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/elf.h> 73 #include <sys/kernel.h> 74 #include <sys/ktr.h> 75 #include <sys/lock.h> 76 #include <sys/mutex.h> 77 #include <sys/proc.h> 78 #include <sys/vmmeter.h> 79 #include <sys/mman.h> 80 #include <sys/vnode.h> 81 #include <sys/racct.h> 82 #include <sys/resourcevar.h> 83 #include <sys/rwlock.h> 84 #include <sys/file.h> 85 #include <sys/sysctl.h> 86 #include <sys/sysent.h> 87 #include <sys/shm.h> 88 89 #include <vm/vm.h> 90 #include <vm/vm_param.h> 91 #include <vm/pmap.h> 92 #include <vm/vm_map.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_object.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vm_kern.h> 98 #include <vm/vm_extern.h> 99 #include <vm/vnode_pager.h> 100 #include <vm/swap_pager.h> 101 #include <vm/uma.h> 102 103 /* 104 * Virtual memory maps provide for the mapping, protection, 105 * and sharing of virtual memory objects. In addition, 106 * this module provides for an efficient virtual copy of 107 * memory from one map to another. 108 * 109 * Synchronization is required prior to most operations. 110 * 111 * Maps consist of an ordered doubly-linked list of simple 112 * entries; a self-adjusting binary search tree of these 113 * entries is used to speed up lookups. 114 * 115 * Since portions of maps are specified by start/end addresses, 116 * which may not align with existing map entries, all 117 * routines merely "clip" entries to these start/end values. 118 * [That is, an entry is split into two, bordering at a 119 * start or end value.] Note that these clippings may not 120 * always be necessary (as the two resulting entries are then 121 * not changed); however, the clipping is done for convenience. 122 * 123 * As mentioned above, virtual copy operations are performed 124 * by copying VM object references from one map to 125 * another, and then marking both regions as copy-on-write. 126 */ 127 128 static struct mtx map_sleep_mtx; 129 static uma_zone_t mapentzone; 130 static uma_zone_t kmapentzone; 131 static uma_zone_t mapzone; 132 static uma_zone_t vmspace_zone; 133 static int vmspace_zinit(void *mem, int size, int flags); 134 static int vm_map_zinit(void *mem, int ize, int flags); 135 static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, 136 vm_offset_t max); 137 static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map); 138 static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry); 139 static void vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry); 140 static int vm_map_growstack(vm_map_t map, vm_offset_t addr, 141 vm_map_entry_t gap_entry); 142 static void vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 143 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags); 144 #ifdef INVARIANTS 145 static void vm_map_zdtor(void *mem, int size, void *arg); 146 static void vmspace_zdtor(void *mem, int size, void *arg); 147 #endif 148 static int vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, 149 vm_size_t max_ssize, vm_size_t growsize, vm_prot_t prot, vm_prot_t max, 150 int cow); 151 static void vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 152 vm_offset_t failed_addr); 153 154 #define ENTRY_CHARGED(e) ((e)->cred != NULL || \ 155 ((e)->object.vm_object != NULL && (e)->object.vm_object->cred != NULL && \ 156 !((e)->eflags & MAP_ENTRY_NEEDS_COPY))) 157 158 /* 159 * PROC_VMSPACE_{UN,}LOCK() can be a noop as long as vmspaces are type 160 * stable. 161 */ 162 #define PROC_VMSPACE_LOCK(p) do { } while (0) 163 #define PROC_VMSPACE_UNLOCK(p) do { } while (0) 164 165 /* 166 * VM_MAP_RANGE_CHECK: [ internal use only ] 167 * 168 * Asserts that the starting and ending region 169 * addresses fall within the valid range of the map. 170 */ 171 #define VM_MAP_RANGE_CHECK(map, start, end) \ 172 { \ 173 if (start < vm_map_min(map)) \ 174 start = vm_map_min(map); \ 175 if (end > vm_map_max(map)) \ 176 end = vm_map_max(map); \ 177 if (start > end) \ 178 start = end; \ 179 } 180 181 /* 182 * vm_map_startup: 183 * 184 * Initialize the vm_map module. Must be called before 185 * any other vm_map routines. 186 * 187 * Map and entry structures are allocated from the general 188 * purpose memory pool with some exceptions: 189 * 190 * - The kernel map and kmem submap are allocated statically. 191 * - Kernel map entries are allocated out of a static pool. 192 * 193 * These restrictions are necessary since malloc() uses the 194 * maps and requires map entries. 195 */ 196 197 void 198 vm_map_startup(void) 199 { 200 mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF); 201 mapzone = uma_zcreate("MAP", sizeof(struct vm_map), NULL, 202 #ifdef INVARIANTS 203 vm_map_zdtor, 204 #else 205 NULL, 206 #endif 207 vm_map_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 208 uma_prealloc(mapzone, MAX_KMAP); 209 kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry), 210 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 211 UMA_ZONE_MTXCLASS | UMA_ZONE_VM); 212 mapentzone = uma_zcreate("MAP ENTRY", sizeof(struct vm_map_entry), 213 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 214 vmspace_zone = uma_zcreate("VMSPACE", sizeof(struct vmspace), NULL, 215 #ifdef INVARIANTS 216 vmspace_zdtor, 217 #else 218 NULL, 219 #endif 220 vmspace_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 221 } 222 223 static int 224 vmspace_zinit(void *mem, int size, int flags) 225 { 226 struct vmspace *vm; 227 228 vm = (struct vmspace *)mem; 229 230 vm->vm_map.pmap = NULL; 231 (void)vm_map_zinit(&vm->vm_map, sizeof(vm->vm_map), flags); 232 PMAP_LOCK_INIT(vmspace_pmap(vm)); 233 return (0); 234 } 235 236 static int 237 vm_map_zinit(void *mem, int size, int flags) 238 { 239 vm_map_t map; 240 241 map = (vm_map_t)mem; 242 memset(map, 0, sizeof(*map)); 243 mtx_init(&map->system_mtx, "vm map (system)", NULL, MTX_DEF | MTX_DUPOK); 244 sx_init(&map->lock, "vm map (user)"); 245 return (0); 246 } 247 248 #ifdef INVARIANTS 249 static void 250 vmspace_zdtor(void *mem, int size, void *arg) 251 { 252 struct vmspace *vm; 253 254 vm = (struct vmspace *)mem; 255 256 vm_map_zdtor(&vm->vm_map, sizeof(vm->vm_map), arg); 257 } 258 static void 259 vm_map_zdtor(void *mem, int size, void *arg) 260 { 261 vm_map_t map; 262 263 map = (vm_map_t)mem; 264 KASSERT(map->nentries == 0, 265 ("map %p nentries == %d on free.", 266 map, map->nentries)); 267 KASSERT(map->size == 0, 268 ("map %p size == %lu on free.", 269 map, (unsigned long)map->size)); 270 } 271 #endif /* INVARIANTS */ 272 273 /* 274 * Allocate a vmspace structure, including a vm_map and pmap, 275 * and initialize those structures. The refcnt is set to 1. 276 * 277 * If 'pinit' is NULL then the embedded pmap is initialized via pmap_pinit(). 278 */ 279 struct vmspace * 280 vmspace_alloc(vm_offset_t min, vm_offset_t max, pmap_pinit_t pinit) 281 { 282 struct vmspace *vm; 283 284 vm = uma_zalloc(vmspace_zone, M_WAITOK); 285 KASSERT(vm->vm_map.pmap == NULL, ("vm_map.pmap must be NULL")); 286 if (!pinit(vmspace_pmap(vm))) { 287 uma_zfree(vmspace_zone, vm); 288 return (NULL); 289 } 290 CTR1(KTR_VM, "vmspace_alloc: %p", vm); 291 _vm_map_init(&vm->vm_map, vmspace_pmap(vm), min, max); 292 vm->vm_refcnt = 1; 293 vm->vm_shm = NULL; 294 vm->vm_swrss = 0; 295 vm->vm_tsize = 0; 296 vm->vm_dsize = 0; 297 vm->vm_ssize = 0; 298 vm->vm_taddr = 0; 299 vm->vm_daddr = 0; 300 vm->vm_maxsaddr = 0; 301 return (vm); 302 } 303 304 #ifdef RACCT 305 static void 306 vmspace_container_reset(struct proc *p) 307 { 308 309 PROC_LOCK(p); 310 racct_set(p, RACCT_DATA, 0); 311 racct_set(p, RACCT_STACK, 0); 312 racct_set(p, RACCT_RSS, 0); 313 racct_set(p, RACCT_MEMLOCK, 0); 314 racct_set(p, RACCT_VMEM, 0); 315 PROC_UNLOCK(p); 316 } 317 #endif 318 319 static inline void 320 vmspace_dofree(struct vmspace *vm) 321 { 322 323 CTR1(KTR_VM, "vmspace_free: %p", vm); 324 325 /* 326 * Make sure any SysV shm is freed, it might not have been in 327 * exit1(). 328 */ 329 shmexit(vm); 330 331 /* 332 * Lock the map, to wait out all other references to it. 333 * Delete all of the mappings and pages they hold, then call 334 * the pmap module to reclaim anything left. 335 */ 336 (void)vm_map_remove(&vm->vm_map, vm_map_min(&vm->vm_map), 337 vm_map_max(&vm->vm_map)); 338 339 pmap_release(vmspace_pmap(vm)); 340 vm->vm_map.pmap = NULL; 341 uma_zfree(vmspace_zone, vm); 342 } 343 344 void 345 vmspace_free(struct vmspace *vm) 346 { 347 348 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 349 "vmspace_free() called"); 350 351 if (vm->vm_refcnt == 0) 352 panic("vmspace_free: attempt to free already freed vmspace"); 353 354 if (atomic_fetchadd_int(&vm->vm_refcnt, -1) == 1) 355 vmspace_dofree(vm); 356 } 357 358 void 359 vmspace_exitfree(struct proc *p) 360 { 361 struct vmspace *vm; 362 363 PROC_VMSPACE_LOCK(p); 364 vm = p->p_vmspace; 365 p->p_vmspace = NULL; 366 PROC_VMSPACE_UNLOCK(p); 367 KASSERT(vm == &vmspace0, ("vmspace_exitfree: wrong vmspace")); 368 vmspace_free(vm); 369 } 370 371 void 372 vmspace_exit(struct thread *td) 373 { 374 int refcnt; 375 struct vmspace *vm; 376 struct proc *p; 377 378 /* 379 * Release user portion of address space. 380 * This releases references to vnodes, 381 * which could cause I/O if the file has been unlinked. 382 * Need to do this early enough that we can still sleep. 383 * 384 * The last exiting process to reach this point releases as 385 * much of the environment as it can. vmspace_dofree() is the 386 * slower fallback in case another process had a temporary 387 * reference to the vmspace. 388 */ 389 390 p = td->td_proc; 391 vm = p->p_vmspace; 392 atomic_add_int(&vmspace0.vm_refcnt, 1); 393 refcnt = vm->vm_refcnt; 394 do { 395 if (refcnt > 1 && p->p_vmspace != &vmspace0) { 396 /* Switch now since other proc might free vmspace */ 397 PROC_VMSPACE_LOCK(p); 398 p->p_vmspace = &vmspace0; 399 PROC_VMSPACE_UNLOCK(p); 400 pmap_activate(td); 401 } 402 } while (!atomic_fcmpset_int(&vm->vm_refcnt, &refcnt, refcnt - 1)); 403 if (refcnt == 1) { 404 if (p->p_vmspace != vm) { 405 /* vmspace not yet freed, switch back */ 406 PROC_VMSPACE_LOCK(p); 407 p->p_vmspace = vm; 408 PROC_VMSPACE_UNLOCK(p); 409 pmap_activate(td); 410 } 411 pmap_remove_pages(vmspace_pmap(vm)); 412 /* Switch now since this proc will free vmspace */ 413 PROC_VMSPACE_LOCK(p); 414 p->p_vmspace = &vmspace0; 415 PROC_VMSPACE_UNLOCK(p); 416 pmap_activate(td); 417 vmspace_dofree(vm); 418 } 419 #ifdef RACCT 420 if (racct_enable) 421 vmspace_container_reset(p); 422 #endif 423 } 424 425 /* Acquire reference to vmspace owned by another process. */ 426 427 struct vmspace * 428 vmspace_acquire_ref(struct proc *p) 429 { 430 struct vmspace *vm; 431 int refcnt; 432 433 PROC_VMSPACE_LOCK(p); 434 vm = p->p_vmspace; 435 if (vm == NULL) { 436 PROC_VMSPACE_UNLOCK(p); 437 return (NULL); 438 } 439 refcnt = vm->vm_refcnt; 440 do { 441 if (refcnt <= 0) { /* Avoid 0->1 transition */ 442 PROC_VMSPACE_UNLOCK(p); 443 return (NULL); 444 } 445 } while (!atomic_fcmpset_int(&vm->vm_refcnt, &refcnt, refcnt + 1)); 446 if (vm != p->p_vmspace) { 447 PROC_VMSPACE_UNLOCK(p); 448 vmspace_free(vm); 449 return (NULL); 450 } 451 PROC_VMSPACE_UNLOCK(p); 452 return (vm); 453 } 454 455 /* 456 * Switch between vmspaces in an AIO kernel process. 457 * 458 * The new vmspace is either the vmspace of a user process obtained 459 * from an active AIO request or the initial vmspace of the AIO kernel 460 * process (when it is idling). Because user processes will block to 461 * drain any active AIO requests before proceeding in exit() or 462 * execve(), the reference count for vmspaces from AIO requests can 463 * never be 0. Similarly, AIO kernel processes hold an extra 464 * reference on their initial vmspace for the life of the process. As 465 * a result, the 'newvm' vmspace always has a non-zero reference 466 * count. This permits an additional reference on 'newvm' to be 467 * acquired via a simple atomic increment rather than the loop in 468 * vmspace_acquire_ref() above. 469 */ 470 void 471 vmspace_switch_aio(struct vmspace *newvm) 472 { 473 struct vmspace *oldvm; 474 475 /* XXX: Need some way to assert that this is an aio daemon. */ 476 477 KASSERT(newvm->vm_refcnt > 0, 478 ("vmspace_switch_aio: newvm unreferenced")); 479 480 oldvm = curproc->p_vmspace; 481 if (oldvm == newvm) 482 return; 483 484 /* 485 * Point to the new address space and refer to it. 486 */ 487 curproc->p_vmspace = newvm; 488 atomic_add_int(&newvm->vm_refcnt, 1); 489 490 /* Activate the new mapping. */ 491 pmap_activate(curthread); 492 493 vmspace_free(oldvm); 494 } 495 496 void 497 _vm_map_lock(vm_map_t map, const char *file, int line) 498 { 499 500 if (map->system_map) 501 mtx_lock_flags_(&map->system_mtx, 0, file, line); 502 else 503 sx_xlock_(&map->lock, file, line); 504 map->timestamp++; 505 } 506 507 void 508 vm_map_entry_set_vnode_text(vm_map_entry_t entry, bool add) 509 { 510 vm_object_t object, object1; 511 struct vnode *vp; 512 513 if ((entry->eflags & MAP_ENTRY_VN_EXEC) == 0) 514 return; 515 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 516 ("Submap with execs")); 517 object = entry->object.vm_object; 518 KASSERT(object != NULL, ("No object for text, entry %p", entry)); 519 VM_OBJECT_RLOCK(object); 520 while ((object1 = object->backing_object) != NULL) { 521 VM_OBJECT_RLOCK(object1); 522 VM_OBJECT_RUNLOCK(object); 523 object = object1; 524 } 525 526 vp = NULL; 527 if (object->type == OBJT_DEAD) { 528 /* 529 * For OBJT_DEAD objects, v_writecount was handled in 530 * vnode_pager_dealloc(). 531 */ 532 } else if (object->type == OBJT_VNODE) { 533 vp = object->handle; 534 } else if (object->type == OBJT_SWAP) { 535 KASSERT((object->flags & OBJ_TMPFS_NODE) != 0, 536 ("vm_map_entry_set_vnode_text: swap and !TMPFS " 537 "entry %p, object %p, add %d", entry, object, add)); 538 /* 539 * Tmpfs VREG node, which was reclaimed, has 540 * OBJ_TMPFS_NODE flag set, but not OBJ_TMPFS. In 541 * this case there is no v_writecount to adjust. 542 */ 543 if ((object->flags & OBJ_TMPFS) != 0) 544 vp = object->un_pager.swp.swp_tmpfs; 545 } else { 546 KASSERT(0, 547 ("vm_map_entry_set_vnode_text: wrong object type, " 548 "entry %p, object %p, add %d", entry, object, add)); 549 } 550 if (vp != NULL) { 551 if (add) { 552 VOP_SET_TEXT_CHECKED(vp); 553 VM_OBJECT_RUNLOCK(object); 554 } else { 555 vhold(vp); 556 VM_OBJECT_RUNLOCK(object); 557 vn_lock(vp, LK_SHARED | LK_RETRY); 558 VOP_UNSET_TEXT_CHECKED(vp); 559 VOP_UNLOCK(vp, 0); 560 vdrop(vp); 561 } 562 } else { 563 VM_OBJECT_RUNLOCK(object); 564 } 565 } 566 567 /* 568 * Use a different name for this vm_map_entry field when it's use 569 * is not consistent with its use as part of an ordered search tree. 570 */ 571 #define defer_next right 572 573 static void 574 vm_map_process_deferred(void) 575 { 576 struct thread *td; 577 vm_map_entry_t entry, next; 578 vm_object_t object; 579 580 td = curthread; 581 entry = td->td_map_def_user; 582 td->td_map_def_user = NULL; 583 while (entry != NULL) { 584 next = entry->defer_next; 585 MPASS((entry->eflags & (MAP_ENTRY_WRITECNT | 586 MAP_ENTRY_VN_EXEC)) != (MAP_ENTRY_WRITECNT | 587 MAP_ENTRY_VN_EXEC)); 588 if ((entry->eflags & MAP_ENTRY_WRITECNT) != 0) { 589 /* 590 * Decrement the object's writemappings and 591 * possibly the vnode's v_writecount. 592 */ 593 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 594 ("Submap with writecount")); 595 object = entry->object.vm_object; 596 KASSERT(object != NULL, ("No object for writecount")); 597 vm_pager_release_writecount(object, entry->start, 598 entry->end); 599 } 600 vm_map_entry_set_vnode_text(entry, false); 601 vm_map_entry_deallocate(entry, FALSE); 602 entry = next; 603 } 604 } 605 606 #ifdef INVARIANTS 607 static void 608 _vm_map_assert_locked(vm_map_t map, const char *file, int line) 609 { 610 611 if (map->system_map) 612 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 613 else 614 sx_assert_(&map->lock, SA_XLOCKED, file, line); 615 } 616 617 #define VM_MAP_ASSERT_LOCKED(map) \ 618 _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE) 619 620 enum { VMMAP_CHECK_NONE, VMMAP_CHECK_UNLOCK, VMMAP_CHECK_ALL }; 621 #ifdef DIAGNOSTIC 622 static int enable_vmmap_check = VMMAP_CHECK_UNLOCK; 623 #else 624 static int enable_vmmap_check = VMMAP_CHECK_NONE; 625 #endif 626 SYSCTL_INT(_debug, OID_AUTO, vmmap_check, CTLFLAG_RWTUN, 627 &enable_vmmap_check, 0, "Enable vm map consistency checking"); 628 629 static void _vm_map_assert_consistent(vm_map_t map, int check); 630 631 #define VM_MAP_ASSERT_CONSISTENT(map) \ 632 _vm_map_assert_consistent(map, VMMAP_CHECK_ALL) 633 #ifdef DIAGNOSTIC 634 #define VM_MAP_UNLOCK_CONSISTENT(map) do { \ 635 if (map->nupdates > map->nentries) { \ 636 _vm_map_assert_consistent(map, VMMAP_CHECK_UNLOCK); \ 637 map->nupdates = 0; \ 638 } \ 639 } while (0) 640 #else 641 #define VM_MAP_UNLOCK_CONSISTENT(map) 642 #endif 643 #else 644 #define VM_MAP_ASSERT_LOCKED(map) 645 #define VM_MAP_ASSERT_CONSISTENT(map) 646 #define VM_MAP_UNLOCK_CONSISTENT(map) 647 #endif /* INVARIANTS */ 648 649 void 650 _vm_map_unlock(vm_map_t map, const char *file, int line) 651 { 652 653 VM_MAP_UNLOCK_CONSISTENT(map); 654 if (map->system_map) 655 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 656 else { 657 sx_xunlock_(&map->lock, file, line); 658 vm_map_process_deferred(); 659 } 660 } 661 662 void 663 _vm_map_lock_read(vm_map_t map, const char *file, int line) 664 { 665 666 if (map->system_map) 667 mtx_lock_flags_(&map->system_mtx, 0, file, line); 668 else 669 sx_slock_(&map->lock, file, line); 670 } 671 672 void 673 _vm_map_unlock_read(vm_map_t map, const char *file, int line) 674 { 675 676 if (map->system_map) 677 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 678 else { 679 sx_sunlock_(&map->lock, file, line); 680 vm_map_process_deferred(); 681 } 682 } 683 684 int 685 _vm_map_trylock(vm_map_t map, const char *file, int line) 686 { 687 int error; 688 689 error = map->system_map ? 690 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 691 !sx_try_xlock_(&map->lock, file, line); 692 if (error == 0) 693 map->timestamp++; 694 return (error == 0); 695 } 696 697 int 698 _vm_map_trylock_read(vm_map_t map, const char *file, int line) 699 { 700 int error; 701 702 error = map->system_map ? 703 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 704 !sx_try_slock_(&map->lock, file, line); 705 return (error == 0); 706 } 707 708 /* 709 * _vm_map_lock_upgrade: [ internal use only ] 710 * 711 * Tries to upgrade a read (shared) lock on the specified map to a write 712 * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a 713 * non-zero value if the upgrade fails. If the upgrade fails, the map is 714 * returned without a read or write lock held. 715 * 716 * Requires that the map be read locked. 717 */ 718 int 719 _vm_map_lock_upgrade(vm_map_t map, const char *file, int line) 720 { 721 unsigned int last_timestamp; 722 723 if (map->system_map) { 724 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 725 } else { 726 if (!sx_try_upgrade_(&map->lock, file, line)) { 727 last_timestamp = map->timestamp; 728 sx_sunlock_(&map->lock, file, line); 729 vm_map_process_deferred(); 730 /* 731 * If the map's timestamp does not change while the 732 * map is unlocked, then the upgrade succeeds. 733 */ 734 sx_xlock_(&map->lock, file, line); 735 if (last_timestamp != map->timestamp) { 736 sx_xunlock_(&map->lock, file, line); 737 return (1); 738 } 739 } 740 } 741 map->timestamp++; 742 return (0); 743 } 744 745 void 746 _vm_map_lock_downgrade(vm_map_t map, const char *file, int line) 747 { 748 749 if (map->system_map) { 750 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 751 } else { 752 VM_MAP_UNLOCK_CONSISTENT(map); 753 sx_downgrade_(&map->lock, file, line); 754 } 755 } 756 757 /* 758 * vm_map_locked: 759 * 760 * Returns a non-zero value if the caller holds a write (exclusive) lock 761 * on the specified map and the value "0" otherwise. 762 */ 763 int 764 vm_map_locked(vm_map_t map) 765 { 766 767 if (map->system_map) 768 return (mtx_owned(&map->system_mtx)); 769 else 770 return (sx_xlocked(&map->lock)); 771 } 772 773 /* 774 * _vm_map_unlock_and_wait: 775 * 776 * Atomically releases the lock on the specified map and puts the calling 777 * thread to sleep. The calling thread will remain asleep until either 778 * vm_map_wakeup() is performed on the map or the specified timeout is 779 * exceeded. 780 * 781 * WARNING! This function does not perform deferred deallocations of 782 * objects and map entries. Therefore, the calling thread is expected to 783 * reacquire the map lock after reawakening and later perform an ordinary 784 * unlock operation, such as vm_map_unlock(), before completing its 785 * operation on the map. 786 */ 787 int 788 _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line) 789 { 790 791 VM_MAP_UNLOCK_CONSISTENT(map); 792 mtx_lock(&map_sleep_mtx); 793 if (map->system_map) 794 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 795 else 796 sx_xunlock_(&map->lock, file, line); 797 return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps", 798 timo)); 799 } 800 801 /* 802 * vm_map_wakeup: 803 * 804 * Awaken any threads that have slept on the map using 805 * vm_map_unlock_and_wait(). 806 */ 807 void 808 vm_map_wakeup(vm_map_t map) 809 { 810 811 /* 812 * Acquire and release map_sleep_mtx to prevent a wakeup() 813 * from being performed (and lost) between the map unlock 814 * and the msleep() in _vm_map_unlock_and_wait(). 815 */ 816 mtx_lock(&map_sleep_mtx); 817 mtx_unlock(&map_sleep_mtx); 818 wakeup(&map->root); 819 } 820 821 void 822 vm_map_busy(vm_map_t map) 823 { 824 825 VM_MAP_ASSERT_LOCKED(map); 826 map->busy++; 827 } 828 829 void 830 vm_map_unbusy(vm_map_t map) 831 { 832 833 VM_MAP_ASSERT_LOCKED(map); 834 KASSERT(map->busy, ("vm_map_unbusy: not busy")); 835 if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) { 836 vm_map_modflags(map, 0, MAP_BUSY_WAKEUP); 837 wakeup(&map->busy); 838 } 839 } 840 841 void 842 vm_map_wait_busy(vm_map_t map) 843 { 844 845 VM_MAP_ASSERT_LOCKED(map); 846 while (map->busy) { 847 vm_map_modflags(map, MAP_BUSY_WAKEUP, 0); 848 if (map->system_map) 849 msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0); 850 else 851 sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0); 852 } 853 map->timestamp++; 854 } 855 856 long 857 vmspace_resident_count(struct vmspace *vmspace) 858 { 859 return pmap_resident_count(vmspace_pmap(vmspace)); 860 } 861 862 /* 863 * vm_map_create: 864 * 865 * Creates and returns a new empty VM map with 866 * the given physical map structure, and having 867 * the given lower and upper address bounds. 868 */ 869 vm_map_t 870 vm_map_create(pmap_t pmap, vm_offset_t min, vm_offset_t max) 871 { 872 vm_map_t result; 873 874 result = uma_zalloc(mapzone, M_WAITOK); 875 CTR1(KTR_VM, "vm_map_create: %p", result); 876 _vm_map_init(result, pmap, min, max); 877 return (result); 878 } 879 880 /* 881 * Initialize an existing vm_map structure 882 * such as that in the vmspace structure. 883 */ 884 static void 885 _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 886 { 887 888 map->header.next = map->header.prev = &map->header; 889 map->header.eflags = MAP_ENTRY_HEADER; 890 map->needs_wakeup = FALSE; 891 map->system_map = 0; 892 map->pmap = pmap; 893 map->header.end = min; 894 map->header.start = max; 895 map->flags = 0; 896 map->root = NULL; 897 map->timestamp = 0; 898 map->busy = 0; 899 map->anon_loc = 0; 900 #ifdef DIAGNOSTIC 901 map->nupdates = 0; 902 #endif 903 } 904 905 void 906 vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 907 { 908 909 _vm_map_init(map, pmap, min, max); 910 mtx_init(&map->system_mtx, "system map", NULL, MTX_DEF | MTX_DUPOK); 911 sx_init(&map->lock, "user map"); 912 } 913 914 /* 915 * vm_map_entry_dispose: [ internal use only ] 916 * 917 * Inverse of vm_map_entry_create. 918 */ 919 static void 920 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry) 921 { 922 uma_zfree(map->system_map ? kmapentzone : mapentzone, entry); 923 } 924 925 /* 926 * vm_map_entry_create: [ internal use only ] 927 * 928 * Allocates a VM map entry for insertion. 929 * No entry fields are filled in. 930 */ 931 static vm_map_entry_t 932 vm_map_entry_create(vm_map_t map) 933 { 934 vm_map_entry_t new_entry; 935 936 if (map->system_map) 937 new_entry = uma_zalloc(kmapentzone, M_NOWAIT); 938 else 939 new_entry = uma_zalloc(mapentzone, M_WAITOK); 940 if (new_entry == NULL) 941 panic("vm_map_entry_create: kernel resources exhausted"); 942 return (new_entry); 943 } 944 945 /* 946 * vm_map_entry_set_behavior: 947 * 948 * Set the expected access behavior, either normal, random, or 949 * sequential. 950 */ 951 static inline void 952 vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior) 953 { 954 entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) | 955 (behavior & MAP_ENTRY_BEHAV_MASK); 956 } 957 958 /* 959 * vm_map_entry_max_free_{left,right}: 960 * 961 * Compute the size of the largest free gap between two entries, 962 * one the root of a tree and the other the ancestor of that root 963 * that is the least or greatest ancestor found on the search path. 964 */ 965 static inline vm_size_t 966 vm_map_entry_max_free_left(vm_map_entry_t root, vm_map_entry_t left_ancestor) 967 { 968 969 return (root->left != NULL ? 970 root->left->max_free : root->start - left_ancestor->end); 971 } 972 973 static inline vm_size_t 974 vm_map_entry_max_free_right(vm_map_entry_t root, vm_map_entry_t right_ancestor) 975 { 976 977 return (root->right != NULL ? 978 root->right->max_free : right_ancestor->start - root->end); 979 } 980 981 /* 982 * vm_map_entry_{pred,succ}: 983 * 984 * Find the {predecessor, successor} of the entry by taking one step 985 * in the appropriate direction and backtracking as much as necessary. 986 */ 987 static inline vm_map_entry_t 988 vm_map_entry_pred(vm_map_entry_t entry) 989 { 990 991 return (entry->prev); 992 } 993 994 /* vm_map_entry_succ is defined in vm_map.h. */ 995 996 static inline vm_size_t 997 vm_size_max(vm_size_t a, vm_size_t b) 998 { 999 1000 return (a > b ? a : b); 1001 } 1002 1003 #define SPLAY_LEFT_STEP(root, y, rlist, test) do { \ 1004 vm_size_t max_free; \ 1005 \ 1006 /* \ 1007 * Infer root->right->max_free == root->max_free when \ 1008 * y->max_free < root->max_free || root->max_free == 0. \ 1009 * Otherwise, look right to find it. \ 1010 */ \ 1011 y = root->left; \ 1012 max_free = root->max_free; \ 1013 KASSERT(max_free >= vm_map_entry_max_free_right(root, rlist), \ 1014 ("%s: max_free invariant fails", __func__)); \ 1015 if (y == NULL ? max_free > 0 : max_free - 1 < y->max_free) \ 1016 max_free = vm_map_entry_max_free_right(root, rlist); \ 1017 if (y != NULL && (test)) { \ 1018 /* Rotate right and make y root. */ \ 1019 root->left = y->right; \ 1020 y->right = root; \ 1021 if (max_free < y->max_free) \ 1022 root->max_free = max_free = \ 1023 vm_size_max(max_free, \ 1024 vm_map_entry_max_free_left(root, y)); \ 1025 root = y; \ 1026 y = root->left; \ 1027 } \ 1028 /* Copy right->max_free. Put root on rlist. */ \ 1029 root->max_free = max_free; \ 1030 KASSERT(max_free == vm_map_entry_max_free_right(root, rlist), \ 1031 ("%s: max_free not copied from right", __func__)); \ 1032 root->left = rlist; \ 1033 rlist = root; \ 1034 root = y; \ 1035 } while (0) 1036 1037 #define SPLAY_RIGHT_STEP(root, y, llist, test) do { \ 1038 vm_size_t max_free; \ 1039 \ 1040 /* \ 1041 * Infer root->left->max_free == root->max_free when \ 1042 * y->max_free < root->max_free || root->max_free == 0. \ 1043 * Otherwise, look left to find it. \ 1044 */ \ 1045 y = root->right; \ 1046 max_free = root->max_free; \ 1047 KASSERT(max_free >= vm_map_entry_max_free_left(root, llist), \ 1048 ("%s: max_free invariant fails", __func__)); \ 1049 if (y == NULL ? max_free > 0 : max_free - 1 < y->max_free) \ 1050 max_free = vm_map_entry_max_free_left(root, llist); \ 1051 if (y != NULL && (test)) { \ 1052 /* Rotate left and make y root. */ \ 1053 root->right = y->left; \ 1054 y->left = root; \ 1055 if (max_free < y->max_free) \ 1056 root->max_free = max_free = \ 1057 vm_size_max(max_free, \ 1058 vm_map_entry_max_free_right(root, y)); \ 1059 root = y; \ 1060 y = root->right; \ 1061 } \ 1062 /* Copy left->max_free. Put root on llist. */ \ 1063 root->max_free = max_free; \ 1064 KASSERT(max_free == vm_map_entry_max_free_left(root, llist), \ 1065 ("%s: max_free not copied from left", __func__)); \ 1066 root->right = llist; \ 1067 llist = root; \ 1068 root = y; \ 1069 } while (0) 1070 1071 /* 1072 * Walk down the tree until we find addr or a NULL pointer where addr would go, 1073 * breaking off left and right subtrees of nodes less than, or greater than 1074 * addr. Treat pointers to nodes with max_free < length as NULL pointers. 1075 * llist and rlist are the two sides in reverse order (bottom-up), with llist 1076 * linked by the right pointer and rlist linked by the left pointer in the 1077 * vm_map_entry, and both lists terminated by &map->header. This function, and 1078 * the subsequent call to vm_map_splay_merge, rely on the start and end address 1079 * values in &map->header. 1080 */ 1081 static __always_inline vm_map_entry_t 1082 vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length, 1083 vm_map_entry_t *llist, vm_map_entry_t *rlist) 1084 { 1085 vm_map_entry_t root, y; 1086 1087 *llist = *rlist = &map->header; 1088 root = map->root; 1089 while (root != NULL && root->max_free >= length) { 1090 KASSERT((*llist)->end <= root->start && 1091 root->end <= (*rlist)->start, 1092 ("%s: root not within tree bounds", __func__)); 1093 if (addr < root->start) { 1094 SPLAY_LEFT_STEP(root, y, *rlist, 1095 y->max_free >= length && addr < y->start); 1096 } else if (addr >= root->end) { 1097 SPLAY_RIGHT_STEP(root, y, *llist, 1098 y->max_free >= length && addr >= y->end); 1099 } else 1100 break; 1101 } 1102 return (root); 1103 } 1104 1105 static __always_inline void 1106 vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *rlist) 1107 { 1108 vm_map_entry_t hi, y; 1109 1110 hi = root->right; 1111 while (hi != NULL) 1112 SPLAY_LEFT_STEP(hi, y, *rlist, true); 1113 } 1114 1115 static __always_inline void 1116 vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *llist) 1117 { 1118 vm_map_entry_t lo, y; 1119 1120 lo = root->left; 1121 while (lo != NULL) 1122 SPLAY_RIGHT_STEP(lo, y, *llist, true); 1123 } 1124 1125 static inline void 1126 vm_map_entry_swap(vm_map_entry_t *a, vm_map_entry_t *b) 1127 { 1128 vm_map_entry_t tmp; 1129 1130 tmp = *b; 1131 *b = *a; 1132 *a = tmp; 1133 } 1134 1135 /* 1136 * Walk back up the two spines, flip the pointers and set max_free. The 1137 * subtrees of the root go at the bottom of llist and rlist. 1138 */ 1139 static vm_size_t 1140 vm_map_splay_merge_left_walk(vm_map_entry_t header, vm_map_entry_t root, 1141 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t llist) 1142 { 1143 do { 1144 /* 1145 * The max_free values of the children of llist are in 1146 * llist->max_free and max_free. Update with the 1147 * max value. 1148 */ 1149 llist->max_free = max_free = 1150 vm_size_max(llist->max_free, max_free); 1151 vm_map_entry_swap(&llist->right, &tail); 1152 vm_map_entry_swap(&tail, &llist); 1153 } while (llist != header); 1154 root->left = tail; 1155 return (max_free); 1156 } 1157 1158 /* 1159 * When llist is known to be the predecessor of root. 1160 */ 1161 static inline vm_size_t 1162 vm_map_splay_merge_pred(vm_map_entry_t header, vm_map_entry_t root, 1163 vm_map_entry_t llist) 1164 { 1165 vm_size_t max_free; 1166 1167 max_free = root->start - llist->end; 1168 if (llist != header) { 1169 max_free = vm_map_splay_merge_left_walk(header, root, 1170 NULL, max_free, llist); 1171 } else { 1172 root->left = NULL; 1173 } 1174 return (max_free); 1175 } 1176 1177 /* 1178 * When llist may or may not be the predecessor of root. 1179 */ 1180 static inline vm_size_t 1181 vm_map_splay_merge_left(vm_map_entry_t header, vm_map_entry_t root, 1182 vm_map_entry_t llist) 1183 { 1184 vm_size_t max_free; 1185 1186 max_free = vm_map_entry_max_free_left(root, llist); 1187 if (llist != header) { 1188 max_free = vm_map_splay_merge_left_walk(header, root, 1189 root->left, max_free, llist); 1190 } 1191 return (max_free); 1192 } 1193 1194 static vm_size_t 1195 vm_map_splay_merge_right_walk(vm_map_entry_t header, vm_map_entry_t root, 1196 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t rlist) 1197 { 1198 do { 1199 /* 1200 * The max_free values of the children of rlist are in 1201 * rlist->max_free and max_free. Update with the 1202 * max value. 1203 */ 1204 rlist->max_free = max_free = 1205 vm_size_max(rlist->max_free, max_free); 1206 vm_map_entry_swap(&rlist->left, &tail); 1207 vm_map_entry_swap(&tail, &rlist); 1208 } while (rlist != header); 1209 root->right = tail; 1210 return (max_free); 1211 } 1212 1213 /* 1214 * When rlist is known to be the succecessor of root. 1215 */ 1216 static inline vm_size_t 1217 vm_map_splay_merge_succ(vm_map_entry_t header, vm_map_entry_t root, 1218 vm_map_entry_t rlist) 1219 { 1220 vm_size_t max_free; 1221 1222 max_free = rlist->start - root->end; 1223 if (rlist != header) { 1224 max_free = vm_map_splay_merge_right_walk(header, root, 1225 NULL, max_free, rlist); 1226 } else { 1227 root->right = NULL; 1228 } 1229 return (max_free); 1230 } 1231 1232 /* 1233 * When rlist may or may not be the succecessor of root. 1234 */ 1235 static inline vm_size_t 1236 vm_map_splay_merge_right(vm_map_entry_t header, vm_map_entry_t root, 1237 vm_map_entry_t rlist) 1238 { 1239 vm_size_t max_free; 1240 1241 max_free = vm_map_entry_max_free_right(root, rlist); 1242 if (rlist != header) { 1243 max_free = vm_map_splay_merge_right_walk(header, root, 1244 root->right, max_free, rlist); 1245 } 1246 return (max_free); 1247 } 1248 1249 /* 1250 * vm_map_splay: 1251 * 1252 * The Sleator and Tarjan top-down splay algorithm with the 1253 * following variation. Max_free must be computed bottom-up, so 1254 * on the downward pass, maintain the left and right spines in 1255 * reverse order. Then, make a second pass up each side to fix 1256 * the pointers and compute max_free. The time bound is O(log n) 1257 * amortized. 1258 * 1259 * The new root is the vm_map_entry containing "addr", or else an 1260 * adjacent entry (lower if possible) if addr is not in the tree. 1261 * 1262 * The map must be locked, and leaves it so. 1263 * 1264 * Returns: the new root. 1265 */ 1266 static vm_map_entry_t 1267 vm_map_splay(vm_map_t map, vm_offset_t addr) 1268 { 1269 vm_map_entry_t header, llist, rlist, root; 1270 vm_size_t max_free_left, max_free_right; 1271 1272 header = &map->header; 1273 root = vm_map_splay_split(map, addr, 0, &llist, &rlist); 1274 if (root != NULL) { 1275 max_free_left = vm_map_splay_merge_left(header, root, llist); 1276 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1277 } else if (llist != header) { 1278 /* 1279 * Recover the greatest node in the left 1280 * subtree and make it the root. 1281 */ 1282 root = llist; 1283 llist = root->right; 1284 max_free_left = vm_map_splay_merge_left(header, root, llist); 1285 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1286 } else if (rlist != header) { 1287 /* 1288 * Recover the least node in the right 1289 * subtree and make it the root. 1290 */ 1291 root = rlist; 1292 rlist = root->left; 1293 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1294 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1295 } else { 1296 /* There is no root. */ 1297 return (NULL); 1298 } 1299 root->max_free = vm_size_max(max_free_left, max_free_right); 1300 map->root = root; 1301 VM_MAP_ASSERT_CONSISTENT(map); 1302 return (root); 1303 } 1304 1305 /* 1306 * vm_map_entry_{un,}link: 1307 * 1308 * Insert/remove entries from maps. 1309 */ 1310 static void 1311 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) 1312 { 1313 vm_map_entry_t header, llist, rlist, root; 1314 1315 CTR3(KTR_VM, 1316 "vm_map_entry_link: map %p, nentries %d, entry %p", map, 1317 map->nentries, entry); 1318 VM_MAP_ASSERT_LOCKED(map); 1319 map->nentries++; 1320 header = &map->header; 1321 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1322 KASSERT(root == NULL, 1323 ("vm_map_entry_link: link object already mapped")); 1324 entry->prev = llist; 1325 entry->next = rlist; 1326 llist->next = rlist->prev = entry; 1327 root = entry; 1328 root->max_free = vm_size_max( 1329 vm_map_splay_merge_pred(header, root, llist), 1330 vm_map_splay_merge_succ(header, root, rlist)); 1331 map->root = root; 1332 VM_MAP_ASSERT_CONSISTENT(map); 1333 } 1334 1335 enum unlink_merge_type { 1336 UNLINK_MERGE_NONE, 1337 UNLINK_MERGE_NEXT 1338 }; 1339 1340 static void 1341 vm_map_entry_unlink(vm_map_t map, vm_map_entry_t entry, 1342 enum unlink_merge_type op) 1343 { 1344 vm_map_entry_t header, llist, rlist, root, y; 1345 vm_size_t max_free_left, max_free_right; 1346 1347 VM_MAP_ASSERT_LOCKED(map); 1348 header = &map->header; 1349 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1350 KASSERT(root != NULL, 1351 ("vm_map_entry_unlink: unlink object not mapped")); 1352 1353 vm_map_splay_findprev(root, &llist); 1354 vm_map_splay_findnext(root, &rlist); 1355 if (op == UNLINK_MERGE_NEXT) { 1356 rlist->start = root->start; 1357 rlist->offset = root->offset; 1358 } 1359 if (llist != header) { 1360 root = llist; 1361 llist = root->right; 1362 max_free_left = vm_map_splay_merge_left(header, root, llist); 1363 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1364 } else if (rlist != header) { 1365 root = rlist; 1366 rlist = root->left; 1367 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1368 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1369 } else 1370 root = NULL; 1371 y = entry->next; 1372 y->prev = entry->prev; 1373 y->prev->next = y; 1374 if (root != NULL) 1375 root->max_free = vm_size_max(max_free_left, max_free_right); 1376 map->root = root; 1377 VM_MAP_ASSERT_CONSISTENT(map); 1378 map->nentries--; 1379 CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map, 1380 map->nentries, entry); 1381 } 1382 1383 /* 1384 * vm_map_entry_resize: 1385 * 1386 * Resize a vm_map_entry, recompute the amount of free space that 1387 * follows it and propagate that value up the tree. 1388 * 1389 * The map must be locked, and leaves it so. 1390 */ 1391 static void 1392 vm_map_entry_resize(vm_map_t map, vm_map_entry_t entry, vm_size_t grow_amount) 1393 { 1394 vm_map_entry_t header, llist, rlist, root; 1395 1396 VM_MAP_ASSERT_LOCKED(map); 1397 header = &map->header; 1398 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1399 KASSERT(root != NULL, ("%s: resize object not mapped", __func__)); 1400 vm_map_splay_findnext(root, &rlist); 1401 entry->end += grow_amount; 1402 root->max_free = vm_size_max( 1403 vm_map_splay_merge_left(header, root, llist), 1404 vm_map_splay_merge_succ(header, root, rlist)); 1405 map->root = root; 1406 VM_MAP_ASSERT_CONSISTENT(map); 1407 CTR4(KTR_VM, "%s: map %p, nentries %d, entry %p", 1408 __func__, map, map->nentries, entry); 1409 } 1410 1411 /* 1412 * vm_map_lookup_entry: [ internal use only ] 1413 * 1414 * Finds the map entry containing (or 1415 * immediately preceding) the specified address 1416 * in the given map; the entry is returned 1417 * in the "entry" parameter. The boolean 1418 * result indicates whether the address is 1419 * actually contained in the map. 1420 */ 1421 boolean_t 1422 vm_map_lookup_entry( 1423 vm_map_t map, 1424 vm_offset_t address, 1425 vm_map_entry_t *entry) /* OUT */ 1426 { 1427 vm_map_entry_t cur, header, lbound; 1428 boolean_t locked; 1429 1430 /* 1431 * If the map is empty, then the map entry immediately preceding 1432 * "address" is the map's header. 1433 */ 1434 header = &map->header; 1435 cur = map->root; 1436 if (cur == NULL) { 1437 *entry = header; 1438 return (FALSE); 1439 } 1440 if (address >= cur->start && cur->end > address) { 1441 *entry = cur; 1442 return (TRUE); 1443 } 1444 if ((locked = vm_map_locked(map)) || 1445 sx_try_upgrade(&map->lock)) { 1446 /* 1447 * Splay requires a write lock on the map. However, it only 1448 * restructures the binary search tree; it does not otherwise 1449 * change the map. Thus, the map's timestamp need not change 1450 * on a temporary upgrade. 1451 */ 1452 cur = vm_map_splay(map, address); 1453 if (!locked) { 1454 VM_MAP_UNLOCK_CONSISTENT(map); 1455 sx_downgrade(&map->lock); 1456 } 1457 1458 /* 1459 * If "address" is contained within a map entry, the new root 1460 * is that map entry. Otherwise, the new root is a map entry 1461 * immediately before or after "address". 1462 */ 1463 if (address < cur->start) { 1464 *entry = header; 1465 return (FALSE); 1466 } 1467 *entry = cur; 1468 return (address < cur->end); 1469 } 1470 /* 1471 * Since the map is only locked for read access, perform a 1472 * standard binary search tree lookup for "address". 1473 */ 1474 lbound = header; 1475 do { 1476 if (address < cur->start) { 1477 cur = cur->left; 1478 } else if (cur->end <= address) { 1479 lbound = cur; 1480 cur = cur->right; 1481 } else { 1482 *entry = cur; 1483 return (TRUE); 1484 } 1485 } while (cur != NULL); 1486 *entry = lbound; 1487 return (FALSE); 1488 } 1489 1490 /* 1491 * vm_map_insert: 1492 * 1493 * Inserts the given whole VM object into the target 1494 * map at the specified address range. The object's 1495 * size should match that of the address range. 1496 * 1497 * Requires that the map be locked, and leaves it so. 1498 * 1499 * If object is non-NULL, ref count must be bumped by caller 1500 * prior to making call to account for the new entry. 1501 */ 1502 int 1503 vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1504 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow) 1505 { 1506 vm_map_entry_t new_entry, next_entry, prev_entry; 1507 struct ucred *cred; 1508 vm_eflags_t protoeflags; 1509 vm_inherit_t inheritance; 1510 1511 VM_MAP_ASSERT_LOCKED(map); 1512 KASSERT(object != kernel_object || 1513 (cow & MAP_COPY_ON_WRITE) == 0, 1514 ("vm_map_insert: kernel object and COW")); 1515 KASSERT(object == NULL || (cow & MAP_NOFAULT) == 0, 1516 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 1517 KASSERT((prot & ~max) == 0, 1518 ("prot %#x is not subset of max_prot %#x", prot, max)); 1519 1520 /* 1521 * Check that the start and end points are not bogus. 1522 */ 1523 if (start < vm_map_min(map) || end > vm_map_max(map) || 1524 start >= end) 1525 return (KERN_INVALID_ADDRESS); 1526 1527 /* 1528 * Find the entry prior to the proposed starting address; if it's part 1529 * of an existing entry, this range is bogus. 1530 */ 1531 if (vm_map_lookup_entry(map, start, &prev_entry)) 1532 return (KERN_NO_SPACE); 1533 1534 /* 1535 * Assert that the next entry doesn't overlap the end point. 1536 */ 1537 next_entry = vm_map_entry_succ(prev_entry); 1538 if (next_entry->start < end) 1539 return (KERN_NO_SPACE); 1540 1541 if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL || 1542 max != VM_PROT_NONE)) 1543 return (KERN_INVALID_ARGUMENT); 1544 1545 protoeflags = 0; 1546 if (cow & MAP_COPY_ON_WRITE) 1547 protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; 1548 if (cow & MAP_NOFAULT) 1549 protoeflags |= MAP_ENTRY_NOFAULT; 1550 if (cow & MAP_DISABLE_SYNCER) 1551 protoeflags |= MAP_ENTRY_NOSYNC; 1552 if (cow & MAP_DISABLE_COREDUMP) 1553 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1554 if (cow & MAP_STACK_GROWS_DOWN) 1555 protoeflags |= MAP_ENTRY_GROWS_DOWN; 1556 if (cow & MAP_STACK_GROWS_UP) 1557 protoeflags |= MAP_ENTRY_GROWS_UP; 1558 if (cow & MAP_WRITECOUNT) 1559 protoeflags |= MAP_ENTRY_WRITECNT; 1560 if (cow & MAP_VN_EXEC) 1561 protoeflags |= MAP_ENTRY_VN_EXEC; 1562 if ((cow & MAP_CREATE_GUARD) != 0) 1563 protoeflags |= MAP_ENTRY_GUARD; 1564 if ((cow & MAP_CREATE_STACK_GAP_DN) != 0) 1565 protoeflags |= MAP_ENTRY_STACK_GAP_DN; 1566 if ((cow & MAP_CREATE_STACK_GAP_UP) != 0) 1567 protoeflags |= MAP_ENTRY_STACK_GAP_UP; 1568 if (cow & MAP_INHERIT_SHARE) 1569 inheritance = VM_INHERIT_SHARE; 1570 else 1571 inheritance = VM_INHERIT_DEFAULT; 1572 1573 cred = NULL; 1574 if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0) 1575 goto charged; 1576 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) && 1577 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) { 1578 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start)) 1579 return (KERN_RESOURCE_SHORTAGE); 1580 KASSERT(object == NULL || 1581 (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 || 1582 object->cred == NULL, 1583 ("overcommit: vm_map_insert o %p", object)); 1584 cred = curthread->td_ucred; 1585 } 1586 1587 charged: 1588 /* Expand the kernel pmap, if necessary. */ 1589 if (map == kernel_map && end > kernel_vm_end) 1590 pmap_growkernel(end); 1591 if (object != NULL) { 1592 /* 1593 * OBJ_ONEMAPPING must be cleared unless this mapping 1594 * is trivially proven to be the only mapping for any 1595 * of the object's pages. (Object granularity 1596 * reference counting is insufficient to recognize 1597 * aliases with precision.) 1598 */ 1599 if ((object->flags & OBJ_ANON) != 0) { 1600 VM_OBJECT_WLOCK(object); 1601 if (object->ref_count > 1 || object->shadow_count != 0) 1602 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1603 VM_OBJECT_WUNLOCK(object); 1604 } 1605 } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) == 1606 protoeflags && 1607 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP | 1608 MAP_VN_EXEC)) == 0 && 1609 prev_entry->end == start && (prev_entry->cred == cred || 1610 (prev_entry->object.vm_object != NULL && 1611 prev_entry->object.vm_object->cred == cred)) && 1612 vm_object_coalesce(prev_entry->object.vm_object, 1613 prev_entry->offset, 1614 (vm_size_t)(prev_entry->end - prev_entry->start), 1615 (vm_size_t)(end - prev_entry->end), cred != NULL && 1616 (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) { 1617 /* 1618 * We were able to extend the object. Determine if we 1619 * can extend the previous map entry to include the 1620 * new range as well. 1621 */ 1622 if (prev_entry->inheritance == inheritance && 1623 prev_entry->protection == prot && 1624 prev_entry->max_protection == max && 1625 prev_entry->wired_count == 0) { 1626 KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) == 1627 0, ("prev_entry %p has incoherent wiring", 1628 prev_entry)); 1629 if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0) 1630 map->size += end - prev_entry->end; 1631 vm_map_entry_resize(map, prev_entry, 1632 end - prev_entry->end); 1633 vm_map_try_merge_entries(map, prev_entry, next_entry); 1634 return (KERN_SUCCESS); 1635 } 1636 1637 /* 1638 * If we can extend the object but cannot extend the 1639 * map entry, we have to create a new map entry. We 1640 * must bump the ref count on the extended object to 1641 * account for it. object may be NULL. 1642 */ 1643 object = prev_entry->object.vm_object; 1644 offset = prev_entry->offset + 1645 (prev_entry->end - prev_entry->start); 1646 vm_object_reference(object); 1647 if (cred != NULL && object != NULL && object->cred != NULL && 1648 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 1649 /* Object already accounts for this uid. */ 1650 cred = NULL; 1651 } 1652 } 1653 if (cred != NULL) 1654 crhold(cred); 1655 1656 /* 1657 * Create a new entry 1658 */ 1659 new_entry = vm_map_entry_create(map); 1660 new_entry->start = start; 1661 new_entry->end = end; 1662 new_entry->cred = NULL; 1663 1664 new_entry->eflags = protoeflags; 1665 new_entry->object.vm_object = object; 1666 new_entry->offset = offset; 1667 1668 new_entry->inheritance = inheritance; 1669 new_entry->protection = prot; 1670 new_entry->max_protection = max; 1671 new_entry->wired_count = 0; 1672 new_entry->wiring_thread = NULL; 1673 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT; 1674 new_entry->next_read = start; 1675 1676 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry), 1677 ("overcommit: vm_map_insert leaks vm_map %p", new_entry)); 1678 new_entry->cred = cred; 1679 1680 /* 1681 * Insert the new entry into the list 1682 */ 1683 vm_map_entry_link(map, new_entry); 1684 if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0) 1685 map->size += new_entry->end - new_entry->start; 1686 1687 /* 1688 * Try to coalesce the new entry with both the previous and next 1689 * entries in the list. Previously, we only attempted to coalesce 1690 * with the previous entry when object is NULL. Here, we handle the 1691 * other cases, which are less common. 1692 */ 1693 vm_map_try_merge_entries(map, prev_entry, new_entry); 1694 vm_map_try_merge_entries(map, new_entry, next_entry); 1695 1696 if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) { 1697 vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset), 1698 end - start, cow & MAP_PREFAULT_PARTIAL); 1699 } 1700 1701 return (KERN_SUCCESS); 1702 } 1703 1704 /* 1705 * vm_map_findspace: 1706 * 1707 * Find the first fit (lowest VM address) for "length" free bytes 1708 * beginning at address >= start in the given map. 1709 * 1710 * In a vm_map_entry, "max_free" is the maximum amount of 1711 * contiguous free space between an entry in its subtree and a 1712 * neighbor of that entry. This allows finding a free region in 1713 * one path down the tree, so O(log n) amortized with splay 1714 * trees. 1715 * 1716 * The map must be locked, and leaves it so. 1717 * 1718 * Returns: starting address if sufficient space, 1719 * vm_map_max(map)-length+1 if insufficient space. 1720 */ 1721 vm_offset_t 1722 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length) 1723 { 1724 vm_map_entry_t header, llist, rlist, root, y; 1725 vm_size_t left_length, max_free_left, max_free_right; 1726 vm_offset_t gap_end; 1727 1728 /* 1729 * Request must fit within min/max VM address and must avoid 1730 * address wrap. 1731 */ 1732 start = MAX(start, vm_map_min(map)); 1733 if (start >= vm_map_max(map) || length > vm_map_max(map) - start) 1734 return (vm_map_max(map) - length + 1); 1735 1736 /* Empty tree means wide open address space. */ 1737 if (map->root == NULL) 1738 return (start); 1739 1740 /* 1741 * After splay_split, if start is within an entry, push it to the start 1742 * of the following gap. If rlist is at the end of the gap containing 1743 * start, save the end of that gap in gap_end to see if the gap is big 1744 * enough; otherwise set gap_end to start skip gap-checking and move 1745 * directly to a search of the right subtree. 1746 */ 1747 header = &map->header; 1748 root = vm_map_splay_split(map, start, length, &llist, &rlist); 1749 gap_end = rlist->start; 1750 if (root != NULL) { 1751 start = root->end; 1752 if (root->right != NULL) 1753 gap_end = start; 1754 max_free_left = vm_map_splay_merge_left(header, root, llist); 1755 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1756 } else if (rlist != header) { 1757 root = rlist; 1758 rlist = root->left; 1759 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1760 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1761 } else { 1762 root = llist; 1763 llist = root->right; 1764 max_free_left = vm_map_splay_merge_left(header, root, llist); 1765 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1766 } 1767 root->max_free = vm_size_max(max_free_left, max_free_right); 1768 map->root = root; 1769 VM_MAP_ASSERT_CONSISTENT(map); 1770 if (length <= gap_end - start) 1771 return (start); 1772 1773 /* With max_free, can immediately tell if no solution. */ 1774 if (root->right == NULL || length > root->right->max_free) 1775 return (vm_map_max(map) - length + 1); 1776 1777 /* 1778 * Splay for the least large-enough gap in the right subtree. 1779 */ 1780 llist = rlist = header; 1781 for (left_length = 0;; 1782 left_length = vm_map_entry_max_free_left(root, llist)) { 1783 if (length <= left_length) 1784 SPLAY_LEFT_STEP(root, y, rlist, 1785 length <= vm_map_entry_max_free_left(y, llist)); 1786 else 1787 SPLAY_RIGHT_STEP(root, y, llist, 1788 length > vm_map_entry_max_free_left(y, root)); 1789 if (root == NULL) 1790 break; 1791 } 1792 root = llist; 1793 llist = root->right; 1794 max_free_left = vm_map_splay_merge_left(header, root, llist); 1795 if (rlist == header) { 1796 root->max_free = vm_size_max(max_free_left, 1797 vm_map_splay_merge_succ(header, root, rlist)); 1798 } else { 1799 y = rlist; 1800 rlist = y->left; 1801 y->max_free = vm_size_max( 1802 vm_map_splay_merge_pred(root, y, root), 1803 vm_map_splay_merge_right(header, y, rlist)); 1804 root->right = y; 1805 root->max_free = vm_size_max(max_free_left, y->max_free); 1806 } 1807 map->root = root; 1808 VM_MAP_ASSERT_CONSISTENT(map); 1809 return (root->end); 1810 } 1811 1812 int 1813 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1814 vm_offset_t start, vm_size_t length, vm_prot_t prot, 1815 vm_prot_t max, int cow) 1816 { 1817 vm_offset_t end; 1818 int result; 1819 1820 end = start + length; 1821 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 1822 object == NULL, 1823 ("vm_map_fixed: non-NULL backing object for stack")); 1824 vm_map_lock(map); 1825 VM_MAP_RANGE_CHECK(map, start, end); 1826 if ((cow & MAP_CHECK_EXCL) == 0) 1827 vm_map_delete(map, start, end); 1828 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 1829 result = vm_map_stack_locked(map, start, length, sgrowsiz, 1830 prot, max, cow); 1831 } else { 1832 result = vm_map_insert(map, object, offset, start, end, 1833 prot, max, cow); 1834 } 1835 vm_map_unlock(map); 1836 return (result); 1837 } 1838 1839 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10}; 1840 static const int aslr_pages_rnd_32[2] = {0x100, 0x4}; 1841 1842 static int cluster_anon = 1; 1843 SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW, 1844 &cluster_anon, 0, 1845 "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always"); 1846 1847 static bool 1848 clustering_anon_allowed(vm_offset_t addr) 1849 { 1850 1851 switch (cluster_anon) { 1852 case 0: 1853 return (false); 1854 case 1: 1855 return (addr == 0); 1856 case 2: 1857 default: 1858 return (true); 1859 } 1860 } 1861 1862 static long aslr_restarts; 1863 SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD, 1864 &aslr_restarts, 0, 1865 "Number of aslr failures"); 1866 1867 #define MAP_32BIT_MAX_ADDR ((vm_offset_t)1 << 31) 1868 1869 /* 1870 * Searches for the specified amount of free space in the given map with the 1871 * specified alignment. Performs an address-ordered, first-fit search from 1872 * the given address "*addr", with an optional upper bound "max_addr". If the 1873 * parameter "alignment" is zero, then the alignment is computed from the 1874 * given (object, offset) pair so as to enable the greatest possible use of 1875 * superpage mappings. Returns KERN_SUCCESS and the address of the free space 1876 * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE. 1877 * 1878 * The map must be locked. Initially, there must be at least "length" bytes 1879 * of free space at the given address. 1880 */ 1881 static int 1882 vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1883 vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, 1884 vm_offset_t alignment) 1885 { 1886 vm_offset_t aligned_addr, free_addr; 1887 1888 VM_MAP_ASSERT_LOCKED(map); 1889 free_addr = *addr; 1890 KASSERT(free_addr == vm_map_findspace(map, free_addr, length), 1891 ("caller failed to provide space %#jx at address %p", 1892 (uintmax_t)length, (void *)free_addr)); 1893 for (;;) { 1894 /* 1895 * At the start of every iteration, the free space at address 1896 * "*addr" is at least "length" bytes. 1897 */ 1898 if (alignment == 0) 1899 pmap_align_superpage(object, offset, addr, length); 1900 else if ((*addr & (alignment - 1)) != 0) { 1901 *addr &= ~(alignment - 1); 1902 *addr += alignment; 1903 } 1904 aligned_addr = *addr; 1905 if (aligned_addr == free_addr) { 1906 /* 1907 * Alignment did not change "*addr", so "*addr" must 1908 * still provide sufficient free space. 1909 */ 1910 return (KERN_SUCCESS); 1911 } 1912 1913 /* 1914 * Test for address wrap on "*addr". A wrapped "*addr" could 1915 * be a valid address, in which case vm_map_findspace() cannot 1916 * be relied upon to fail. 1917 */ 1918 if (aligned_addr < free_addr) 1919 return (KERN_NO_SPACE); 1920 *addr = vm_map_findspace(map, aligned_addr, length); 1921 if (*addr + length > vm_map_max(map) || 1922 (max_addr != 0 && *addr + length > max_addr)) 1923 return (KERN_NO_SPACE); 1924 free_addr = *addr; 1925 if (free_addr == aligned_addr) { 1926 /* 1927 * If a successful call to vm_map_findspace() did not 1928 * change "*addr", then "*addr" must still be aligned 1929 * and provide sufficient free space. 1930 */ 1931 return (KERN_SUCCESS); 1932 } 1933 } 1934 } 1935 1936 /* 1937 * vm_map_find finds an unallocated region in the target address 1938 * map with the given length. The search is defined to be 1939 * first-fit from the specified address; the region found is 1940 * returned in the same parameter. 1941 * 1942 * If object is non-NULL, ref count must be bumped by caller 1943 * prior to making call to account for the new entry. 1944 */ 1945 int 1946 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1947 vm_offset_t *addr, /* IN/OUT */ 1948 vm_size_t length, vm_offset_t max_addr, int find_space, 1949 vm_prot_t prot, vm_prot_t max, int cow) 1950 { 1951 vm_offset_t alignment, curr_min_addr, min_addr; 1952 int gap, pidx, rv, try; 1953 bool cluster, en_aslr, update_anon; 1954 1955 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 1956 object == NULL, 1957 ("vm_map_find: non-NULL backing object for stack")); 1958 MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE && 1959 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0)); 1960 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL || 1961 (object->flags & OBJ_COLORED) == 0)) 1962 find_space = VMFS_ANY_SPACE; 1963 if (find_space >> 8 != 0) { 1964 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags")); 1965 alignment = (vm_offset_t)1 << (find_space >> 8); 1966 } else 1967 alignment = 0; 1968 en_aslr = (map->flags & MAP_ASLR) != 0; 1969 update_anon = cluster = clustering_anon_allowed(*addr) && 1970 (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 && 1971 find_space != VMFS_NO_SPACE && object == NULL && 1972 (cow & (MAP_INHERIT_SHARE | MAP_STACK_GROWS_UP | 1973 MAP_STACK_GROWS_DOWN)) == 0 && prot != PROT_NONE; 1974 curr_min_addr = min_addr = *addr; 1975 if (en_aslr && min_addr == 0 && !cluster && 1976 find_space != VMFS_NO_SPACE && 1977 (map->flags & MAP_ASLR_IGNSTART) != 0) 1978 curr_min_addr = min_addr = vm_map_min(map); 1979 try = 0; 1980 vm_map_lock(map); 1981 if (cluster) { 1982 curr_min_addr = map->anon_loc; 1983 if (curr_min_addr == 0) 1984 cluster = false; 1985 } 1986 if (find_space != VMFS_NO_SPACE) { 1987 KASSERT(find_space == VMFS_ANY_SPACE || 1988 find_space == VMFS_OPTIMAL_SPACE || 1989 find_space == VMFS_SUPER_SPACE || 1990 alignment != 0, ("unexpected VMFS flag")); 1991 again: 1992 /* 1993 * When creating an anonymous mapping, try clustering 1994 * with an existing anonymous mapping first. 1995 * 1996 * We make up to two attempts to find address space 1997 * for a given find_space value. The first attempt may 1998 * apply randomization or may cluster with an existing 1999 * anonymous mapping. If this first attempt fails, 2000 * perform a first-fit search of the available address 2001 * space. 2002 * 2003 * If all tries failed, and find_space is 2004 * VMFS_OPTIMAL_SPACE, fallback to VMFS_ANY_SPACE. 2005 * Again enable clustering and randomization. 2006 */ 2007 try++; 2008 MPASS(try <= 2); 2009 2010 if (try == 2) { 2011 /* 2012 * Second try: we failed either to find a 2013 * suitable region for randomizing the 2014 * allocation, or to cluster with an existing 2015 * mapping. Retry with free run. 2016 */ 2017 curr_min_addr = (map->flags & MAP_ASLR_IGNSTART) != 0 ? 2018 vm_map_min(map) : min_addr; 2019 atomic_add_long(&aslr_restarts, 1); 2020 } 2021 2022 if (try == 1 && en_aslr && !cluster) { 2023 /* 2024 * Find space for allocation, including 2025 * gap needed for later randomization. 2026 */ 2027 pidx = MAXPAGESIZES > 1 && pagesizes[1] != 0 && 2028 (find_space == VMFS_SUPER_SPACE || find_space == 2029 VMFS_OPTIMAL_SPACE) ? 1 : 0; 2030 gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR && 2031 (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ? 2032 aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx]; 2033 *addr = vm_map_findspace(map, curr_min_addr, 2034 length + gap * pagesizes[pidx]); 2035 if (*addr + length + gap * pagesizes[pidx] > 2036 vm_map_max(map)) 2037 goto again; 2038 /* And randomize the start address. */ 2039 *addr += (arc4random() % gap) * pagesizes[pidx]; 2040 if (max_addr != 0 && *addr + length > max_addr) 2041 goto again; 2042 } else { 2043 *addr = vm_map_findspace(map, curr_min_addr, length); 2044 if (*addr + length > vm_map_max(map) || 2045 (max_addr != 0 && *addr + length > max_addr)) { 2046 if (cluster) { 2047 cluster = false; 2048 MPASS(try == 1); 2049 goto again; 2050 } 2051 rv = KERN_NO_SPACE; 2052 goto done; 2053 } 2054 } 2055 2056 if (find_space != VMFS_ANY_SPACE && 2057 (rv = vm_map_alignspace(map, object, offset, addr, length, 2058 max_addr, alignment)) != KERN_SUCCESS) { 2059 if (find_space == VMFS_OPTIMAL_SPACE) { 2060 find_space = VMFS_ANY_SPACE; 2061 curr_min_addr = min_addr; 2062 cluster = update_anon; 2063 try = 0; 2064 goto again; 2065 } 2066 goto done; 2067 } 2068 } else if ((cow & MAP_REMAP) != 0) { 2069 if (*addr < vm_map_min(map) || 2070 *addr + length > vm_map_max(map) || 2071 *addr + length <= length) { 2072 rv = KERN_INVALID_ADDRESS; 2073 goto done; 2074 } 2075 vm_map_delete(map, *addr, *addr + length); 2076 } 2077 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 2078 rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot, 2079 max, cow); 2080 } else { 2081 rv = vm_map_insert(map, object, offset, *addr, *addr + length, 2082 prot, max, cow); 2083 } 2084 if (rv == KERN_SUCCESS && update_anon) 2085 map->anon_loc = *addr + length; 2086 done: 2087 vm_map_unlock(map); 2088 return (rv); 2089 } 2090 2091 /* 2092 * vm_map_find_min() is a variant of vm_map_find() that takes an 2093 * additional parameter (min_addr) and treats the given address 2094 * (*addr) differently. Specifically, it treats *addr as a hint 2095 * and not as the minimum address where the mapping is created. 2096 * 2097 * This function works in two phases. First, it tries to 2098 * allocate above the hint. If that fails and the hint is 2099 * greater than min_addr, it performs a second pass, replacing 2100 * the hint with min_addr as the minimum address for the 2101 * allocation. 2102 */ 2103 int 2104 vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2105 vm_offset_t *addr, vm_size_t length, vm_offset_t min_addr, 2106 vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max, 2107 int cow) 2108 { 2109 vm_offset_t hint; 2110 int rv; 2111 2112 hint = *addr; 2113 for (;;) { 2114 rv = vm_map_find(map, object, offset, addr, length, max_addr, 2115 find_space, prot, max, cow); 2116 if (rv == KERN_SUCCESS || min_addr >= hint) 2117 return (rv); 2118 *addr = hint = min_addr; 2119 } 2120 } 2121 2122 /* 2123 * A map entry with any of the following flags set must not be merged with 2124 * another entry. 2125 */ 2126 #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP | \ 2127 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC) 2128 2129 static bool 2130 vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry) 2131 { 2132 2133 KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 || 2134 (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0, 2135 ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable", 2136 prev, entry)); 2137 return (prev->end == entry->start && 2138 prev->object.vm_object == entry->object.vm_object && 2139 (prev->object.vm_object == NULL || 2140 prev->offset + (prev->end - prev->start) == entry->offset) && 2141 prev->eflags == entry->eflags && 2142 prev->protection == entry->protection && 2143 prev->max_protection == entry->max_protection && 2144 prev->inheritance == entry->inheritance && 2145 prev->wired_count == entry->wired_count && 2146 prev->cred == entry->cred); 2147 } 2148 2149 static void 2150 vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry) 2151 { 2152 2153 /* 2154 * If the backing object is a vnode object, vm_object_deallocate() 2155 * calls vrele(). However, vrele() does not lock the vnode because 2156 * the vnode has additional references. Thus, the map lock can be 2157 * kept without causing a lock-order reversal with the vnode lock. 2158 * 2159 * Since we count the number of virtual page mappings in 2160 * object->un_pager.vnp.writemappings, the writemappings value 2161 * should not be adjusted when the entry is disposed of. 2162 */ 2163 if (entry->object.vm_object != NULL) 2164 vm_object_deallocate(entry->object.vm_object); 2165 if (entry->cred != NULL) 2166 crfree(entry->cred); 2167 vm_map_entry_dispose(map, entry); 2168 } 2169 2170 /* 2171 * vm_map_try_merge_entries: 2172 * 2173 * Compare the given map entry to its predecessor, and merge its precessor 2174 * into it if possible. The entry remains valid, and may be extended. 2175 * The predecessor may be deleted. 2176 * 2177 * The map must be locked. 2178 */ 2179 void 2180 vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev_entry, 2181 vm_map_entry_t entry) 2182 { 2183 2184 VM_MAP_ASSERT_LOCKED(map); 2185 if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 && 2186 vm_map_mergeable_neighbors(prev_entry, entry)) { 2187 vm_map_entry_unlink(map, prev_entry, UNLINK_MERGE_NEXT); 2188 vm_map_merged_neighbor_dispose(map, prev_entry); 2189 } 2190 } 2191 2192 /* 2193 * vm_map_entry_back: 2194 * 2195 * Allocate an object to back a map entry. 2196 */ 2197 static inline void 2198 vm_map_entry_back(vm_map_entry_t entry) 2199 { 2200 vm_object_t object; 2201 2202 KASSERT(entry->object.vm_object == NULL, 2203 ("map entry %p has backing object", entry)); 2204 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2205 ("map entry %p is a submap", entry)); 2206 object = vm_object_allocate_anon(atop(entry->end - entry->start)); 2207 entry->object.vm_object = object; 2208 entry->offset = 0; 2209 if (entry->cred != NULL) { 2210 object->cred = entry->cred; 2211 object->charge = entry->end - entry->start; 2212 entry->cred = NULL; 2213 } 2214 } 2215 2216 /* 2217 * vm_map_entry_charge_object 2218 * 2219 * If there is no object backing this entry, create one. Otherwise, if 2220 * the entry has cred, give it to the backing object. 2221 */ 2222 static inline void 2223 vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry) 2224 { 2225 2226 VM_MAP_ASSERT_LOCKED(map); 2227 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2228 ("map entry %p is a submap", entry)); 2229 if (entry->object.vm_object == NULL && !map->system_map && 2230 (entry->eflags & MAP_ENTRY_GUARD) == 0) 2231 vm_map_entry_back(entry); 2232 else if (entry->object.vm_object != NULL && 2233 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) && 2234 entry->cred != NULL) { 2235 VM_OBJECT_WLOCK(entry->object.vm_object); 2236 KASSERT(entry->object.vm_object->cred == NULL, 2237 ("OVERCOMMIT: %s: both cred e %p", __func__, entry)); 2238 entry->object.vm_object->cred = entry->cred; 2239 entry->object.vm_object->charge = entry->end - entry->start; 2240 VM_OBJECT_WUNLOCK(entry->object.vm_object); 2241 entry->cred = NULL; 2242 } 2243 } 2244 2245 /* 2246 * vm_map_clip_start: [ internal use only ] 2247 * 2248 * Asserts that the given entry begins at or after 2249 * the specified address; if necessary, 2250 * it splits the entry into two. 2251 */ 2252 #define vm_map_clip_start(map, entry, startaddr) \ 2253 { \ 2254 if (startaddr > entry->start) \ 2255 _vm_map_clip_start(map, entry, startaddr); \ 2256 } 2257 2258 /* 2259 * This routine is called only when it is known that 2260 * the entry must be split. 2261 */ 2262 static void 2263 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start) 2264 { 2265 vm_map_entry_t new_entry; 2266 2267 VM_MAP_ASSERT_LOCKED(map); 2268 KASSERT(entry->end > start && entry->start < start, 2269 ("_vm_map_clip_start: invalid clip of entry %p", entry)); 2270 2271 /* 2272 * Create a backing object now, if none exists, so that more individual 2273 * objects won't be created after the map entry is split. 2274 */ 2275 vm_map_entry_charge_object(map, entry); 2276 2277 /* Clone the entry. */ 2278 new_entry = vm_map_entry_create(map); 2279 *new_entry = *entry; 2280 2281 /* 2282 * Split off the front portion. Insert the new entry BEFORE this one, 2283 * so that this entry has the specified starting address. 2284 */ 2285 new_entry->end = start; 2286 entry->offset += (start - entry->start); 2287 entry->start = start; 2288 if (new_entry->cred != NULL) 2289 crhold(entry->cred); 2290 2291 vm_map_entry_link(map, new_entry); 2292 2293 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2294 vm_object_reference(new_entry->object.vm_object); 2295 vm_map_entry_set_vnode_text(new_entry, true); 2296 /* 2297 * The object->un_pager.vnp.writemappings for the 2298 * object of MAP_ENTRY_WRITECNT type entry shall be 2299 * kept as is here. The virtual pages are 2300 * re-distributed among the clipped entries, so the sum is 2301 * left the same. 2302 */ 2303 } 2304 } 2305 2306 /* 2307 * vm_map_clip_end: [ internal use only ] 2308 * 2309 * Asserts that the given entry ends at or before 2310 * the specified address; if necessary, 2311 * it splits the entry into two. 2312 */ 2313 #define vm_map_clip_end(map, entry, endaddr) \ 2314 { \ 2315 if ((endaddr) < (entry->end)) \ 2316 _vm_map_clip_end((map), (entry), (endaddr)); \ 2317 } 2318 2319 /* 2320 * This routine is called only when it is known that 2321 * the entry must be split. 2322 */ 2323 static void 2324 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end) 2325 { 2326 vm_map_entry_t new_entry; 2327 2328 VM_MAP_ASSERT_LOCKED(map); 2329 KASSERT(entry->start < end && entry->end > end, 2330 ("_vm_map_clip_end: invalid clip of entry %p", entry)); 2331 2332 /* 2333 * Create a backing object now, if none exists, so that more individual 2334 * objects won't be created after the map entry is split. 2335 */ 2336 vm_map_entry_charge_object(map, entry); 2337 2338 /* Clone the entry. */ 2339 new_entry = vm_map_entry_create(map); 2340 *new_entry = *entry; 2341 2342 /* 2343 * Split off the back portion. Insert the new entry AFTER this one, 2344 * so that this entry has the specified ending address. 2345 */ 2346 new_entry->start = entry->end = end; 2347 new_entry->offset += (end - entry->start); 2348 if (new_entry->cred != NULL) 2349 crhold(entry->cred); 2350 2351 vm_map_entry_link(map, new_entry); 2352 2353 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2354 vm_object_reference(new_entry->object.vm_object); 2355 vm_map_entry_set_vnode_text(new_entry, true); 2356 } 2357 } 2358 2359 /* 2360 * vm_map_submap: [ kernel use only ] 2361 * 2362 * Mark the given range as handled by a subordinate map. 2363 * 2364 * This range must have been created with vm_map_find, 2365 * and no other operations may have been performed on this 2366 * range prior to calling vm_map_submap. 2367 * 2368 * Only a limited number of operations can be performed 2369 * within this rage after calling vm_map_submap: 2370 * vm_fault 2371 * [Don't try vm_map_copy!] 2372 * 2373 * To remove a submapping, one must first remove the 2374 * range from the superior map, and then destroy the 2375 * submap (if desired). [Better yet, don't try it.] 2376 */ 2377 int 2378 vm_map_submap( 2379 vm_map_t map, 2380 vm_offset_t start, 2381 vm_offset_t end, 2382 vm_map_t submap) 2383 { 2384 vm_map_entry_t entry; 2385 int result; 2386 2387 result = KERN_INVALID_ARGUMENT; 2388 2389 vm_map_lock(submap); 2390 submap->flags |= MAP_IS_SUB_MAP; 2391 vm_map_unlock(submap); 2392 2393 vm_map_lock(map); 2394 2395 VM_MAP_RANGE_CHECK(map, start, end); 2396 2397 if (vm_map_lookup_entry(map, start, &entry)) { 2398 vm_map_clip_start(map, entry, start); 2399 } else 2400 entry = vm_map_entry_succ(entry); 2401 2402 vm_map_clip_end(map, entry, end); 2403 2404 if ((entry->start == start) && (entry->end == end) && 2405 ((entry->eflags & MAP_ENTRY_COW) == 0) && 2406 (entry->object.vm_object == NULL)) { 2407 entry->object.sub_map = submap; 2408 entry->eflags |= MAP_ENTRY_IS_SUB_MAP; 2409 result = KERN_SUCCESS; 2410 } 2411 vm_map_unlock(map); 2412 2413 if (result != KERN_SUCCESS) { 2414 vm_map_lock(submap); 2415 submap->flags &= ~MAP_IS_SUB_MAP; 2416 vm_map_unlock(submap); 2417 } 2418 return (result); 2419 } 2420 2421 /* 2422 * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified 2423 */ 2424 #define MAX_INIT_PT 96 2425 2426 /* 2427 * vm_map_pmap_enter: 2428 * 2429 * Preload the specified map's pmap with mappings to the specified 2430 * object's memory-resident pages. No further physical pages are 2431 * allocated, and no further virtual pages are retrieved from secondary 2432 * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a 2433 * limited number of page mappings are created at the low-end of the 2434 * specified address range. (For this purpose, a superpage mapping 2435 * counts as one page mapping.) Otherwise, all resident pages within 2436 * the specified address range are mapped. 2437 */ 2438 static void 2439 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 2440 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags) 2441 { 2442 vm_offset_t start; 2443 vm_page_t p, p_start; 2444 vm_pindex_t mask, psize, threshold, tmpidx; 2445 2446 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL) 2447 return; 2448 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2449 VM_OBJECT_WLOCK(object); 2450 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2451 pmap_object_init_pt(map->pmap, addr, object, pindex, 2452 size); 2453 VM_OBJECT_WUNLOCK(object); 2454 return; 2455 } 2456 VM_OBJECT_LOCK_DOWNGRADE(object); 2457 } else 2458 VM_OBJECT_RLOCK(object); 2459 2460 psize = atop(size); 2461 if (psize + pindex > object->size) { 2462 if (object->size < pindex) { 2463 VM_OBJECT_RUNLOCK(object); 2464 return; 2465 } 2466 psize = object->size - pindex; 2467 } 2468 2469 start = 0; 2470 p_start = NULL; 2471 threshold = MAX_INIT_PT; 2472 2473 p = vm_page_find_least(object, pindex); 2474 /* 2475 * Assert: the variable p is either (1) the page with the 2476 * least pindex greater than or equal to the parameter pindex 2477 * or (2) NULL. 2478 */ 2479 for (; 2480 p != NULL && (tmpidx = p->pindex - pindex) < psize; 2481 p = TAILQ_NEXT(p, listq)) { 2482 /* 2483 * don't allow an madvise to blow away our really 2484 * free pages allocating pv entries. 2485 */ 2486 if (((flags & MAP_PREFAULT_MADVISE) != 0 && 2487 vm_page_count_severe()) || 2488 ((flags & MAP_PREFAULT_PARTIAL) != 0 && 2489 tmpidx >= threshold)) { 2490 psize = tmpidx; 2491 break; 2492 } 2493 if (vm_page_all_valid(p)) { 2494 if (p_start == NULL) { 2495 start = addr + ptoa(tmpidx); 2496 p_start = p; 2497 } 2498 /* Jump ahead if a superpage mapping is possible. */ 2499 if (p->psind > 0 && ((addr + ptoa(tmpidx)) & 2500 (pagesizes[p->psind] - 1)) == 0) { 2501 mask = atop(pagesizes[p->psind]) - 1; 2502 if (tmpidx + mask < psize && 2503 vm_page_ps_test(p, PS_ALL_VALID, NULL)) { 2504 p += mask; 2505 threshold += mask; 2506 } 2507 } 2508 } else if (p_start != NULL) { 2509 pmap_enter_object(map->pmap, start, addr + 2510 ptoa(tmpidx), p_start, prot); 2511 p_start = NULL; 2512 } 2513 } 2514 if (p_start != NULL) 2515 pmap_enter_object(map->pmap, start, addr + ptoa(psize), 2516 p_start, prot); 2517 VM_OBJECT_RUNLOCK(object); 2518 } 2519 2520 /* 2521 * vm_map_protect: 2522 * 2523 * Sets the protection of the specified address 2524 * region in the target map. If "set_max" is 2525 * specified, the maximum protection is to be set; 2526 * otherwise, only the current protection is affected. 2527 */ 2528 int 2529 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 2530 vm_prot_t new_prot, boolean_t set_max) 2531 { 2532 vm_map_entry_t entry, first_entry, in_tran, prev_entry; 2533 vm_object_t obj; 2534 struct ucred *cred; 2535 vm_prot_t old_prot; 2536 int rv; 2537 2538 if (start == end) 2539 return (KERN_SUCCESS); 2540 2541 again: 2542 in_tran = NULL; 2543 vm_map_lock(map); 2544 2545 /* 2546 * Ensure that we are not concurrently wiring pages. vm_map_wire() may 2547 * need to fault pages into the map and will drop the map lock while 2548 * doing so, and the VM object may end up in an inconsistent state if we 2549 * update the protection on the map entry in between faults. 2550 */ 2551 vm_map_wait_busy(map); 2552 2553 VM_MAP_RANGE_CHECK(map, start, end); 2554 2555 if (!vm_map_lookup_entry(map, start, &first_entry)) 2556 first_entry = vm_map_entry_succ(first_entry); 2557 2558 /* 2559 * Make a first pass to check for protection violations. 2560 */ 2561 for (entry = first_entry; entry->start < end; 2562 entry = vm_map_entry_succ(entry)) { 2563 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 2564 continue; 2565 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 2566 vm_map_unlock(map); 2567 return (KERN_INVALID_ARGUMENT); 2568 } 2569 if ((new_prot & entry->max_protection) != new_prot) { 2570 vm_map_unlock(map); 2571 return (KERN_PROTECTION_FAILURE); 2572 } 2573 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0) 2574 in_tran = entry; 2575 } 2576 2577 /* 2578 * Postpone the operation until all in-transition map entries have 2579 * stabilized. An in-transition entry might already have its pages 2580 * wired and wired_count incremented, but not yet have its 2581 * MAP_ENTRY_USER_WIRED flag set. In which case, we would fail to call 2582 * vm_fault_copy_entry() in the final loop below. 2583 */ 2584 if (in_tran != NULL) { 2585 in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2586 vm_map_unlock_and_wait(map, 0); 2587 goto again; 2588 } 2589 2590 /* 2591 * Before changing the protections, try to reserve swap space for any 2592 * private (i.e., copy-on-write) mappings that are transitioning from 2593 * read-only to read/write access. If a reservation fails, break out 2594 * of this loop early and let the next loop simplify the entries, since 2595 * some may now be mergeable. 2596 */ 2597 rv = KERN_SUCCESS; 2598 vm_map_clip_start(map, first_entry, start); 2599 for (entry = first_entry; entry->start < end; 2600 entry = vm_map_entry_succ(entry)) { 2601 vm_map_clip_end(map, entry, end); 2602 2603 if (set_max || 2604 ((new_prot & ~entry->protection) & VM_PROT_WRITE) == 0 || 2605 ENTRY_CHARGED(entry) || 2606 (entry->eflags & MAP_ENTRY_GUARD) != 0) { 2607 continue; 2608 } 2609 2610 cred = curthread->td_ucred; 2611 obj = entry->object.vm_object; 2612 2613 if (obj == NULL || 2614 (entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0) { 2615 if (!swap_reserve(entry->end - entry->start)) { 2616 rv = KERN_RESOURCE_SHORTAGE; 2617 end = entry->end; 2618 break; 2619 } 2620 crhold(cred); 2621 entry->cred = cred; 2622 continue; 2623 } 2624 2625 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) 2626 continue; 2627 VM_OBJECT_WLOCK(obj); 2628 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) { 2629 VM_OBJECT_WUNLOCK(obj); 2630 continue; 2631 } 2632 2633 /* 2634 * Charge for the whole object allocation now, since 2635 * we cannot distinguish between non-charged and 2636 * charged clipped mapping of the same object later. 2637 */ 2638 KASSERT(obj->charge == 0, 2639 ("vm_map_protect: object %p overcharged (entry %p)", 2640 obj, entry)); 2641 if (!swap_reserve(ptoa(obj->size))) { 2642 VM_OBJECT_WUNLOCK(obj); 2643 rv = KERN_RESOURCE_SHORTAGE; 2644 end = entry->end; 2645 break; 2646 } 2647 2648 crhold(cred); 2649 obj->cred = cred; 2650 obj->charge = ptoa(obj->size); 2651 VM_OBJECT_WUNLOCK(obj); 2652 } 2653 2654 /* 2655 * If enough swap space was available, go back and fix up protections. 2656 * Otherwise, just simplify entries, since some may have been modified. 2657 * [Note that clipping is not necessary the second time.] 2658 */ 2659 for (prev_entry = vm_map_entry_pred(first_entry), entry = first_entry; 2660 entry->start < end; 2661 vm_map_try_merge_entries(map, prev_entry, entry), 2662 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 2663 if (rv != KERN_SUCCESS || 2664 (entry->eflags & MAP_ENTRY_GUARD) != 0) 2665 continue; 2666 2667 old_prot = entry->protection; 2668 2669 if (set_max) 2670 entry->protection = 2671 (entry->max_protection = new_prot) & 2672 old_prot; 2673 else 2674 entry->protection = new_prot; 2675 2676 /* 2677 * For user wired map entries, the normal lazy evaluation of 2678 * write access upgrades through soft page faults is 2679 * undesirable. Instead, immediately copy any pages that are 2680 * copy-on-write and enable write access in the physical map. 2681 */ 2682 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0 && 2683 (entry->protection & VM_PROT_WRITE) != 0 && 2684 (old_prot & VM_PROT_WRITE) == 0) 2685 vm_fault_copy_entry(map, map, entry, entry, NULL); 2686 2687 /* 2688 * When restricting access, update the physical map. Worry 2689 * about copy-on-write here. 2690 */ 2691 if ((old_prot & ~entry->protection) != 0) { 2692 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 2693 VM_PROT_ALL) 2694 pmap_protect(map->pmap, entry->start, 2695 entry->end, 2696 entry->protection & MASK(entry)); 2697 #undef MASK 2698 } 2699 } 2700 vm_map_try_merge_entries(map, prev_entry, entry); 2701 vm_map_unlock(map); 2702 return (rv); 2703 } 2704 2705 /* 2706 * vm_map_madvise: 2707 * 2708 * This routine traverses a processes map handling the madvise 2709 * system call. Advisories are classified as either those effecting 2710 * the vm_map_entry structure, or those effecting the underlying 2711 * objects. 2712 */ 2713 int 2714 vm_map_madvise( 2715 vm_map_t map, 2716 vm_offset_t start, 2717 vm_offset_t end, 2718 int behav) 2719 { 2720 vm_map_entry_t entry, prev_entry; 2721 bool modify_map; 2722 2723 /* 2724 * Some madvise calls directly modify the vm_map_entry, in which case 2725 * we need to use an exclusive lock on the map and we need to perform 2726 * various clipping operations. Otherwise we only need a read-lock 2727 * on the map. 2728 */ 2729 switch(behav) { 2730 case MADV_NORMAL: 2731 case MADV_SEQUENTIAL: 2732 case MADV_RANDOM: 2733 case MADV_NOSYNC: 2734 case MADV_AUTOSYNC: 2735 case MADV_NOCORE: 2736 case MADV_CORE: 2737 if (start == end) 2738 return (0); 2739 modify_map = true; 2740 vm_map_lock(map); 2741 break; 2742 case MADV_WILLNEED: 2743 case MADV_DONTNEED: 2744 case MADV_FREE: 2745 if (start == end) 2746 return (0); 2747 modify_map = false; 2748 vm_map_lock_read(map); 2749 break; 2750 default: 2751 return (EINVAL); 2752 } 2753 2754 /* 2755 * Locate starting entry and clip if necessary. 2756 */ 2757 VM_MAP_RANGE_CHECK(map, start, end); 2758 2759 if (vm_map_lookup_entry(map, start, &entry)) { 2760 if (modify_map) 2761 vm_map_clip_start(map, entry, start); 2762 prev_entry = vm_map_entry_pred(entry); 2763 } else { 2764 prev_entry = entry; 2765 entry = vm_map_entry_succ(entry); 2766 } 2767 2768 if (modify_map) { 2769 /* 2770 * madvise behaviors that are implemented in the vm_map_entry. 2771 * 2772 * We clip the vm_map_entry so that behavioral changes are 2773 * limited to the specified address range. 2774 */ 2775 for (; entry->start < end; 2776 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 2777 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2778 continue; 2779 2780 vm_map_clip_end(map, entry, end); 2781 2782 switch (behav) { 2783 case MADV_NORMAL: 2784 vm_map_entry_set_behavior(entry, 2785 MAP_ENTRY_BEHAV_NORMAL); 2786 break; 2787 case MADV_SEQUENTIAL: 2788 vm_map_entry_set_behavior(entry, 2789 MAP_ENTRY_BEHAV_SEQUENTIAL); 2790 break; 2791 case MADV_RANDOM: 2792 vm_map_entry_set_behavior(entry, 2793 MAP_ENTRY_BEHAV_RANDOM); 2794 break; 2795 case MADV_NOSYNC: 2796 entry->eflags |= MAP_ENTRY_NOSYNC; 2797 break; 2798 case MADV_AUTOSYNC: 2799 entry->eflags &= ~MAP_ENTRY_NOSYNC; 2800 break; 2801 case MADV_NOCORE: 2802 entry->eflags |= MAP_ENTRY_NOCOREDUMP; 2803 break; 2804 case MADV_CORE: 2805 entry->eflags &= ~MAP_ENTRY_NOCOREDUMP; 2806 break; 2807 default: 2808 break; 2809 } 2810 vm_map_try_merge_entries(map, prev_entry, entry); 2811 } 2812 vm_map_try_merge_entries(map, prev_entry, entry); 2813 vm_map_unlock(map); 2814 } else { 2815 vm_pindex_t pstart, pend; 2816 2817 /* 2818 * madvise behaviors that are implemented in the underlying 2819 * vm_object. 2820 * 2821 * Since we don't clip the vm_map_entry, we have to clip 2822 * the vm_object pindex and count. 2823 */ 2824 for (; entry->start < end; 2825 entry = vm_map_entry_succ(entry)) { 2826 vm_offset_t useEnd, useStart; 2827 2828 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2829 continue; 2830 2831 /* 2832 * MADV_FREE would otherwise rewind time to 2833 * the creation of the shadow object. Because 2834 * we hold the VM map read-locked, neither the 2835 * entry's object nor the presence of a 2836 * backing object can change. 2837 */ 2838 if (behav == MADV_FREE && 2839 entry->object.vm_object != NULL && 2840 entry->object.vm_object->backing_object != NULL) 2841 continue; 2842 2843 pstart = OFF_TO_IDX(entry->offset); 2844 pend = pstart + atop(entry->end - entry->start); 2845 useStart = entry->start; 2846 useEnd = entry->end; 2847 2848 if (entry->start < start) { 2849 pstart += atop(start - entry->start); 2850 useStart = start; 2851 } 2852 if (entry->end > end) { 2853 pend -= atop(entry->end - end); 2854 useEnd = end; 2855 } 2856 2857 if (pstart >= pend) 2858 continue; 2859 2860 /* 2861 * Perform the pmap_advise() before clearing 2862 * PGA_REFERENCED in vm_page_advise(). Otherwise, a 2863 * concurrent pmap operation, such as pmap_remove(), 2864 * could clear a reference in the pmap and set 2865 * PGA_REFERENCED on the page before the pmap_advise() 2866 * had completed. Consequently, the page would appear 2867 * referenced based upon an old reference that 2868 * occurred before this pmap_advise() ran. 2869 */ 2870 if (behav == MADV_DONTNEED || behav == MADV_FREE) 2871 pmap_advise(map->pmap, useStart, useEnd, 2872 behav); 2873 2874 vm_object_madvise(entry->object.vm_object, pstart, 2875 pend, behav); 2876 2877 /* 2878 * Pre-populate paging structures in the 2879 * WILLNEED case. For wired entries, the 2880 * paging structures are already populated. 2881 */ 2882 if (behav == MADV_WILLNEED && 2883 entry->wired_count == 0) { 2884 vm_map_pmap_enter(map, 2885 useStart, 2886 entry->protection, 2887 entry->object.vm_object, 2888 pstart, 2889 ptoa(pend - pstart), 2890 MAP_PREFAULT_MADVISE 2891 ); 2892 } 2893 } 2894 vm_map_unlock_read(map); 2895 } 2896 return (0); 2897 } 2898 2899 2900 /* 2901 * vm_map_inherit: 2902 * 2903 * Sets the inheritance of the specified address 2904 * range in the target map. Inheritance 2905 * affects how the map will be shared with 2906 * child maps at the time of vmspace_fork. 2907 */ 2908 int 2909 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 2910 vm_inherit_t new_inheritance) 2911 { 2912 vm_map_entry_t entry, prev_entry; 2913 2914 switch (new_inheritance) { 2915 case VM_INHERIT_NONE: 2916 case VM_INHERIT_COPY: 2917 case VM_INHERIT_SHARE: 2918 case VM_INHERIT_ZERO: 2919 break; 2920 default: 2921 return (KERN_INVALID_ARGUMENT); 2922 } 2923 if (start == end) 2924 return (KERN_SUCCESS); 2925 vm_map_lock(map); 2926 VM_MAP_RANGE_CHECK(map, start, end); 2927 if (vm_map_lookup_entry(map, start, &prev_entry)) { 2928 entry = prev_entry; 2929 vm_map_clip_start(map, entry, start); 2930 prev_entry = vm_map_entry_pred(entry); 2931 } else 2932 entry = vm_map_entry_succ(prev_entry); 2933 for (; entry->start < end; 2934 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 2935 vm_map_clip_end(map, entry, end); 2936 if ((entry->eflags & MAP_ENTRY_GUARD) == 0 || 2937 new_inheritance != VM_INHERIT_ZERO) 2938 entry->inheritance = new_inheritance; 2939 vm_map_try_merge_entries(map, prev_entry, entry); 2940 } 2941 vm_map_try_merge_entries(map, prev_entry, entry); 2942 vm_map_unlock(map); 2943 return (KERN_SUCCESS); 2944 } 2945 2946 /* 2947 * vm_map_entry_in_transition: 2948 * 2949 * Release the map lock, and sleep until the entry is no longer in 2950 * transition. Awake and acquire the map lock. If the map changed while 2951 * another held the lock, lookup a possibly-changed entry at or after the 2952 * 'start' position of the old entry. 2953 */ 2954 static vm_map_entry_t 2955 vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start, 2956 vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry) 2957 { 2958 vm_map_entry_t entry; 2959 vm_offset_t start; 2960 u_int last_timestamp; 2961 2962 VM_MAP_ASSERT_LOCKED(map); 2963 KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 2964 ("not in-tranition map entry %p", in_entry)); 2965 /* 2966 * We have not yet clipped the entry. 2967 */ 2968 start = MAX(in_start, in_entry->start); 2969 in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2970 last_timestamp = map->timestamp; 2971 if (vm_map_unlock_and_wait(map, 0)) { 2972 /* 2973 * Allow interruption of user wiring/unwiring? 2974 */ 2975 } 2976 vm_map_lock(map); 2977 if (last_timestamp + 1 == map->timestamp) 2978 return (in_entry); 2979 2980 /* 2981 * Look again for the entry because the map was modified while it was 2982 * unlocked. Specifically, the entry may have been clipped, merged, or 2983 * deleted. 2984 */ 2985 if (!vm_map_lookup_entry(map, start, &entry)) { 2986 if (!holes_ok) { 2987 *io_end = start; 2988 return (NULL); 2989 } 2990 entry = vm_map_entry_succ(entry); 2991 } 2992 return (entry); 2993 } 2994 2995 /* 2996 * vm_map_unwire: 2997 * 2998 * Implements both kernel and user unwiring. 2999 */ 3000 int 3001 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 3002 int flags) 3003 { 3004 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3005 int rv; 3006 bool holes_ok, need_wakeup, user_unwire; 3007 3008 if (start == end) 3009 return (KERN_SUCCESS); 3010 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3011 user_unwire = (flags & VM_MAP_WIRE_USER) != 0; 3012 vm_map_lock(map); 3013 VM_MAP_RANGE_CHECK(map, start, end); 3014 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3015 if (holes_ok) 3016 first_entry = vm_map_entry_succ(first_entry); 3017 else { 3018 vm_map_unlock(map); 3019 return (KERN_INVALID_ADDRESS); 3020 } 3021 } 3022 rv = KERN_SUCCESS; 3023 for (entry = first_entry; entry->start < end; entry = next_entry) { 3024 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3025 /* 3026 * We have not yet clipped the entry. 3027 */ 3028 next_entry = vm_map_entry_in_transition(map, start, 3029 &end, holes_ok, entry); 3030 if (next_entry == NULL) { 3031 if (entry == first_entry) { 3032 vm_map_unlock(map); 3033 return (KERN_INVALID_ADDRESS); 3034 } 3035 rv = KERN_INVALID_ADDRESS; 3036 break; 3037 } 3038 first_entry = (entry == first_entry) ? 3039 next_entry : NULL; 3040 continue; 3041 } 3042 vm_map_clip_start(map, entry, start); 3043 vm_map_clip_end(map, entry, end); 3044 /* 3045 * Mark the entry in case the map lock is released. (See 3046 * above.) 3047 */ 3048 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3049 entry->wiring_thread == NULL, 3050 ("owned map entry %p", entry)); 3051 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3052 entry->wiring_thread = curthread; 3053 next_entry = vm_map_entry_succ(entry); 3054 /* 3055 * Check the map for holes in the specified region. 3056 * If holes_ok, skip this check. 3057 */ 3058 if (!holes_ok && 3059 entry->end < end && next_entry->start > entry->end) { 3060 end = entry->end; 3061 rv = KERN_INVALID_ADDRESS; 3062 break; 3063 } 3064 /* 3065 * If system unwiring, require that the entry is system wired. 3066 */ 3067 if (!user_unwire && 3068 vm_map_entry_system_wired_count(entry) == 0) { 3069 end = entry->end; 3070 rv = KERN_INVALID_ARGUMENT; 3071 break; 3072 } 3073 } 3074 need_wakeup = false; 3075 if (first_entry == NULL && 3076 !vm_map_lookup_entry(map, start, &first_entry)) { 3077 KASSERT(holes_ok, ("vm_map_unwire: lookup failed")); 3078 prev_entry = first_entry; 3079 entry = vm_map_entry_succ(first_entry); 3080 } else { 3081 prev_entry = vm_map_entry_pred(first_entry); 3082 entry = first_entry; 3083 } 3084 for (; entry->start < end; 3085 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3086 /* 3087 * If holes_ok was specified, an empty 3088 * space in the unwired region could have been mapped 3089 * while the map lock was dropped for draining 3090 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread 3091 * could be simultaneously wiring this new mapping 3092 * entry. Detect these cases and skip any entries 3093 * marked as in transition by us. 3094 */ 3095 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3096 entry->wiring_thread != curthread) { 3097 KASSERT(holes_ok, 3098 ("vm_map_unwire: !HOLESOK and new/changed entry")); 3099 continue; 3100 } 3101 3102 if (rv == KERN_SUCCESS && (!user_unwire || 3103 (entry->eflags & MAP_ENTRY_USER_WIRED))) { 3104 if (entry->wired_count == 1) 3105 vm_map_entry_unwire(map, entry); 3106 else 3107 entry->wired_count--; 3108 if (user_unwire) 3109 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3110 } 3111 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3112 ("vm_map_unwire: in-transition flag missing %p", entry)); 3113 KASSERT(entry->wiring_thread == curthread, 3114 ("vm_map_unwire: alien wire %p", entry)); 3115 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 3116 entry->wiring_thread = NULL; 3117 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3118 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3119 need_wakeup = true; 3120 } 3121 vm_map_try_merge_entries(map, prev_entry, entry); 3122 } 3123 vm_map_try_merge_entries(map, prev_entry, entry); 3124 vm_map_unlock(map); 3125 if (need_wakeup) 3126 vm_map_wakeup(map); 3127 return (rv); 3128 } 3129 3130 static void 3131 vm_map_wire_user_count_sub(u_long npages) 3132 { 3133 3134 atomic_subtract_long(&vm_user_wire_count, npages); 3135 } 3136 3137 static bool 3138 vm_map_wire_user_count_add(u_long npages) 3139 { 3140 u_long wired; 3141 3142 wired = vm_user_wire_count; 3143 do { 3144 if (npages + wired > vm_page_max_user_wired) 3145 return (false); 3146 } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired, 3147 npages + wired)); 3148 3149 return (true); 3150 } 3151 3152 /* 3153 * vm_map_wire_entry_failure: 3154 * 3155 * Handle a wiring failure on the given entry. 3156 * 3157 * The map should be locked. 3158 */ 3159 static void 3160 vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 3161 vm_offset_t failed_addr) 3162 { 3163 3164 VM_MAP_ASSERT_LOCKED(map); 3165 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 && 3166 entry->wired_count == 1, 3167 ("vm_map_wire_entry_failure: entry %p isn't being wired", entry)); 3168 KASSERT(failed_addr < entry->end, 3169 ("vm_map_wire_entry_failure: entry %p was fully wired", entry)); 3170 3171 /* 3172 * If any pages at the start of this entry were successfully wired, 3173 * then unwire them. 3174 */ 3175 if (failed_addr > entry->start) { 3176 pmap_unwire(map->pmap, entry->start, failed_addr); 3177 vm_object_unwire(entry->object.vm_object, entry->offset, 3178 failed_addr - entry->start, PQ_ACTIVE); 3179 } 3180 3181 /* 3182 * Assign an out-of-range value to represent the failure to wire this 3183 * entry. 3184 */ 3185 entry->wired_count = -1; 3186 } 3187 3188 int 3189 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3190 { 3191 int rv; 3192 3193 vm_map_lock(map); 3194 rv = vm_map_wire_locked(map, start, end, flags); 3195 vm_map_unlock(map); 3196 return (rv); 3197 } 3198 3199 3200 /* 3201 * vm_map_wire_locked: 3202 * 3203 * Implements both kernel and user wiring. Returns with the map locked, 3204 * the map lock may be dropped. 3205 */ 3206 int 3207 vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3208 { 3209 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3210 vm_offset_t faddr, saved_end, saved_start; 3211 u_long npages; 3212 u_int last_timestamp; 3213 int rv; 3214 bool holes_ok, need_wakeup, user_wire; 3215 vm_prot_t prot; 3216 3217 VM_MAP_ASSERT_LOCKED(map); 3218 3219 if (start == end) 3220 return (KERN_SUCCESS); 3221 prot = 0; 3222 if (flags & VM_MAP_WIRE_WRITE) 3223 prot |= VM_PROT_WRITE; 3224 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3225 user_wire = (flags & VM_MAP_WIRE_USER) != 0; 3226 VM_MAP_RANGE_CHECK(map, start, end); 3227 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3228 if (holes_ok) 3229 first_entry = vm_map_entry_succ(first_entry); 3230 else 3231 return (KERN_INVALID_ADDRESS); 3232 } 3233 for (entry = first_entry; entry->start < end; entry = next_entry) { 3234 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3235 /* 3236 * We have not yet clipped the entry. 3237 */ 3238 next_entry = vm_map_entry_in_transition(map, start, 3239 &end, holes_ok, entry); 3240 if (next_entry == NULL) { 3241 if (entry == first_entry) 3242 return (KERN_INVALID_ADDRESS); 3243 rv = KERN_INVALID_ADDRESS; 3244 goto done; 3245 } 3246 first_entry = (entry == first_entry) ? 3247 next_entry : NULL; 3248 continue; 3249 } 3250 vm_map_clip_start(map, entry, start); 3251 vm_map_clip_end(map, entry, end); 3252 /* 3253 * Mark the entry in case the map lock is released. (See 3254 * above.) 3255 */ 3256 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3257 entry->wiring_thread == NULL, 3258 ("owned map entry %p", entry)); 3259 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3260 entry->wiring_thread = curthread; 3261 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 3262 || (entry->protection & prot) != prot) { 3263 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED; 3264 if (!holes_ok) { 3265 end = entry->end; 3266 rv = KERN_INVALID_ADDRESS; 3267 goto done; 3268 } 3269 } else if (entry->wired_count == 0) { 3270 entry->wired_count++; 3271 3272 npages = atop(entry->end - entry->start); 3273 if (user_wire && !vm_map_wire_user_count_add(npages)) { 3274 vm_map_wire_entry_failure(map, entry, 3275 entry->start); 3276 end = entry->end; 3277 rv = KERN_RESOURCE_SHORTAGE; 3278 goto done; 3279 } 3280 3281 /* 3282 * Release the map lock, relying on the in-transition 3283 * mark. Mark the map busy for fork. 3284 */ 3285 saved_start = entry->start; 3286 saved_end = entry->end; 3287 last_timestamp = map->timestamp; 3288 vm_map_busy(map); 3289 vm_map_unlock(map); 3290 3291 faddr = saved_start; 3292 do { 3293 /* 3294 * Simulate a fault to get the page and enter 3295 * it into the physical map. 3296 */ 3297 if ((rv = vm_fault(map, faddr, 3298 VM_PROT_NONE, VM_FAULT_WIRE, NULL)) != 3299 KERN_SUCCESS) 3300 break; 3301 } while ((faddr += PAGE_SIZE) < saved_end); 3302 vm_map_lock(map); 3303 vm_map_unbusy(map); 3304 if (last_timestamp + 1 != map->timestamp) { 3305 /* 3306 * Look again for the entry because the map was 3307 * modified while it was unlocked. The entry 3308 * may have been clipped, but NOT merged or 3309 * deleted. 3310 */ 3311 if (!vm_map_lookup_entry(map, saved_start, 3312 &next_entry)) 3313 KASSERT(false, 3314 ("vm_map_wire: lookup failed")); 3315 first_entry = (entry == first_entry) ? 3316 next_entry : NULL; 3317 for (entry = next_entry; entry->end < saved_end; 3318 entry = vm_map_entry_succ(entry)) { 3319 /* 3320 * In case of failure, handle entries 3321 * that were not fully wired here; 3322 * fully wired entries are handled 3323 * later. 3324 */ 3325 if (rv != KERN_SUCCESS && 3326 faddr < entry->end) 3327 vm_map_wire_entry_failure(map, 3328 entry, faddr); 3329 } 3330 } 3331 if (rv != KERN_SUCCESS) { 3332 vm_map_wire_entry_failure(map, entry, faddr); 3333 if (user_wire) 3334 vm_map_wire_user_count_sub(npages); 3335 end = entry->end; 3336 goto done; 3337 } 3338 } else if (!user_wire || 3339 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3340 entry->wired_count++; 3341 } 3342 /* 3343 * Check the map for holes in the specified region. 3344 * If holes_ok was specified, skip this check. 3345 */ 3346 next_entry = vm_map_entry_succ(entry); 3347 if (!holes_ok && 3348 entry->end < end && next_entry->start > entry->end) { 3349 end = entry->end; 3350 rv = KERN_INVALID_ADDRESS; 3351 goto done; 3352 } 3353 } 3354 rv = KERN_SUCCESS; 3355 done: 3356 need_wakeup = false; 3357 if (first_entry == NULL && 3358 !vm_map_lookup_entry(map, start, &first_entry)) { 3359 KASSERT(holes_ok, ("vm_map_wire: lookup failed")); 3360 prev_entry = first_entry; 3361 entry = vm_map_entry_succ(first_entry); 3362 } else { 3363 prev_entry = vm_map_entry_pred(first_entry); 3364 entry = first_entry; 3365 } 3366 for (; entry->start < end; 3367 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3368 /* 3369 * If holes_ok was specified, an empty 3370 * space in the unwired region could have been mapped 3371 * while the map lock was dropped for faulting in the 3372 * pages or draining MAP_ENTRY_IN_TRANSITION. 3373 * Moreover, another thread could be simultaneously 3374 * wiring this new mapping entry. Detect these cases 3375 * and skip any entries marked as in transition not by us. 3376 */ 3377 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3378 entry->wiring_thread != curthread) { 3379 KASSERT(holes_ok, 3380 ("vm_map_wire: !HOLESOK and new/changed entry")); 3381 continue; 3382 } 3383 3384 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) { 3385 /* do nothing */ 3386 } else if (rv == KERN_SUCCESS) { 3387 if (user_wire) 3388 entry->eflags |= MAP_ENTRY_USER_WIRED; 3389 } else if (entry->wired_count == -1) { 3390 /* 3391 * Wiring failed on this entry. Thus, unwiring is 3392 * unnecessary. 3393 */ 3394 entry->wired_count = 0; 3395 } else if (!user_wire || 3396 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3397 /* 3398 * Undo the wiring. Wiring succeeded on this entry 3399 * but failed on a later entry. 3400 */ 3401 if (entry->wired_count == 1) { 3402 vm_map_entry_unwire(map, entry); 3403 if (user_wire) 3404 vm_map_wire_user_count_sub( 3405 atop(entry->end - entry->start)); 3406 } else 3407 entry->wired_count--; 3408 } 3409 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3410 ("vm_map_wire: in-transition flag missing %p", entry)); 3411 KASSERT(entry->wiring_thread == curthread, 3412 ("vm_map_wire: alien wire %p", entry)); 3413 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION | 3414 MAP_ENTRY_WIRE_SKIPPED); 3415 entry->wiring_thread = NULL; 3416 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3417 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3418 need_wakeup = true; 3419 } 3420 vm_map_try_merge_entries(map, prev_entry, entry); 3421 } 3422 vm_map_try_merge_entries(map, prev_entry, entry); 3423 if (need_wakeup) 3424 vm_map_wakeup(map); 3425 return (rv); 3426 } 3427 3428 /* 3429 * vm_map_sync 3430 * 3431 * Push any dirty cached pages in the address range to their pager. 3432 * If syncio is TRUE, dirty pages are written synchronously. 3433 * If invalidate is TRUE, any cached pages are freed as well. 3434 * 3435 * If the size of the region from start to end is zero, we are 3436 * supposed to flush all modified pages within the region containing 3437 * start. Unfortunately, a region can be split or coalesced with 3438 * neighboring regions, making it difficult to determine what the 3439 * original region was. Therefore, we approximate this requirement by 3440 * flushing the current region containing start. 3441 * 3442 * Returns an error if any part of the specified range is not mapped. 3443 */ 3444 int 3445 vm_map_sync( 3446 vm_map_t map, 3447 vm_offset_t start, 3448 vm_offset_t end, 3449 boolean_t syncio, 3450 boolean_t invalidate) 3451 { 3452 vm_map_entry_t entry, first_entry, next_entry; 3453 vm_size_t size; 3454 vm_object_t object; 3455 vm_ooffset_t offset; 3456 unsigned int last_timestamp; 3457 boolean_t failed; 3458 3459 vm_map_lock_read(map); 3460 VM_MAP_RANGE_CHECK(map, start, end); 3461 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3462 vm_map_unlock_read(map); 3463 return (KERN_INVALID_ADDRESS); 3464 } else if (start == end) { 3465 start = first_entry->start; 3466 end = first_entry->end; 3467 } 3468 /* 3469 * Make a first pass to check for user-wired memory and holes. 3470 */ 3471 for (entry = first_entry; entry->start < end; entry = next_entry) { 3472 if (invalidate && 3473 (entry->eflags & MAP_ENTRY_USER_WIRED) != 0) { 3474 vm_map_unlock_read(map); 3475 return (KERN_INVALID_ARGUMENT); 3476 } 3477 next_entry = vm_map_entry_succ(entry); 3478 if (end > entry->end && 3479 entry->end != next_entry->start) { 3480 vm_map_unlock_read(map); 3481 return (KERN_INVALID_ADDRESS); 3482 } 3483 } 3484 3485 if (invalidate) 3486 pmap_remove(map->pmap, start, end); 3487 failed = FALSE; 3488 3489 /* 3490 * Make a second pass, cleaning/uncaching pages from the indicated 3491 * objects as we go. 3492 */ 3493 for (entry = first_entry; entry->start < end;) { 3494 offset = entry->offset + (start - entry->start); 3495 size = (end <= entry->end ? end : entry->end) - start; 3496 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 3497 vm_map_t smap; 3498 vm_map_entry_t tentry; 3499 vm_size_t tsize; 3500 3501 smap = entry->object.sub_map; 3502 vm_map_lock_read(smap); 3503 (void) vm_map_lookup_entry(smap, offset, &tentry); 3504 tsize = tentry->end - offset; 3505 if (tsize < size) 3506 size = tsize; 3507 object = tentry->object.vm_object; 3508 offset = tentry->offset + (offset - tentry->start); 3509 vm_map_unlock_read(smap); 3510 } else { 3511 object = entry->object.vm_object; 3512 } 3513 vm_object_reference(object); 3514 last_timestamp = map->timestamp; 3515 vm_map_unlock_read(map); 3516 if (!vm_object_sync(object, offset, size, syncio, invalidate)) 3517 failed = TRUE; 3518 start += size; 3519 vm_object_deallocate(object); 3520 vm_map_lock_read(map); 3521 if (last_timestamp == map->timestamp || 3522 !vm_map_lookup_entry(map, start, &entry)) 3523 entry = vm_map_entry_succ(entry); 3524 } 3525 3526 vm_map_unlock_read(map); 3527 return (failed ? KERN_FAILURE : KERN_SUCCESS); 3528 } 3529 3530 /* 3531 * vm_map_entry_unwire: [ internal use only ] 3532 * 3533 * Make the region specified by this entry pageable. 3534 * 3535 * The map in question should be locked. 3536 * [This is the reason for this routine's existence.] 3537 */ 3538 static void 3539 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 3540 { 3541 vm_size_t size; 3542 3543 VM_MAP_ASSERT_LOCKED(map); 3544 KASSERT(entry->wired_count > 0, 3545 ("vm_map_entry_unwire: entry %p isn't wired", entry)); 3546 3547 size = entry->end - entry->start; 3548 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) 3549 vm_map_wire_user_count_sub(atop(size)); 3550 pmap_unwire(map->pmap, entry->start, entry->end); 3551 vm_object_unwire(entry->object.vm_object, entry->offset, size, 3552 PQ_ACTIVE); 3553 entry->wired_count = 0; 3554 } 3555 3556 static void 3557 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map) 3558 { 3559 3560 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) 3561 vm_object_deallocate(entry->object.vm_object); 3562 uma_zfree(system_map ? kmapentzone : mapentzone, entry); 3563 } 3564 3565 /* 3566 * vm_map_entry_delete: [ internal use only ] 3567 * 3568 * Deallocate the given entry from the target map. 3569 */ 3570 static void 3571 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry) 3572 { 3573 vm_object_t object; 3574 vm_pindex_t offidxstart, offidxend, count, size1; 3575 vm_size_t size; 3576 3577 vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE); 3578 object = entry->object.vm_object; 3579 3580 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { 3581 MPASS(entry->cred == NULL); 3582 MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0); 3583 MPASS(object == NULL); 3584 vm_map_entry_deallocate(entry, map->system_map); 3585 return; 3586 } 3587 3588 size = entry->end - entry->start; 3589 map->size -= size; 3590 3591 if (entry->cred != NULL) { 3592 swap_release_by_cred(size, entry->cred); 3593 crfree(entry->cred); 3594 } 3595 3596 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || object == NULL) { 3597 entry->object.vm_object = NULL; 3598 } else if ((object->flags & OBJ_ANON) != 0 || 3599 object == kernel_object) { 3600 KASSERT(entry->cred == NULL || object->cred == NULL || 3601 (entry->eflags & MAP_ENTRY_NEEDS_COPY), 3602 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry)); 3603 count = atop(size); 3604 offidxstart = OFF_TO_IDX(entry->offset); 3605 offidxend = offidxstart + count; 3606 VM_OBJECT_WLOCK(object); 3607 if (object->ref_count != 1 && 3608 ((object->flags & OBJ_ONEMAPPING) != 0 || 3609 object == kernel_object)) { 3610 vm_object_collapse(object); 3611 3612 /* 3613 * The option OBJPR_NOTMAPPED can be passed here 3614 * because vm_map_delete() already performed 3615 * pmap_remove() on the only mapping to this range 3616 * of pages. 3617 */ 3618 vm_object_page_remove(object, offidxstart, offidxend, 3619 OBJPR_NOTMAPPED); 3620 if (object->type == OBJT_SWAP) 3621 swap_pager_freespace(object, offidxstart, 3622 count); 3623 if (offidxend >= object->size && 3624 offidxstart < object->size) { 3625 size1 = object->size; 3626 object->size = offidxstart; 3627 if (object->cred != NULL) { 3628 size1 -= object->size; 3629 KASSERT(object->charge >= ptoa(size1), 3630 ("object %p charge < 0", object)); 3631 swap_release_by_cred(ptoa(size1), 3632 object->cred); 3633 object->charge -= ptoa(size1); 3634 } 3635 } 3636 } 3637 VM_OBJECT_WUNLOCK(object); 3638 } 3639 if (map->system_map) 3640 vm_map_entry_deallocate(entry, TRUE); 3641 else { 3642 entry->defer_next = curthread->td_map_def_user; 3643 curthread->td_map_def_user = entry; 3644 } 3645 } 3646 3647 /* 3648 * vm_map_delete: [ internal use only ] 3649 * 3650 * Deallocates the given address range from the target 3651 * map. 3652 */ 3653 int 3654 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end) 3655 { 3656 vm_map_entry_t entry; 3657 vm_map_entry_t first_entry; 3658 3659 VM_MAP_ASSERT_LOCKED(map); 3660 if (start == end) 3661 return (KERN_SUCCESS); 3662 3663 /* 3664 * Find the start of the region, and clip it 3665 */ 3666 if (!vm_map_lookup_entry(map, start, &first_entry)) 3667 entry = vm_map_entry_succ(first_entry); 3668 else { 3669 entry = first_entry; 3670 vm_map_clip_start(map, entry, start); 3671 } 3672 3673 /* 3674 * Step through all entries in this region 3675 */ 3676 while (entry->start < end) { 3677 vm_map_entry_t next; 3678 3679 /* 3680 * Wait for wiring or unwiring of an entry to complete. 3681 * Also wait for any system wirings to disappear on 3682 * user maps. 3683 */ 3684 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 || 3685 (vm_map_pmap(map) != kernel_pmap && 3686 vm_map_entry_system_wired_count(entry) != 0)) { 3687 unsigned int last_timestamp; 3688 vm_offset_t saved_start; 3689 vm_map_entry_t tmp_entry; 3690 3691 saved_start = entry->start; 3692 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3693 last_timestamp = map->timestamp; 3694 (void) vm_map_unlock_and_wait(map, 0); 3695 vm_map_lock(map); 3696 if (last_timestamp + 1 != map->timestamp) { 3697 /* 3698 * Look again for the entry because the map was 3699 * modified while it was unlocked. 3700 * Specifically, the entry may have been 3701 * clipped, merged, or deleted. 3702 */ 3703 if (!vm_map_lookup_entry(map, saved_start, 3704 &tmp_entry)) 3705 entry = vm_map_entry_succ(tmp_entry); 3706 else { 3707 entry = tmp_entry; 3708 vm_map_clip_start(map, entry, 3709 saved_start); 3710 } 3711 } 3712 continue; 3713 } 3714 vm_map_clip_end(map, entry, end); 3715 3716 next = vm_map_entry_succ(entry); 3717 3718 /* 3719 * Unwire before removing addresses from the pmap; otherwise, 3720 * unwiring will put the entries back in the pmap. 3721 */ 3722 if (entry->wired_count != 0) 3723 vm_map_entry_unwire(map, entry); 3724 3725 /* 3726 * Remove mappings for the pages, but only if the 3727 * mappings could exist. For instance, it does not 3728 * make sense to call pmap_remove() for guard entries. 3729 */ 3730 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 3731 entry->object.vm_object != NULL) 3732 pmap_remove(map->pmap, entry->start, entry->end); 3733 3734 if (entry->end == map->anon_loc) 3735 map->anon_loc = entry->start; 3736 3737 /* 3738 * Delete the entry only after removing all pmap 3739 * entries pointing to its pages. (Otherwise, its 3740 * page frames may be reallocated, and any modify bits 3741 * will be set in the wrong object!) 3742 */ 3743 vm_map_entry_delete(map, entry); 3744 entry = next; 3745 } 3746 return (KERN_SUCCESS); 3747 } 3748 3749 /* 3750 * vm_map_remove: 3751 * 3752 * Remove the given address range from the target map. 3753 * This is the exported form of vm_map_delete. 3754 */ 3755 int 3756 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 3757 { 3758 int result; 3759 3760 vm_map_lock(map); 3761 VM_MAP_RANGE_CHECK(map, start, end); 3762 result = vm_map_delete(map, start, end); 3763 vm_map_unlock(map); 3764 return (result); 3765 } 3766 3767 /* 3768 * vm_map_check_protection: 3769 * 3770 * Assert that the target map allows the specified privilege on the 3771 * entire address region given. The entire region must be allocated. 3772 * 3773 * WARNING! This code does not and should not check whether the 3774 * contents of the region is accessible. For example a smaller file 3775 * might be mapped into a larger address space. 3776 * 3777 * NOTE! This code is also called by munmap(). 3778 * 3779 * The map must be locked. A read lock is sufficient. 3780 */ 3781 boolean_t 3782 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 3783 vm_prot_t protection) 3784 { 3785 vm_map_entry_t entry; 3786 vm_map_entry_t tmp_entry; 3787 3788 if (!vm_map_lookup_entry(map, start, &tmp_entry)) 3789 return (FALSE); 3790 entry = tmp_entry; 3791 3792 while (start < end) { 3793 /* 3794 * No holes allowed! 3795 */ 3796 if (start < entry->start) 3797 return (FALSE); 3798 /* 3799 * Check protection associated with entry. 3800 */ 3801 if ((entry->protection & protection) != protection) 3802 return (FALSE); 3803 /* go to next entry */ 3804 start = entry->end; 3805 entry = vm_map_entry_succ(entry); 3806 } 3807 return (TRUE); 3808 } 3809 3810 3811 /* 3812 * 3813 * vm_map_copy_swap_object: 3814 * 3815 * Copies a swap-backed object from an existing map entry to a 3816 * new one. Carries forward the swap charge. May change the 3817 * src object on return. 3818 */ 3819 static void 3820 vm_map_copy_swap_object(vm_map_entry_t src_entry, vm_map_entry_t dst_entry, 3821 vm_offset_t size, vm_ooffset_t *fork_charge) 3822 { 3823 vm_object_t src_object; 3824 struct ucred *cred; 3825 int charged; 3826 3827 src_object = src_entry->object.vm_object; 3828 VM_OBJECT_WLOCK(src_object); 3829 charged = ENTRY_CHARGED(src_entry); 3830 vm_object_collapse(src_object); 3831 if ((src_object->flags & OBJ_ONEMAPPING) != 0) { 3832 vm_object_split(src_entry); 3833 src_object = src_entry->object.vm_object; 3834 } 3835 vm_object_reference_locked(src_object); 3836 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 3837 if (src_entry->cred != NULL && 3838 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 3839 KASSERT(src_object->cred == NULL, 3840 ("OVERCOMMIT: vm_map_copy_anon_entry: cred %p", 3841 src_object)); 3842 src_object->cred = src_entry->cred; 3843 src_object->charge = size; 3844 } 3845 VM_OBJECT_WUNLOCK(src_object); 3846 dst_entry->object.vm_object = src_object; 3847 if (charged) { 3848 cred = curthread->td_ucred; 3849 crhold(cred); 3850 dst_entry->cred = cred; 3851 *fork_charge += size; 3852 if (!(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 3853 crhold(cred); 3854 src_entry->cred = cred; 3855 *fork_charge += size; 3856 } 3857 } 3858 } 3859 3860 /* 3861 * vm_map_copy_entry: 3862 * 3863 * Copies the contents of the source entry to the destination 3864 * entry. The entries *must* be aligned properly. 3865 */ 3866 static void 3867 vm_map_copy_entry( 3868 vm_map_t src_map, 3869 vm_map_t dst_map, 3870 vm_map_entry_t src_entry, 3871 vm_map_entry_t dst_entry, 3872 vm_ooffset_t *fork_charge) 3873 { 3874 vm_object_t src_object; 3875 vm_map_entry_t fake_entry; 3876 vm_offset_t size; 3877 3878 VM_MAP_ASSERT_LOCKED(dst_map); 3879 3880 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) 3881 return; 3882 3883 if (src_entry->wired_count == 0 || 3884 (src_entry->protection & VM_PROT_WRITE) == 0) { 3885 /* 3886 * If the source entry is marked needs_copy, it is already 3887 * write-protected. 3888 */ 3889 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && 3890 (src_entry->protection & VM_PROT_WRITE) != 0) { 3891 pmap_protect(src_map->pmap, 3892 src_entry->start, 3893 src_entry->end, 3894 src_entry->protection & ~VM_PROT_WRITE); 3895 } 3896 3897 /* 3898 * Make a copy of the object. 3899 */ 3900 size = src_entry->end - src_entry->start; 3901 if ((src_object = src_entry->object.vm_object) != NULL) { 3902 if (src_object->type == OBJT_DEFAULT || 3903 src_object->type == OBJT_SWAP) { 3904 vm_map_copy_swap_object(src_entry, dst_entry, 3905 size, fork_charge); 3906 /* May have split/collapsed, reload obj. */ 3907 src_object = src_entry->object.vm_object; 3908 } else { 3909 vm_object_reference(src_object); 3910 dst_entry->object.vm_object = src_object; 3911 } 3912 src_entry->eflags |= MAP_ENTRY_COW | 3913 MAP_ENTRY_NEEDS_COPY; 3914 dst_entry->eflags |= MAP_ENTRY_COW | 3915 MAP_ENTRY_NEEDS_COPY; 3916 dst_entry->offset = src_entry->offset; 3917 if (src_entry->eflags & MAP_ENTRY_WRITECNT) { 3918 /* 3919 * MAP_ENTRY_WRITECNT cannot 3920 * indicate write reference from 3921 * src_entry, since the entry is 3922 * marked as needs copy. Allocate a 3923 * fake entry that is used to 3924 * decrement object->un_pager writecount 3925 * at the appropriate time. Attach 3926 * fake_entry to the deferred list. 3927 */ 3928 fake_entry = vm_map_entry_create(dst_map); 3929 fake_entry->eflags = MAP_ENTRY_WRITECNT; 3930 src_entry->eflags &= ~MAP_ENTRY_WRITECNT; 3931 vm_object_reference(src_object); 3932 fake_entry->object.vm_object = src_object; 3933 fake_entry->start = src_entry->start; 3934 fake_entry->end = src_entry->end; 3935 fake_entry->defer_next = 3936 curthread->td_map_def_user; 3937 curthread->td_map_def_user = fake_entry; 3938 } 3939 3940 pmap_copy(dst_map->pmap, src_map->pmap, 3941 dst_entry->start, dst_entry->end - dst_entry->start, 3942 src_entry->start); 3943 } else { 3944 dst_entry->object.vm_object = NULL; 3945 dst_entry->offset = 0; 3946 if (src_entry->cred != NULL) { 3947 dst_entry->cred = curthread->td_ucred; 3948 crhold(dst_entry->cred); 3949 *fork_charge += size; 3950 } 3951 } 3952 } else { 3953 /* 3954 * We don't want to make writeable wired pages copy-on-write. 3955 * Immediately copy these pages into the new map by simulating 3956 * page faults. The new pages are pageable. 3957 */ 3958 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry, 3959 fork_charge); 3960 } 3961 } 3962 3963 /* 3964 * vmspace_map_entry_forked: 3965 * Update the newly-forked vmspace each time a map entry is inherited 3966 * or copied. The values for vm_dsize and vm_tsize are approximate 3967 * (and mostly-obsolete ideas in the face of mmap(2) et al.) 3968 */ 3969 static void 3970 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2, 3971 vm_map_entry_t entry) 3972 { 3973 vm_size_t entrysize; 3974 vm_offset_t newend; 3975 3976 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 3977 return; 3978 entrysize = entry->end - entry->start; 3979 vm2->vm_map.size += entrysize; 3980 if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) { 3981 vm2->vm_ssize += btoc(entrysize); 3982 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr && 3983 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) { 3984 newend = MIN(entry->end, 3985 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)); 3986 vm2->vm_dsize += btoc(newend - entry->start); 3987 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr && 3988 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) { 3989 newend = MIN(entry->end, 3990 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)); 3991 vm2->vm_tsize += btoc(newend - entry->start); 3992 } 3993 } 3994 3995 /* 3996 * vmspace_fork: 3997 * Create a new process vmspace structure and vm_map 3998 * based on those of an existing process. The new map 3999 * is based on the old map, according to the inheritance 4000 * values on the regions in that map. 4001 * 4002 * XXX It might be worth coalescing the entries added to the new vmspace. 4003 * 4004 * The source map must not be locked. 4005 */ 4006 struct vmspace * 4007 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge) 4008 { 4009 struct vmspace *vm2; 4010 vm_map_t new_map, old_map; 4011 vm_map_entry_t new_entry, old_entry; 4012 vm_object_t object; 4013 int error, locked; 4014 vm_inherit_t inh; 4015 4016 old_map = &vm1->vm_map; 4017 /* Copy immutable fields of vm1 to vm2. */ 4018 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map), 4019 pmap_pinit); 4020 if (vm2 == NULL) 4021 return (NULL); 4022 4023 vm2->vm_taddr = vm1->vm_taddr; 4024 vm2->vm_daddr = vm1->vm_daddr; 4025 vm2->vm_maxsaddr = vm1->vm_maxsaddr; 4026 vm_map_lock(old_map); 4027 if (old_map->busy) 4028 vm_map_wait_busy(old_map); 4029 new_map = &vm2->vm_map; 4030 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */ 4031 KASSERT(locked, ("vmspace_fork: lock failed")); 4032 4033 error = pmap_vmspace_copy(new_map->pmap, old_map->pmap); 4034 if (error != 0) { 4035 sx_xunlock(&old_map->lock); 4036 sx_xunlock(&new_map->lock); 4037 vm_map_process_deferred(); 4038 vmspace_free(vm2); 4039 return (NULL); 4040 } 4041 4042 new_map->anon_loc = old_map->anon_loc; 4043 4044 VM_MAP_ENTRY_FOREACH(old_entry, old_map) { 4045 if ((old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 4046 panic("vm_map_fork: encountered a submap"); 4047 4048 inh = old_entry->inheritance; 4049 if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 && 4050 inh != VM_INHERIT_NONE) 4051 inh = VM_INHERIT_COPY; 4052 4053 switch (inh) { 4054 case VM_INHERIT_NONE: 4055 break; 4056 4057 case VM_INHERIT_SHARE: 4058 /* 4059 * Clone the entry, creating the shared object if 4060 * necessary. 4061 */ 4062 object = old_entry->object.vm_object; 4063 if (object == NULL) { 4064 vm_map_entry_back(old_entry); 4065 object = old_entry->object.vm_object; 4066 } 4067 4068 /* 4069 * Add the reference before calling vm_object_shadow 4070 * to insure that a shadow object is created. 4071 */ 4072 vm_object_reference(object); 4073 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4074 vm_object_shadow(&old_entry->object.vm_object, 4075 &old_entry->offset, 4076 old_entry->end - old_entry->start); 4077 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 4078 /* Transfer the second reference too. */ 4079 vm_object_reference( 4080 old_entry->object.vm_object); 4081 4082 /* 4083 * As in vm_map_merged_neighbor_dispose(), 4084 * the vnode lock will not be acquired in 4085 * this call to vm_object_deallocate(). 4086 */ 4087 vm_object_deallocate(object); 4088 object = old_entry->object.vm_object; 4089 } 4090 VM_OBJECT_WLOCK(object); 4091 vm_object_clear_flag(object, OBJ_ONEMAPPING); 4092 if (old_entry->cred != NULL) { 4093 KASSERT(object->cred == NULL, ("vmspace_fork both cred")); 4094 object->cred = old_entry->cred; 4095 object->charge = old_entry->end - old_entry->start; 4096 old_entry->cred = NULL; 4097 } 4098 4099 /* 4100 * Assert the correct state of the vnode 4101 * v_writecount while the object is locked, to 4102 * not relock it later for the assertion 4103 * correctness. 4104 */ 4105 if (old_entry->eflags & MAP_ENTRY_WRITECNT && 4106 object->type == OBJT_VNODE) { 4107 KASSERT(((struct vnode *)object->handle)-> 4108 v_writecount > 0, 4109 ("vmspace_fork: v_writecount %p", object)); 4110 KASSERT(object->un_pager.vnp.writemappings > 0, 4111 ("vmspace_fork: vnp.writecount %p", 4112 object)); 4113 } 4114 VM_OBJECT_WUNLOCK(object); 4115 4116 /* 4117 * Clone the entry, referencing the shared object. 4118 */ 4119 new_entry = vm_map_entry_create(new_map); 4120 *new_entry = *old_entry; 4121 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4122 MAP_ENTRY_IN_TRANSITION); 4123 new_entry->wiring_thread = NULL; 4124 new_entry->wired_count = 0; 4125 if (new_entry->eflags & MAP_ENTRY_WRITECNT) { 4126 vm_pager_update_writecount(object, 4127 new_entry->start, new_entry->end); 4128 } 4129 vm_map_entry_set_vnode_text(new_entry, true); 4130 4131 /* 4132 * Insert the entry into the new map -- we know we're 4133 * inserting at the end of the new map. 4134 */ 4135 vm_map_entry_link(new_map, new_entry); 4136 vmspace_map_entry_forked(vm1, vm2, new_entry); 4137 4138 /* 4139 * Update the physical map 4140 */ 4141 pmap_copy(new_map->pmap, old_map->pmap, 4142 new_entry->start, 4143 (old_entry->end - old_entry->start), 4144 old_entry->start); 4145 break; 4146 4147 case VM_INHERIT_COPY: 4148 /* 4149 * Clone the entry and link into the map. 4150 */ 4151 new_entry = vm_map_entry_create(new_map); 4152 *new_entry = *old_entry; 4153 /* 4154 * Copied entry is COW over the old object. 4155 */ 4156 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4157 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT); 4158 new_entry->wiring_thread = NULL; 4159 new_entry->wired_count = 0; 4160 new_entry->object.vm_object = NULL; 4161 new_entry->cred = NULL; 4162 vm_map_entry_link(new_map, new_entry); 4163 vmspace_map_entry_forked(vm1, vm2, new_entry); 4164 vm_map_copy_entry(old_map, new_map, old_entry, 4165 new_entry, fork_charge); 4166 vm_map_entry_set_vnode_text(new_entry, true); 4167 break; 4168 4169 case VM_INHERIT_ZERO: 4170 /* 4171 * Create a new anonymous mapping entry modelled from 4172 * the old one. 4173 */ 4174 new_entry = vm_map_entry_create(new_map); 4175 memset(new_entry, 0, sizeof(*new_entry)); 4176 4177 new_entry->start = old_entry->start; 4178 new_entry->end = old_entry->end; 4179 new_entry->eflags = old_entry->eflags & 4180 ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION | 4181 MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC); 4182 new_entry->protection = old_entry->protection; 4183 new_entry->max_protection = old_entry->max_protection; 4184 new_entry->inheritance = VM_INHERIT_ZERO; 4185 4186 vm_map_entry_link(new_map, new_entry); 4187 vmspace_map_entry_forked(vm1, vm2, new_entry); 4188 4189 new_entry->cred = curthread->td_ucred; 4190 crhold(new_entry->cred); 4191 *fork_charge += (new_entry->end - new_entry->start); 4192 4193 break; 4194 } 4195 } 4196 /* 4197 * Use inlined vm_map_unlock() to postpone handling the deferred 4198 * map entries, which cannot be done until both old_map and 4199 * new_map locks are released. 4200 */ 4201 sx_xunlock(&old_map->lock); 4202 sx_xunlock(&new_map->lock); 4203 vm_map_process_deferred(); 4204 4205 return (vm2); 4206 } 4207 4208 /* 4209 * Create a process's stack for exec_new_vmspace(). This function is never 4210 * asked to wire the newly created stack. 4211 */ 4212 int 4213 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4214 vm_prot_t prot, vm_prot_t max, int cow) 4215 { 4216 vm_size_t growsize, init_ssize; 4217 rlim_t vmemlim; 4218 int rv; 4219 4220 MPASS((map->flags & MAP_WIREFUTURE) == 0); 4221 growsize = sgrowsiz; 4222 init_ssize = (max_ssize < growsize) ? max_ssize : growsize; 4223 vm_map_lock(map); 4224 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4225 /* If we would blow our VMEM resource limit, no go */ 4226 if (map->size + init_ssize > vmemlim) { 4227 rv = KERN_NO_SPACE; 4228 goto out; 4229 } 4230 rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot, 4231 max, cow); 4232 out: 4233 vm_map_unlock(map); 4234 return (rv); 4235 } 4236 4237 static int stack_guard_page = 1; 4238 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN, 4239 &stack_guard_page, 0, 4240 "Specifies the number of guard pages for a stack that grows"); 4241 4242 static int 4243 vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4244 vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow) 4245 { 4246 vm_map_entry_t new_entry, prev_entry; 4247 vm_offset_t bot, gap_bot, gap_top, top; 4248 vm_size_t init_ssize, sgp; 4249 int orient, rv; 4250 4251 /* 4252 * The stack orientation is piggybacked with the cow argument. 4253 * Extract it into orient and mask the cow argument so that we 4254 * don't pass it around further. 4255 */ 4256 orient = cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP); 4257 KASSERT(orient != 0, ("No stack grow direction")); 4258 KASSERT(orient != (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP), 4259 ("bi-dir stack")); 4260 4261 if (addrbos < vm_map_min(map) || 4262 addrbos + max_ssize > vm_map_max(map) || 4263 addrbos + max_ssize <= addrbos) 4264 return (KERN_INVALID_ADDRESS); 4265 sgp = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4266 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4267 (vm_size_t)stack_guard_page * PAGE_SIZE; 4268 if (sgp >= max_ssize) 4269 return (KERN_INVALID_ARGUMENT); 4270 4271 init_ssize = growsize; 4272 if (max_ssize < init_ssize + sgp) 4273 init_ssize = max_ssize - sgp; 4274 4275 /* If addr is already mapped, no go */ 4276 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) 4277 return (KERN_NO_SPACE); 4278 4279 /* 4280 * If we can't accommodate max_ssize in the current mapping, no go. 4281 */ 4282 if (vm_map_entry_succ(prev_entry)->start < addrbos + max_ssize) 4283 return (KERN_NO_SPACE); 4284 4285 /* 4286 * We initially map a stack of only init_ssize. We will grow as 4287 * needed later. Depending on the orientation of the stack (i.e. 4288 * the grow direction) we either map at the top of the range, the 4289 * bottom of the range or in the middle. 4290 * 4291 * Note: we would normally expect prot and max to be VM_PROT_ALL, 4292 * and cow to be 0. Possibly we should eliminate these as input 4293 * parameters, and just pass these values here in the insert call. 4294 */ 4295 if (orient == MAP_STACK_GROWS_DOWN) { 4296 bot = addrbos + max_ssize - init_ssize; 4297 top = bot + init_ssize; 4298 gap_bot = addrbos; 4299 gap_top = bot; 4300 } else /* if (orient == MAP_STACK_GROWS_UP) */ { 4301 bot = addrbos; 4302 top = bot + init_ssize; 4303 gap_bot = top; 4304 gap_top = addrbos + max_ssize; 4305 } 4306 rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow); 4307 if (rv != KERN_SUCCESS) 4308 return (rv); 4309 new_entry = vm_map_entry_succ(prev_entry); 4310 KASSERT(new_entry->end == top || new_entry->start == bot, 4311 ("Bad entry start/end for new stack entry")); 4312 KASSERT((orient & MAP_STACK_GROWS_DOWN) == 0 || 4313 (new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0, 4314 ("new entry lacks MAP_ENTRY_GROWS_DOWN")); 4315 KASSERT((orient & MAP_STACK_GROWS_UP) == 0 || 4316 (new_entry->eflags & MAP_ENTRY_GROWS_UP) != 0, 4317 ("new entry lacks MAP_ENTRY_GROWS_UP")); 4318 if (gap_bot == gap_top) 4319 return (KERN_SUCCESS); 4320 rv = vm_map_insert(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE, 4321 VM_PROT_NONE, MAP_CREATE_GUARD | (orient == MAP_STACK_GROWS_DOWN ? 4322 MAP_CREATE_STACK_GAP_DN : MAP_CREATE_STACK_GAP_UP)); 4323 if (rv == KERN_SUCCESS) { 4324 /* 4325 * Gap can never successfully handle a fault, so 4326 * read-ahead logic is never used for it. Re-use 4327 * next_read of the gap entry to store 4328 * stack_guard_page for vm_map_growstack(). 4329 */ 4330 if (orient == MAP_STACK_GROWS_DOWN) 4331 vm_map_entry_pred(new_entry)->next_read = sgp; 4332 else 4333 vm_map_entry_succ(new_entry)->next_read = sgp; 4334 } else { 4335 (void)vm_map_delete(map, bot, top); 4336 } 4337 return (rv); 4338 } 4339 4340 /* 4341 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we 4342 * successfully grow the stack. 4343 */ 4344 static int 4345 vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry) 4346 { 4347 vm_map_entry_t stack_entry; 4348 struct proc *p; 4349 struct vmspace *vm; 4350 struct ucred *cred; 4351 vm_offset_t gap_end, gap_start, grow_start; 4352 vm_size_t grow_amount, guard, max_grow; 4353 rlim_t lmemlim, stacklim, vmemlim; 4354 int rv, rv1; 4355 bool gap_deleted, grow_down, is_procstack; 4356 #ifdef notyet 4357 uint64_t limit; 4358 #endif 4359 #ifdef RACCT 4360 int error; 4361 #endif 4362 4363 p = curproc; 4364 vm = p->p_vmspace; 4365 4366 /* 4367 * Disallow stack growth when the access is performed by a 4368 * debugger or AIO daemon. The reason is that the wrong 4369 * resource limits are applied. 4370 */ 4371 if (p != initproc && (map != &p->p_vmspace->vm_map || 4372 p->p_textvp == NULL)) 4373 return (KERN_FAILURE); 4374 4375 MPASS(!map->system_map); 4376 4377 lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK); 4378 stacklim = lim_cur(curthread, RLIMIT_STACK); 4379 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4380 retry: 4381 /* If addr is not in a hole for a stack grow area, no need to grow. */ 4382 if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry)) 4383 return (KERN_FAILURE); 4384 if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0) 4385 return (KERN_SUCCESS); 4386 if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_DN) != 0) { 4387 stack_entry = vm_map_entry_succ(gap_entry); 4388 if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 || 4389 stack_entry->start != gap_entry->end) 4390 return (KERN_FAILURE); 4391 grow_amount = round_page(stack_entry->start - addr); 4392 grow_down = true; 4393 } else if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_UP) != 0) { 4394 stack_entry = vm_map_entry_pred(gap_entry); 4395 if ((stack_entry->eflags & MAP_ENTRY_GROWS_UP) == 0 || 4396 stack_entry->end != gap_entry->start) 4397 return (KERN_FAILURE); 4398 grow_amount = round_page(addr + 1 - stack_entry->end); 4399 grow_down = false; 4400 } else { 4401 return (KERN_FAILURE); 4402 } 4403 guard = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4404 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4405 gap_entry->next_read; 4406 max_grow = gap_entry->end - gap_entry->start; 4407 if (guard > max_grow) 4408 return (KERN_NO_SPACE); 4409 max_grow -= guard; 4410 if (grow_amount > max_grow) 4411 return (KERN_NO_SPACE); 4412 4413 /* 4414 * If this is the main process stack, see if we're over the stack 4415 * limit. 4416 */ 4417 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr && 4418 addr < (vm_offset_t)p->p_sysent->sv_usrstack; 4419 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) 4420 return (KERN_NO_SPACE); 4421 4422 #ifdef RACCT 4423 if (racct_enable) { 4424 PROC_LOCK(p); 4425 if (is_procstack && racct_set(p, RACCT_STACK, 4426 ctob(vm->vm_ssize) + grow_amount)) { 4427 PROC_UNLOCK(p); 4428 return (KERN_NO_SPACE); 4429 } 4430 PROC_UNLOCK(p); 4431 } 4432 #endif 4433 4434 grow_amount = roundup(grow_amount, sgrowsiz); 4435 if (grow_amount > max_grow) 4436 grow_amount = max_grow; 4437 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) { 4438 grow_amount = trunc_page((vm_size_t)stacklim) - 4439 ctob(vm->vm_ssize); 4440 } 4441 4442 #ifdef notyet 4443 PROC_LOCK(p); 4444 limit = racct_get_available(p, RACCT_STACK); 4445 PROC_UNLOCK(p); 4446 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit)) 4447 grow_amount = limit - ctob(vm->vm_ssize); 4448 #endif 4449 4450 if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) { 4451 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) { 4452 rv = KERN_NO_SPACE; 4453 goto out; 4454 } 4455 #ifdef RACCT 4456 if (racct_enable) { 4457 PROC_LOCK(p); 4458 if (racct_set(p, RACCT_MEMLOCK, 4459 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) { 4460 PROC_UNLOCK(p); 4461 rv = KERN_NO_SPACE; 4462 goto out; 4463 } 4464 PROC_UNLOCK(p); 4465 } 4466 #endif 4467 } 4468 4469 /* If we would blow our VMEM resource limit, no go */ 4470 if (map->size + grow_amount > vmemlim) { 4471 rv = KERN_NO_SPACE; 4472 goto out; 4473 } 4474 #ifdef RACCT 4475 if (racct_enable) { 4476 PROC_LOCK(p); 4477 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) { 4478 PROC_UNLOCK(p); 4479 rv = KERN_NO_SPACE; 4480 goto out; 4481 } 4482 PROC_UNLOCK(p); 4483 } 4484 #endif 4485 4486 if (vm_map_lock_upgrade(map)) { 4487 gap_entry = NULL; 4488 vm_map_lock_read(map); 4489 goto retry; 4490 } 4491 4492 if (grow_down) { 4493 grow_start = gap_entry->end - grow_amount; 4494 if (gap_entry->start + grow_amount == gap_entry->end) { 4495 gap_start = gap_entry->start; 4496 gap_end = gap_entry->end; 4497 vm_map_entry_delete(map, gap_entry); 4498 gap_deleted = true; 4499 } else { 4500 MPASS(gap_entry->start < gap_entry->end - grow_amount); 4501 vm_map_entry_resize(map, gap_entry, -grow_amount); 4502 gap_deleted = false; 4503 } 4504 rv = vm_map_insert(map, NULL, 0, grow_start, 4505 grow_start + grow_amount, 4506 stack_entry->protection, stack_entry->max_protection, 4507 MAP_STACK_GROWS_DOWN); 4508 if (rv != KERN_SUCCESS) { 4509 if (gap_deleted) { 4510 rv1 = vm_map_insert(map, NULL, 0, gap_start, 4511 gap_end, VM_PROT_NONE, VM_PROT_NONE, 4512 MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP_DN); 4513 MPASS(rv1 == KERN_SUCCESS); 4514 } else 4515 vm_map_entry_resize(map, gap_entry, 4516 grow_amount); 4517 } 4518 } else { 4519 grow_start = stack_entry->end; 4520 cred = stack_entry->cred; 4521 if (cred == NULL && stack_entry->object.vm_object != NULL) 4522 cred = stack_entry->object.vm_object->cred; 4523 if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred)) 4524 rv = KERN_NO_SPACE; 4525 /* Grow the underlying object if applicable. */ 4526 else if (stack_entry->object.vm_object == NULL || 4527 vm_object_coalesce(stack_entry->object.vm_object, 4528 stack_entry->offset, 4529 (vm_size_t)(stack_entry->end - stack_entry->start), 4530 grow_amount, cred != NULL)) { 4531 if (gap_entry->start + grow_amount == gap_entry->end) { 4532 vm_map_entry_delete(map, gap_entry); 4533 vm_map_entry_resize(map, stack_entry, 4534 grow_amount); 4535 } else { 4536 gap_entry->start += grow_amount; 4537 stack_entry->end += grow_amount; 4538 } 4539 map->size += grow_amount; 4540 rv = KERN_SUCCESS; 4541 } else 4542 rv = KERN_FAILURE; 4543 } 4544 if (rv == KERN_SUCCESS && is_procstack) 4545 vm->vm_ssize += btoc(grow_amount); 4546 4547 /* 4548 * Heed the MAP_WIREFUTURE flag if it was set for this process. 4549 */ 4550 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) { 4551 rv = vm_map_wire_locked(map, grow_start, 4552 grow_start + grow_amount, 4553 VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); 4554 } 4555 vm_map_lock_downgrade(map); 4556 4557 out: 4558 #ifdef RACCT 4559 if (racct_enable && rv != KERN_SUCCESS) { 4560 PROC_LOCK(p); 4561 error = racct_set(p, RACCT_VMEM, map->size); 4562 KASSERT(error == 0, ("decreasing RACCT_VMEM failed")); 4563 if (!old_mlock) { 4564 error = racct_set(p, RACCT_MEMLOCK, 4565 ptoa(pmap_wired_count(map->pmap))); 4566 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed")); 4567 } 4568 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize)); 4569 KASSERT(error == 0, ("decreasing RACCT_STACK failed")); 4570 PROC_UNLOCK(p); 4571 } 4572 #endif 4573 4574 return (rv); 4575 } 4576 4577 /* 4578 * Unshare the specified VM space for exec. If other processes are 4579 * mapped to it, then create a new one. The new vmspace is null. 4580 */ 4581 int 4582 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser) 4583 { 4584 struct vmspace *oldvmspace = p->p_vmspace; 4585 struct vmspace *newvmspace; 4586 4587 KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0, 4588 ("vmspace_exec recursed")); 4589 newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit); 4590 if (newvmspace == NULL) 4591 return (ENOMEM); 4592 newvmspace->vm_swrss = oldvmspace->vm_swrss; 4593 /* 4594 * This code is written like this for prototype purposes. The 4595 * goal is to avoid running down the vmspace here, but let the 4596 * other process's that are still using the vmspace to finally 4597 * run it down. Even though there is little or no chance of blocking 4598 * here, it is a good idea to keep this form for future mods. 4599 */ 4600 PROC_VMSPACE_LOCK(p); 4601 p->p_vmspace = newvmspace; 4602 PROC_VMSPACE_UNLOCK(p); 4603 if (p == curthread->td_proc) 4604 pmap_activate(curthread); 4605 curthread->td_pflags |= TDP_EXECVMSPC; 4606 return (0); 4607 } 4608 4609 /* 4610 * Unshare the specified VM space for forcing COW. This 4611 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 4612 */ 4613 int 4614 vmspace_unshare(struct proc *p) 4615 { 4616 struct vmspace *oldvmspace = p->p_vmspace; 4617 struct vmspace *newvmspace; 4618 vm_ooffset_t fork_charge; 4619 4620 if (oldvmspace->vm_refcnt == 1) 4621 return (0); 4622 fork_charge = 0; 4623 newvmspace = vmspace_fork(oldvmspace, &fork_charge); 4624 if (newvmspace == NULL) 4625 return (ENOMEM); 4626 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) { 4627 vmspace_free(newvmspace); 4628 return (ENOMEM); 4629 } 4630 PROC_VMSPACE_LOCK(p); 4631 p->p_vmspace = newvmspace; 4632 PROC_VMSPACE_UNLOCK(p); 4633 if (p == curthread->td_proc) 4634 pmap_activate(curthread); 4635 vmspace_free(oldvmspace); 4636 return (0); 4637 } 4638 4639 /* 4640 * vm_map_lookup: 4641 * 4642 * Finds the VM object, offset, and 4643 * protection for a given virtual address in the 4644 * specified map, assuming a page fault of the 4645 * type specified. 4646 * 4647 * Leaves the map in question locked for read; return 4648 * values are guaranteed until a vm_map_lookup_done 4649 * call is performed. Note that the map argument 4650 * is in/out; the returned map must be used in 4651 * the call to vm_map_lookup_done. 4652 * 4653 * A handle (out_entry) is returned for use in 4654 * vm_map_lookup_done, to make that fast. 4655 * 4656 * If a lookup is requested with "write protection" 4657 * specified, the map may be changed to perform virtual 4658 * copying operations, although the data referenced will 4659 * remain the same. 4660 */ 4661 int 4662 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4663 vm_offset_t vaddr, 4664 vm_prot_t fault_typea, 4665 vm_map_entry_t *out_entry, /* OUT */ 4666 vm_object_t *object, /* OUT */ 4667 vm_pindex_t *pindex, /* OUT */ 4668 vm_prot_t *out_prot, /* OUT */ 4669 boolean_t *wired) /* OUT */ 4670 { 4671 vm_map_entry_t entry; 4672 vm_map_t map = *var_map; 4673 vm_prot_t prot; 4674 vm_prot_t fault_type = fault_typea; 4675 vm_object_t eobject; 4676 vm_size_t size; 4677 struct ucred *cred; 4678 4679 RetryLookup: 4680 4681 vm_map_lock_read(map); 4682 4683 RetryLookupLocked: 4684 /* 4685 * Lookup the faulting address. 4686 */ 4687 if (!vm_map_lookup_entry(map, vaddr, out_entry)) { 4688 vm_map_unlock_read(map); 4689 return (KERN_INVALID_ADDRESS); 4690 } 4691 4692 entry = *out_entry; 4693 4694 /* 4695 * Handle submaps. 4696 */ 4697 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 4698 vm_map_t old_map = map; 4699 4700 *var_map = map = entry->object.sub_map; 4701 vm_map_unlock_read(old_map); 4702 goto RetryLookup; 4703 } 4704 4705 /* 4706 * Check whether this task is allowed to have this page. 4707 */ 4708 prot = entry->protection; 4709 if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) { 4710 fault_typea &= ~VM_PROT_FAULT_LOOKUP; 4711 if (prot == VM_PROT_NONE && map != kernel_map && 4712 (entry->eflags & MAP_ENTRY_GUARD) != 0 && 4713 (entry->eflags & (MAP_ENTRY_STACK_GAP_DN | 4714 MAP_ENTRY_STACK_GAP_UP)) != 0 && 4715 vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS) 4716 goto RetryLookupLocked; 4717 } 4718 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 4719 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) { 4720 vm_map_unlock_read(map); 4721 return (KERN_PROTECTION_FAILURE); 4722 } 4723 KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags & 4724 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) != 4725 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY), 4726 ("entry %p flags %x", entry, entry->eflags)); 4727 if ((fault_typea & VM_PROT_COPY) != 0 && 4728 (entry->max_protection & VM_PROT_WRITE) == 0 && 4729 (entry->eflags & MAP_ENTRY_COW) == 0) { 4730 vm_map_unlock_read(map); 4731 return (KERN_PROTECTION_FAILURE); 4732 } 4733 4734 /* 4735 * If this page is not pageable, we have to get it for all possible 4736 * accesses. 4737 */ 4738 *wired = (entry->wired_count != 0); 4739 if (*wired) 4740 fault_type = entry->protection; 4741 size = entry->end - entry->start; 4742 /* 4743 * If the entry was copy-on-write, we either ... 4744 */ 4745 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4746 /* 4747 * If we want to write the page, we may as well handle that 4748 * now since we've got the map locked. 4749 * 4750 * If we don't need to write the page, we just demote the 4751 * permissions allowed. 4752 */ 4753 if ((fault_type & VM_PROT_WRITE) != 0 || 4754 (fault_typea & VM_PROT_COPY) != 0) { 4755 /* 4756 * Make a new object, and place it in the object 4757 * chain. Note that no new references have appeared 4758 * -- one just moved from the map to the new 4759 * object. 4760 */ 4761 if (vm_map_lock_upgrade(map)) 4762 goto RetryLookup; 4763 4764 if (entry->cred == NULL) { 4765 /* 4766 * The debugger owner is charged for 4767 * the memory. 4768 */ 4769 cred = curthread->td_ucred; 4770 crhold(cred); 4771 if (!swap_reserve_by_cred(size, cred)) { 4772 crfree(cred); 4773 vm_map_unlock(map); 4774 return (KERN_RESOURCE_SHORTAGE); 4775 } 4776 entry->cred = cred; 4777 } 4778 vm_object_shadow(&entry->object.vm_object, 4779 &entry->offset, size); 4780 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 4781 eobject = entry->object.vm_object; 4782 if (eobject->cred != NULL) { 4783 /* 4784 * The object was not shadowed. 4785 */ 4786 swap_release_by_cred(size, entry->cred); 4787 crfree(entry->cred); 4788 entry->cred = NULL; 4789 } else if (entry->cred != NULL) { 4790 VM_OBJECT_WLOCK(eobject); 4791 eobject->cred = entry->cred; 4792 eobject->charge = size; 4793 VM_OBJECT_WUNLOCK(eobject); 4794 entry->cred = NULL; 4795 } 4796 4797 vm_map_lock_downgrade(map); 4798 } else { 4799 /* 4800 * We're attempting to read a copy-on-write page -- 4801 * don't allow writes. 4802 */ 4803 prot &= ~VM_PROT_WRITE; 4804 } 4805 } 4806 4807 /* 4808 * Create an object if necessary. 4809 */ 4810 if (entry->object.vm_object == NULL && 4811 !map->system_map) { 4812 if (vm_map_lock_upgrade(map)) 4813 goto RetryLookup; 4814 entry->object.vm_object = vm_object_allocate_anon(atop(size)); 4815 entry->offset = 0; 4816 if (entry->cred != NULL) { 4817 VM_OBJECT_WLOCK(entry->object.vm_object); 4818 entry->object.vm_object->cred = entry->cred; 4819 entry->object.vm_object->charge = size; 4820 VM_OBJECT_WUNLOCK(entry->object.vm_object); 4821 entry->cred = NULL; 4822 } 4823 vm_map_lock_downgrade(map); 4824 } 4825 4826 /* 4827 * Return the object/offset from this entry. If the entry was 4828 * copy-on-write or empty, it has been fixed up. 4829 */ 4830 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 4831 *object = entry->object.vm_object; 4832 4833 *out_prot = prot; 4834 return (KERN_SUCCESS); 4835 } 4836 4837 /* 4838 * vm_map_lookup_locked: 4839 * 4840 * Lookup the faulting address. A version of vm_map_lookup that returns 4841 * KERN_FAILURE instead of blocking on map lock or memory allocation. 4842 */ 4843 int 4844 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */ 4845 vm_offset_t vaddr, 4846 vm_prot_t fault_typea, 4847 vm_map_entry_t *out_entry, /* OUT */ 4848 vm_object_t *object, /* OUT */ 4849 vm_pindex_t *pindex, /* OUT */ 4850 vm_prot_t *out_prot, /* OUT */ 4851 boolean_t *wired) /* OUT */ 4852 { 4853 vm_map_entry_t entry; 4854 vm_map_t map = *var_map; 4855 vm_prot_t prot; 4856 vm_prot_t fault_type = fault_typea; 4857 4858 /* 4859 * Lookup the faulting address. 4860 */ 4861 if (!vm_map_lookup_entry(map, vaddr, out_entry)) 4862 return (KERN_INVALID_ADDRESS); 4863 4864 entry = *out_entry; 4865 4866 /* 4867 * Fail if the entry refers to a submap. 4868 */ 4869 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) 4870 return (KERN_FAILURE); 4871 4872 /* 4873 * Check whether this task is allowed to have this page. 4874 */ 4875 prot = entry->protection; 4876 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 4877 if ((fault_type & prot) != fault_type) 4878 return (KERN_PROTECTION_FAILURE); 4879 4880 /* 4881 * If this page is not pageable, we have to get it for all possible 4882 * accesses. 4883 */ 4884 *wired = (entry->wired_count != 0); 4885 if (*wired) 4886 fault_type = entry->protection; 4887 4888 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4889 /* 4890 * Fail if the entry was copy-on-write for a write fault. 4891 */ 4892 if (fault_type & VM_PROT_WRITE) 4893 return (KERN_FAILURE); 4894 /* 4895 * We're attempting to read a copy-on-write page -- 4896 * don't allow writes. 4897 */ 4898 prot &= ~VM_PROT_WRITE; 4899 } 4900 4901 /* 4902 * Fail if an object should be created. 4903 */ 4904 if (entry->object.vm_object == NULL && !map->system_map) 4905 return (KERN_FAILURE); 4906 4907 /* 4908 * Return the object/offset from this entry. If the entry was 4909 * copy-on-write or empty, it has been fixed up. 4910 */ 4911 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 4912 *object = entry->object.vm_object; 4913 4914 *out_prot = prot; 4915 return (KERN_SUCCESS); 4916 } 4917 4918 /* 4919 * vm_map_lookup_done: 4920 * 4921 * Releases locks acquired by a vm_map_lookup 4922 * (according to the handle returned by that lookup). 4923 */ 4924 void 4925 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry) 4926 { 4927 /* 4928 * Unlock the main-level map 4929 */ 4930 vm_map_unlock_read(map); 4931 } 4932 4933 vm_offset_t 4934 vm_map_max_KBI(const struct vm_map *map) 4935 { 4936 4937 return (vm_map_max(map)); 4938 } 4939 4940 vm_offset_t 4941 vm_map_min_KBI(const struct vm_map *map) 4942 { 4943 4944 return (vm_map_min(map)); 4945 } 4946 4947 pmap_t 4948 vm_map_pmap_KBI(vm_map_t map) 4949 { 4950 4951 return (map->pmap); 4952 } 4953 4954 #ifdef INVARIANTS 4955 static void 4956 _vm_map_assert_consistent(vm_map_t map, int check) 4957 { 4958 vm_map_entry_t entry, prev; 4959 vm_size_t max_left, max_right; 4960 4961 #ifdef DIAGNOSTIC 4962 ++map->nupdates; 4963 #endif 4964 if (enable_vmmap_check != check) 4965 return; 4966 4967 prev = &map->header; 4968 VM_MAP_ENTRY_FOREACH(entry, map) { 4969 KASSERT(prev->end <= entry->start, 4970 ("map %p prev->end = %jx, start = %jx", map, 4971 (uintmax_t)prev->end, (uintmax_t)entry->start)); 4972 KASSERT(entry->start < entry->end, 4973 ("map %p start = %jx, end = %jx", map, 4974 (uintmax_t)entry->start, (uintmax_t)entry->end)); 4975 KASSERT(entry->end <= vm_map_entry_succ(entry)->start, 4976 ("map %p end = %jx, next->start = %jx", map, 4977 (uintmax_t)entry->end, 4978 (uintmax_t)vm_map_entry_succ(entry)->start)); 4979 KASSERT(entry->left == NULL || 4980 entry->left->start < entry->start, 4981 ("map %p left->start = %jx, start = %jx", map, 4982 (uintmax_t)entry->left->start, (uintmax_t)entry->start)); 4983 KASSERT(entry->right == NULL || 4984 entry->start < entry->right->start, 4985 ("map %p start = %jx, right->start = %jx", map, 4986 (uintmax_t)entry->start, (uintmax_t)entry->right->start)); 4987 max_left = vm_map_entry_max_free_left(entry, 4988 vm_map_entry_pred(entry)); 4989 max_right = vm_map_entry_max_free_right(entry, 4990 vm_map_entry_succ(entry)); 4991 KASSERT(entry->max_free == MAX(max_left, max_right), 4992 ("map %p max = %jx, max_left = %jx, max_right = %jx", map, 4993 (uintmax_t)entry->max_free, 4994 (uintmax_t)max_left, (uintmax_t)max_right)); 4995 prev = entry; 4996 } 4997 KASSERT(prev->end <= entry->start, 4998 ("map %p prev->end = %jx, start = %jx", map, 4999 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5000 } 5001 #endif 5002 5003 #include "opt_ddb.h" 5004 #ifdef DDB 5005 #include <sys/kernel.h> 5006 5007 #include <ddb/ddb.h> 5008 5009 static void 5010 vm_map_print(vm_map_t map) 5011 { 5012 vm_map_entry_t entry, prev; 5013 5014 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 5015 (void *)map, 5016 (void *)map->pmap, map->nentries, map->timestamp); 5017 5018 db_indent += 2; 5019 prev = &map->header; 5020 VM_MAP_ENTRY_FOREACH(entry, map) { 5021 db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n", 5022 (void *)entry, (void *)entry->start, (void *)entry->end, 5023 entry->eflags); 5024 { 5025 static char *inheritance_name[4] = 5026 {"share", "copy", "none", "donate_copy"}; 5027 5028 db_iprintf(" prot=%x/%x/%s", 5029 entry->protection, 5030 entry->max_protection, 5031 inheritance_name[(int)(unsigned char) 5032 entry->inheritance]); 5033 if (entry->wired_count != 0) 5034 db_printf(", wired"); 5035 } 5036 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 5037 db_printf(", share=%p, offset=0x%jx\n", 5038 (void *)entry->object.sub_map, 5039 (uintmax_t)entry->offset); 5040 if (prev == &map->header || 5041 prev->object.sub_map != 5042 entry->object.sub_map) { 5043 db_indent += 2; 5044 vm_map_print((vm_map_t)entry->object.sub_map); 5045 db_indent -= 2; 5046 } 5047 } else { 5048 if (entry->cred != NULL) 5049 db_printf(", ruid %d", entry->cred->cr_ruid); 5050 db_printf(", object=%p, offset=0x%jx", 5051 (void *)entry->object.vm_object, 5052 (uintmax_t)entry->offset); 5053 if (entry->object.vm_object && entry->object.vm_object->cred) 5054 db_printf(", obj ruid %d charge %jx", 5055 entry->object.vm_object->cred->cr_ruid, 5056 (uintmax_t)entry->object.vm_object->charge); 5057 if (entry->eflags & MAP_ENTRY_COW) 5058 db_printf(", copy (%s)", 5059 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 5060 db_printf("\n"); 5061 5062 if (prev == &map->header || 5063 prev->object.vm_object != 5064 entry->object.vm_object) { 5065 db_indent += 2; 5066 vm_object_print((db_expr_t)(intptr_t) 5067 entry->object.vm_object, 5068 0, 0, (char *)0); 5069 db_indent -= 2; 5070 } 5071 } 5072 prev = entry; 5073 } 5074 db_indent -= 2; 5075 } 5076 5077 DB_SHOW_COMMAND(map, map) 5078 { 5079 5080 if (!have_addr) { 5081 db_printf("usage: show map <addr>\n"); 5082 return; 5083 } 5084 vm_map_print((vm_map_t)addr); 5085 } 5086 5087 DB_SHOW_COMMAND(procvm, procvm) 5088 { 5089 struct proc *p; 5090 5091 if (have_addr) { 5092 p = db_lookup_proc(addr); 5093 } else { 5094 p = curproc; 5095 } 5096 5097 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 5098 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 5099 (void *)vmspace_pmap(p->p_vmspace)); 5100 5101 vm_map_print((vm_map_t)&p->p_vmspace->vm_map); 5102 } 5103 5104 #endif /* DDB */ 5105