offidxstart + count; 3507 VM_OBJECT_WLOCK(object); 3508 if (object->ref_count != 1 && ((object->flags & (OBJ_NOSPLIT | 3509 OBJ_ONEMAPPING)) == OBJ_ONEMAPPING || 3510 object == kernel_object)) { 3511 vm_object_collapse(object); 3512 3513 /* 3514 * The option OBJPR_NOTMAPPED can be passed here 3515 * because vm_map_delete() already performed 3516 * pmap_remove() on the only mapping to this range 3517 * of pages. 3518 */ 3519 vm_object_page_remove(object, offidxstart, offidxend, 3520 OBJPR_NOTMAPPED); 3521 if (object->type == OBJT_SWAP) 3522 swap_pager_freespace(object, offidxstart, 3523 count); 3524 if (offidxend >= object->size && 3525 offidxstart < object->size) { 3526 size1 = object->size; 3527 object->size = offidxstart; 3528 if (object->cred != NULL) { 3529 size1 -= object->size; 3530 KASSERT(object->charge >= ptoa(size1), 3531 ("object %p charge < 0", object)); 3532 swap_release_by_cred(ptoa(size1), 3533 object->cred); 3534 object->charge -= ptoa(size1); 3535 } 3536 } 3537 } 3538 VM_OBJECT_WUNLOCK(object); 3539 } else 3540 entry->object.vm_object = NULL; 3541 if (map->system_map) 3542 vm_map_entry_deallocate(entry, TRUE); 3543 else { 3544 entry->next = curthread->td_map_def_user; 3545 curthread->td_map_def_user = entry; 3546 } 3547 } 3548 3549 /* 3550 * vm_map_delete: [ internal use only ] 3551 * 3552 * Deallocates the given address range from the target 3553 * map. 3554 */ 3555 int 3556 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end) 3557 { 3558 vm_map_entry_t entry; 3559 vm_map_entry_t first_entry; 3560 3561 VM_MAP_ASSERT_LOCKED(map); 3562 if (start == end) 3563 return (KERN_SUCCESS); 3564 3565 /* 3566 * Find the start of the region, and clip it 3567 */ 3568 if (!vm_map_lookup_entry(map, start, &first_entry)) 3569 entry = first_entry->next; 3570 else { 3571 entry = first_entry; 3572 vm_map_clip_start(map, entry, start); 3573 } 3574 3575 /* 3576 * Step through all entries in this region 3577 */ 3578 while (entry->start < end) { 3579 vm_map_entry_t next; 3580 3581 /* 3582 * Wait for wiring or unwiring of an entry to complete. 3583 * Also wait for any system wirings to disappear on 3584 * user maps. 3585 */ 3586 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 || 3587 (vm_map_pmap(map) != kernel_pmap && 3588 vm_map_entry_system_wired_count(entry) != 0)) { 3589 unsigned int last_timestamp; 3590 vm_offset_t saved_start; 3591 vm_map_entry_t tmp_entry; 3592 3593 saved_start = entry->start; 3594 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3595 last_timestamp = map->timestamp; 3596 (void) vm_map_unlock_and_wait(map, 0); 3597 vm_map_lock(map); 3598 if (last_timestamp + 1 != map->timestamp) { 3599 /* 3600 * Look again for the entry because the map was 3601 * modified while it was unlocked. 3602 * Specifically, the entry may have been 3603 * clipped, merged, or deleted. 3604 */ 3605 if (!vm_map_lookup_entry(map, saved_start, 3606 &tmp_entry)) 3607 entry = tmp_entry->next; 3608 else { 3609 entry = tmp_entry; 3610 vm_map_clip_start(map, entry, 3611 saved_start); 3612 } 3613 } 3614 continue; 3615 } 3616 vm_map_clip_end(map, entry, end); 3617 3618 next = entry->next; 3619 3620 /* 3621 * Unwire before removing addresses from the pmap; otherwise, 3622 * unwiring will put the entries back in the pmap. 3623 */ 3624 if (entry->wired_count != 0) 3625 vm_map_entry_unwire(map, entry); 3626 3627 /* 3628 * Remove mappings for the pages, but only if the 3629 * mappings could exist. For instance, it does not 3630 * make sense to call pmap_remove() for guard entries. 3631 */ 3632 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 3633 entry->object.vm_object != NULL) 3634 pmap_remove(map->pmap, entry->start, entry->end); 3635 3636 if (entry->end == map->anon_loc) 3637 map->anon_loc = entry->start; 3638 3639 /* 3640 * Delete the entry only after removing all pmap 3641 * entries pointing to its pages. (Otherwise, its 3642 * page frames may be reallocated, and any modify bits 3643 * will be set in the wrong object!) 3644 */ 3645 vm_map_entry_delete(map, entry); 3646 entry = next; 3647 } 3648 return (KERN_SUCCESS); 3649 } 3650 3651 /* 3652 * vm_map_remove: 3653 * 3654 * Remove the given address range from the target map. 3655 * This is the exported form of vm_map_delete. 3656 */ 3657 int 3658 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 3659 { 3660 int result; 3661 3662 vm_map_lock(map); 3663 VM_MAP_RANGE_CHECK(map, start, end); 3664 result = vm_map_delete(map, start, end); 3665 vm_map_unlock(map); 3666 return (result); 3667 } 3668 3669 /* 3670 * vm_map_check_protection: 3671 * 3672 * Assert that the target map allows the specified privilege on the 3673 * entire address region given. The entire region must be allocated. 3674 * 3675 * WARNING! This code does not and should not check whether the 3676 * contents of the region is accessible. For example a smaller file 3677 * might be mapped into a larger address space. 3678 * 3679 * NOTE! This code is also called by munmap(). 3680 * 3681 * The map must be locked. A read lock is sufficient. 3682 */ 3683 boolean_t 3684 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 3685 vm_prot_t protection) 3686 { 3687 vm_map_entry_t entry; 3688 vm_map_entry_t tmp_entry; 3689 3690 if (!vm_map_lookup_entry(map, start, &tmp_entry)) 3691 return (FALSE); 3692 entry = tmp_entry; 3693 3694 while (start < end) { 3695 /* 3696 * No holes allowed! 3697 */ 3698 if (start < entry->start) 3699 return (FALSE); 3700 /* 3701 * Check protection associated with entry. 3702 */ 3703 if ((entry->protection & protection) != protection) 3704 return (FALSE); 3705 /* go to next entry */ 3706 start = entry->end; 3707 entry = entry->next; 3708 } 3709 return (TRUE); 3710 } 3711 3712 /* 3713 * vm_map_copy_entry: 3714 * 3715 * Copies the contents of the source entry to the destination 3716 * entry. The entries *must* be aligned properly. 3717 */ 3718 static void 3719 vm_map_copy_entry( 3720 vm_map_t src_map, 3721 vm_map_t dst_map, 3722 vm_map_entry_t src_entry, 3723 vm_map_entry_t dst_entry, 3724 vm_ooffset_t *fork_charge) 3725 { 3726 vm_object_t src_object; 3727 vm_map_entry_t fake_entry; 3728 vm_offset_t size; 3729 struct ucred *cred; 3730 int charged; 3731 3732 VM_MAP_ASSERT_LOCKED(dst_map); 3733 3734 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) 3735 return; 3736 3737 if (src_entry->wired_count == 0 || 3738 (src_entry->protection & VM_PROT_WRITE) == 0) { 3739 /* 3740 * If the source entry is marked needs_copy, it is already 3741 * write-protected. 3742 */ 3743 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && 3744 (src_entry->protection & VM_PROT_WRITE) != 0) { 3745 pmap_protect(src_map->pmap, 3746 src_entry->start, 3747 src_entry->end, 3748 src_entry->protection & ~VM_PROT_WRITE); 3749 } 3750 3751 /* 3752 * Make a copy of the object. 3753 */ 3754 size = src_entry->end - src_entry->start; 3755 if ((src_object = src_entry->object.vm_object) != NULL) { 3756 VM_OBJECT_WLOCK(src_object); 3757 charged = ENTRY_CHARGED(src_entry); 3758 if (src_object->handle == NULL && 3759 (src_object->type == OBJT_DEFAULT || 3760 src_object->type == OBJT_SWAP)) { 3761 vm_object_collapse(src_object); 3762 if ((src_object->flags & (OBJ_NOSPLIT | 3763 OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) { 3764 vm_object_split(src_entry); 3765 src_object = 3766 src_entry->object.vm_object; 3767 } 3768 } 3769 vm_object_reference_locked(src_object); 3770 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 3771 if (src_entry->cred != NULL && 3772 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 3773 KASSERT(src_object->cred == NULL, 3774 ("OVERCOMMIT: vm_map_copy_entry: cred %p", 3775 src_object)); 3776 src_object->cred = src_entry->cred; 3777 src_object->charge = size; 3778 } 3779 VM_OBJECT_WUNLOCK(src_object); 3780 dst_entry->object.vm_object = src_object; 3781 if (charged) { 3782 cred = curthread->td_ucred; 3783 crhold(cred); 3784 dst_entry->cred = cred; 3785 *fork_charge += size; 3786 if (!(src_entry->eflags & 3787 MAP_ENTRY_NEEDS_COPY)) { 3788 crhold(cred); 3789 src_entry->cred = cred; 3790 *fork_charge += size; 3791 } 3792 } 3793 src_entry->eflags |= MAP_ENTRY_COW | 3794 MAP_ENTRY_NEEDS_COPY; 3795 dst_entry->eflags |= MAP_ENTRY_COW | 3796 MAP_ENTRY_NEEDS_COPY; 3797 dst_entry->offset = src_entry->offset; 3798 if (src_entry->eflags & MAP_ENTRY_WRITECNT) { 3799 /* 3800 * MAP_ENTRY_WRITECNT cannot 3801 * indicate write reference from 3802 * src_entry, since the entry is 3803 * marked as needs copy. Allocate a 3804 * fake entry that is used to 3805 * decrement object->un_pager writecount 3806 * at the appropriate time. Attach 3807 * fake_entry to the deferred list. 3808 */ 3809 fake_entry = vm_map_entry_create(dst_map); 3810 fake_entry->eflags = MAP_ENTRY_WRITECNT; 3811 src_entry->eflags &= ~MAP_ENTRY_WRITECNT; 3812 vm_object_reference(src_object); 3813 fake_entry->object.vm_object = src_object; 3814 fake_entry->start = src_entry->start; 3815 fake_entry->end = src_entry->end; 3816 fake_entry->next = curthread->td_map_def_user; 3817 curthread->td_map_def_user = fake_entry; 3818 } 3819 3820 pmap_copy(dst_map->pmap, src_map->pmap, 3821 dst_entry->start, dst_entry->end - dst_entry->start, 3822 src_entry->start); 3823 } else { 3824 dst_entry->object.vm_object = NULL; 3825 dst_entry->offset = 0; 3826 if (src_entry->cred != NULL) { 3827 dst_entry->cred = curthread->td_ucred; 3828 crhold(dst_entry->cred); 3829 *fork_charge += size; 3830 } 3831 } 3832 } else { 3833 /* 3834 * We don't want to make writeable wired pages copy-on-write. 3835 * Immediately copy these pages into the new map by simulating 3836 * page faults. The new pages are pageable. 3837 */ 3838 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry, 3839 fork_charge); 3840 } 3841 } 3842 3843 /* 3844 * vmspace_map_entry_forked: 3845 * Update the newly-forked vmspace each time a map entry is inherited 3846 * or copied. The values for vm_dsize and vm_tsize are approximate 3847 * (and mostly-obsolete ideas in the face of mmap(2) et al.) 3848 */ 3849 static void 3850 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2, 3851 vm_map_entry_t entry) 3852 { 3853 vm_size_t entrysize; 3854 vm_offset_t newend; 3855 3856 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 3857 return; 3858 entrysize = entry->end - entry->start; 3859 vm2->vm_map.size += entrysize; 3860 if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) { 3861 vm2->vm_ssize += btoc(entrysize); 3862 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr && 3863 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) { 3864 newend = MIN(entry->end, 3865 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)); 3866 vm2->vm_dsize += btoc(newend - entry->start); 3867 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr && 3868 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) { 3869 newend = MIN(entry->end, 3870 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)); 3871 vm2->vm_tsize += btoc(newend - entry->start); 3872 } 3873 } 3874 3875 /* 3876 * vmspace_fork: 3877 * Create a new process vmspace structure and vm_map 3878 * based on those of an existing process. The new map 3879 * is based on the old map, according to the inheritance 3880 * values on the regions in that map. 3881 * 3882 * XXX It might be worth coalescing the entries added to the new vmspace. 3883 * 3884 * The source map must not be locked. 3885 */ 3886 struct vmspace * 3887 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge) 3888 { 3889 struct vmspace *vm2; 3890 vm_map_t new_map, old_map; 3891 vm_map_entry_t new_entry, old_entry; 3892 vm_object_t object; 3893 int error, locked; 3894 vm_inherit_t inh; 3895 3896 old_map = &vm1->vm_map; 3897 /* Copy immutable fields of vm1 to vm2. */ 3898 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map), 3899 pmap_pinit); 3900 if (vm2 == NULL) 3901 return (NULL); 3902 3903 vm2->vm_taddr = vm1->vm_taddr; 3904 vm2->vm_daddr = vm1->vm_daddr; 3905 vm2->vm_maxsaddr = vm1->vm_maxsaddr; 3906 vm_map_lock(old_map); 3907 if (old_map->busy) 3908 vm_map_wait_busy(old_map); 3909 new_map = &vm2->vm_map; 3910 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */ 3911 KASSERT(locked, ("vmspace_fork: lock failed")); 3912 3913 error = pmap_vmspace_copy(new_map->pmap, old_map->pmap); 3914 if (error != 0) { 3915 sx_xunlock(&old_map->lock); 3916 sx_xunlock(&new_map->lock); 3917 vm_map_process_deferred(); 3918 vmspace_free(vm2); 3919 return (NULL); 3920 } 3921 3922 new_map->anon_loc = old_map->anon_loc; 3923 3924 old_entry = old_map->header.next; 3925 3926 while (old_entry != &old_map->header) { 3927 if (old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) 3928 panic("vm_map_fork: encountered a submap"); 3929 3930 inh = old_entry->inheritance; 3931 if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 && 3932 inh != VM_INHERIT_NONE) 3933 inh = VM_INHERIT_COPY; 3934 3935 switch (inh) { 3936 case VM_INHERIT_NONE: 3937 break; 3938 3939 case VM_INHERIT_SHARE: 3940 /* 3941 * Clone the entry, creating the shared object if necessary. 3942 */ 3943 object = old_entry->object.vm_object; 3944 if (object == NULL) { 3945 vm_map_entry_back(old_entry); 3946 object = old_entry->object.vm_object; 3947 } 3948 3949 /* 3950 * Add the reference before calling vm_object_shadow 3951 * to insure that a shadow object is created. 3952 */ 3953 vm_object_reference(object); 3954 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3955 vm_object_shadow(&old_entry->object.vm_object, 3956 &old_entry->offset, 3957 old_entry->end - old_entry->start); 3958 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 3959 /* Transfer the second reference too. */ 3960 vm_object_reference( 3961 old_entry->object.vm_object); 3962 3963 /* 3964 * As in vm_map_merged_neighbor_dispose(), 3965 * the vnode lock will not be acquired in 3966 * this call to vm_object_deallocate(). 3967 */ 3968 vm_object_deallocate(object); 3969 object = old_entry->object.vm_object; 3970 } 3971 VM_OBJECT_WLOCK(object); 3972 vm_object_clear_flag(object, OBJ_ONEMAPPING); 3973 if (old_entry->cred != NULL) { 3974 KASSERT(object->cred == NULL, ("vmspace_fork both cred")); 3975 object->cred = old_entry->cred; 3976 object->charge = old_entry->end - old_entry->start; 3977 old_entry->cred = NULL; 3978 } 3979 3980 /* 3981 * Assert the correct state of the vnode 3982 * v_writecount while the object is locked, to 3983 * not relock it later for the assertion 3984 * correctness. 3985 */ 3986 if (old_entry->eflags & MAP_ENTRY_WRITECNT && 3987 object->type == OBJT_VNODE) { 3988 KASSERT(((struct vnode *)object->handle)-> 3989 v_writecount > 0, 3990 ("vmspace_fork: v_writecount %p", object)); 3991 KASSERT(object->un_pager.vnp.writemappings > 0, 3992 ("vmspace_fork: vnp.writecount %p", 3993 object)); 3994 } 3995 VM_OBJECT_WUNLOCK(object); 3996 3997 /* 3998 * Clone the entry, referencing the shared object. 3999 */ 4000 new_entry = vm_map_entry_create(new_map); 4001 *new_entry = *old_entry; 4002 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4003 MAP_ENTRY_IN_TRANSITION); 4004 new_entry->wiring_thread = NULL; 4005 new_entry->wired_count = 0; 4006 if (new_entry->eflags & MAP_ENTRY_WRITECNT) { 4007 vm_pager_update_writecount(object, 4008 new_entry->start, new_entry->end); 4009 } 4010 vm_map_entry_set_vnode_text(new_entry, true); 4011 4012 /* 4013 * Insert the entry into the new map -- we know we're 4014 * inserting at the end of the new map. 4015 */ 4016 vm_map_entry_link(new_map, new_entry); 4017 vmspace_map_entry_forked(vm1, vm2, new_entry); 4018 4019 /* 4020 * Update the physical map 4021 */ 4022 pmap_copy(new_map->pmap, old_map->pmap, 4023 new_entry->start, 4024 (old_entry->end - old_entry->start), 4025 old_entry->start); 4026 break; 4027 4028 case VM_INHERIT_COPY: 4029 /* 4030 * Clone the entry and link into the map. 4031 */ 4032 new_entry = vm_map_entry_create(new_map); 4033 *new_entry = *old_entry; 4034 /* 4035 * Copied entry is COW over the old object. 4036 */ 4037 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4038 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT); 4039 new_entry->wiring_thread = NULL; 4040 new_entry->wired_count = 0; 4041 new_entry->object.vm_object = NULL; 4042 new_entry->cred = NULL; 4043 vm_map_entry_link(new_map, new_entry); 4044 vmspace_map_entry_forked(vm1, vm2, new_entry); 4045 vm_map_copy_entry(old_map, new_map, old_entry, 4046 new_entry, fork_charge); 4047 vm_map_entry_set_vnode_text(new_entry, true); 4048 break; 4049 4050 case VM_INHERIT_ZERO: 4051 /* 4052 * Create a new anonymous mapping entry modelled from 4053 * the old one. 4054 */ 4055 new_entry = vm_map_entry_create(new_map); 4056 memset(new_entry, 0, sizeof(*new_entry)); 4057 4058 new_entry->start = old_entry->start; 4059 new_entry->end = old_entry->end; 4060 new_entry->eflags = old_entry->eflags & 4061 ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION | 4062 MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC); 4063 new_entry->protection = old_entry->protection; 4064 new_entry->max_protection = old_entry->max_protection; 4065 new_entry->inheritance = VM_INHERIT_ZERO; 4066 4067 vm_map_entry_link(new_map, new_entry); 4068 vmspace_map_entry_forked(vm1, vm2, new_entry); 4069 4070 new_entry->cred = curthread->td_ucred; 4071 crhold(new_entry->cred); 4072 *fork_charge += (new_entry->end - new_entry->start); 4073 4074 break; 4075 } 4076 old_entry = old_entry->next; 4077 } 4078 /* 4079 * Use inlined vm_map_unlock() to postpone handling the deferred 4080 * map entries, which cannot be done until both old_map and 4081 * new_map locks are released. 4082 */ 4083 sx_xunlock(&old_map->lock); 4084 sx_xunlock(&new_map->lock); 4085 vm_map_process_deferred(); 4086 4087 return (vm2); 4088 } 4089 4090 /* 4091 * Create a process's stack for exec_new_vmspace(). This function is never 4092 * asked to wire the newly created stack. 4093 */ 4094 int 4095 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4096 vm_prot_t prot, vm_prot_t max, int cow) 4097 { 4098 vm_size_t growsize, init_ssize; 4099 rlim_t vmemlim; 4100 int rv; 4101 4102 MPASS((map->flags & MAP_WIREFUTURE) == 0); 4103 growsize = sgrowsiz; 4104 init_ssize = (max_ssize < growsize) ? max_ssize : growsize; 4105 vm_map_lock(map); 4106 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4107 /* If we would blow our VMEM resource limit, no go */ 4108 if (map->size + init_ssize > vmemlim) { 4109 rv = KERN_NO_SPACE; 4110 goto out; 4111 } 4112 rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot, 4113 max, cow); 4114 out: 4115 vm_map_unlock(map); 4116 return (rv); 4117 } 4118 4119 static int stack_guard_page = 1; 4120 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN, 4121 &stack_guard_page, 0, 4122 "Specifies the number of guard pages for a stack that grows"); 4123 4124 static int 4125 vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4126 vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow) 4127 { 4128 vm_map_entry_t new_entry, prev_entry; 4129 vm_offset_t bot, gap_bot, gap_top, top; 4130 vm_size_t init_ssize, sgp; 4131 int orient, rv; 4132 4133 /* 4134 * The stack orientation is piggybacked with the cow argument. 4135 * Extract it into orient and mask the cow argument so that we 4136 * don't pass it around further. 4137 */ 4138 orient = cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP); 4139 KASSERT(orient != 0, ("No stack grow direction")); 4140 KASSERT(orient != (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP), 4141 ("bi-dir stack")); 4142 4143 if (addrbos < vm_map_min(map) || 4144 addrbos + max_ssize > vm_map_max(map) || 4145 addrbos + max_ssize <= addrbos) 4146 return (KERN_INVALID_ADDRESS); 4147 sgp = (curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 ? 0 : 4148 (vm_size_t)stack_guard_page * PAGE_SIZE; 4149 if (sgp >= max_ssize) 4150 return (KERN_INVALID_ARGUMENT); 4151 4152 init_ssize = growsize; 4153 if (max_ssize < init_ssize + sgp) 4154 init_ssize = max_ssize - sgp; 4155 4156 /* If addr is already mapped, no go */ 4157 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) 4158 return (KERN_NO_SPACE); 4159 4160 /* 4161 * If we can't accommodate max_ssize in the current mapping, no go. 4162 */ 4163 if (prev_entry->next->start < addrbos + max_ssize) 4164 return (KERN_NO_SPACE); 4165 4166 /* 4167 * We initially map a stack of only init_ssize. We will grow as 4168 * needed later. Depending on the orientation of the stack (i.e. 4169 * the grow direction) we either map at the top of the range, the 4170 * bottom of the range or in the middle. 4171 * 4172 * Note: we would normally expect prot and max to be VM_PROT_ALL, 4173 * and cow to be 0. Possibly we should eliminate these as input 4174 * parameters, and just pass these values here in the insert call. 4175 */ 4176 if (orient == MAP_STACK_GROWS_DOWN) { 4177 bot = addrbos + max_ssize - init_ssize; 4178 top = bot + init_ssize; 4179 gap_bot = addrbos; 4180 gap_top = bot; 4181 } else /* if (orient == MAP_STACK_GROWS_UP) */ { 4182 bot = addrbos; 4183 top = bot + init_ssize; 4184 gap_bot = top; 4185 gap_top = addrbos + max_ssize; 4186 } 4187 rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow); 4188 if (rv != KERN_SUCCESS) 4189 return (rv); 4190 new_entry = prev_entry->next; 4191 KASSERT(new_entry->end == top || new_entry->start == bot, 4192 ("Bad entry start/end for new stack entry")); 4193 KASSERT((orient & MAP_STACK_GROWS_DOWN) == 0 || 4194 (new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0, 4195 ("new entry lacks MAP_ENTRY_GROWS_DOWN")); 4196 KASSERT((orient & MAP_STACK_GROWS_UP) == 0 || 4197 (new_entry->eflags & MAP_ENTRY_GROWS_UP) != 0, 4198 ("new entry lacks MAP_ENTRY_GROWS_UP")); 4199 if (gap_bot == gap_top) 4200 return (KERN_SUCCESS); 4201 rv = vm_map_insert(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE, 4202 VM_PROT_NONE, MAP_CREATE_GUARD | (orient == MAP_STACK_GROWS_DOWN ? 4203 MAP_CREATE_STACK_GAP_DN : MAP_CREATE_STACK_GAP_UP)); 4204 if (rv == KERN_SUCCESS) { 4205 /* 4206 * Gap can never successfully handle a fault, so 4207 * read-ahead logic is never used for it. Re-use 4208 * next_read of the gap entry to store 4209 * stack_guard_page for vm_map_growstack(). 4210 */ 4211 if (orient == MAP_STACK_GROWS_DOWN) 4212 new_entry->prev->next_read = sgp; 4213 else 4214 new_entry->next->next_read = sgp; 4215 } else { 4216 (void)vm_map_delete(map, bot, top); 4217 } 4218 return (rv); 4219 } 4220 4221 /* 4222 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we 4223 * successfully grow the stack. 4224 */ 4225 static int 4226 vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry) 4227 { 4228 vm_map_entry_t stack_entry; 4229 struct proc *p; 4230 struct vmspace *vm; 4231 struct ucred *cred; 4232 vm_offset_t gap_end, gap_start, grow_start; 4233 vm_size_t grow_amount, guard, max_grow; 4234 rlim_t lmemlim, stacklim, vmemlim; 4235 int rv, rv1; 4236 bool gap_deleted, grow_down, is_procstack; 4237 #ifdef notyet 4238 uint64_t limit; 4239 #endif 4240 #ifdef RACCT 4241 int error; 4242 #endif 4243 4244 p = curproc; 4245 vm = p->p_vmspace; 4246 4247 /* 4248 * Disallow stack growth when the access is performed by a 4249 * debugger or AIO daemon. The reason is that the wrong 4250 * resource limits are applied. 4251 */ 4252 if (p != initproc && (map != &p->p_vmspace->vm_map || 4253 p->p_textvp == NULL)) 4254 return (KERN_FAILURE); 4255 4256 MPASS(!map->system_map); 4257 4258 lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK); 4259 stacklim = lim_cur(curthread, RLIMIT_STACK); 4260 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4261 retry: 4262 /* If addr is not in a hole for a stack grow area, no need to grow. */ 4263 if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry)) 4264 return (KERN_FAILURE); 4265 if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0) 4266 return (KERN_SUCCESS); 4267 if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_DN) != 0) { 4268 stack_entry = gap_entry->next; 4269 if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 || 4270 stack_entry->start != gap_entry->end) 4271 return (KERN_FAILURE); 4272 grow_amount = round_page(stack_entry->start - addr); 4273 grow_down = true; 4274 } else if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_UP) != 0) { 4275 stack_entry = gap_entry->prev; 4276 if ((stack_entry->eflags & MAP_ENTRY_GROWS_UP) == 0 || 4277 stack_entry->end != gap_entry->start) 4278 return (KERN_FAILURE); 4279 grow_amount = round_page(addr + 1 - stack_entry->end); 4280 grow_down = false; 4281 } else { 4282 return (KERN_FAILURE); 4283 } 4284 guard = (curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 ? 0 : 4285 gap_entry->next_read; 4286 max_grow = gap_entry->end - gap_entry->start; 4287 if (guard > max_grow) 4288 return (KERN_NO_SPACE); 4289 max_grow -= guard; 4290 if (grow_amount > max_grow) 4291 return (KERN_NO_SPACE); 4292 4293 /* 4294 * If this is the main process stack, see if we're over the stack 4295 * limit. 4296 */ 4297 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr && 4298 addr < (vm_offset_t)p->p_sysent->sv_usrstack; 4299 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) 4300 return (KERN_NO_SPACE); 4301 4302 #ifdef RACCT 4303 if (racct_enable) { 4304 PROC_LOCK(p); 4305 if (is_procstack && racct_set(p, RACCT_STACK, 4306 ctob(vm->vm_ssize) + grow_amount)) { 4307 PROC_UNLOCK(p); 4308 return (KERN_NO_SPACE); 4309 } 4310 PROC_UNLOCK(p); 4311 } 4312 #endif 4313 4314 grow_amount = roundup(grow_amount, sgrowsiz); 4315 if (grow_amount > max_grow) 4316 grow_amount = max_grow; 4317 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) { 4318 grow_amount = trunc_page((vm_size_t)stacklim) - 4319 ctob(vm->vm_ssize); 4320 } 4321 4322 #ifdef notyet 4323 PROC_LOCK(p); 4324 limit = racct_get_available(p, RACCT_STACK); 4325 PROC_UNLOCK(p); 4326 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit)) 4327 grow_amount = limit - ctob(vm->vm_ssize); 4328 #endif 4329 4330 if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) { 4331 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) { 4332 rv = KERN_NO_SPACE; 4333 goto out; 4334 } 4335 #ifdef RACCT 4336 if (racct_enable) { 4337 PROC_LOCK(p); 4338 if (racct_set(p, RACCT_MEMLOCK, 4339 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) { 4340 PROC_UNLOCK(p); 4341 rv = KERN_NO_SPACE; 4342 goto out; 4343 } 4344 PROC_UNLOCK(p); 4345 } 4346 #endif 4347 } 4348 4349 /* If we would blow our VMEM resource limit, no go */ 4350 if (map->size + grow_amount > vmemlim) { 4351 rv = KERN_NO_SPACE; 4352 goto out; 4353 } 4354 #ifdef RACCT 4355 if (racct_enable) { 4356 PROC_LOCK(p); 4357 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) { 4358 PROC_UNLOCK(p); 4359 rv = KERN_NO_SPACE; 4360 goto out; 4361 } 4362 PROC_UNLOCK(p); 4363 } 4364 #endif 4365 4366 if (vm_map_lock_upgrade(map)) { 4367 gap_entry = NULL; 4368 vm_map_lock_read(map); 4369 goto retry; 4370 } 4371 4372 if (grow_down) { 4373 grow_start = gap_entry->end - grow_amount; 4374 if (gap_entry->start + grow_amount == gap_entry->end) { 4375 gap_start = gap_entry->start; 4376 gap_end = gap_entry->end; 4377 vm_map_entry_delete(map, gap_entry); 4378 gap_deleted = true; 4379 } else { 4380 MPASS(gap_entry->start < gap_entry->end - grow_amount); 4381 vm_map_entry_resize(map, gap_entry, -grow_amount); 4382 gap_deleted = false; 4383 } 4384 rv = vm_map_insert(map, NULL, 0, grow_start, 4385 grow_start + grow_amount, 4386 stack_entry->protection, stack_entry->max_protection, 4387 MAP_STACK_GROWS_DOWN); 4388 if (rv != KERN_SUCCESS) { 4389 if (gap_deleted) { 4390 rv1 = vm_map_insert(map, NULL, 0, gap_start, 4391 gap_end, VM_PROT_NONE, VM_PROT_NONE, 4392 MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP_DN); 4393 MPASS(rv1 == KERN_SUCCESS); 4394 } else 4395 vm_map_entry_resize(map, gap_entry, 4396 grow_amount); 4397 } 4398 } else { 4399 grow_start = stack_entry->end; 4400 cred = stack_entry->cred; 4401 if (cred == NULL && stack_entry->object.vm_object != NULL) 4402 cred = stack_entry->object.vm_object->cred; 4403 if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred)) 4404 rv = KERN_NO_SPACE; 4405 /* Grow the underlying object if applicable. */ 4406 else if (stack_entry->object.vm_object == NULL || 4407 vm_object_coalesce(stack_entry->object.vm_object, 4408 stack_entry->offset, 4409 (vm_size_t)(stack_entry->end - stack_entry->start), 4410 grow_amount, cred != NULL)) { 4411 if (gap_entry->start + grow_amount == gap_entry->end) { 4412 vm_map_entry_delete(map, gap_entry); 4413 vm_map_entry_resize(map, stack_entry, 4414 grow_amount); 4415 } else { 4416 gap_entry->start += grow_amount; 4417 stack_entry->end += grow_amount; 4418 } 4419 map->size += grow_amount; 4420 rv = KERN_SUCCESS; 4421 } else 4422 rv = KERN_FAILURE; 4423 } 4424 if (rv == KERN_SUCCESS && is_procstack) 4425 vm->vm_ssize += btoc(grow_amount); 4426 4427 /* 4428 * Heed the MAP_WIREFUTURE flag if it was set for this process. 4429 */ 4430 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) { 4431 rv = vm_map_wire_locked(map, grow_start, 4432 grow_start + grow_amount, 4433 VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); 4434 } 4435 vm_map_lock_downgrade(map); 4436 4437 out: 4438 #ifdef RACCT 4439 if (racct_enable && rv != KERN_SUCCESS) { 4440 PROC_LOCK(p); 4441 error = racct_set(p, RACCT_VMEM, map->size); 4442 KASSERT(error == 0, ("decreasing RACCT_VMEM failed")); 4443 if (!old_mlock) { 4444 error = racct_set(p, RACCT_MEMLOCK, 4445 ptoa(pmap_wired_count(map->pmap))); 4446 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed")); 4447 } 4448 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize)); 4449 KASSERT(error == 0, ("decreasing RACCT_STACK failed")); 4450 PROC_UNLOCK(p); 4451 } 4452 #endif 4453 4454 return (rv); 4455 } 4456 4457 /* 4458 * Unshare the specified VM space for exec. If other processes are 4459 * mapped to it, then create a new one. The new vmspace is null. 4460 */ 4461 int 4462 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser) 4463 { 4464 struct vmspace *oldvmspace = p->p_vmspace; 4465 struct vmspace *newvmspace; 4466 4467 KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0, 4468 ("vmspace_exec recursed")); 4469 newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit); 4470 if (newvmspace == NULL) 4471 return (ENOMEM); 4472 newvmspace->vm_swrss = oldvmspace->vm_swrss; 4473 /* 4474 * This code is written like this for prototype purposes. The 4475 * goal is to avoid running down the vmspace here, but let the 4476 * other process's that are still using the vmspace to finally 4477 * run it down. Even though there is little or no chance of blocking 4478 * here, it is a good idea to keep this form for future mods. 4479 */ 4480 PROC_VMSPACE_LOCK(p); 4481 p->p_vmspace = newvmspace; 4482 PROC_VMSPACE_UNLOCK(p); 4483 if (p == curthread->td_proc) 4484 pmap_activate(curthread); 4485 curthread->td_pflags |= TDP_EXECVMSPC; 4486 return (0); 4487 } 4488 4489 /* 4490 * Unshare the specified VM space for forcing COW. This 4491 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 4492 */ 4493 int 4494 vmspace_unshare(struct proc *p) 4495 { 4496 struct vmspace *oldvmspace = p->p_vmspace; 4497 struct vmspace *newvmspace; 4498 vm_ooffset_t fork_charge; 4499 4500 if (oldvmspace->vm_refcnt == 1) 4501 return (0); 4502 fork_charge = 0; 4503 newvmspace = vmspace_fork(oldvmspace, &fork_charge); 4504 if (newvmspace == NULL) 4505 return (ENOMEM); 4506 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) { 4507 vmspace_free(newvmspace); 4508 return (ENOMEM); 4509 } 4510 PROC_VMSPACE_LOCK(p); 4511 p->p_vmspace = newvmspace; 4512 PROC_VMSPACE_UNLOCK(p); 4513 if (p == curthread->td_proc) 4514 pmap_activate(curthread); 4515 vmspace_free(oldvmspace); 4516 return (0); 4517 } 4518 4519 /* 4520 * vm_map_lookup: 4521 * 4522 * Finds the VM object, offset, and 4523 * protection for a given virtual address in the 4524 * specified map, assuming a page fault of the 4525 * type specified. 4526 * 4527 * Leaves the map in question locked for read; return 4528 * values are guaranteed until a vm_map_lookup_done 4529 * call is performed. Note that the map argument 4530 * is in/out; the returned map must be used in 4531 * the call to vm_map_lookup_done. 4532 * 4533 * A handle (out_entry) is returned for use in 4534 * vm_map_lookup_done, to make that fast. 4535 * 4536 * If a lookup is requested with "write protection" 4537 * specified, the map may be changed to perform virtual 4538 * copying operations, although the data referenced will 4539 * remain the same. 4540 */ 4541 int 4542 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4543 vm_offset_t vaddr, 4544 vm_prot_t fault_typea, 4545 vm_map_entry_t *out_entry, /* OUT */ 4546 vm_object_t *object, /* OUT */ 4547 vm_pindex_t *pindex, /* OUT */ 4548 vm_prot_t *out_prot, /* OUT */ 4549 boolean_t *wired) /* OUT */ 4550 { 4551 vm_map_entry_t entry; 4552 vm_map_t map = *var_map; 4553 vm_prot_t prot; 4554 vm_prot_t fault_type = fault_typea; 4555 vm_object_t eobject; 4556 vm_size_t size; 4557 struct ucred *cred; 4558 4559 RetryLookup: 4560 4561 vm_map_lock_read(map); 4562 4563 RetryLookupLocked: 4564 /* 4565 * Lookup the faulting address. 4566 */ 4567 if (!vm_map_lookup_entry(map, vaddr, out_entry)) { 4568 vm_map_unlock_read(map); 4569 return (KERN_INVALID_ADDRESS); 4570 } 4571 4572 entry = *out_entry; 4573 4574 /* 4575 * Handle submaps. 4576 */ 4577 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 4578 vm_map_t old_map = map; 4579 4580 *var_map = map = entry->object.sub_map; 4581 vm_map_unlock_read(old_map); 4582 goto RetryLookup; 4583 } 4584 4585 /* 4586 * Check whether this task is allowed to have this page. 4587 */ 4588 prot = entry->protection; 4589 if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) { 4590 fault_typea &= ~VM_PROT_FAULT_LOOKUP; 4591 if (prot == VM_PROT_NONE && map != kernel_map && 4592 (entry->eflags & MAP_ENTRY_GUARD) != 0 && 4593 (entry->eflags & (MAP_ENTRY_STACK_GAP_DN | 4594 MAP_ENTRY_STACK_GAP_UP)) != 0 && 4595 vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS) 4596 goto RetryLookupLocked; 4597 } 4598 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 4599 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) { 4600 vm_map_unlock_read(map); 4601 return (KERN_PROTECTION_FAILURE); 4602 } 4603 KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags & 4604 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) != 4605 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY), 4606 ("entry %p flags %x", entry, entry->eflags)); 4607 if ((fault_typea & VM_PROT_COPY) != 0 && 4608 (entry->max_protection & VM_PROT_WRITE) == 0 && 4609 (entry->eflags & MAP_ENTRY_COW) == 0) { 4610 vm_map_unlock_read(map); 4611 return (KERN_PROTECTION_FAILURE); 4612 } 4613 4614 /* 4615 * If this page is not pageable, we have to get it for all possible 4616 * accesses. 4617 */ 4618 *wired = (entry->wired_count != 0); 4619 if (*wired) 4620 fault_type = entry->protection; 4621 size = entry->end - entry->start; 4622 /* 4623 * If the entry was copy-on-write, we either ... 4624 */ 4625 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4626 /* 4627 * If we want to write the page, we may as well handle that 4628 * now since we've got the map locked. 4629 * 4630 * If we don't need to write the page, we just demote the 4631 * permissions allowed. 4632 */ 4633 if ((fault_type & VM_PROT_WRITE) != 0 || 4634 (fault_typea & VM_PROT_COPY) != 0) { 4635 /* 4636 * Make a new object, and place it in the object 4637 * chain. Note that no new references have appeared 4638 * -- one just moved from the map to the new 4639 * object. 4640 */ 4641 if (vm_map_lock_upgrade(map)) 4642 goto RetryLookup; 4643 4644 if (entry->cred == NULL) { 4645 /* 4646 * The debugger owner is charged for 4647 * the memory. 4648 */ 4649 cred = curthread->td_ucred; 4650 crhold(cred); 4651 if (!swap_reserve_by_cred(size, cred)) { 4652 crfree(cred); 4653 vm_map_unlock(map); 4654 return (KERN_RESOURCE_SHORTAGE); 4655 } 4656 entry->cred = cred; 4657 } 4658 vm_object_shadow(&entry->object.vm_object, 4659 &entry->offset, size); 4660 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 4661 eobject = entry->object.vm_object; 4662 if (eobject->cred != NULL) { 4663 /* 4664 * The object was not shadowed. 4665 */ 4666 swap_release_by_cred(size, entry->cred); 4667 crfree(entry->cred); 4668 entry->cred = NULL; 4669 } else if (entry->cred != NULL) { 4670 VM_OBJECT_WLOCK(eobject); 4671 eobject->cred = entry->cred; 4672 eobject->charge = size; 4673 VM_OBJECT_WUNLOCK(eobject); 4674 entry->cred = NULL; 4675 } 4676 4677 vm_map_lock_downgrade(map); 4678 } else { 4679 /* 4680 * We're attempting to read a copy-on-write page -- 4681 * don't allow writes. 4682 */ 4683 prot &= ~VM_PROT_WRITE; 4684 } 4685 } 4686 4687 /* 4688 * Create an object if necessary. 4689 */ 4690 if (entry->object.vm_object == NULL && 4691 !map->system_map) { 4692 if (vm_map_lock_upgrade(map)) 4693 goto RetryLookup; 4694 entry->object.vm_object = vm_object_allocate(OBJT_DEFAULT, 4695 atop(size)); 4696 entry->offset = 0; 4697 if (entry->cred != NULL) { 4698 VM_OBJECT_WLOCK(entry->object.vm_object); 4699 entry->object.vm_object->cred = entry->cred; 4700 entry->object.vm_object->charge = size; 4701 VM_OBJECT_WUNLOCK(entry->object.vm_object); 4702 entry->cred = NULL; 4703 } 4704 vm_map_lock_downgrade(map); 4705 } 4706 4707 /* 4708 * Return the object/offset from this entry. If the entry was 4709 * copy-on-write or empty, it has been fixed up. 4710 */ 4711 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 4712 *object = entry->object.vm_object; 4713 4714 *out_prot = prot; 4715 return (KERN_SUCCESS); 4716 } 4717 4718 /* 4719 * vm_map_lookup_locked: 4720 * 4721 * Lookup the faulting address. A version of vm_map_lookup that returns 4722 * KERN_FAILURE instead of blocking on map lock or memory allocation. 4723 */ 4724 int 4725 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */ 4726 vm_offset_t vaddr, 4727 vm_prot_t fault_typea, 4728 vm_map_entry_t *out_entry, /* OUT */ 4729 vm_object_t *object, /* OUT */ 4730 vm_pindex_t *pindex, /* OUT */ 4731 vm_prot_t *out_prot, /* OUT */ 4732 boolean_t *wired) /* OUT */ 4733 { 4734 vm_map_entry_t entry; 4735 vm_map_t map = *var_map; 4736 vm_prot_t prot; 4737 vm_prot_t fault_type = fault_typea; 4738 4739 /* 4740 * Lookup the faulting address. 4741 */ 4742 if (!vm_map_lookup_entry(map, vaddr, out_entry)) 4743 return (KERN_INVALID_ADDRESS); 4744 4745 entry = *out_entry; 4746 4747 /* 4748 * Fail if the entry refers to a submap. 4749 */ 4750 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) 4751 return (KERN_FAILURE); 4752 4753 /* 4754 * Check whether this task is allowed to have this page. 4755 */ 4756 prot = entry->protection; 4757 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 4758 if ((fault_type & prot) != fault_type) 4759 return (KERN_PROTECTION_FAILURE); 4760 4761 /* 4762 * If this page is not pageable, we have to get it for all possible 4763 * accesses. 4764 */ 4765 *wired = (entry->wired_count != 0); 4766 if (*wired) 4767 fault_type = entry->protection; 4768 4769 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4770 /* 4771 * Fail if the entry was copy-on-write for a write fault. 4772 */ 4773 if (fault_type & VM_PROT_WRITE) 4774 return (KERN_FAILURE); 4775 /* 4776 * We're attempting to read a copy-on-write page -- 4777 * don't allow writes. 4778 */ 4779 prot &= ~VM_PROT_WRITE; 4780 } 4781 4782 /* 4783 * Fail if an object should be created. 4784 */ 4785 if (entry->object.vm_object == NULL && !map->system_map) 4786 return (KERN_FAILURE); 4787 4788 /* 4789 * Return the object/offset from this entry. If the entry was 4790 * copy-on-write or empty, it has been fixed up. 4791 */ 4792 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 4793 *object = entry->object.vm_object; 4794 4795 *out_prot = prot; 4796 return (KERN_SUCCESS); 4797 } 4798 4799 /* 4800 * vm_map_lookup_done: 4801 * 4802 * Releases locks acquired by a vm_map_lookup 4803 * (according to the handle returned by that lookup). 4804 */ 4805 void 4806 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry) 4807 { 4808 /* 4809 * Unlock the main-level map 4810 */ 4811 vm_map_unlock_read(map); 4812 } 4813 4814 vm_offset_t 4815 vm_map_max_KBI(const struct vm_map *map) 4816 { 4817 4818 return (vm_map_max(map)); 4819 } 4820 4821 vm_offset_t 4822 vm_map_min_KBI(const struct vm_map *map) 4823 { 4824 4825 return (vm_map_min(map)); 4826 } 4827 4828 pmap_t 4829 vm_map_pmap_KBI(vm_map_t map) 4830 { 4831 4832 return (map->pmap); 4833 } 4834 4835 #include "opt_ddb.h" 4836 #ifdef DDB 4837 #include <sys/kernel.h> 4838 4839 #include <ddb/ddb.h> 4840 4841 static void 4842 vm_map_print(vm_map_t map) 4843 { 4844 vm_map_entry_t entry, prev; 4845 4846 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 4847 (void *)map, 4848 (void *)map->pmap, map->nentries, map->timestamp); 4849 4850 db_indent += 2; 4851 for (prev = &map->header; (entry = prev->next) != &map->header; 4852 prev = entry) { 4853 db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n", 4854 (void *)entry, (void *)entry->start, (void *)entry->end, 4855 entry->eflags); 4856 { 4857 static char *inheritance_name[4] = 4858 {"share", "copy", "none", "donate_copy"}; 4859 4860 db_iprintf(" prot=%x/%x/%s", 4861 entry->protection, 4862 entry->max_protection, 4863 inheritance_name[(int)(unsigned char) 4864 entry->inheritance]); 4865 if (entry->wired_count != 0) 4866 db_printf(", wired"); 4867 } 4868 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 4869 db_printf(", share=%p, offset=0x%jx\n", 4870 (void *)entry->object.sub_map, 4871 (uintmax_t)entry->offset); 4872 if (prev == &map->header || 4873 prev->object.sub_map != 4874 entry->object.sub_map) { 4875 db_indent += 2; 4876 vm_map_print((vm_map_t)entry->object.sub_map); 4877 db_indent -= 2; 4878 } 4879 } else { 4880 if (entry->cred != NULL) 4881 db_printf(", ruid %d", entry->cred->cr_ruid); 4882 db_printf(", object=%p, offset=0x%jx", 4883 (void *)entry->object.vm_object, 4884 (uintmax_t)entry->offset); 4885 if (entry->object.vm_object && entry->object.vm_object->cred) 4886 db_printf(", obj ruid %d charge %jx", 4887 entry->object.vm_object->cred->cr_ruid, 4888 (uintmax_t)entry->object.vm_object->charge); 4889 if (entry->eflags & MAP_ENTRY_COW) 4890 db_printf(", copy (%s)", 4891 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4892 db_printf("\n"); 4893 4894 if (prev == &map->header || 4895 prev->object.vm_object != 4896 entry->object.vm_object) { 4897 db_indent += 2; 4898 vm_object_print((db_expr_t)(intptr_t) 4899 entry->object.vm_object, 4900 0, 0, (char *)0); 4901 db_indent -= 2; 4902 } 4903 } 4904 } 4905 db_indent -= 2; 4906 } 4907 4908 DB_SHOW_COMMAND(map, map) 4909 { 4910 4911 if (!have_addr) { 4912 db_printf("usage: show map <addr>\n"); 4913 return; 4914 } 4915 vm_map_print((vm_map_t)addr); 4916 } 4917 4918 DB_SHOW_COMMAND(procvm, procvm) 4919 { 4920 struct proc *p; 4921 4922 if (have_addr) { 4923 p = db_lookup_proc(addr); 4924 } else { 4925 p = curproc; 4926 } 4927 4928 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 4929 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 4930 (void *)vmspace_pmap(p->p_vmspace)); 4931 4932 vm_map_print((vm_map_t)&p->p_vmspace->vm_map); 4933 } 4934 4935 #endif /* DDB */ 4936