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 vmspace_zone; 132 static int vmspace_zinit(void *mem, int size, int flags); 133 static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, 134 vm_offset_t max); 135 static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map); 136 static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry); 137 static void vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry); 138 static int vm_map_growstack(vm_map_t map, vm_offset_t addr, 139 vm_map_entry_t gap_entry); 140 static void vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 141 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags); 142 #ifdef INVARIANTS 143 static void vmspace_zdtor(void *mem, int size, void *arg); 144 #endif 145 static int vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, 146 vm_size_t max_ssize, vm_size_t growsize, vm_prot_t prot, vm_prot_t max, 147 int cow); 148 static void vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 149 vm_offset_t failed_addr); 150 151 #define ENTRY_CHARGED(e) ((e)->cred != NULL || \ 152 ((e)->object.vm_object != NULL && (e)->object.vm_object->cred != NULL && \ 153 !((e)->eflags & MAP_ENTRY_NEEDS_COPY))) 154 155 /* 156 * PROC_VMSPACE_{UN,}LOCK() can be a noop as long as vmspaces are type 157 * stable. 158 */ 159 #define PROC_VMSPACE_LOCK(p) do { } while (0) 160 #define PROC_VMSPACE_UNLOCK(p) do { } while (0) 161 162 /* 163 * VM_MAP_RANGE_CHECK: [ internal use only ] 164 * 165 * Asserts that the starting and ending region 166 * addresses fall within the valid range of the map. 167 */ 168 #define VM_MAP_RANGE_CHECK(map, start, end) \ 169 { \ 170 if (start < vm_map_min(map)) \ 171 start = vm_map_min(map); \ 172 if (end > vm_map_max(map)) \ 173 end = vm_map_max(map); \ 174 if (start > end) \ 175 start = end; \ 176 } 177 178 #ifndef UMA_MD_SMALL_ALLOC 179 180 /* 181 * Allocate a new slab for kernel map entries. The kernel map may be locked or 182 * unlocked, depending on whether the request is coming from the kernel map or a 183 * submap. This function allocates a virtual address range directly from the 184 * kernel map instead of the kmem_* layer to avoid recursion on the kernel map 185 * lock and also to avoid triggering allocator recursion in the vmem boundary 186 * tag allocator. 187 */ 188 static void * 189 kmapent_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 190 int wait) 191 { 192 vm_offset_t addr; 193 int error, locked; 194 195 *pflag = UMA_SLAB_PRIV; 196 197 if (!(locked = vm_map_locked(kernel_map))) 198 vm_map_lock(kernel_map); 199 addr = vm_map_findspace(kernel_map, vm_map_min(kernel_map), bytes); 200 if (addr + bytes < addr || addr + bytes > vm_map_max(kernel_map)) 201 panic("%s: kernel map is exhausted", __func__); 202 error = vm_map_insert(kernel_map, NULL, 0, addr, addr + bytes, 203 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 204 if (error != KERN_SUCCESS) 205 panic("%s: vm_map_insert() failed: %d", __func__, error); 206 if (!locked) 207 vm_map_unlock(kernel_map); 208 error = kmem_back_domain(domain, kernel_object, addr, bytes, M_NOWAIT | 209 M_USE_RESERVE | (wait & M_ZERO)); 210 if (error == KERN_SUCCESS) { 211 return ((void *)addr); 212 } else { 213 if (!locked) 214 vm_map_lock(kernel_map); 215 vm_map_delete(kernel_map, addr, bytes); 216 if (!locked) 217 vm_map_unlock(kernel_map); 218 return (NULL); 219 } 220 } 221 222 static void 223 kmapent_free(void *item, vm_size_t size, uint8_t pflag) 224 { 225 vm_offset_t addr; 226 int error __diagused; 227 228 if ((pflag & UMA_SLAB_PRIV) == 0) 229 /* XXX leaked */ 230 return; 231 232 addr = (vm_offset_t)item; 233 kmem_unback(kernel_object, addr, size); 234 error = vm_map_remove(kernel_map, addr, addr + size); 235 KASSERT(error == KERN_SUCCESS, 236 ("%s: vm_map_remove failed: %d", __func__, error)); 237 } 238 239 /* 240 * The worst-case upper bound on the number of kernel map entries that may be 241 * created before the zone must be replenished in _vm_map_unlock(). 242 */ 243 #define KMAPENT_RESERVE 1 244 245 #endif /* !UMD_MD_SMALL_ALLOC */ 246 247 /* 248 * vm_map_startup: 249 * 250 * Initialize the vm_map module. Must be called before any other vm_map 251 * routines. 252 * 253 * User map and entry structures are allocated from the general purpose 254 * memory pool. Kernel maps are statically defined. Kernel map entries 255 * require special handling to avoid recursion; see the comments above 256 * kmapent_alloc() and in vm_map_entry_create(). 257 */ 258 void 259 vm_map_startup(void) 260 { 261 mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF); 262 263 /* 264 * Disable the use of per-CPU buckets: map entry allocation is 265 * serialized by the kernel map lock. 266 */ 267 kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry), 268 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 269 UMA_ZONE_VM | UMA_ZONE_NOBUCKET); 270 #ifndef UMA_MD_SMALL_ALLOC 271 /* Reserve an extra map entry for use when replenishing the reserve. */ 272 uma_zone_reserve(kmapentzone, KMAPENT_RESERVE + 1); 273 uma_prealloc(kmapentzone, KMAPENT_RESERVE + 1); 274 uma_zone_set_allocf(kmapentzone, kmapent_alloc); 275 uma_zone_set_freef(kmapentzone, kmapent_free); 276 #endif 277 278 mapentzone = uma_zcreate("MAP ENTRY", sizeof(struct vm_map_entry), 279 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 280 vmspace_zone = uma_zcreate("VMSPACE", sizeof(struct vmspace), NULL, 281 #ifdef INVARIANTS 282 vmspace_zdtor, 283 #else 284 NULL, 285 #endif 286 vmspace_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 287 } 288 289 static int 290 vmspace_zinit(void *mem, int size, int flags) 291 { 292 struct vmspace *vm; 293 vm_map_t map; 294 295 vm = (struct vmspace *)mem; 296 map = &vm->vm_map; 297 298 memset(map, 0, sizeof(*map)); 299 mtx_init(&map->system_mtx, "vm map (system)", NULL, 300 MTX_DEF | MTX_DUPOK); 301 sx_init(&map->lock, "vm map (user)"); 302 PMAP_LOCK_INIT(vmspace_pmap(vm)); 303 return (0); 304 } 305 306 #ifdef INVARIANTS 307 static void 308 vmspace_zdtor(void *mem, int size, void *arg) 309 { 310 struct vmspace *vm; 311 312 vm = (struct vmspace *)mem; 313 KASSERT(vm->vm_map.nentries == 0, 314 ("vmspace %p nentries == %d on free", vm, vm->vm_map.nentries)); 315 KASSERT(vm->vm_map.size == 0, 316 ("vmspace %p size == %ju on free", vm, (uintmax_t)vm->vm_map.size)); 317 } 318 #endif /* INVARIANTS */ 319 320 /* 321 * Allocate a vmspace structure, including a vm_map and pmap, 322 * and initialize those structures. The refcnt is set to 1. 323 */ 324 struct vmspace * 325 vmspace_alloc(vm_offset_t min, vm_offset_t max, pmap_pinit_t pinit) 326 { 327 struct vmspace *vm; 328 329 vm = uma_zalloc(vmspace_zone, M_WAITOK); 330 KASSERT(vm->vm_map.pmap == NULL, ("vm_map.pmap must be NULL")); 331 if (!pinit(vmspace_pmap(vm))) { 332 uma_zfree(vmspace_zone, vm); 333 return (NULL); 334 } 335 CTR1(KTR_VM, "vmspace_alloc: %p", vm); 336 _vm_map_init(&vm->vm_map, vmspace_pmap(vm), min, max); 337 refcount_init(&vm->vm_refcnt, 1); 338 vm->vm_shm = NULL; 339 vm->vm_swrss = 0; 340 vm->vm_tsize = 0; 341 vm->vm_dsize = 0; 342 vm->vm_ssize = 0; 343 vm->vm_taddr = 0; 344 vm->vm_daddr = 0; 345 vm->vm_maxsaddr = 0; 346 return (vm); 347 } 348 349 #ifdef RACCT 350 static void 351 vmspace_container_reset(struct proc *p) 352 { 353 354 PROC_LOCK(p); 355 racct_set(p, RACCT_DATA, 0); 356 racct_set(p, RACCT_STACK, 0); 357 racct_set(p, RACCT_RSS, 0); 358 racct_set(p, RACCT_MEMLOCK, 0); 359 racct_set(p, RACCT_VMEM, 0); 360 PROC_UNLOCK(p); 361 } 362 #endif 363 364 static inline void 365 vmspace_dofree(struct vmspace *vm) 366 { 367 368 CTR1(KTR_VM, "vmspace_free: %p", vm); 369 370 /* 371 * Make sure any SysV shm is freed, it might not have been in 372 * exit1(). 373 */ 374 shmexit(vm); 375 376 /* 377 * Lock the map, to wait out all other references to it. 378 * Delete all of the mappings and pages they hold, then call 379 * the pmap module to reclaim anything left. 380 */ 381 (void)vm_map_remove(&vm->vm_map, vm_map_min(&vm->vm_map), 382 vm_map_max(&vm->vm_map)); 383 384 pmap_release(vmspace_pmap(vm)); 385 vm->vm_map.pmap = NULL; 386 uma_zfree(vmspace_zone, vm); 387 } 388 389 void 390 vmspace_free(struct vmspace *vm) 391 { 392 393 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 394 "vmspace_free() called"); 395 396 if (refcount_release(&vm->vm_refcnt)) 397 vmspace_dofree(vm); 398 } 399 400 void 401 vmspace_exitfree(struct proc *p) 402 { 403 struct vmspace *vm; 404 405 PROC_VMSPACE_LOCK(p); 406 vm = p->p_vmspace; 407 p->p_vmspace = NULL; 408 PROC_VMSPACE_UNLOCK(p); 409 KASSERT(vm == &vmspace0, ("vmspace_exitfree: wrong vmspace")); 410 vmspace_free(vm); 411 } 412 413 void 414 vmspace_exit(struct thread *td) 415 { 416 struct vmspace *vm; 417 struct proc *p; 418 bool released; 419 420 p = td->td_proc; 421 vm = p->p_vmspace; 422 423 /* 424 * Prepare to release the vmspace reference. The thread that releases 425 * the last reference is responsible for tearing down the vmspace. 426 * However, threads not releasing the final reference must switch to the 427 * kernel's vmspace0 before the decrement so that the subsequent pmap 428 * deactivation does not modify a freed vmspace. 429 */ 430 refcount_acquire(&vmspace0.vm_refcnt); 431 if (!(released = refcount_release_if_last(&vm->vm_refcnt))) { 432 if (p->p_vmspace != &vmspace0) { 433 PROC_VMSPACE_LOCK(p); 434 p->p_vmspace = &vmspace0; 435 PROC_VMSPACE_UNLOCK(p); 436 pmap_activate(td); 437 } 438 released = refcount_release(&vm->vm_refcnt); 439 } 440 if (released) { 441 /* 442 * pmap_remove_pages() expects the pmap to be active, so switch 443 * back first if necessary. 444 */ 445 if (p->p_vmspace != vm) { 446 PROC_VMSPACE_LOCK(p); 447 p->p_vmspace = vm; 448 PROC_VMSPACE_UNLOCK(p); 449 pmap_activate(td); 450 } 451 pmap_remove_pages(vmspace_pmap(vm)); 452 PROC_VMSPACE_LOCK(p); 453 p->p_vmspace = &vmspace0; 454 PROC_VMSPACE_UNLOCK(p); 455 pmap_activate(td); 456 vmspace_dofree(vm); 457 } 458 #ifdef RACCT 459 if (racct_enable) 460 vmspace_container_reset(p); 461 #endif 462 } 463 464 /* Acquire reference to vmspace owned by another process. */ 465 466 struct vmspace * 467 vmspace_acquire_ref(struct proc *p) 468 { 469 struct vmspace *vm; 470 471 PROC_VMSPACE_LOCK(p); 472 vm = p->p_vmspace; 473 if (vm == NULL || !refcount_acquire_if_not_zero(&vm->vm_refcnt)) { 474 PROC_VMSPACE_UNLOCK(p); 475 return (NULL); 476 } 477 if (vm != p->p_vmspace) { 478 PROC_VMSPACE_UNLOCK(p); 479 vmspace_free(vm); 480 return (NULL); 481 } 482 PROC_VMSPACE_UNLOCK(p); 483 return (vm); 484 } 485 486 /* 487 * Switch between vmspaces in an AIO kernel process. 488 * 489 * The new vmspace is either the vmspace of a user process obtained 490 * from an active AIO request or the initial vmspace of the AIO kernel 491 * process (when it is idling). Because user processes will block to 492 * drain any active AIO requests before proceeding in exit() or 493 * execve(), the reference count for vmspaces from AIO requests can 494 * never be 0. Similarly, AIO kernel processes hold an extra 495 * reference on their initial vmspace for the life of the process. As 496 * a result, the 'newvm' vmspace always has a non-zero reference 497 * count. This permits an additional reference on 'newvm' to be 498 * acquired via a simple atomic increment rather than the loop in 499 * vmspace_acquire_ref() above. 500 */ 501 void 502 vmspace_switch_aio(struct vmspace *newvm) 503 { 504 struct vmspace *oldvm; 505 506 /* XXX: Need some way to assert that this is an aio daemon. */ 507 508 KASSERT(refcount_load(&newvm->vm_refcnt) > 0, 509 ("vmspace_switch_aio: newvm unreferenced")); 510 511 oldvm = curproc->p_vmspace; 512 if (oldvm == newvm) 513 return; 514 515 /* 516 * Point to the new address space and refer to it. 517 */ 518 curproc->p_vmspace = newvm; 519 refcount_acquire(&newvm->vm_refcnt); 520 521 /* Activate the new mapping. */ 522 pmap_activate(curthread); 523 524 vmspace_free(oldvm); 525 } 526 527 void 528 _vm_map_lock(vm_map_t map, const char *file, int line) 529 { 530 531 if (map->system_map) 532 mtx_lock_flags_(&map->system_mtx, 0, file, line); 533 else 534 sx_xlock_(&map->lock, file, line); 535 map->timestamp++; 536 } 537 538 void 539 vm_map_entry_set_vnode_text(vm_map_entry_t entry, bool add) 540 { 541 vm_object_t object; 542 struct vnode *vp; 543 bool vp_held; 544 545 if ((entry->eflags & MAP_ENTRY_VN_EXEC) == 0) 546 return; 547 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 548 ("Submap with execs")); 549 object = entry->object.vm_object; 550 KASSERT(object != NULL, ("No object for text, entry %p", entry)); 551 if ((object->flags & OBJ_ANON) != 0) 552 object = object->handle; 553 else 554 KASSERT(object->backing_object == NULL, 555 ("non-anon object %p shadows", object)); 556 KASSERT(object != NULL, ("No content object for text, entry %p obj %p", 557 entry, entry->object.vm_object)); 558 559 /* 560 * Mostly, we do not lock the backing object. It is 561 * referenced by the entry we are processing, so it cannot go 562 * away. 563 */ 564 vm_pager_getvp(object, &vp, &vp_held); 565 if (vp != NULL) { 566 if (add) { 567 VOP_SET_TEXT_CHECKED(vp); 568 } else { 569 vn_lock(vp, LK_SHARED | LK_RETRY); 570 VOP_UNSET_TEXT_CHECKED(vp); 571 VOP_UNLOCK(vp); 572 } 573 if (vp_held) 574 vdrop(vp); 575 } 576 } 577 578 /* 579 * Use a different name for this vm_map_entry field when it's use 580 * is not consistent with its use as part of an ordered search tree. 581 */ 582 #define defer_next right 583 584 static void 585 vm_map_process_deferred(void) 586 { 587 struct thread *td; 588 vm_map_entry_t entry, next; 589 vm_object_t object; 590 591 td = curthread; 592 entry = td->td_map_def_user; 593 td->td_map_def_user = NULL; 594 while (entry != NULL) { 595 next = entry->defer_next; 596 MPASS((entry->eflags & (MAP_ENTRY_WRITECNT | 597 MAP_ENTRY_VN_EXEC)) != (MAP_ENTRY_WRITECNT | 598 MAP_ENTRY_VN_EXEC)); 599 if ((entry->eflags & MAP_ENTRY_WRITECNT) != 0) { 600 /* 601 * Decrement the object's writemappings and 602 * possibly the vnode's v_writecount. 603 */ 604 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 605 ("Submap with writecount")); 606 object = entry->object.vm_object; 607 KASSERT(object != NULL, ("No object for writecount")); 608 vm_pager_release_writecount(object, entry->start, 609 entry->end); 610 } 611 vm_map_entry_set_vnode_text(entry, false); 612 vm_map_entry_deallocate(entry, FALSE); 613 entry = next; 614 } 615 } 616 617 #ifdef INVARIANTS 618 static void 619 _vm_map_assert_locked(vm_map_t map, const char *file, int line) 620 { 621 622 if (map->system_map) 623 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 624 else 625 sx_assert_(&map->lock, SA_XLOCKED, file, line); 626 } 627 628 #define VM_MAP_ASSERT_LOCKED(map) \ 629 _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE) 630 631 enum { VMMAP_CHECK_NONE, VMMAP_CHECK_UNLOCK, VMMAP_CHECK_ALL }; 632 #ifdef DIAGNOSTIC 633 static int enable_vmmap_check = VMMAP_CHECK_UNLOCK; 634 #else 635 static int enable_vmmap_check = VMMAP_CHECK_NONE; 636 #endif 637 SYSCTL_INT(_debug, OID_AUTO, vmmap_check, CTLFLAG_RWTUN, 638 &enable_vmmap_check, 0, "Enable vm map consistency checking"); 639 640 static void _vm_map_assert_consistent(vm_map_t map, int check); 641 642 #define VM_MAP_ASSERT_CONSISTENT(map) \ 643 _vm_map_assert_consistent(map, VMMAP_CHECK_ALL) 644 #ifdef DIAGNOSTIC 645 #define VM_MAP_UNLOCK_CONSISTENT(map) do { \ 646 if (map->nupdates > map->nentries) { \ 647 _vm_map_assert_consistent(map, VMMAP_CHECK_UNLOCK); \ 648 map->nupdates = 0; \ 649 } \ 650 } while (0) 651 #else 652 #define VM_MAP_UNLOCK_CONSISTENT(map) 653 #endif 654 #else 655 #define VM_MAP_ASSERT_LOCKED(map) 656 #define VM_MAP_ASSERT_CONSISTENT(map) 657 #define VM_MAP_UNLOCK_CONSISTENT(map) 658 #endif /* INVARIANTS */ 659 660 void 661 _vm_map_unlock(vm_map_t map, const char *file, int line) 662 { 663 664 VM_MAP_UNLOCK_CONSISTENT(map); 665 if (map->system_map) { 666 #ifndef UMA_MD_SMALL_ALLOC 667 if (map == kernel_map && (map->flags & MAP_REPLENISH) != 0) { 668 uma_prealloc(kmapentzone, 1); 669 map->flags &= ~MAP_REPLENISH; 670 } 671 #endif 672 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 673 } else { 674 sx_xunlock_(&map->lock, file, line); 675 vm_map_process_deferred(); 676 } 677 } 678 679 void 680 _vm_map_lock_read(vm_map_t map, const char *file, int line) 681 { 682 683 if (map->system_map) 684 mtx_lock_flags_(&map->system_mtx, 0, file, line); 685 else 686 sx_slock_(&map->lock, file, line); 687 } 688 689 void 690 _vm_map_unlock_read(vm_map_t map, const char *file, int line) 691 { 692 693 if (map->system_map) { 694 KASSERT((map->flags & MAP_REPLENISH) == 0, 695 ("%s: MAP_REPLENISH leaked", __func__)); 696 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 697 } else { 698 sx_sunlock_(&map->lock, file, line); 699 vm_map_process_deferred(); 700 } 701 } 702 703 int 704 _vm_map_trylock(vm_map_t map, const char *file, int line) 705 { 706 int error; 707 708 error = map->system_map ? 709 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 710 !sx_try_xlock_(&map->lock, file, line); 711 if (error == 0) 712 map->timestamp++; 713 return (error == 0); 714 } 715 716 int 717 _vm_map_trylock_read(vm_map_t map, const char *file, int line) 718 { 719 int error; 720 721 error = map->system_map ? 722 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 723 !sx_try_slock_(&map->lock, file, line); 724 return (error == 0); 725 } 726 727 /* 728 * _vm_map_lock_upgrade: [ internal use only ] 729 * 730 * Tries to upgrade a read (shared) lock on the specified map to a write 731 * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a 732 * non-zero value if the upgrade fails. If the upgrade fails, the map is 733 * returned without a read or write lock held. 734 * 735 * Requires that the map be read locked. 736 */ 737 int 738 _vm_map_lock_upgrade(vm_map_t map, const char *file, int line) 739 { 740 unsigned int last_timestamp; 741 742 if (map->system_map) { 743 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 744 } else { 745 if (!sx_try_upgrade_(&map->lock, file, line)) { 746 last_timestamp = map->timestamp; 747 sx_sunlock_(&map->lock, file, line); 748 vm_map_process_deferred(); 749 /* 750 * If the map's timestamp does not change while the 751 * map is unlocked, then the upgrade succeeds. 752 */ 753 sx_xlock_(&map->lock, file, line); 754 if (last_timestamp != map->timestamp) { 755 sx_xunlock_(&map->lock, file, line); 756 return (1); 757 } 758 } 759 } 760 map->timestamp++; 761 return (0); 762 } 763 764 void 765 _vm_map_lock_downgrade(vm_map_t map, const char *file, int line) 766 { 767 768 if (map->system_map) { 769 KASSERT((map->flags & MAP_REPLENISH) == 0, 770 ("%s: MAP_REPLENISH leaked", __func__)); 771 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 772 } else { 773 VM_MAP_UNLOCK_CONSISTENT(map); 774 sx_downgrade_(&map->lock, file, line); 775 } 776 } 777 778 /* 779 * vm_map_locked: 780 * 781 * Returns a non-zero value if the caller holds a write (exclusive) lock 782 * on the specified map and the value "0" otherwise. 783 */ 784 int 785 vm_map_locked(vm_map_t map) 786 { 787 788 if (map->system_map) 789 return (mtx_owned(&map->system_mtx)); 790 else 791 return (sx_xlocked(&map->lock)); 792 } 793 794 /* 795 * _vm_map_unlock_and_wait: 796 * 797 * Atomically releases the lock on the specified map and puts the calling 798 * thread to sleep. The calling thread will remain asleep until either 799 * vm_map_wakeup() is performed on the map or the specified timeout is 800 * exceeded. 801 * 802 * WARNING! This function does not perform deferred deallocations of 803 * objects and map entries. Therefore, the calling thread is expected to 804 * reacquire the map lock after reawakening and later perform an ordinary 805 * unlock operation, such as vm_map_unlock(), before completing its 806 * operation on the map. 807 */ 808 int 809 _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line) 810 { 811 812 VM_MAP_UNLOCK_CONSISTENT(map); 813 mtx_lock(&map_sleep_mtx); 814 if (map->system_map) { 815 KASSERT((map->flags & MAP_REPLENISH) == 0, 816 ("%s: MAP_REPLENISH leaked", __func__)); 817 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 818 } else { 819 sx_xunlock_(&map->lock, file, line); 820 } 821 return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps", 822 timo)); 823 } 824 825 /* 826 * vm_map_wakeup: 827 * 828 * Awaken any threads that have slept on the map using 829 * vm_map_unlock_and_wait(). 830 */ 831 void 832 vm_map_wakeup(vm_map_t map) 833 { 834 835 /* 836 * Acquire and release map_sleep_mtx to prevent a wakeup() 837 * from being performed (and lost) between the map unlock 838 * and the msleep() in _vm_map_unlock_and_wait(). 839 */ 840 mtx_lock(&map_sleep_mtx); 841 mtx_unlock(&map_sleep_mtx); 842 wakeup(&map->root); 843 } 844 845 void 846 vm_map_busy(vm_map_t map) 847 { 848 849 VM_MAP_ASSERT_LOCKED(map); 850 map->busy++; 851 } 852 853 void 854 vm_map_unbusy(vm_map_t map) 855 { 856 857 VM_MAP_ASSERT_LOCKED(map); 858 KASSERT(map->busy, ("vm_map_unbusy: not busy")); 859 if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) { 860 vm_map_modflags(map, 0, MAP_BUSY_WAKEUP); 861 wakeup(&map->busy); 862 } 863 } 864 865 void 866 vm_map_wait_busy(vm_map_t map) 867 { 868 869 VM_MAP_ASSERT_LOCKED(map); 870 while (map->busy) { 871 vm_map_modflags(map, MAP_BUSY_WAKEUP, 0); 872 if (map->system_map) 873 msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0); 874 else 875 sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0); 876 } 877 map->timestamp++; 878 } 879 880 long 881 vmspace_resident_count(struct vmspace *vmspace) 882 { 883 return pmap_resident_count(vmspace_pmap(vmspace)); 884 } 885 886 /* 887 * Initialize an existing vm_map structure 888 * such as that in the vmspace structure. 889 */ 890 static void 891 _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 892 { 893 894 map->header.eflags = MAP_ENTRY_HEADER; 895 map->needs_wakeup = FALSE; 896 map->system_map = 0; 897 map->pmap = pmap; 898 map->header.end = min; 899 map->header.start = max; 900 map->flags = 0; 901 map->header.left = map->header.right = &map->header; 902 map->root = NULL; 903 map->timestamp = 0; 904 map->busy = 0; 905 map->anon_loc = 0; 906 #ifdef DIAGNOSTIC 907 map->nupdates = 0; 908 #endif 909 } 910 911 void 912 vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 913 { 914 915 _vm_map_init(map, pmap, min, max); 916 mtx_init(&map->system_mtx, "vm map (system)", NULL, 917 MTX_DEF | MTX_DUPOK); 918 sx_init(&map->lock, "vm map (user)"); 919 } 920 921 /* 922 * vm_map_entry_dispose: [ internal use only ] 923 * 924 * Inverse of vm_map_entry_create. 925 */ 926 static void 927 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry) 928 { 929 uma_zfree(map->system_map ? kmapentzone : mapentzone, entry); 930 } 931 932 /* 933 * vm_map_entry_create: [ internal use only ] 934 * 935 * Allocates a VM map entry for insertion. 936 * No entry fields are filled in. 937 */ 938 static vm_map_entry_t 939 vm_map_entry_create(vm_map_t map) 940 { 941 vm_map_entry_t new_entry; 942 943 #ifndef UMA_MD_SMALL_ALLOC 944 if (map == kernel_map) { 945 VM_MAP_ASSERT_LOCKED(map); 946 947 /* 948 * A new slab of kernel map entries cannot be allocated at this 949 * point because the kernel map has not yet been updated to 950 * reflect the caller's request. Therefore, we allocate a new 951 * map entry, dipping into the reserve if necessary, and set a 952 * flag indicating that the reserve must be replenished before 953 * the map is unlocked. 954 */ 955 new_entry = uma_zalloc(kmapentzone, M_NOWAIT | M_NOVM); 956 if (new_entry == NULL) { 957 new_entry = uma_zalloc(kmapentzone, 958 M_NOWAIT | M_NOVM | M_USE_RESERVE); 959 kernel_map->flags |= MAP_REPLENISH; 960 } 961 } else 962 #endif 963 if (map->system_map) { 964 new_entry = uma_zalloc(kmapentzone, M_NOWAIT); 965 } else { 966 new_entry = uma_zalloc(mapentzone, M_WAITOK); 967 } 968 KASSERT(new_entry != NULL, 969 ("vm_map_entry_create: kernel resources exhausted")); 970 return (new_entry); 971 } 972 973 /* 974 * vm_map_entry_set_behavior: 975 * 976 * Set the expected access behavior, either normal, random, or 977 * sequential. 978 */ 979 static inline void 980 vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior) 981 { 982 entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) | 983 (behavior & MAP_ENTRY_BEHAV_MASK); 984 } 985 986 /* 987 * vm_map_entry_max_free_{left,right}: 988 * 989 * Compute the size of the largest free gap between two entries, 990 * one the root of a tree and the other the ancestor of that root 991 * that is the least or greatest ancestor found on the search path. 992 */ 993 static inline vm_size_t 994 vm_map_entry_max_free_left(vm_map_entry_t root, vm_map_entry_t left_ancestor) 995 { 996 997 return (root->left != left_ancestor ? 998 root->left->max_free : root->start - left_ancestor->end); 999 } 1000 1001 static inline vm_size_t 1002 vm_map_entry_max_free_right(vm_map_entry_t root, vm_map_entry_t right_ancestor) 1003 { 1004 1005 return (root->right != right_ancestor ? 1006 root->right->max_free : right_ancestor->start - root->end); 1007 } 1008 1009 /* 1010 * vm_map_entry_{pred,succ}: 1011 * 1012 * Find the {predecessor, successor} of the entry by taking one step 1013 * in the appropriate direction and backtracking as much as necessary. 1014 * vm_map_entry_succ is defined in vm_map.h. 1015 */ 1016 static inline vm_map_entry_t 1017 vm_map_entry_pred(vm_map_entry_t entry) 1018 { 1019 vm_map_entry_t prior; 1020 1021 prior = entry->left; 1022 if (prior->right->start < entry->start) { 1023 do 1024 prior = prior->right; 1025 while (prior->right != entry); 1026 } 1027 return (prior); 1028 } 1029 1030 static inline vm_size_t 1031 vm_size_max(vm_size_t a, vm_size_t b) 1032 { 1033 1034 return (a > b ? a : b); 1035 } 1036 1037 #define SPLAY_LEFT_STEP(root, y, llist, rlist, test) do { \ 1038 vm_map_entry_t z; \ 1039 vm_size_t max_free; \ 1040 \ 1041 /* \ 1042 * Infer root->right->max_free == root->max_free when \ 1043 * y->max_free < root->max_free || root->max_free == 0. \ 1044 * Otherwise, look right to find it. \ 1045 */ \ 1046 y = root->left; \ 1047 max_free = root->max_free; \ 1048 KASSERT(max_free == vm_size_max( \ 1049 vm_map_entry_max_free_left(root, llist), \ 1050 vm_map_entry_max_free_right(root, rlist)), \ 1051 ("%s: max_free invariant fails", __func__)); \ 1052 if (max_free - 1 < vm_map_entry_max_free_left(root, llist)) \ 1053 max_free = vm_map_entry_max_free_right(root, rlist); \ 1054 if (y != llist && (test)) { \ 1055 /* Rotate right and make y root. */ \ 1056 z = y->right; \ 1057 if (z != root) { \ 1058 root->left = z; \ 1059 y->right = root; \ 1060 if (max_free < y->max_free) \ 1061 root->max_free = max_free = \ 1062 vm_size_max(max_free, z->max_free); \ 1063 } else if (max_free < y->max_free) \ 1064 root->max_free = max_free = \ 1065 vm_size_max(max_free, root->start - y->end);\ 1066 root = y; \ 1067 y = root->left; \ 1068 } \ 1069 /* Copy right->max_free. Put root on rlist. */ \ 1070 root->max_free = max_free; \ 1071 KASSERT(max_free == vm_map_entry_max_free_right(root, rlist), \ 1072 ("%s: max_free not copied from right", __func__)); \ 1073 root->left = rlist; \ 1074 rlist = root; \ 1075 root = y != llist ? y : NULL; \ 1076 } while (0) 1077 1078 #define SPLAY_RIGHT_STEP(root, y, llist, rlist, test) do { \ 1079 vm_map_entry_t z; \ 1080 vm_size_t max_free; \ 1081 \ 1082 /* \ 1083 * Infer root->left->max_free == root->max_free when \ 1084 * y->max_free < root->max_free || root->max_free == 0. \ 1085 * Otherwise, look left to find it. \ 1086 */ \ 1087 y = root->right; \ 1088 max_free = root->max_free; \ 1089 KASSERT(max_free == vm_size_max( \ 1090 vm_map_entry_max_free_left(root, llist), \ 1091 vm_map_entry_max_free_right(root, rlist)), \ 1092 ("%s: max_free invariant fails", __func__)); \ 1093 if (max_free - 1 < vm_map_entry_max_free_right(root, rlist)) \ 1094 max_free = vm_map_entry_max_free_left(root, llist); \ 1095 if (y != rlist && (test)) { \ 1096 /* Rotate left and make y root. */ \ 1097 z = y->left; \ 1098 if (z != root) { \ 1099 root->right = z; \ 1100 y->left = root; \ 1101 if (max_free < y->max_free) \ 1102 root->max_free = max_free = \ 1103 vm_size_max(max_free, z->max_free); \ 1104 } else if (max_free < y->max_free) \ 1105 root->max_free = max_free = \ 1106 vm_size_max(max_free, y->start - root->end);\ 1107 root = y; \ 1108 y = root->right; \ 1109 } \ 1110 /* Copy left->max_free. Put root on llist. */ \ 1111 root->max_free = max_free; \ 1112 KASSERT(max_free == vm_map_entry_max_free_left(root, llist), \ 1113 ("%s: max_free not copied from left", __func__)); \ 1114 root->right = llist; \ 1115 llist = root; \ 1116 root = y != rlist ? y : NULL; \ 1117 } while (0) 1118 1119 /* 1120 * Walk down the tree until we find addr or a gap where addr would go, breaking 1121 * off left and right subtrees of nodes less than, or greater than addr. Treat 1122 * subtrees with root->max_free < length as empty trees. llist and rlist are 1123 * the two sides in reverse order (bottom-up), with llist linked by the right 1124 * pointer and rlist linked by the left pointer in the vm_map_entry, and both 1125 * lists terminated by &map->header. This function, and the subsequent call to 1126 * vm_map_splay_merge_{left,right,pred,succ}, rely on the start and end address 1127 * values in &map->header. 1128 */ 1129 static __always_inline vm_map_entry_t 1130 vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length, 1131 vm_map_entry_t *llist, vm_map_entry_t *rlist) 1132 { 1133 vm_map_entry_t left, right, root, y; 1134 1135 left = right = &map->header; 1136 root = map->root; 1137 while (root != NULL && root->max_free >= length) { 1138 KASSERT(left->end <= root->start && 1139 root->end <= right->start, 1140 ("%s: root not within tree bounds", __func__)); 1141 if (addr < root->start) { 1142 SPLAY_LEFT_STEP(root, y, left, right, 1143 y->max_free >= length && addr < y->start); 1144 } else if (addr >= root->end) { 1145 SPLAY_RIGHT_STEP(root, y, left, right, 1146 y->max_free >= length && addr >= y->end); 1147 } else 1148 break; 1149 } 1150 *llist = left; 1151 *rlist = right; 1152 return (root); 1153 } 1154 1155 static __always_inline void 1156 vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *rlist) 1157 { 1158 vm_map_entry_t hi, right, y; 1159 1160 right = *rlist; 1161 hi = root->right == right ? NULL : root->right; 1162 if (hi == NULL) 1163 return; 1164 do 1165 SPLAY_LEFT_STEP(hi, y, root, right, true); 1166 while (hi != NULL); 1167 *rlist = right; 1168 } 1169 1170 static __always_inline void 1171 vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *llist) 1172 { 1173 vm_map_entry_t left, lo, y; 1174 1175 left = *llist; 1176 lo = root->left == left ? NULL : root->left; 1177 if (lo == NULL) 1178 return; 1179 do 1180 SPLAY_RIGHT_STEP(lo, y, left, root, true); 1181 while (lo != NULL); 1182 *llist = left; 1183 } 1184 1185 static inline void 1186 vm_map_entry_swap(vm_map_entry_t *a, vm_map_entry_t *b) 1187 { 1188 vm_map_entry_t tmp; 1189 1190 tmp = *b; 1191 *b = *a; 1192 *a = tmp; 1193 } 1194 1195 /* 1196 * Walk back up the two spines, flip the pointers and set max_free. The 1197 * subtrees of the root go at the bottom of llist and rlist. 1198 */ 1199 static vm_size_t 1200 vm_map_splay_merge_left_walk(vm_map_entry_t header, vm_map_entry_t root, 1201 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t llist) 1202 { 1203 do { 1204 /* 1205 * The max_free values of the children of llist are in 1206 * llist->max_free and max_free. Update with the 1207 * max value. 1208 */ 1209 llist->max_free = max_free = 1210 vm_size_max(llist->max_free, max_free); 1211 vm_map_entry_swap(&llist->right, &tail); 1212 vm_map_entry_swap(&tail, &llist); 1213 } while (llist != header); 1214 root->left = tail; 1215 return (max_free); 1216 } 1217 1218 /* 1219 * When llist is known to be the predecessor of root. 1220 */ 1221 static inline vm_size_t 1222 vm_map_splay_merge_pred(vm_map_entry_t header, vm_map_entry_t root, 1223 vm_map_entry_t llist) 1224 { 1225 vm_size_t max_free; 1226 1227 max_free = root->start - llist->end; 1228 if (llist != header) { 1229 max_free = vm_map_splay_merge_left_walk(header, root, 1230 root, max_free, llist); 1231 } else { 1232 root->left = header; 1233 header->right = root; 1234 } 1235 return (max_free); 1236 } 1237 1238 /* 1239 * When llist may or may not be the predecessor of root. 1240 */ 1241 static inline vm_size_t 1242 vm_map_splay_merge_left(vm_map_entry_t header, vm_map_entry_t root, 1243 vm_map_entry_t llist) 1244 { 1245 vm_size_t max_free; 1246 1247 max_free = vm_map_entry_max_free_left(root, llist); 1248 if (llist != header) { 1249 max_free = vm_map_splay_merge_left_walk(header, root, 1250 root->left == llist ? root : root->left, 1251 max_free, llist); 1252 } 1253 return (max_free); 1254 } 1255 1256 static vm_size_t 1257 vm_map_splay_merge_right_walk(vm_map_entry_t header, vm_map_entry_t root, 1258 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t rlist) 1259 { 1260 do { 1261 /* 1262 * The max_free values of the children of rlist are in 1263 * rlist->max_free and max_free. Update with the 1264 * max value. 1265 */ 1266 rlist->max_free = max_free = 1267 vm_size_max(rlist->max_free, max_free); 1268 vm_map_entry_swap(&rlist->left, &tail); 1269 vm_map_entry_swap(&tail, &rlist); 1270 } while (rlist != header); 1271 root->right = tail; 1272 return (max_free); 1273 } 1274 1275 /* 1276 * When rlist is known to be the succecessor of root. 1277 */ 1278 static inline vm_size_t 1279 vm_map_splay_merge_succ(vm_map_entry_t header, vm_map_entry_t root, 1280 vm_map_entry_t rlist) 1281 { 1282 vm_size_t max_free; 1283 1284 max_free = rlist->start - root->end; 1285 if (rlist != header) { 1286 max_free = vm_map_splay_merge_right_walk(header, root, 1287 root, max_free, rlist); 1288 } else { 1289 root->right = header; 1290 header->left = root; 1291 } 1292 return (max_free); 1293 } 1294 1295 /* 1296 * When rlist may or may not be the succecessor of root. 1297 */ 1298 static inline vm_size_t 1299 vm_map_splay_merge_right(vm_map_entry_t header, vm_map_entry_t root, 1300 vm_map_entry_t rlist) 1301 { 1302 vm_size_t max_free; 1303 1304 max_free = vm_map_entry_max_free_right(root, rlist); 1305 if (rlist != header) { 1306 max_free = vm_map_splay_merge_right_walk(header, root, 1307 root->right == rlist ? root : root->right, 1308 max_free, rlist); 1309 } 1310 return (max_free); 1311 } 1312 1313 /* 1314 * vm_map_splay: 1315 * 1316 * The Sleator and Tarjan top-down splay algorithm with the 1317 * following variation. Max_free must be computed bottom-up, so 1318 * on the downward pass, maintain the left and right spines in 1319 * reverse order. Then, make a second pass up each side to fix 1320 * the pointers and compute max_free. The time bound is O(log n) 1321 * amortized. 1322 * 1323 * The tree is threaded, which means that there are no null pointers. 1324 * When a node has no left child, its left pointer points to its 1325 * predecessor, which the last ancestor on the search path from the root 1326 * where the search branched right. Likewise, when a node has no right 1327 * child, its right pointer points to its successor. The map header node 1328 * is the predecessor of the first map entry, and the successor of the 1329 * last. 1330 * 1331 * The new root is the vm_map_entry containing "addr", or else an 1332 * adjacent entry (lower if possible) if addr is not in the tree. 1333 * 1334 * The map must be locked, and leaves it so. 1335 * 1336 * Returns: the new root. 1337 */ 1338 static vm_map_entry_t 1339 vm_map_splay(vm_map_t map, vm_offset_t addr) 1340 { 1341 vm_map_entry_t header, llist, rlist, root; 1342 vm_size_t max_free_left, max_free_right; 1343 1344 header = &map->header; 1345 root = vm_map_splay_split(map, addr, 0, &llist, &rlist); 1346 if (root != NULL) { 1347 max_free_left = vm_map_splay_merge_left(header, root, llist); 1348 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1349 } else if (llist != header) { 1350 /* 1351 * Recover the greatest node in the left 1352 * subtree and make it the root. 1353 */ 1354 root = llist; 1355 llist = root->right; 1356 max_free_left = vm_map_splay_merge_left(header, root, llist); 1357 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1358 } else if (rlist != header) { 1359 /* 1360 * Recover the least node in the right 1361 * subtree and make it the root. 1362 */ 1363 root = rlist; 1364 rlist = root->left; 1365 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1366 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1367 } else { 1368 /* There is no root. */ 1369 return (NULL); 1370 } 1371 root->max_free = vm_size_max(max_free_left, max_free_right); 1372 map->root = root; 1373 VM_MAP_ASSERT_CONSISTENT(map); 1374 return (root); 1375 } 1376 1377 /* 1378 * vm_map_entry_{un,}link: 1379 * 1380 * Insert/remove entries from maps. On linking, if new entry clips 1381 * existing entry, trim existing entry to avoid overlap, and manage 1382 * offsets. On unlinking, merge disappearing entry with neighbor, if 1383 * called for, and manage offsets. Callers should not modify fields in 1384 * entries already mapped. 1385 */ 1386 static void 1387 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) 1388 { 1389 vm_map_entry_t header, llist, rlist, root; 1390 vm_size_t max_free_left, max_free_right; 1391 1392 CTR3(KTR_VM, 1393 "vm_map_entry_link: map %p, nentries %d, entry %p", map, 1394 map->nentries, entry); 1395 VM_MAP_ASSERT_LOCKED(map); 1396 map->nentries++; 1397 header = &map->header; 1398 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1399 if (root == NULL) { 1400 /* 1401 * The new entry does not overlap any existing entry in the 1402 * map, so it becomes the new root of the map tree. 1403 */ 1404 max_free_left = vm_map_splay_merge_pred(header, entry, llist); 1405 max_free_right = vm_map_splay_merge_succ(header, entry, rlist); 1406 } else if (entry->start == root->start) { 1407 /* 1408 * The new entry is a clone of root, with only the end field 1409 * changed. The root entry will be shrunk to abut the new 1410 * entry, and will be the right child of the new root entry in 1411 * the modified map. 1412 */ 1413 KASSERT(entry->end < root->end, 1414 ("%s: clip_start not within entry", __func__)); 1415 vm_map_splay_findprev(root, &llist); 1416 root->offset += entry->end - root->start; 1417 root->start = entry->end; 1418 max_free_left = vm_map_splay_merge_pred(header, entry, llist); 1419 max_free_right = root->max_free = vm_size_max( 1420 vm_map_splay_merge_pred(entry, root, entry), 1421 vm_map_splay_merge_right(header, root, rlist)); 1422 } else { 1423 /* 1424 * The new entry is a clone of root, with only the start field 1425 * changed. The root entry will be shrunk to abut the new 1426 * entry, and will be the left child of the new root entry in 1427 * the modified map. 1428 */ 1429 KASSERT(entry->end == root->end, 1430 ("%s: clip_start not within entry", __func__)); 1431 vm_map_splay_findnext(root, &rlist); 1432 entry->offset += entry->start - root->start; 1433 root->end = entry->start; 1434 max_free_left = root->max_free = vm_size_max( 1435 vm_map_splay_merge_left(header, root, llist), 1436 vm_map_splay_merge_succ(entry, root, entry)); 1437 max_free_right = vm_map_splay_merge_succ(header, entry, rlist); 1438 } 1439 entry->max_free = vm_size_max(max_free_left, max_free_right); 1440 map->root = entry; 1441 VM_MAP_ASSERT_CONSISTENT(map); 1442 } 1443 1444 enum unlink_merge_type { 1445 UNLINK_MERGE_NONE, 1446 UNLINK_MERGE_NEXT 1447 }; 1448 1449 static void 1450 vm_map_entry_unlink(vm_map_t map, vm_map_entry_t entry, 1451 enum unlink_merge_type op) 1452 { 1453 vm_map_entry_t header, llist, rlist, root; 1454 vm_size_t max_free_left, max_free_right; 1455 1456 VM_MAP_ASSERT_LOCKED(map); 1457 header = &map->header; 1458 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1459 KASSERT(root != NULL, 1460 ("vm_map_entry_unlink: unlink object not mapped")); 1461 1462 vm_map_splay_findprev(root, &llist); 1463 vm_map_splay_findnext(root, &rlist); 1464 if (op == UNLINK_MERGE_NEXT) { 1465 rlist->start = root->start; 1466 rlist->offset = root->offset; 1467 } 1468 if (llist != header) { 1469 root = llist; 1470 llist = root->right; 1471 max_free_left = vm_map_splay_merge_left(header, root, llist); 1472 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1473 } else if (rlist != header) { 1474 root = rlist; 1475 rlist = root->left; 1476 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1477 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1478 } else { 1479 header->left = header->right = header; 1480 root = NULL; 1481 } 1482 if (root != NULL) 1483 root->max_free = vm_size_max(max_free_left, max_free_right); 1484 map->root = root; 1485 VM_MAP_ASSERT_CONSISTENT(map); 1486 map->nentries--; 1487 CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map, 1488 map->nentries, entry); 1489 } 1490 1491 /* 1492 * vm_map_entry_resize: 1493 * 1494 * Resize a vm_map_entry, recompute the amount of free space that 1495 * follows it and propagate that value up the tree. 1496 * 1497 * The map must be locked, and leaves it so. 1498 */ 1499 static void 1500 vm_map_entry_resize(vm_map_t map, vm_map_entry_t entry, vm_size_t grow_amount) 1501 { 1502 vm_map_entry_t header, llist, rlist, root; 1503 1504 VM_MAP_ASSERT_LOCKED(map); 1505 header = &map->header; 1506 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1507 KASSERT(root != NULL, ("%s: resize object not mapped", __func__)); 1508 vm_map_splay_findnext(root, &rlist); 1509 entry->end += grow_amount; 1510 root->max_free = vm_size_max( 1511 vm_map_splay_merge_left(header, root, llist), 1512 vm_map_splay_merge_succ(header, root, rlist)); 1513 map->root = root; 1514 VM_MAP_ASSERT_CONSISTENT(map); 1515 CTR4(KTR_VM, "%s: map %p, nentries %d, entry %p", 1516 __func__, map, map->nentries, entry); 1517 } 1518 1519 /* 1520 * vm_map_lookup_entry: [ internal use only ] 1521 * 1522 * Finds the map entry containing (or 1523 * immediately preceding) the specified address 1524 * in the given map; the entry is returned 1525 * in the "entry" parameter. The boolean 1526 * result indicates whether the address is 1527 * actually contained in the map. 1528 */ 1529 boolean_t 1530 vm_map_lookup_entry( 1531 vm_map_t map, 1532 vm_offset_t address, 1533 vm_map_entry_t *entry) /* OUT */ 1534 { 1535 vm_map_entry_t cur, header, lbound, ubound; 1536 boolean_t locked; 1537 1538 /* 1539 * If the map is empty, then the map entry immediately preceding 1540 * "address" is the map's header. 1541 */ 1542 header = &map->header; 1543 cur = map->root; 1544 if (cur == NULL) { 1545 *entry = header; 1546 return (FALSE); 1547 } 1548 if (address >= cur->start && cur->end > address) { 1549 *entry = cur; 1550 return (TRUE); 1551 } 1552 if ((locked = vm_map_locked(map)) || 1553 sx_try_upgrade(&map->lock)) { 1554 /* 1555 * Splay requires a write lock on the map. However, it only 1556 * restructures the binary search tree; it does not otherwise 1557 * change the map. Thus, the map's timestamp need not change 1558 * on a temporary upgrade. 1559 */ 1560 cur = vm_map_splay(map, address); 1561 if (!locked) { 1562 VM_MAP_UNLOCK_CONSISTENT(map); 1563 sx_downgrade(&map->lock); 1564 } 1565 1566 /* 1567 * If "address" is contained within a map entry, the new root 1568 * is that map entry. Otherwise, the new root is a map entry 1569 * immediately before or after "address". 1570 */ 1571 if (address < cur->start) { 1572 *entry = header; 1573 return (FALSE); 1574 } 1575 *entry = cur; 1576 return (address < cur->end); 1577 } 1578 /* 1579 * Since the map is only locked for read access, perform a 1580 * standard binary search tree lookup for "address". 1581 */ 1582 lbound = ubound = header; 1583 for (;;) { 1584 if (address < cur->start) { 1585 ubound = cur; 1586 cur = cur->left; 1587 if (cur == lbound) 1588 break; 1589 } else if (cur->end <= address) { 1590 lbound = cur; 1591 cur = cur->right; 1592 if (cur == ubound) 1593 break; 1594 } else { 1595 *entry = cur; 1596 return (TRUE); 1597 } 1598 } 1599 *entry = lbound; 1600 return (FALSE); 1601 } 1602 1603 /* 1604 * vm_map_insert: 1605 * 1606 * Inserts the given whole VM object into the target 1607 * map at the specified address range. The object's 1608 * size should match that of the address range. 1609 * 1610 * Requires that the map be locked, and leaves it so. 1611 * 1612 * If object is non-NULL, ref count must be bumped by caller 1613 * prior to making call to account for the new entry. 1614 */ 1615 int 1616 vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1617 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow) 1618 { 1619 vm_map_entry_t new_entry, next_entry, prev_entry; 1620 struct ucred *cred; 1621 vm_eflags_t protoeflags; 1622 vm_inherit_t inheritance; 1623 u_long bdry; 1624 u_int bidx; 1625 1626 VM_MAP_ASSERT_LOCKED(map); 1627 KASSERT(object != kernel_object || 1628 (cow & MAP_COPY_ON_WRITE) == 0, 1629 ("vm_map_insert: kernel object and COW")); 1630 KASSERT(object == NULL || (cow & MAP_NOFAULT) == 0 || 1631 (cow & MAP_SPLIT_BOUNDARY_MASK) != 0, 1632 ("vm_map_insert: paradoxical MAP_NOFAULT request, obj %p cow %#x", 1633 object, cow)); 1634 KASSERT((prot & ~max) == 0, 1635 ("prot %#x is not subset of max_prot %#x", prot, max)); 1636 1637 /* 1638 * Check that the start and end points are not bogus. 1639 */ 1640 if (start == end || !vm_map_range_valid(map, start, end)) 1641 return (KERN_INVALID_ADDRESS); 1642 1643 if ((map->flags & MAP_WXORX) != 0 && (prot & (VM_PROT_WRITE | 1644 VM_PROT_EXECUTE)) == (VM_PROT_WRITE | VM_PROT_EXECUTE)) 1645 return (KERN_PROTECTION_FAILURE); 1646 1647 /* 1648 * Find the entry prior to the proposed starting address; if it's part 1649 * of an existing entry, this range is bogus. 1650 */ 1651 if (vm_map_lookup_entry(map, start, &prev_entry)) 1652 return (KERN_NO_SPACE); 1653 1654 /* 1655 * Assert that the next entry doesn't overlap the end point. 1656 */ 1657 next_entry = vm_map_entry_succ(prev_entry); 1658 if (next_entry->start < end) 1659 return (KERN_NO_SPACE); 1660 1661 if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL || 1662 max != VM_PROT_NONE)) 1663 return (KERN_INVALID_ARGUMENT); 1664 1665 protoeflags = 0; 1666 if (cow & MAP_COPY_ON_WRITE) 1667 protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; 1668 if (cow & MAP_NOFAULT) 1669 protoeflags |= MAP_ENTRY_NOFAULT; 1670 if (cow & MAP_DISABLE_SYNCER) 1671 protoeflags |= MAP_ENTRY_NOSYNC; 1672 if (cow & MAP_DISABLE_COREDUMP) 1673 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1674 if (cow & MAP_STACK_GROWS_DOWN) 1675 protoeflags |= MAP_ENTRY_GROWS_DOWN; 1676 if (cow & MAP_STACK_GROWS_UP) 1677 protoeflags |= MAP_ENTRY_GROWS_UP; 1678 if (cow & MAP_WRITECOUNT) 1679 protoeflags |= MAP_ENTRY_WRITECNT; 1680 if (cow & MAP_VN_EXEC) 1681 protoeflags |= MAP_ENTRY_VN_EXEC; 1682 if ((cow & MAP_CREATE_GUARD) != 0) 1683 protoeflags |= MAP_ENTRY_GUARD; 1684 if ((cow & MAP_CREATE_STACK_GAP_DN) != 0) 1685 protoeflags |= MAP_ENTRY_STACK_GAP_DN; 1686 if ((cow & MAP_CREATE_STACK_GAP_UP) != 0) 1687 protoeflags |= MAP_ENTRY_STACK_GAP_UP; 1688 if (cow & MAP_INHERIT_SHARE) 1689 inheritance = VM_INHERIT_SHARE; 1690 else 1691 inheritance = VM_INHERIT_DEFAULT; 1692 if ((cow & MAP_SPLIT_BOUNDARY_MASK) != 0) { 1693 /* This magically ignores index 0, for usual page size. */ 1694 bidx = (cow & MAP_SPLIT_BOUNDARY_MASK) >> 1695 MAP_SPLIT_BOUNDARY_SHIFT; 1696 if (bidx >= MAXPAGESIZES) 1697 return (KERN_INVALID_ARGUMENT); 1698 bdry = pagesizes[bidx] - 1; 1699 if ((start & bdry) != 0 || (end & bdry) != 0) 1700 return (KERN_INVALID_ARGUMENT); 1701 protoeflags |= bidx << MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 1702 } 1703 1704 cred = NULL; 1705 if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0) 1706 goto charged; 1707 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) && 1708 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) { 1709 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start)) 1710 return (KERN_RESOURCE_SHORTAGE); 1711 KASSERT(object == NULL || 1712 (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 || 1713 object->cred == NULL, 1714 ("overcommit: vm_map_insert o %p", object)); 1715 cred = curthread->td_ucred; 1716 } 1717 1718 charged: 1719 /* Expand the kernel pmap, if necessary. */ 1720 if (map == kernel_map && end > kernel_vm_end) 1721 pmap_growkernel(end); 1722 if (object != NULL) { 1723 /* 1724 * OBJ_ONEMAPPING must be cleared unless this mapping 1725 * is trivially proven to be the only mapping for any 1726 * of the object's pages. (Object granularity 1727 * reference counting is insufficient to recognize 1728 * aliases with precision.) 1729 */ 1730 if ((object->flags & OBJ_ANON) != 0) { 1731 VM_OBJECT_WLOCK(object); 1732 if (object->ref_count > 1 || object->shadow_count != 0) 1733 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1734 VM_OBJECT_WUNLOCK(object); 1735 } 1736 } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) == 1737 protoeflags && 1738 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP | 1739 MAP_VN_EXEC)) == 0 && 1740 prev_entry->end == start && (prev_entry->cred == cred || 1741 (prev_entry->object.vm_object != NULL && 1742 prev_entry->object.vm_object->cred == cred)) && 1743 vm_object_coalesce(prev_entry->object.vm_object, 1744 prev_entry->offset, 1745 (vm_size_t)(prev_entry->end - prev_entry->start), 1746 (vm_size_t)(end - prev_entry->end), cred != NULL && 1747 (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) { 1748 /* 1749 * We were able to extend the object. Determine if we 1750 * can extend the previous map entry to include the 1751 * new range as well. 1752 */ 1753 if (prev_entry->inheritance == inheritance && 1754 prev_entry->protection == prot && 1755 prev_entry->max_protection == max && 1756 prev_entry->wired_count == 0) { 1757 KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) == 1758 0, ("prev_entry %p has incoherent wiring", 1759 prev_entry)); 1760 if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0) 1761 map->size += end - prev_entry->end; 1762 vm_map_entry_resize(map, prev_entry, 1763 end - prev_entry->end); 1764 vm_map_try_merge_entries(map, prev_entry, next_entry); 1765 return (KERN_SUCCESS); 1766 } 1767 1768 /* 1769 * If we can extend the object but cannot extend the 1770 * map entry, we have to create a new map entry. We 1771 * must bump the ref count on the extended object to 1772 * account for it. object may be NULL. 1773 */ 1774 object = prev_entry->object.vm_object; 1775 offset = prev_entry->offset + 1776 (prev_entry->end - prev_entry->start); 1777 vm_object_reference(object); 1778 if (cred != NULL && object != NULL && object->cred != NULL && 1779 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 1780 /* Object already accounts for this uid. */ 1781 cred = NULL; 1782 } 1783 } 1784 if (cred != NULL) 1785 crhold(cred); 1786 1787 /* 1788 * Create a new entry 1789 */ 1790 new_entry = vm_map_entry_create(map); 1791 new_entry->start = start; 1792 new_entry->end = end; 1793 new_entry->cred = NULL; 1794 1795 new_entry->eflags = protoeflags; 1796 new_entry->object.vm_object = object; 1797 new_entry->offset = offset; 1798 1799 new_entry->inheritance = inheritance; 1800 new_entry->protection = prot; 1801 new_entry->max_protection = max; 1802 new_entry->wired_count = 0; 1803 new_entry->wiring_thread = NULL; 1804 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT; 1805 new_entry->next_read = start; 1806 1807 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry), 1808 ("overcommit: vm_map_insert leaks vm_map %p", new_entry)); 1809 new_entry->cred = cred; 1810 1811 /* 1812 * Insert the new entry into the list 1813 */ 1814 vm_map_entry_link(map, new_entry); 1815 if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0) 1816 map->size += new_entry->end - new_entry->start; 1817 1818 /* 1819 * Try to coalesce the new entry with both the previous and next 1820 * entries in the list. Previously, we only attempted to coalesce 1821 * with the previous entry when object is NULL. Here, we handle the 1822 * other cases, which are less common. 1823 */ 1824 vm_map_try_merge_entries(map, prev_entry, new_entry); 1825 vm_map_try_merge_entries(map, new_entry, next_entry); 1826 1827 if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) { 1828 vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset), 1829 end - start, cow & MAP_PREFAULT_PARTIAL); 1830 } 1831 1832 return (KERN_SUCCESS); 1833 } 1834 1835 /* 1836 * vm_map_findspace: 1837 * 1838 * Find the first fit (lowest VM address) for "length" free bytes 1839 * beginning at address >= start in the given map. 1840 * 1841 * In a vm_map_entry, "max_free" is the maximum amount of 1842 * contiguous free space between an entry in its subtree and a 1843 * neighbor of that entry. This allows finding a free region in 1844 * one path down the tree, so O(log n) amortized with splay 1845 * trees. 1846 * 1847 * The map must be locked, and leaves it so. 1848 * 1849 * Returns: starting address if sufficient space, 1850 * vm_map_max(map)-length+1 if insufficient space. 1851 */ 1852 vm_offset_t 1853 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length) 1854 { 1855 vm_map_entry_t header, llist, rlist, root, y; 1856 vm_size_t left_length, max_free_left, max_free_right; 1857 vm_offset_t gap_end; 1858 1859 VM_MAP_ASSERT_LOCKED(map); 1860 1861 /* 1862 * Request must fit within min/max VM address and must avoid 1863 * address wrap. 1864 */ 1865 start = MAX(start, vm_map_min(map)); 1866 if (start >= vm_map_max(map) || length > vm_map_max(map) - start) 1867 return (vm_map_max(map) - length + 1); 1868 1869 /* Empty tree means wide open address space. */ 1870 if (map->root == NULL) 1871 return (start); 1872 1873 /* 1874 * After splay_split, if start is within an entry, push it to the start 1875 * of the following gap. If rlist is at the end of the gap containing 1876 * start, save the end of that gap in gap_end to see if the gap is big 1877 * enough; otherwise set gap_end to start skip gap-checking and move 1878 * directly to a search of the right subtree. 1879 */ 1880 header = &map->header; 1881 root = vm_map_splay_split(map, start, length, &llist, &rlist); 1882 gap_end = rlist->start; 1883 if (root != NULL) { 1884 start = root->end; 1885 if (root->right != rlist) 1886 gap_end = start; 1887 max_free_left = vm_map_splay_merge_left(header, root, llist); 1888 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1889 } else if (rlist != header) { 1890 root = rlist; 1891 rlist = root->left; 1892 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1893 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1894 } else { 1895 root = llist; 1896 llist = root->right; 1897 max_free_left = vm_map_splay_merge_left(header, root, llist); 1898 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1899 } 1900 root->max_free = vm_size_max(max_free_left, max_free_right); 1901 map->root = root; 1902 VM_MAP_ASSERT_CONSISTENT(map); 1903 if (length <= gap_end - start) 1904 return (start); 1905 1906 /* With max_free, can immediately tell if no solution. */ 1907 if (root->right == header || length > root->right->max_free) 1908 return (vm_map_max(map) - length + 1); 1909 1910 /* 1911 * Splay for the least large-enough gap in the right subtree. 1912 */ 1913 llist = rlist = header; 1914 for (left_length = 0;; 1915 left_length = vm_map_entry_max_free_left(root, llist)) { 1916 if (length <= left_length) 1917 SPLAY_LEFT_STEP(root, y, llist, rlist, 1918 length <= vm_map_entry_max_free_left(y, llist)); 1919 else 1920 SPLAY_RIGHT_STEP(root, y, llist, rlist, 1921 length > vm_map_entry_max_free_left(y, root)); 1922 if (root == NULL) 1923 break; 1924 } 1925 root = llist; 1926 llist = root->right; 1927 max_free_left = vm_map_splay_merge_left(header, root, llist); 1928 if (rlist == header) { 1929 root->max_free = vm_size_max(max_free_left, 1930 vm_map_splay_merge_succ(header, root, rlist)); 1931 } else { 1932 y = rlist; 1933 rlist = y->left; 1934 y->max_free = vm_size_max( 1935 vm_map_splay_merge_pred(root, y, root), 1936 vm_map_splay_merge_right(header, y, rlist)); 1937 root->max_free = vm_size_max(max_free_left, y->max_free); 1938 } 1939 map->root = root; 1940 VM_MAP_ASSERT_CONSISTENT(map); 1941 return (root->end); 1942 } 1943 1944 int 1945 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1946 vm_offset_t start, vm_size_t length, vm_prot_t prot, 1947 vm_prot_t max, int cow) 1948 { 1949 vm_offset_t end; 1950 int result; 1951 1952 end = start + length; 1953 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 1954 object == NULL, 1955 ("vm_map_fixed: non-NULL backing object for stack")); 1956 vm_map_lock(map); 1957 VM_MAP_RANGE_CHECK(map, start, end); 1958 if ((cow & MAP_CHECK_EXCL) == 0) { 1959 result = vm_map_delete(map, start, end); 1960 if (result != KERN_SUCCESS) 1961 goto out; 1962 } 1963 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 1964 result = vm_map_stack_locked(map, start, length, sgrowsiz, 1965 prot, max, cow); 1966 } else { 1967 result = vm_map_insert(map, object, offset, start, end, 1968 prot, max, cow); 1969 } 1970 out: 1971 vm_map_unlock(map); 1972 return (result); 1973 } 1974 1975 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10}; 1976 static const int aslr_pages_rnd_32[2] = {0x100, 0x4}; 1977 1978 static int cluster_anon = 1; 1979 SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW, 1980 &cluster_anon, 0, 1981 "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always"); 1982 1983 static bool 1984 clustering_anon_allowed(vm_offset_t addr) 1985 { 1986 1987 switch (cluster_anon) { 1988 case 0: 1989 return (false); 1990 case 1: 1991 return (addr == 0); 1992 case 2: 1993 default: 1994 return (true); 1995 } 1996 } 1997 1998 static long aslr_restarts; 1999 SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD, 2000 &aslr_restarts, 0, 2001 "Number of aslr failures"); 2002 2003 /* 2004 * Searches for the specified amount of free space in the given map with the 2005 * specified alignment. Performs an address-ordered, first-fit search from 2006 * the given address "*addr", with an optional upper bound "max_addr". If the 2007 * parameter "alignment" is zero, then the alignment is computed from the 2008 * given (object, offset) pair so as to enable the greatest possible use of 2009 * superpage mappings. Returns KERN_SUCCESS and the address of the free space 2010 * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE. 2011 * 2012 * The map must be locked. Initially, there must be at least "length" bytes 2013 * of free space at the given address. 2014 */ 2015 static int 2016 vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2017 vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, 2018 vm_offset_t alignment) 2019 { 2020 vm_offset_t aligned_addr, free_addr; 2021 2022 VM_MAP_ASSERT_LOCKED(map); 2023 free_addr = *addr; 2024 KASSERT(free_addr == vm_map_findspace(map, free_addr, length), 2025 ("caller failed to provide space %#jx at address %p", 2026 (uintmax_t)length, (void *)free_addr)); 2027 for (;;) { 2028 /* 2029 * At the start of every iteration, the free space at address 2030 * "*addr" is at least "length" bytes. 2031 */ 2032 if (alignment == 0) 2033 pmap_align_superpage(object, offset, addr, length); 2034 else 2035 *addr = roundup2(*addr, alignment); 2036 aligned_addr = *addr; 2037 if (aligned_addr == free_addr) { 2038 /* 2039 * Alignment did not change "*addr", so "*addr" must 2040 * still provide sufficient free space. 2041 */ 2042 return (KERN_SUCCESS); 2043 } 2044 2045 /* 2046 * Test for address wrap on "*addr". A wrapped "*addr" could 2047 * be a valid address, in which case vm_map_findspace() cannot 2048 * be relied upon to fail. 2049 */ 2050 if (aligned_addr < free_addr) 2051 return (KERN_NO_SPACE); 2052 *addr = vm_map_findspace(map, aligned_addr, length); 2053 if (*addr + length > vm_map_max(map) || 2054 (max_addr != 0 && *addr + length > max_addr)) 2055 return (KERN_NO_SPACE); 2056 free_addr = *addr; 2057 if (free_addr == aligned_addr) { 2058 /* 2059 * If a successful call to vm_map_findspace() did not 2060 * change "*addr", then "*addr" must still be aligned 2061 * and provide sufficient free space. 2062 */ 2063 return (KERN_SUCCESS); 2064 } 2065 } 2066 } 2067 2068 int 2069 vm_map_find_aligned(vm_map_t map, vm_offset_t *addr, vm_size_t length, 2070 vm_offset_t max_addr, vm_offset_t alignment) 2071 { 2072 /* XXXKIB ASLR eh ? */ 2073 *addr = vm_map_findspace(map, *addr, length); 2074 if (*addr + length > vm_map_max(map) || 2075 (max_addr != 0 && *addr + length > max_addr)) 2076 return (KERN_NO_SPACE); 2077 return (vm_map_alignspace(map, NULL, 0, addr, length, max_addr, 2078 alignment)); 2079 } 2080 2081 /* 2082 * vm_map_find finds an unallocated region in the target address 2083 * map with the given length. The search is defined to be 2084 * first-fit from the specified address; the region found is 2085 * returned in the same parameter. 2086 * 2087 * If object is non-NULL, ref count must be bumped by caller 2088 * prior to making call to account for the new entry. 2089 */ 2090 int 2091 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2092 vm_offset_t *addr, /* IN/OUT */ 2093 vm_size_t length, vm_offset_t max_addr, int find_space, 2094 vm_prot_t prot, vm_prot_t max, int cow) 2095 { 2096 vm_offset_t alignment, curr_min_addr, min_addr; 2097 int gap, pidx, rv, try; 2098 bool cluster, en_aslr, update_anon; 2099 2100 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 2101 object == NULL, 2102 ("vm_map_find: non-NULL backing object for stack")); 2103 MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE && 2104 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0)); 2105 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL || 2106 (object->flags & OBJ_COLORED) == 0)) 2107 find_space = VMFS_ANY_SPACE; 2108 if (find_space >> 8 != 0) { 2109 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags")); 2110 alignment = (vm_offset_t)1 << (find_space >> 8); 2111 } else 2112 alignment = 0; 2113 en_aslr = (map->flags & MAP_ASLR) != 0; 2114 update_anon = cluster = clustering_anon_allowed(*addr) && 2115 (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 && 2116 find_space != VMFS_NO_SPACE && object == NULL && 2117 (cow & (MAP_INHERIT_SHARE | MAP_STACK_GROWS_UP | 2118 MAP_STACK_GROWS_DOWN)) == 0 && prot != PROT_NONE; 2119 curr_min_addr = min_addr = *addr; 2120 if (en_aslr && min_addr == 0 && !cluster && 2121 find_space != VMFS_NO_SPACE && 2122 (map->flags & MAP_ASLR_IGNSTART) != 0) 2123 curr_min_addr = min_addr = vm_map_min(map); 2124 try = 0; 2125 vm_map_lock(map); 2126 if (cluster) { 2127 curr_min_addr = map->anon_loc; 2128 if (curr_min_addr == 0) 2129 cluster = false; 2130 } 2131 if (find_space != VMFS_NO_SPACE) { 2132 KASSERT(find_space == VMFS_ANY_SPACE || 2133 find_space == VMFS_OPTIMAL_SPACE || 2134 find_space == VMFS_SUPER_SPACE || 2135 alignment != 0, ("unexpected VMFS flag")); 2136 again: 2137 /* 2138 * When creating an anonymous mapping, try clustering 2139 * with an existing anonymous mapping first. 2140 * 2141 * We make up to two attempts to find address space 2142 * for a given find_space value. The first attempt may 2143 * apply randomization or may cluster with an existing 2144 * anonymous mapping. If this first attempt fails, 2145 * perform a first-fit search of the available address 2146 * space. 2147 * 2148 * If all tries failed, and find_space is 2149 * VMFS_OPTIMAL_SPACE, fallback to VMFS_ANY_SPACE. 2150 * Again enable clustering and randomization. 2151 */ 2152 try++; 2153 MPASS(try <= 2); 2154 2155 if (try == 2) { 2156 /* 2157 * Second try: we failed either to find a 2158 * suitable region for randomizing the 2159 * allocation, or to cluster with an existing 2160 * mapping. Retry with free run. 2161 */ 2162 curr_min_addr = (map->flags & MAP_ASLR_IGNSTART) != 0 ? 2163 vm_map_min(map) : min_addr; 2164 atomic_add_long(&aslr_restarts, 1); 2165 } 2166 2167 if (try == 1 && en_aslr && !cluster) { 2168 /* 2169 * Find space for allocation, including 2170 * gap needed for later randomization. 2171 */ 2172 pidx = MAXPAGESIZES > 1 && pagesizes[1] != 0 && 2173 (find_space == VMFS_SUPER_SPACE || find_space == 2174 VMFS_OPTIMAL_SPACE) ? 1 : 0; 2175 gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR && 2176 (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ? 2177 aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx]; 2178 *addr = vm_map_findspace(map, curr_min_addr, 2179 length + gap * pagesizes[pidx]); 2180 if (*addr + length + gap * pagesizes[pidx] > 2181 vm_map_max(map)) 2182 goto again; 2183 /* And randomize the start address. */ 2184 *addr += (arc4random() % gap) * pagesizes[pidx]; 2185 if (max_addr != 0 && *addr + length > max_addr) 2186 goto again; 2187 } else { 2188 *addr = vm_map_findspace(map, curr_min_addr, length); 2189 if (*addr + length > vm_map_max(map) || 2190 (max_addr != 0 && *addr + length > max_addr)) { 2191 if (cluster) { 2192 cluster = false; 2193 MPASS(try == 1); 2194 goto again; 2195 } 2196 rv = KERN_NO_SPACE; 2197 goto done; 2198 } 2199 } 2200 2201 if (find_space != VMFS_ANY_SPACE && 2202 (rv = vm_map_alignspace(map, object, offset, addr, length, 2203 max_addr, alignment)) != KERN_SUCCESS) { 2204 if (find_space == VMFS_OPTIMAL_SPACE) { 2205 find_space = VMFS_ANY_SPACE; 2206 curr_min_addr = min_addr; 2207 cluster = update_anon; 2208 try = 0; 2209 goto again; 2210 } 2211 goto done; 2212 } 2213 } else if ((cow & MAP_REMAP) != 0) { 2214 if (!vm_map_range_valid(map, *addr, *addr + length)) { 2215 rv = KERN_INVALID_ADDRESS; 2216 goto done; 2217 } 2218 rv = vm_map_delete(map, *addr, *addr + length); 2219 if (rv != KERN_SUCCESS) 2220 goto done; 2221 } 2222 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 2223 rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot, 2224 max, cow); 2225 } else { 2226 rv = vm_map_insert(map, object, offset, *addr, *addr + length, 2227 prot, max, cow); 2228 } 2229 if (rv == KERN_SUCCESS && update_anon) 2230 map->anon_loc = *addr + length; 2231 done: 2232 vm_map_unlock(map); 2233 return (rv); 2234 } 2235 2236 /* 2237 * vm_map_find_min() is a variant of vm_map_find() that takes an 2238 * additional parameter (min_addr) and treats the given address 2239 * (*addr) differently. Specifically, it treats *addr as a hint 2240 * and not as the minimum address where the mapping is created. 2241 * 2242 * This function works in two phases. First, it tries to 2243 * allocate above the hint. If that fails and the hint is 2244 * greater than min_addr, it performs a second pass, replacing 2245 * the hint with min_addr as the minimum address for the 2246 * allocation. 2247 */ 2248 int 2249 vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2250 vm_offset_t *addr, vm_size_t length, vm_offset_t min_addr, 2251 vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max, 2252 int cow) 2253 { 2254 vm_offset_t hint; 2255 int rv; 2256 2257 hint = *addr; 2258 for (;;) { 2259 rv = vm_map_find(map, object, offset, addr, length, max_addr, 2260 find_space, prot, max, cow); 2261 if (rv == KERN_SUCCESS || min_addr >= hint) 2262 return (rv); 2263 *addr = hint = min_addr; 2264 } 2265 } 2266 2267 /* 2268 * A map entry with any of the following flags set must not be merged with 2269 * another entry. 2270 */ 2271 #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP | \ 2272 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC) 2273 2274 static bool 2275 vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry) 2276 { 2277 2278 KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 || 2279 (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0, 2280 ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable", 2281 prev, entry)); 2282 return (prev->end == entry->start && 2283 prev->object.vm_object == entry->object.vm_object && 2284 (prev->object.vm_object == NULL || 2285 prev->offset + (prev->end - prev->start) == entry->offset) && 2286 prev->eflags == entry->eflags && 2287 prev->protection == entry->protection && 2288 prev->max_protection == entry->max_protection && 2289 prev->inheritance == entry->inheritance && 2290 prev->wired_count == entry->wired_count && 2291 prev->cred == entry->cred); 2292 } 2293 2294 static void 2295 vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry) 2296 { 2297 2298 /* 2299 * If the backing object is a vnode object, vm_object_deallocate() 2300 * calls vrele(). However, vrele() does not lock the vnode because 2301 * the vnode has additional references. Thus, the map lock can be 2302 * kept without causing a lock-order reversal with the vnode lock. 2303 * 2304 * Since we count the number of virtual page mappings in 2305 * object->un_pager.vnp.writemappings, the writemappings value 2306 * should not be adjusted when the entry is disposed of. 2307 */ 2308 if (entry->object.vm_object != NULL) 2309 vm_object_deallocate(entry->object.vm_object); 2310 if (entry->cred != NULL) 2311 crfree(entry->cred); 2312 vm_map_entry_dispose(map, entry); 2313 } 2314 2315 /* 2316 * vm_map_try_merge_entries: 2317 * 2318 * Compare the given map entry to its predecessor, and merge its precessor 2319 * into it if possible. The entry remains valid, and may be extended. 2320 * The predecessor may be deleted. 2321 * 2322 * The map must be locked. 2323 */ 2324 void 2325 vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev_entry, 2326 vm_map_entry_t entry) 2327 { 2328 2329 VM_MAP_ASSERT_LOCKED(map); 2330 if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 && 2331 vm_map_mergeable_neighbors(prev_entry, entry)) { 2332 vm_map_entry_unlink(map, prev_entry, UNLINK_MERGE_NEXT); 2333 vm_map_merged_neighbor_dispose(map, prev_entry); 2334 } 2335 } 2336 2337 /* 2338 * vm_map_entry_back: 2339 * 2340 * Allocate an object to back a map entry. 2341 */ 2342 static inline void 2343 vm_map_entry_back(vm_map_entry_t entry) 2344 { 2345 vm_object_t object; 2346 2347 KASSERT(entry->object.vm_object == NULL, 2348 ("map entry %p has backing object", entry)); 2349 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2350 ("map entry %p is a submap", entry)); 2351 object = vm_object_allocate_anon(atop(entry->end - entry->start), NULL, 2352 entry->cred, entry->end - entry->start); 2353 entry->object.vm_object = object; 2354 entry->offset = 0; 2355 entry->cred = NULL; 2356 } 2357 2358 /* 2359 * vm_map_entry_charge_object 2360 * 2361 * If there is no object backing this entry, create one. Otherwise, if 2362 * the entry has cred, give it to the backing object. 2363 */ 2364 static inline void 2365 vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry) 2366 { 2367 2368 VM_MAP_ASSERT_LOCKED(map); 2369 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2370 ("map entry %p is a submap", entry)); 2371 if (entry->object.vm_object == NULL && !map->system_map && 2372 (entry->eflags & MAP_ENTRY_GUARD) == 0) 2373 vm_map_entry_back(entry); 2374 else if (entry->object.vm_object != NULL && 2375 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) && 2376 entry->cred != NULL) { 2377 VM_OBJECT_WLOCK(entry->object.vm_object); 2378 KASSERT(entry->object.vm_object->cred == NULL, 2379 ("OVERCOMMIT: %s: both cred e %p", __func__, entry)); 2380 entry->object.vm_object->cred = entry->cred; 2381 entry->object.vm_object->charge = entry->end - entry->start; 2382 VM_OBJECT_WUNLOCK(entry->object.vm_object); 2383 entry->cred = NULL; 2384 } 2385 } 2386 2387 /* 2388 * vm_map_entry_clone 2389 * 2390 * Create a duplicate map entry for clipping. 2391 */ 2392 static vm_map_entry_t 2393 vm_map_entry_clone(vm_map_t map, vm_map_entry_t entry) 2394 { 2395 vm_map_entry_t new_entry; 2396 2397 VM_MAP_ASSERT_LOCKED(map); 2398 2399 /* 2400 * Create a backing object now, if none exists, so that more individual 2401 * objects won't be created after the map entry is split. 2402 */ 2403 vm_map_entry_charge_object(map, entry); 2404 2405 /* Clone the entry. */ 2406 new_entry = vm_map_entry_create(map); 2407 *new_entry = *entry; 2408 if (new_entry->cred != NULL) 2409 crhold(entry->cred); 2410 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2411 vm_object_reference(new_entry->object.vm_object); 2412 vm_map_entry_set_vnode_text(new_entry, true); 2413 /* 2414 * The object->un_pager.vnp.writemappings for the object of 2415 * MAP_ENTRY_WRITECNT type entry shall be kept as is here. The 2416 * virtual pages are re-distributed among the clipped entries, 2417 * so the sum is left the same. 2418 */ 2419 } 2420 return (new_entry); 2421 } 2422 2423 /* 2424 * vm_map_clip_start: [ internal use only ] 2425 * 2426 * Asserts that the given entry begins at or after 2427 * the specified address; if necessary, 2428 * it splits the entry into two. 2429 */ 2430 static int 2431 vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t startaddr) 2432 { 2433 vm_map_entry_t new_entry; 2434 int bdry_idx; 2435 2436 if (!map->system_map) 2437 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2438 "%s: map %p entry %p start 0x%jx", __func__, map, entry, 2439 (uintmax_t)startaddr); 2440 2441 if (startaddr <= entry->start) 2442 return (KERN_SUCCESS); 2443 2444 VM_MAP_ASSERT_LOCKED(map); 2445 KASSERT(entry->end > startaddr && entry->start < startaddr, 2446 ("%s: invalid clip of entry %p", __func__, entry)); 2447 2448 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 2449 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 2450 if (bdry_idx != 0) { 2451 if ((startaddr & (pagesizes[bdry_idx] - 1)) != 0) 2452 return (KERN_INVALID_ARGUMENT); 2453 } 2454 2455 new_entry = vm_map_entry_clone(map, entry); 2456 2457 /* 2458 * Split off the front portion. Insert the new entry BEFORE this one, 2459 * so that this entry has the specified starting address. 2460 */ 2461 new_entry->end = startaddr; 2462 vm_map_entry_link(map, new_entry); 2463 return (KERN_SUCCESS); 2464 } 2465 2466 /* 2467 * vm_map_lookup_clip_start: 2468 * 2469 * Find the entry at or just after 'start', and clip it if 'start' is in 2470 * the interior of the entry. Return entry after 'start', and in 2471 * prev_entry set the entry before 'start'. 2472 */ 2473 static int 2474 vm_map_lookup_clip_start(vm_map_t map, vm_offset_t start, 2475 vm_map_entry_t *res_entry, vm_map_entry_t *prev_entry) 2476 { 2477 vm_map_entry_t entry; 2478 int rv; 2479 2480 if (!map->system_map) 2481 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2482 "%s: map %p start 0x%jx prev %p", __func__, map, 2483 (uintmax_t)start, prev_entry); 2484 2485 if (vm_map_lookup_entry(map, start, prev_entry)) { 2486 entry = *prev_entry; 2487 rv = vm_map_clip_start(map, entry, start); 2488 if (rv != KERN_SUCCESS) 2489 return (rv); 2490 *prev_entry = vm_map_entry_pred(entry); 2491 } else 2492 entry = vm_map_entry_succ(*prev_entry); 2493 *res_entry = entry; 2494 return (KERN_SUCCESS); 2495 } 2496 2497 /* 2498 * vm_map_clip_end: [ internal use only ] 2499 * 2500 * Asserts that the given entry ends at or before 2501 * the specified address; if necessary, 2502 * it splits the entry into two. 2503 */ 2504 static int 2505 vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t endaddr) 2506 { 2507 vm_map_entry_t new_entry; 2508 int bdry_idx; 2509 2510 if (!map->system_map) 2511 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2512 "%s: map %p entry %p end 0x%jx", __func__, map, entry, 2513 (uintmax_t)endaddr); 2514 2515 if (endaddr >= entry->end) 2516 return (KERN_SUCCESS); 2517 2518 VM_MAP_ASSERT_LOCKED(map); 2519 KASSERT(entry->start < endaddr && entry->end > endaddr, 2520 ("%s: invalid clip of entry %p", __func__, entry)); 2521 2522 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 2523 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 2524 if (bdry_idx != 0) { 2525 if ((endaddr & (pagesizes[bdry_idx] - 1)) != 0) 2526 return (KERN_INVALID_ARGUMENT); 2527 } 2528 2529 new_entry = vm_map_entry_clone(map, entry); 2530 2531 /* 2532 * Split off the back portion. Insert the new entry AFTER this one, 2533 * so that this entry has the specified ending address. 2534 */ 2535 new_entry->start = endaddr; 2536 vm_map_entry_link(map, new_entry); 2537 2538 return (KERN_SUCCESS); 2539 } 2540 2541 /* 2542 * vm_map_submap: [ kernel use only ] 2543 * 2544 * Mark the given range as handled by a subordinate map. 2545 * 2546 * This range must have been created with vm_map_find, 2547 * and no other operations may have been performed on this 2548 * range prior to calling vm_map_submap. 2549 * 2550 * Only a limited number of operations can be performed 2551 * within this rage after calling vm_map_submap: 2552 * vm_fault 2553 * [Don't try vm_map_copy!] 2554 * 2555 * To remove a submapping, one must first remove the 2556 * range from the superior map, and then destroy the 2557 * submap (if desired). [Better yet, don't try it.] 2558 */ 2559 int 2560 vm_map_submap( 2561 vm_map_t map, 2562 vm_offset_t start, 2563 vm_offset_t end, 2564 vm_map_t submap) 2565 { 2566 vm_map_entry_t entry; 2567 int result; 2568 2569 result = KERN_INVALID_ARGUMENT; 2570 2571 vm_map_lock(submap); 2572 submap->flags |= MAP_IS_SUB_MAP; 2573 vm_map_unlock(submap); 2574 2575 vm_map_lock(map); 2576 VM_MAP_RANGE_CHECK(map, start, end); 2577 if (vm_map_lookup_entry(map, start, &entry) && entry->end >= end && 2578 (entry->eflags & MAP_ENTRY_COW) == 0 && 2579 entry->object.vm_object == NULL) { 2580 result = vm_map_clip_start(map, entry, start); 2581 if (result != KERN_SUCCESS) 2582 goto unlock; 2583 result = vm_map_clip_end(map, entry, end); 2584 if (result != KERN_SUCCESS) 2585 goto unlock; 2586 entry->object.sub_map = submap; 2587 entry->eflags |= MAP_ENTRY_IS_SUB_MAP; 2588 result = KERN_SUCCESS; 2589 } 2590 unlock: 2591 vm_map_unlock(map); 2592 2593 if (result != KERN_SUCCESS) { 2594 vm_map_lock(submap); 2595 submap->flags &= ~MAP_IS_SUB_MAP; 2596 vm_map_unlock(submap); 2597 } 2598 return (result); 2599 } 2600 2601 /* 2602 * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified 2603 */ 2604 #define MAX_INIT_PT 96 2605 2606 /* 2607 * vm_map_pmap_enter: 2608 * 2609 * Preload the specified map's pmap with mappings to the specified 2610 * object's memory-resident pages. No further physical pages are 2611 * allocated, and no further virtual pages are retrieved from secondary 2612 * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a 2613 * limited number of page mappings are created at the low-end of the 2614 * specified address range. (For this purpose, a superpage mapping 2615 * counts as one page mapping.) Otherwise, all resident pages within 2616 * the specified address range are mapped. 2617 */ 2618 static void 2619 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 2620 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags) 2621 { 2622 vm_offset_t start; 2623 vm_page_t p, p_start; 2624 vm_pindex_t mask, psize, threshold, tmpidx; 2625 2626 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL) 2627 return; 2628 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2629 VM_OBJECT_WLOCK(object); 2630 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2631 pmap_object_init_pt(map->pmap, addr, object, pindex, 2632 size); 2633 VM_OBJECT_WUNLOCK(object); 2634 return; 2635 } 2636 VM_OBJECT_LOCK_DOWNGRADE(object); 2637 } else 2638 VM_OBJECT_RLOCK(object); 2639 2640 psize = atop(size); 2641 if (psize + pindex > object->size) { 2642 if (pindex >= object->size) { 2643 VM_OBJECT_RUNLOCK(object); 2644 return; 2645 } 2646 psize = object->size - pindex; 2647 } 2648 2649 start = 0; 2650 p_start = NULL; 2651 threshold = MAX_INIT_PT; 2652 2653 p = vm_page_find_least(object, pindex); 2654 /* 2655 * Assert: the variable p is either (1) the page with the 2656 * least pindex greater than or equal to the parameter pindex 2657 * or (2) NULL. 2658 */ 2659 for (; 2660 p != NULL && (tmpidx = p->pindex - pindex) < psize; 2661 p = TAILQ_NEXT(p, listq)) { 2662 /* 2663 * don't allow an madvise to blow away our really 2664 * free pages allocating pv entries. 2665 */ 2666 if (((flags & MAP_PREFAULT_MADVISE) != 0 && 2667 vm_page_count_severe()) || 2668 ((flags & MAP_PREFAULT_PARTIAL) != 0 && 2669 tmpidx >= threshold)) { 2670 psize = tmpidx; 2671 break; 2672 } 2673 if (vm_page_all_valid(p)) { 2674 if (p_start == NULL) { 2675 start = addr + ptoa(tmpidx); 2676 p_start = p; 2677 } 2678 /* Jump ahead if a superpage mapping is possible. */ 2679 if (p->psind > 0 && ((addr + ptoa(tmpidx)) & 2680 (pagesizes[p->psind] - 1)) == 0) { 2681 mask = atop(pagesizes[p->psind]) - 1; 2682 if (tmpidx + mask < psize && 2683 vm_page_ps_test(p, PS_ALL_VALID, NULL)) { 2684 p += mask; 2685 threshold += mask; 2686 } 2687 } 2688 } else if (p_start != NULL) { 2689 pmap_enter_object(map->pmap, start, addr + 2690 ptoa(tmpidx), p_start, prot); 2691 p_start = NULL; 2692 } 2693 } 2694 if (p_start != NULL) 2695 pmap_enter_object(map->pmap, start, addr + ptoa(psize), 2696 p_start, prot); 2697 VM_OBJECT_RUNLOCK(object); 2698 } 2699 2700 /* 2701 * vm_map_protect: 2702 * 2703 * Sets the protection and/or the maximum protection of the 2704 * specified address region in the target map. 2705 */ 2706 int 2707 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 2708 vm_prot_t new_prot, vm_prot_t new_maxprot, int flags) 2709 { 2710 vm_map_entry_t entry, first_entry, in_tran, prev_entry; 2711 vm_object_t obj; 2712 struct ucred *cred; 2713 vm_prot_t old_prot; 2714 int rv; 2715 2716 if (start == end) 2717 return (KERN_SUCCESS); 2718 2719 if ((flags & (VM_MAP_PROTECT_SET_PROT | VM_MAP_PROTECT_SET_MAXPROT)) == 2720 (VM_MAP_PROTECT_SET_PROT | VM_MAP_PROTECT_SET_MAXPROT) && 2721 (new_prot & new_maxprot) != new_prot) 2722 return (KERN_OUT_OF_BOUNDS); 2723 2724 again: 2725 in_tran = NULL; 2726 vm_map_lock(map); 2727 2728 if ((map->flags & MAP_WXORX) != 0 && 2729 (flags & VM_MAP_PROTECT_SET_PROT) != 0 && 2730 (new_prot & (VM_PROT_WRITE | VM_PROT_EXECUTE)) == (VM_PROT_WRITE | 2731 VM_PROT_EXECUTE)) { 2732 vm_map_unlock(map); 2733 return (KERN_PROTECTION_FAILURE); 2734 } 2735 2736 /* 2737 * Ensure that we are not concurrently wiring pages. vm_map_wire() may 2738 * need to fault pages into the map and will drop the map lock while 2739 * doing so, and the VM object may end up in an inconsistent state if we 2740 * update the protection on the map entry in between faults. 2741 */ 2742 vm_map_wait_busy(map); 2743 2744 VM_MAP_RANGE_CHECK(map, start, end); 2745 2746 if (!vm_map_lookup_entry(map, start, &first_entry)) 2747 first_entry = vm_map_entry_succ(first_entry); 2748 2749 /* 2750 * Make a first pass to check for protection violations. 2751 */ 2752 for (entry = first_entry; entry->start < end; 2753 entry = vm_map_entry_succ(entry)) { 2754 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 2755 continue; 2756 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 2757 vm_map_unlock(map); 2758 return (KERN_INVALID_ARGUMENT); 2759 } 2760 if ((flags & VM_MAP_PROTECT_SET_PROT) == 0) 2761 new_prot = entry->protection; 2762 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) == 0) 2763 new_maxprot = entry->max_protection; 2764 if ((new_prot & entry->max_protection) != new_prot || 2765 (new_maxprot & entry->max_protection) != new_maxprot) { 2766 vm_map_unlock(map); 2767 return (KERN_PROTECTION_FAILURE); 2768 } 2769 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0) 2770 in_tran = entry; 2771 } 2772 2773 /* 2774 * Postpone the operation until all in-transition map entries have 2775 * stabilized. An in-transition entry might already have its pages 2776 * wired and wired_count incremented, but not yet have its 2777 * MAP_ENTRY_USER_WIRED flag set. In which case, we would fail to call 2778 * vm_fault_copy_entry() in the final loop below. 2779 */ 2780 if (in_tran != NULL) { 2781 in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2782 vm_map_unlock_and_wait(map, 0); 2783 goto again; 2784 } 2785 2786 /* 2787 * Before changing the protections, try to reserve swap space for any 2788 * private (i.e., copy-on-write) mappings that are transitioning from 2789 * read-only to read/write access. If a reservation fails, break out 2790 * of this loop early and let the next loop simplify the entries, since 2791 * some may now be mergeable. 2792 */ 2793 rv = vm_map_clip_start(map, first_entry, start); 2794 if (rv != KERN_SUCCESS) { 2795 vm_map_unlock(map); 2796 return (rv); 2797 } 2798 for (entry = first_entry; entry->start < end; 2799 entry = vm_map_entry_succ(entry)) { 2800 rv = vm_map_clip_end(map, entry, end); 2801 if (rv != KERN_SUCCESS) { 2802 vm_map_unlock(map); 2803 return (rv); 2804 } 2805 2806 if ((flags & VM_MAP_PROTECT_SET_PROT) == 0 || 2807 ((new_prot & ~entry->protection) & VM_PROT_WRITE) == 0 || 2808 ENTRY_CHARGED(entry) || 2809 (entry->eflags & MAP_ENTRY_GUARD) != 0) 2810 continue; 2811 2812 cred = curthread->td_ucred; 2813 obj = entry->object.vm_object; 2814 2815 if (obj == NULL || 2816 (entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0) { 2817 if (!swap_reserve(entry->end - entry->start)) { 2818 rv = KERN_RESOURCE_SHORTAGE; 2819 end = entry->end; 2820 break; 2821 } 2822 crhold(cred); 2823 entry->cred = cred; 2824 continue; 2825 } 2826 2827 VM_OBJECT_WLOCK(obj); 2828 if (obj->type != OBJT_DEFAULT && 2829 (obj->flags & OBJ_SWAP) == 0) { 2830 VM_OBJECT_WUNLOCK(obj); 2831 continue; 2832 } 2833 2834 /* 2835 * Charge for the whole object allocation now, since 2836 * we cannot distinguish between non-charged and 2837 * charged clipped mapping of the same object later. 2838 */ 2839 KASSERT(obj->charge == 0, 2840 ("vm_map_protect: object %p overcharged (entry %p)", 2841 obj, entry)); 2842 if (!swap_reserve(ptoa(obj->size))) { 2843 VM_OBJECT_WUNLOCK(obj); 2844 rv = KERN_RESOURCE_SHORTAGE; 2845 end = entry->end; 2846 break; 2847 } 2848 2849 crhold(cred); 2850 obj->cred = cred; 2851 obj->charge = ptoa(obj->size); 2852 VM_OBJECT_WUNLOCK(obj); 2853 } 2854 2855 /* 2856 * If enough swap space was available, go back and fix up protections. 2857 * Otherwise, just simplify entries, since some may have been modified. 2858 * [Note that clipping is not necessary the second time.] 2859 */ 2860 for (prev_entry = vm_map_entry_pred(first_entry), entry = first_entry; 2861 entry->start < end; 2862 vm_map_try_merge_entries(map, prev_entry, entry), 2863 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 2864 if (rv != KERN_SUCCESS || 2865 (entry->eflags & MAP_ENTRY_GUARD) != 0) 2866 continue; 2867 2868 old_prot = entry->protection; 2869 2870 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) { 2871 entry->max_protection = new_maxprot; 2872 entry->protection = new_maxprot & old_prot; 2873 } 2874 if ((flags & VM_MAP_PROTECT_SET_PROT) != 0) 2875 entry->protection = new_prot; 2876 2877 /* 2878 * For user wired map entries, the normal lazy evaluation of 2879 * write access upgrades through soft page faults is 2880 * undesirable. Instead, immediately copy any pages that are 2881 * copy-on-write and enable write access in the physical map. 2882 */ 2883 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0 && 2884 (entry->protection & VM_PROT_WRITE) != 0 && 2885 (old_prot & VM_PROT_WRITE) == 0) 2886 vm_fault_copy_entry(map, map, entry, entry, NULL); 2887 2888 /* 2889 * When restricting access, update the physical map. Worry 2890 * about copy-on-write here. 2891 */ 2892 if ((old_prot & ~entry->protection) != 0) { 2893 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 2894 VM_PROT_ALL) 2895 pmap_protect(map->pmap, entry->start, 2896 entry->end, 2897 entry->protection & MASK(entry)); 2898 #undef MASK 2899 } 2900 } 2901 vm_map_try_merge_entries(map, prev_entry, entry); 2902 vm_map_unlock(map); 2903 return (rv); 2904 } 2905 2906 /* 2907 * vm_map_madvise: 2908 * 2909 * This routine traverses a processes map handling the madvise 2910 * system call. Advisories are classified as either those effecting 2911 * the vm_map_entry structure, or those effecting the underlying 2912 * objects. 2913 */ 2914 int 2915 vm_map_madvise( 2916 vm_map_t map, 2917 vm_offset_t start, 2918 vm_offset_t end, 2919 int behav) 2920 { 2921 vm_map_entry_t entry, prev_entry; 2922 int rv; 2923 bool modify_map; 2924 2925 /* 2926 * Some madvise calls directly modify the vm_map_entry, in which case 2927 * we need to use an exclusive lock on the map and we need to perform 2928 * various clipping operations. Otherwise we only need a read-lock 2929 * on the map. 2930 */ 2931 switch(behav) { 2932 case MADV_NORMAL: 2933 case MADV_SEQUENTIAL: 2934 case MADV_RANDOM: 2935 case MADV_NOSYNC: 2936 case MADV_AUTOSYNC: 2937 case MADV_NOCORE: 2938 case MADV_CORE: 2939 if (start == end) 2940 return (0); 2941 modify_map = true; 2942 vm_map_lock(map); 2943 break; 2944 case MADV_WILLNEED: 2945 case MADV_DONTNEED: 2946 case MADV_FREE: 2947 if (start == end) 2948 return (0); 2949 modify_map = false; 2950 vm_map_lock_read(map); 2951 break; 2952 default: 2953 return (EINVAL); 2954 } 2955 2956 /* 2957 * Locate starting entry and clip if necessary. 2958 */ 2959 VM_MAP_RANGE_CHECK(map, start, end); 2960 2961 if (modify_map) { 2962 /* 2963 * madvise behaviors that are implemented in the vm_map_entry. 2964 * 2965 * We clip the vm_map_entry so that behavioral changes are 2966 * limited to the specified address range. 2967 */ 2968 rv = vm_map_lookup_clip_start(map, start, &entry, &prev_entry); 2969 if (rv != KERN_SUCCESS) { 2970 vm_map_unlock(map); 2971 return (vm_mmap_to_errno(rv)); 2972 } 2973 2974 for (; entry->start < end; prev_entry = entry, 2975 entry = vm_map_entry_succ(entry)) { 2976 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2977 continue; 2978 2979 rv = vm_map_clip_end(map, entry, end); 2980 if (rv != KERN_SUCCESS) { 2981 vm_map_unlock(map); 2982 return (vm_mmap_to_errno(rv)); 2983 } 2984 2985 switch (behav) { 2986 case MADV_NORMAL: 2987 vm_map_entry_set_behavior(entry, 2988 MAP_ENTRY_BEHAV_NORMAL); 2989 break; 2990 case MADV_SEQUENTIAL: 2991 vm_map_entry_set_behavior(entry, 2992 MAP_ENTRY_BEHAV_SEQUENTIAL); 2993 break; 2994 case MADV_RANDOM: 2995 vm_map_entry_set_behavior(entry, 2996 MAP_ENTRY_BEHAV_RANDOM); 2997 break; 2998 case MADV_NOSYNC: 2999 entry->eflags |= MAP_ENTRY_NOSYNC; 3000 break; 3001 case MADV_AUTOSYNC: 3002 entry->eflags &= ~MAP_ENTRY_NOSYNC; 3003 break; 3004 case MADV_NOCORE: 3005 entry->eflags |= MAP_ENTRY_NOCOREDUMP; 3006 break; 3007 case MADV_CORE: 3008 entry->eflags &= ~MAP_ENTRY_NOCOREDUMP; 3009 break; 3010 default: 3011 break; 3012 } 3013 vm_map_try_merge_entries(map, prev_entry, entry); 3014 } 3015 vm_map_try_merge_entries(map, prev_entry, entry); 3016 vm_map_unlock(map); 3017 } else { 3018 vm_pindex_t pstart, pend; 3019 3020 /* 3021 * madvise behaviors that are implemented in the underlying 3022 * vm_object. 3023 * 3024 * Since we don't clip the vm_map_entry, we have to clip 3025 * the vm_object pindex and count. 3026 */ 3027 if (!vm_map_lookup_entry(map, start, &entry)) 3028 entry = vm_map_entry_succ(entry); 3029 for (; entry->start < end; 3030 entry = vm_map_entry_succ(entry)) { 3031 vm_offset_t useEnd, useStart; 3032 3033 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 3034 continue; 3035 3036 /* 3037 * MADV_FREE would otherwise rewind time to 3038 * the creation of the shadow object. Because 3039 * we hold the VM map read-locked, neither the 3040 * entry's object nor the presence of a 3041 * backing object can change. 3042 */ 3043 if (behav == MADV_FREE && 3044 entry->object.vm_object != NULL && 3045 entry->object.vm_object->backing_object != NULL) 3046 continue; 3047 3048 pstart = OFF_TO_IDX(entry->offset); 3049 pend = pstart + atop(entry->end - entry->start); 3050 useStart = entry->start; 3051 useEnd = entry->end; 3052 3053 if (entry->start < start) { 3054 pstart += atop(start - entry->start); 3055 useStart = start; 3056 } 3057 if (entry->end > end) { 3058 pend -= atop(entry->end - end); 3059 useEnd = end; 3060 } 3061 3062 if (pstart >= pend) 3063 continue; 3064 3065 /* 3066 * Perform the pmap_advise() before clearing 3067 * PGA_REFERENCED in vm_page_advise(). Otherwise, a 3068 * concurrent pmap operation, such as pmap_remove(), 3069 * could clear a reference in the pmap and set 3070 * PGA_REFERENCED on the page before the pmap_advise() 3071 * had completed. Consequently, the page would appear 3072 * referenced based upon an old reference that 3073 * occurred before this pmap_advise() ran. 3074 */ 3075 if (behav == MADV_DONTNEED || behav == MADV_FREE) 3076 pmap_advise(map->pmap, useStart, useEnd, 3077 behav); 3078 3079 vm_object_madvise(entry->object.vm_object, pstart, 3080 pend, behav); 3081 3082 /* 3083 * Pre-populate paging structures in the 3084 * WILLNEED case. For wired entries, the 3085 * paging structures are already populated. 3086 */ 3087 if (behav == MADV_WILLNEED && 3088 entry->wired_count == 0) { 3089 vm_map_pmap_enter(map, 3090 useStart, 3091 entry->protection, 3092 entry->object.vm_object, 3093 pstart, 3094 ptoa(pend - pstart), 3095 MAP_PREFAULT_MADVISE 3096 ); 3097 } 3098 } 3099 vm_map_unlock_read(map); 3100 } 3101 return (0); 3102 } 3103 3104 /* 3105 * vm_map_inherit: 3106 * 3107 * Sets the inheritance of the specified address 3108 * range in the target map. Inheritance 3109 * affects how the map will be shared with 3110 * child maps at the time of vmspace_fork. 3111 */ 3112 int 3113 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 3114 vm_inherit_t new_inheritance) 3115 { 3116 vm_map_entry_t entry, lentry, prev_entry, start_entry; 3117 int rv; 3118 3119 switch (new_inheritance) { 3120 case VM_INHERIT_NONE: 3121 case VM_INHERIT_COPY: 3122 case VM_INHERIT_SHARE: 3123 case VM_INHERIT_ZERO: 3124 break; 3125 default: 3126 return (KERN_INVALID_ARGUMENT); 3127 } 3128 if (start == end) 3129 return (KERN_SUCCESS); 3130 vm_map_lock(map); 3131 VM_MAP_RANGE_CHECK(map, start, end); 3132 rv = vm_map_lookup_clip_start(map, start, &start_entry, &prev_entry); 3133 if (rv != KERN_SUCCESS) 3134 goto unlock; 3135 if (vm_map_lookup_entry(map, end - 1, &lentry)) { 3136 rv = vm_map_clip_end(map, lentry, end); 3137 if (rv != KERN_SUCCESS) 3138 goto unlock; 3139 } 3140 if (new_inheritance == VM_INHERIT_COPY) { 3141 for (entry = start_entry; entry->start < end; 3142 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3143 if ((entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) 3144 != 0) { 3145 rv = KERN_INVALID_ARGUMENT; 3146 goto unlock; 3147 } 3148 } 3149 } 3150 for (entry = start_entry; entry->start < end; prev_entry = entry, 3151 entry = vm_map_entry_succ(entry)) { 3152 KASSERT(entry->end <= end, ("non-clipped entry %p end %jx %jx", 3153 entry, (uintmax_t)entry->end, (uintmax_t)end)); 3154 if ((entry->eflags & MAP_ENTRY_GUARD) == 0 || 3155 new_inheritance != VM_INHERIT_ZERO) 3156 entry->inheritance = new_inheritance; 3157 vm_map_try_merge_entries(map, prev_entry, entry); 3158 } 3159 vm_map_try_merge_entries(map, prev_entry, entry); 3160 unlock: 3161 vm_map_unlock(map); 3162 return (rv); 3163 } 3164 3165 /* 3166 * vm_map_entry_in_transition: 3167 * 3168 * Release the map lock, and sleep until the entry is no longer in 3169 * transition. Awake and acquire the map lock. If the map changed while 3170 * another held the lock, lookup a possibly-changed entry at or after the 3171 * 'start' position of the old entry. 3172 */ 3173 static vm_map_entry_t 3174 vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start, 3175 vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry) 3176 { 3177 vm_map_entry_t entry; 3178 vm_offset_t start; 3179 u_int last_timestamp; 3180 3181 VM_MAP_ASSERT_LOCKED(map); 3182 KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3183 ("not in-tranition map entry %p", in_entry)); 3184 /* 3185 * We have not yet clipped the entry. 3186 */ 3187 start = MAX(in_start, in_entry->start); 3188 in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3189 last_timestamp = map->timestamp; 3190 if (vm_map_unlock_and_wait(map, 0)) { 3191 /* 3192 * Allow interruption of user wiring/unwiring? 3193 */ 3194 } 3195 vm_map_lock(map); 3196 if (last_timestamp + 1 == map->timestamp) 3197 return (in_entry); 3198 3199 /* 3200 * Look again for the entry because the map was modified while it was 3201 * unlocked. Specifically, the entry may have been clipped, merged, or 3202 * deleted. 3203 */ 3204 if (!vm_map_lookup_entry(map, start, &entry)) { 3205 if (!holes_ok) { 3206 *io_end = start; 3207 return (NULL); 3208 } 3209 entry = vm_map_entry_succ(entry); 3210 } 3211 return (entry); 3212 } 3213 3214 /* 3215 * vm_map_unwire: 3216 * 3217 * Implements both kernel and user unwiring. 3218 */ 3219 int 3220 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 3221 int flags) 3222 { 3223 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3224 int rv; 3225 bool holes_ok, need_wakeup, user_unwire; 3226 3227 if (start == end) 3228 return (KERN_SUCCESS); 3229 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3230 user_unwire = (flags & VM_MAP_WIRE_USER) != 0; 3231 vm_map_lock(map); 3232 VM_MAP_RANGE_CHECK(map, start, end); 3233 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3234 if (holes_ok) 3235 first_entry = vm_map_entry_succ(first_entry); 3236 else { 3237 vm_map_unlock(map); 3238 return (KERN_INVALID_ADDRESS); 3239 } 3240 } 3241 rv = KERN_SUCCESS; 3242 for (entry = first_entry; entry->start < end; entry = next_entry) { 3243 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3244 /* 3245 * We have not yet clipped the entry. 3246 */ 3247 next_entry = vm_map_entry_in_transition(map, start, 3248 &end, holes_ok, entry); 3249 if (next_entry == NULL) { 3250 if (entry == first_entry) { 3251 vm_map_unlock(map); 3252 return (KERN_INVALID_ADDRESS); 3253 } 3254 rv = KERN_INVALID_ADDRESS; 3255 break; 3256 } 3257 first_entry = (entry == first_entry) ? 3258 next_entry : NULL; 3259 continue; 3260 } 3261 rv = vm_map_clip_start(map, entry, start); 3262 if (rv != KERN_SUCCESS) 3263 break; 3264 rv = vm_map_clip_end(map, entry, end); 3265 if (rv != KERN_SUCCESS) 3266 break; 3267 3268 /* 3269 * Mark the entry in case the map lock is released. (See 3270 * above.) 3271 */ 3272 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3273 entry->wiring_thread == NULL, 3274 ("owned map entry %p", entry)); 3275 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3276 entry->wiring_thread = curthread; 3277 next_entry = vm_map_entry_succ(entry); 3278 /* 3279 * Check the map for holes in the specified region. 3280 * If holes_ok, skip this check. 3281 */ 3282 if (!holes_ok && 3283 entry->end < end && next_entry->start > entry->end) { 3284 end = entry->end; 3285 rv = KERN_INVALID_ADDRESS; 3286 break; 3287 } 3288 /* 3289 * If system unwiring, require that the entry is system wired. 3290 */ 3291 if (!user_unwire && 3292 vm_map_entry_system_wired_count(entry) == 0) { 3293 end = entry->end; 3294 rv = KERN_INVALID_ARGUMENT; 3295 break; 3296 } 3297 } 3298 need_wakeup = false; 3299 if (first_entry == NULL && 3300 !vm_map_lookup_entry(map, start, &first_entry)) { 3301 KASSERT(holes_ok, ("vm_map_unwire: lookup failed")); 3302 prev_entry = first_entry; 3303 entry = vm_map_entry_succ(first_entry); 3304 } else { 3305 prev_entry = vm_map_entry_pred(first_entry); 3306 entry = first_entry; 3307 } 3308 for (; entry->start < end; 3309 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3310 /* 3311 * If holes_ok was specified, an empty 3312 * space in the unwired region could have been mapped 3313 * while the map lock was dropped for draining 3314 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread 3315 * could be simultaneously wiring this new mapping 3316 * entry. Detect these cases and skip any entries 3317 * marked as in transition by us. 3318 */ 3319 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3320 entry->wiring_thread != curthread) { 3321 KASSERT(holes_ok, 3322 ("vm_map_unwire: !HOLESOK and new/changed entry")); 3323 continue; 3324 } 3325 3326 if (rv == KERN_SUCCESS && (!user_unwire || 3327 (entry->eflags & MAP_ENTRY_USER_WIRED))) { 3328 if (entry->wired_count == 1) 3329 vm_map_entry_unwire(map, entry); 3330 else 3331 entry->wired_count--; 3332 if (user_unwire) 3333 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3334 } 3335 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3336 ("vm_map_unwire: in-transition flag missing %p", entry)); 3337 KASSERT(entry->wiring_thread == curthread, 3338 ("vm_map_unwire: alien wire %p", entry)); 3339 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 3340 entry->wiring_thread = NULL; 3341 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3342 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3343 need_wakeup = true; 3344 } 3345 vm_map_try_merge_entries(map, prev_entry, entry); 3346 } 3347 vm_map_try_merge_entries(map, prev_entry, entry); 3348 vm_map_unlock(map); 3349 if (need_wakeup) 3350 vm_map_wakeup(map); 3351 return (rv); 3352 } 3353 3354 static void 3355 vm_map_wire_user_count_sub(u_long npages) 3356 { 3357 3358 atomic_subtract_long(&vm_user_wire_count, npages); 3359 } 3360 3361 static bool 3362 vm_map_wire_user_count_add(u_long npages) 3363 { 3364 u_long wired; 3365 3366 wired = vm_user_wire_count; 3367 do { 3368 if (npages + wired > vm_page_max_user_wired) 3369 return (false); 3370 } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired, 3371 npages + wired)); 3372 3373 return (true); 3374 } 3375 3376 /* 3377 * vm_map_wire_entry_failure: 3378 * 3379 * Handle a wiring failure on the given entry. 3380 * 3381 * The map should be locked. 3382 */ 3383 static void 3384 vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 3385 vm_offset_t failed_addr) 3386 { 3387 3388 VM_MAP_ASSERT_LOCKED(map); 3389 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 && 3390 entry->wired_count == 1, 3391 ("vm_map_wire_entry_failure: entry %p isn't being wired", entry)); 3392 KASSERT(failed_addr < entry->end, 3393 ("vm_map_wire_entry_failure: entry %p was fully wired", entry)); 3394 3395 /* 3396 * If any pages at the start of this entry were successfully wired, 3397 * then unwire them. 3398 */ 3399 if (failed_addr > entry->start) { 3400 pmap_unwire(map->pmap, entry->start, failed_addr); 3401 vm_object_unwire(entry->object.vm_object, entry->offset, 3402 failed_addr - entry->start, PQ_ACTIVE); 3403 } 3404 3405 /* 3406 * Assign an out-of-range value to represent the failure to wire this 3407 * entry. 3408 */ 3409 entry->wired_count = -1; 3410 } 3411 3412 int 3413 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3414 { 3415 int rv; 3416 3417 vm_map_lock(map); 3418 rv = vm_map_wire_locked(map, start, end, flags); 3419 vm_map_unlock(map); 3420 return (rv); 3421 } 3422 3423 /* 3424 * vm_map_wire_locked: 3425 * 3426 * Implements both kernel and user wiring. Returns with the map locked, 3427 * the map lock may be dropped. 3428 */ 3429 int 3430 vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3431 { 3432 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3433 vm_offset_t faddr, saved_end, saved_start; 3434 u_long incr, npages; 3435 u_int bidx, last_timestamp; 3436 int rv; 3437 bool holes_ok, need_wakeup, user_wire; 3438 vm_prot_t prot; 3439 3440 VM_MAP_ASSERT_LOCKED(map); 3441 3442 if (start == end) 3443 return (KERN_SUCCESS); 3444 prot = 0; 3445 if (flags & VM_MAP_WIRE_WRITE) 3446 prot |= VM_PROT_WRITE; 3447 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3448 user_wire = (flags & VM_MAP_WIRE_USER) != 0; 3449 VM_MAP_RANGE_CHECK(map, start, end); 3450 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3451 if (holes_ok) 3452 first_entry = vm_map_entry_succ(first_entry); 3453 else 3454 return (KERN_INVALID_ADDRESS); 3455 } 3456 for (entry = first_entry; entry->start < end; entry = next_entry) { 3457 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3458 /* 3459 * We have not yet clipped the entry. 3460 */ 3461 next_entry = vm_map_entry_in_transition(map, start, 3462 &end, holes_ok, entry); 3463 if (next_entry == NULL) { 3464 if (entry == first_entry) 3465 return (KERN_INVALID_ADDRESS); 3466 rv = KERN_INVALID_ADDRESS; 3467 goto done; 3468 } 3469 first_entry = (entry == first_entry) ? 3470 next_entry : NULL; 3471 continue; 3472 } 3473 rv = vm_map_clip_start(map, entry, start); 3474 if (rv != KERN_SUCCESS) 3475 goto done; 3476 rv = vm_map_clip_end(map, entry, end); 3477 if (rv != KERN_SUCCESS) 3478 goto done; 3479 3480 /* 3481 * Mark the entry in case the map lock is released. (See 3482 * above.) 3483 */ 3484 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3485 entry->wiring_thread == NULL, 3486 ("owned map entry %p", entry)); 3487 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3488 entry->wiring_thread = curthread; 3489 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 3490 || (entry->protection & prot) != prot) { 3491 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED; 3492 if (!holes_ok) { 3493 end = entry->end; 3494 rv = KERN_INVALID_ADDRESS; 3495 goto done; 3496 } 3497 } else if (entry->wired_count == 0) { 3498 entry->wired_count++; 3499 3500 npages = atop(entry->end - entry->start); 3501 if (user_wire && !vm_map_wire_user_count_add(npages)) { 3502 vm_map_wire_entry_failure(map, entry, 3503 entry->start); 3504 end = entry->end; 3505 rv = KERN_RESOURCE_SHORTAGE; 3506 goto done; 3507 } 3508 3509 /* 3510 * Release the map lock, relying on the in-transition 3511 * mark. Mark the map busy for fork. 3512 */ 3513 saved_start = entry->start; 3514 saved_end = entry->end; 3515 last_timestamp = map->timestamp; 3516 bidx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) 3517 >> MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 3518 incr = pagesizes[bidx]; 3519 vm_map_busy(map); 3520 vm_map_unlock(map); 3521 3522 for (faddr = saved_start; faddr < saved_end; 3523 faddr += incr) { 3524 /* 3525 * Simulate a fault to get the page and enter 3526 * it into the physical map. 3527 */ 3528 rv = vm_fault(map, faddr, VM_PROT_NONE, 3529 VM_FAULT_WIRE, NULL); 3530 if (rv != KERN_SUCCESS) 3531 break; 3532 } 3533 vm_map_lock(map); 3534 vm_map_unbusy(map); 3535 if (last_timestamp + 1 != map->timestamp) { 3536 /* 3537 * Look again for the entry because the map was 3538 * modified while it was unlocked. The entry 3539 * may have been clipped, but NOT merged or 3540 * deleted. 3541 */ 3542 if (!vm_map_lookup_entry(map, saved_start, 3543 &next_entry)) 3544 KASSERT(false, 3545 ("vm_map_wire: lookup failed")); 3546 first_entry = (entry == first_entry) ? 3547 next_entry : NULL; 3548 for (entry = next_entry; entry->end < saved_end; 3549 entry = vm_map_entry_succ(entry)) { 3550 /* 3551 * In case of failure, handle entries 3552 * that were not fully wired here; 3553 * fully wired entries are handled 3554 * later. 3555 */ 3556 if (rv != KERN_SUCCESS && 3557 faddr < entry->end) 3558 vm_map_wire_entry_failure(map, 3559 entry, faddr); 3560 } 3561 } 3562 if (rv != KERN_SUCCESS) { 3563 vm_map_wire_entry_failure(map, entry, faddr); 3564 if (user_wire) 3565 vm_map_wire_user_count_sub(npages); 3566 end = entry->end; 3567 goto done; 3568 } 3569 } else if (!user_wire || 3570 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3571 entry->wired_count++; 3572 } 3573 /* 3574 * Check the map for holes in the specified region. 3575 * If holes_ok was specified, skip this check. 3576 */ 3577 next_entry = vm_map_entry_succ(entry); 3578 if (!holes_ok && 3579 entry->end < end && next_entry->start > entry->end) { 3580 end = entry->end; 3581 rv = KERN_INVALID_ADDRESS; 3582 goto done; 3583 } 3584 } 3585 rv = KERN_SUCCESS; 3586 done: 3587 need_wakeup = false; 3588 if (first_entry == NULL && 3589 !vm_map_lookup_entry(map, start, &first_entry)) { 3590 KASSERT(holes_ok, ("vm_map_wire: lookup failed")); 3591 prev_entry = first_entry; 3592 entry = vm_map_entry_succ(first_entry); 3593 } else { 3594 prev_entry = vm_map_entry_pred(first_entry); 3595 entry = first_entry; 3596 } 3597 for (; entry->start < end; 3598 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3599 /* 3600 * If holes_ok was specified, an empty 3601 * space in the unwired region could have been mapped 3602 * while the map lock was dropped for faulting in the 3603 * pages or draining MAP_ENTRY_IN_TRANSITION. 3604 * Moreover, another thread could be simultaneously 3605 * wiring this new mapping entry. Detect these cases 3606 * and skip any entries marked as in transition not by us. 3607 * 3608 * Another way to get an entry not marked with 3609 * MAP_ENTRY_IN_TRANSITION is after failed clipping, 3610 * which set rv to KERN_INVALID_ARGUMENT. 3611 */ 3612 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3613 entry->wiring_thread != curthread) { 3614 KASSERT(holes_ok || rv == KERN_INVALID_ARGUMENT, 3615 ("vm_map_wire: !HOLESOK and new/changed entry")); 3616 continue; 3617 } 3618 3619 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) { 3620 /* do nothing */ 3621 } else if (rv == KERN_SUCCESS) { 3622 if (user_wire) 3623 entry->eflags |= MAP_ENTRY_USER_WIRED; 3624 } else if (entry->wired_count == -1) { 3625 /* 3626 * Wiring failed on this entry. Thus, unwiring is 3627 * unnecessary. 3628 */ 3629 entry->wired_count = 0; 3630 } else if (!user_wire || 3631 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3632 /* 3633 * Undo the wiring. Wiring succeeded on this entry 3634 * but failed on a later entry. 3635 */ 3636 if (entry->wired_count == 1) { 3637 vm_map_entry_unwire(map, entry); 3638 if (user_wire) 3639 vm_map_wire_user_count_sub( 3640 atop(entry->end - entry->start)); 3641 } else 3642 entry->wired_count--; 3643 } 3644 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3645 ("vm_map_wire: in-transition flag missing %p", entry)); 3646 KASSERT(entry->wiring_thread == curthread, 3647 ("vm_map_wire: alien wire %p", entry)); 3648 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION | 3649 MAP_ENTRY_WIRE_SKIPPED); 3650 entry->wiring_thread = NULL; 3651 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3652 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3653 need_wakeup = true; 3654 } 3655 vm_map_try_merge_entries(map, prev_entry, entry); 3656 } 3657 vm_map_try_merge_entries(map, prev_entry, entry); 3658 if (need_wakeup) 3659 vm_map_wakeup(map); 3660 return (rv); 3661 } 3662 3663 /* 3664 * vm_map_sync 3665 * 3666 * Push any dirty cached pages in the address range to their pager. 3667 * If syncio is TRUE, dirty pages are written synchronously. 3668 * If invalidate is TRUE, any cached pages are freed as well. 3669 * 3670 * If the size of the region from start to end is zero, we are 3671 * supposed to flush all modified pages within the region containing 3672 * start. Unfortunately, a region can be split or coalesced with 3673 * neighboring regions, making it difficult to determine what the 3674 * original region was. Therefore, we approximate this requirement by 3675 * flushing the current region containing start. 3676 * 3677 * Returns an error if any part of the specified range is not mapped. 3678 */ 3679 int 3680 vm_map_sync( 3681 vm_map_t map, 3682 vm_offset_t start, 3683 vm_offset_t end, 3684 boolean_t syncio, 3685 boolean_t invalidate) 3686 { 3687 vm_map_entry_t entry, first_entry, next_entry; 3688 vm_size_t size; 3689 vm_object_t object; 3690 vm_ooffset_t offset; 3691 unsigned int last_timestamp; 3692 int bdry_idx; 3693 boolean_t failed; 3694 3695 vm_map_lock_read(map); 3696 VM_MAP_RANGE_CHECK(map, start, end); 3697 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3698 vm_map_unlock_read(map); 3699 return (KERN_INVALID_ADDRESS); 3700 } else if (start == end) { 3701 start = first_entry->start; 3702 end = first_entry->end; 3703 } 3704 3705 /* 3706 * Make a first pass to check for user-wired memory, holes, 3707 * and partial invalidation of largepage mappings. 3708 */ 3709 for (entry = first_entry; entry->start < end; entry = next_entry) { 3710 if (invalidate) { 3711 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) { 3712 vm_map_unlock_read(map); 3713 return (KERN_INVALID_ARGUMENT); 3714 } 3715 bdry_idx = (entry->eflags & 3716 MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 3717 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 3718 if (bdry_idx != 0 && 3719 ((start & (pagesizes[bdry_idx] - 1)) != 0 || 3720 (end & (pagesizes[bdry_idx] - 1)) != 0)) { 3721 vm_map_unlock_read(map); 3722 return (KERN_INVALID_ARGUMENT); 3723 } 3724 } 3725 next_entry = vm_map_entry_succ(entry); 3726 if (end > entry->end && 3727 entry->end != next_entry->start) { 3728 vm_map_unlock_read(map); 3729 return (KERN_INVALID_ADDRESS); 3730 } 3731 } 3732 3733 if (invalidate) 3734 pmap_remove(map->pmap, start, end); 3735 failed = FALSE; 3736 3737 /* 3738 * Make a second pass, cleaning/uncaching pages from the indicated 3739 * objects as we go. 3740 */ 3741 for (entry = first_entry; entry->start < end;) { 3742 offset = entry->offset + (start - entry->start); 3743 size = (end <= entry->end ? end : entry->end) - start; 3744 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 3745 vm_map_t smap; 3746 vm_map_entry_t tentry; 3747 vm_size_t tsize; 3748 3749 smap = entry->object.sub_map; 3750 vm_map_lock_read(smap); 3751 (void) vm_map_lookup_entry(smap, offset, &tentry); 3752 tsize = tentry->end - offset; 3753 if (tsize < size) 3754 size = tsize; 3755 object = tentry->object.vm_object; 3756 offset = tentry->offset + (offset - tentry->start); 3757 vm_map_unlock_read(smap); 3758 } else { 3759 object = entry->object.vm_object; 3760 } 3761 vm_object_reference(object); 3762 last_timestamp = map->timestamp; 3763 vm_map_unlock_read(map); 3764 if (!vm_object_sync(object, offset, size, syncio, invalidate)) 3765 failed = TRUE; 3766 start += size; 3767 vm_object_deallocate(object); 3768 vm_map_lock_read(map); 3769 if (last_timestamp == map->timestamp || 3770 !vm_map_lookup_entry(map, start, &entry)) 3771 entry = vm_map_entry_succ(entry); 3772 } 3773 3774 vm_map_unlock_read(map); 3775 return (failed ? KERN_FAILURE : KERN_SUCCESS); 3776 } 3777 3778 /* 3779 * vm_map_entry_unwire: [ internal use only ] 3780 * 3781 * Make the region specified by this entry pageable. 3782 * 3783 * The map in question should be locked. 3784 * [This is the reason for this routine's existence.] 3785 */ 3786 static void 3787 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 3788 { 3789 vm_size_t size; 3790 3791 VM_MAP_ASSERT_LOCKED(map); 3792 KASSERT(entry->wired_count > 0, 3793 ("vm_map_entry_unwire: entry %p isn't wired", entry)); 3794 3795 size = entry->end - entry->start; 3796 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) 3797 vm_map_wire_user_count_sub(atop(size)); 3798 pmap_unwire(map->pmap, entry->start, entry->end); 3799 vm_object_unwire(entry->object.vm_object, entry->offset, size, 3800 PQ_ACTIVE); 3801 entry->wired_count = 0; 3802 } 3803 3804 static void 3805 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map) 3806 { 3807 3808 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) 3809 vm_object_deallocate(entry->object.vm_object); 3810 uma_zfree(system_map ? kmapentzone : mapentzone, entry); 3811 } 3812 3813 /* 3814 * vm_map_entry_delete: [ internal use only ] 3815 * 3816 * Deallocate the given entry from the target map. 3817 */ 3818 static void 3819 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry) 3820 { 3821 vm_object_t object; 3822 vm_pindex_t offidxstart, offidxend, size1; 3823 vm_size_t size; 3824 3825 vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE); 3826 object = entry->object.vm_object; 3827 3828 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { 3829 MPASS(entry->cred == NULL); 3830 MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0); 3831 MPASS(object == NULL); 3832 vm_map_entry_deallocate(entry, map->system_map); 3833 return; 3834 } 3835 3836 size = entry->end - entry->start; 3837 map->size -= size; 3838 3839 if (entry->cred != NULL) { 3840 swap_release_by_cred(size, entry->cred); 3841 crfree(entry->cred); 3842 } 3843 3844 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || object == NULL) { 3845 entry->object.vm_object = NULL; 3846 } else if ((object->flags & OBJ_ANON) != 0 || 3847 object == kernel_object) { 3848 KASSERT(entry->cred == NULL || object->cred == NULL || 3849 (entry->eflags & MAP_ENTRY_NEEDS_COPY), 3850 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry)); 3851 offidxstart = OFF_TO_IDX(entry->offset); 3852 offidxend = offidxstart + atop(size); 3853 VM_OBJECT_WLOCK(object); 3854 if (object->ref_count != 1 && 3855 ((object->flags & OBJ_ONEMAPPING) != 0 || 3856 object == kernel_object)) { 3857 vm_object_collapse(object); 3858 3859 /* 3860 * The option OBJPR_NOTMAPPED can be passed here 3861 * because vm_map_delete() already performed 3862 * pmap_remove() on the only mapping to this range 3863 * of pages. 3864 */ 3865 vm_object_page_remove(object, offidxstart, offidxend, 3866 OBJPR_NOTMAPPED); 3867 if (offidxend >= object->size && 3868 offidxstart < object->size) { 3869 size1 = object->size; 3870 object->size = offidxstart; 3871 if (object->cred != NULL) { 3872 size1 -= object->size; 3873 KASSERT(object->charge >= ptoa(size1), 3874 ("object %p charge < 0", object)); 3875 swap_release_by_cred(ptoa(size1), 3876 object->cred); 3877 object->charge -= ptoa(size1); 3878 } 3879 } 3880 } 3881 VM_OBJECT_WUNLOCK(object); 3882 } 3883 if (map->system_map) 3884 vm_map_entry_deallocate(entry, TRUE); 3885 else { 3886 entry->defer_next = curthread->td_map_def_user; 3887 curthread->td_map_def_user = entry; 3888 } 3889 } 3890 3891 /* 3892 * vm_map_delete: [ internal use only ] 3893 * 3894 * Deallocates the given address range from the target 3895 * map. 3896 */ 3897 int 3898 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end) 3899 { 3900 vm_map_entry_t entry, next_entry, scratch_entry; 3901 int rv; 3902 3903 VM_MAP_ASSERT_LOCKED(map); 3904 3905 if (start == end) 3906 return (KERN_SUCCESS); 3907 3908 /* 3909 * Find the start of the region, and clip it. 3910 * Step through all entries in this region. 3911 */ 3912 rv = vm_map_lookup_clip_start(map, start, &entry, &scratch_entry); 3913 if (rv != KERN_SUCCESS) 3914 return (rv); 3915 for (; entry->start < end; entry = next_entry) { 3916 /* 3917 * Wait for wiring or unwiring of an entry to complete. 3918 * Also wait for any system wirings to disappear on 3919 * user maps. 3920 */ 3921 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 || 3922 (vm_map_pmap(map) != kernel_pmap && 3923 vm_map_entry_system_wired_count(entry) != 0)) { 3924 unsigned int last_timestamp; 3925 vm_offset_t saved_start; 3926 3927 saved_start = entry->start; 3928 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3929 last_timestamp = map->timestamp; 3930 (void) vm_map_unlock_and_wait(map, 0); 3931 vm_map_lock(map); 3932 if (last_timestamp + 1 != map->timestamp) { 3933 /* 3934 * Look again for the entry because the map was 3935 * modified while it was unlocked. 3936 * Specifically, the entry may have been 3937 * clipped, merged, or deleted. 3938 */ 3939 rv = vm_map_lookup_clip_start(map, saved_start, 3940 &next_entry, &scratch_entry); 3941 if (rv != KERN_SUCCESS) 3942 break; 3943 } else 3944 next_entry = entry; 3945 continue; 3946 } 3947 3948 /* XXXKIB or delete to the upper superpage boundary ? */ 3949 rv = vm_map_clip_end(map, entry, end); 3950 if (rv != KERN_SUCCESS) 3951 break; 3952 next_entry = vm_map_entry_succ(entry); 3953 3954 /* 3955 * Unwire before removing addresses from the pmap; otherwise, 3956 * unwiring will put the entries back in the pmap. 3957 */ 3958 if (entry->wired_count != 0) 3959 vm_map_entry_unwire(map, entry); 3960 3961 /* 3962 * Remove mappings for the pages, but only if the 3963 * mappings could exist. For instance, it does not 3964 * make sense to call pmap_remove() for guard entries. 3965 */ 3966 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 3967 entry->object.vm_object != NULL) 3968 pmap_remove(map->pmap, entry->start, entry->end); 3969 3970 if (entry->end == map->anon_loc) 3971 map->anon_loc = entry->start; 3972 3973 /* 3974 * Delete the entry only after removing all pmap 3975 * entries pointing to its pages. (Otherwise, its 3976 * page frames may be reallocated, and any modify bits 3977 * will be set in the wrong object!) 3978 */ 3979 vm_map_entry_delete(map, entry); 3980 } 3981 return (rv); 3982 } 3983 3984 /* 3985 * vm_map_remove: 3986 * 3987 * Remove the given address range from the target map. 3988 * This is the exported form of vm_map_delete. 3989 */ 3990 int 3991 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 3992 { 3993 int result; 3994 3995 vm_map_lock(map); 3996 VM_MAP_RANGE_CHECK(map, start, end); 3997 result = vm_map_delete(map, start, end); 3998 vm_map_unlock(map); 3999 return (result); 4000 } 4001 4002 /* 4003 * vm_map_check_protection: 4004 * 4005 * Assert that the target map allows the specified privilege on the 4006 * entire address region given. The entire region must be allocated. 4007 * 4008 * WARNING! This code does not and should not check whether the 4009 * contents of the region is accessible. For example a smaller file 4010 * might be mapped into a larger address space. 4011 * 4012 * NOTE! This code is also called by munmap(). 4013 * 4014 * The map must be locked. A read lock is sufficient. 4015 */ 4016 boolean_t 4017 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 4018 vm_prot_t protection) 4019 { 4020 vm_map_entry_t entry; 4021 vm_map_entry_t tmp_entry; 4022 4023 if (!vm_map_lookup_entry(map, start, &tmp_entry)) 4024 return (FALSE); 4025 entry = tmp_entry; 4026 4027 while (start < end) { 4028 /* 4029 * No holes allowed! 4030 */ 4031 if (start < entry->start) 4032 return (FALSE); 4033 /* 4034 * Check protection associated with entry. 4035 */ 4036 if ((entry->protection & protection) != protection) 4037 return (FALSE); 4038 /* go to next entry */ 4039 start = entry->end; 4040 entry = vm_map_entry_succ(entry); 4041 } 4042 return (TRUE); 4043 } 4044 4045 /* 4046 * 4047 * vm_map_copy_swap_object: 4048 * 4049 * Copies a swap-backed object from an existing map entry to a 4050 * new one. Carries forward the swap charge. May change the 4051 * src object on return. 4052 */ 4053 static void 4054 vm_map_copy_swap_object(vm_map_entry_t src_entry, vm_map_entry_t dst_entry, 4055 vm_offset_t size, vm_ooffset_t *fork_charge) 4056 { 4057 vm_object_t src_object; 4058 struct ucred *cred; 4059 int charged; 4060 4061 src_object = src_entry->object.vm_object; 4062 charged = ENTRY_CHARGED(src_entry); 4063 if ((src_object->flags & OBJ_ANON) != 0) { 4064 VM_OBJECT_WLOCK(src_object); 4065 vm_object_collapse(src_object); 4066 if ((src_object->flags & OBJ_ONEMAPPING) != 0) { 4067 vm_object_split(src_entry); 4068 src_object = src_entry->object.vm_object; 4069 } 4070 vm_object_reference_locked(src_object); 4071 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 4072 VM_OBJECT_WUNLOCK(src_object); 4073 } else 4074 vm_object_reference(src_object); 4075 if (src_entry->cred != NULL && 4076 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 4077 KASSERT(src_object->cred == NULL, 4078 ("OVERCOMMIT: vm_map_copy_anon_entry: cred %p", 4079 src_object)); 4080 src_object->cred = src_entry->cred; 4081 src_object->charge = size; 4082 } 4083 dst_entry->object.vm_object = src_object; 4084 if (charged) { 4085 cred = curthread->td_ucred; 4086 crhold(cred); 4087 dst_entry->cred = cred; 4088 *fork_charge += size; 4089 if (!(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 4090 crhold(cred); 4091 src_entry->cred = cred; 4092 *fork_charge += size; 4093 } 4094 } 4095 } 4096 4097 /* 4098 * vm_map_copy_entry: 4099 * 4100 * Copies the contents of the source entry to the destination 4101 * entry. The entries *must* be aligned properly. 4102 */ 4103 static void 4104 vm_map_copy_entry( 4105 vm_map_t src_map, 4106 vm_map_t dst_map, 4107 vm_map_entry_t src_entry, 4108 vm_map_entry_t dst_entry, 4109 vm_ooffset_t *fork_charge) 4110 { 4111 vm_object_t src_object; 4112 vm_map_entry_t fake_entry; 4113 vm_offset_t size; 4114 4115 VM_MAP_ASSERT_LOCKED(dst_map); 4116 4117 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) 4118 return; 4119 4120 if (src_entry->wired_count == 0 || 4121 (src_entry->protection & VM_PROT_WRITE) == 0) { 4122 /* 4123 * If the source entry is marked needs_copy, it is already 4124 * write-protected. 4125 */ 4126 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && 4127 (src_entry->protection & VM_PROT_WRITE) != 0) { 4128 pmap_protect(src_map->pmap, 4129 src_entry->start, 4130 src_entry->end, 4131 src_entry->protection & ~VM_PROT_WRITE); 4132 } 4133 4134 /* 4135 * Make a copy of the object. 4136 */ 4137 size = src_entry->end - src_entry->start; 4138 if ((src_object = src_entry->object.vm_object) != NULL) { 4139 /* 4140 * Swap-backed objects need special handling. Note that 4141 * this is an unlocked check, so it is possible to race 4142 * with an OBJT_DEFAULT -> OBJT_SWAP conversion. 4143 */ 4144 if (src_object->type == OBJT_DEFAULT || 4145 src_object->type == OBJT_SWAP || 4146 (src_object->flags & OBJ_SWAP) != 0) { 4147 vm_map_copy_swap_object(src_entry, dst_entry, 4148 size, fork_charge); 4149 /* May have split/collapsed, reload obj. */ 4150 src_object = src_entry->object.vm_object; 4151 } else { 4152 vm_object_reference(src_object); 4153 dst_entry->object.vm_object = src_object; 4154 } 4155 src_entry->eflags |= MAP_ENTRY_COW | 4156 MAP_ENTRY_NEEDS_COPY; 4157 dst_entry->eflags |= MAP_ENTRY_COW | 4158 MAP_ENTRY_NEEDS_COPY; 4159 dst_entry->offset = src_entry->offset; 4160 if (src_entry->eflags & MAP_ENTRY_WRITECNT) { 4161 /* 4162 * MAP_ENTRY_WRITECNT cannot 4163 * indicate write reference from 4164 * src_entry, since the entry is 4165 * marked as needs copy. Allocate a 4166 * fake entry that is used to 4167 * decrement object->un_pager writecount 4168 * at the appropriate time. Attach 4169 * fake_entry to the deferred list. 4170 */ 4171 fake_entry = vm_map_entry_create(dst_map); 4172 fake_entry->eflags = MAP_ENTRY_WRITECNT; 4173 src_entry->eflags &= ~MAP_ENTRY_WRITECNT; 4174 vm_object_reference(src_object); 4175 fake_entry->object.vm_object = src_object; 4176 fake_entry->start = src_entry->start; 4177 fake_entry->end = src_entry->end; 4178 fake_entry->defer_next = 4179 curthread->td_map_def_user; 4180 curthread->td_map_def_user = fake_entry; 4181 } 4182 4183 pmap_copy(dst_map->pmap, src_map->pmap, 4184 dst_entry->start, dst_entry->end - dst_entry->start, 4185 src_entry->start); 4186 } else { 4187 dst_entry->object.vm_object = NULL; 4188 dst_entry->offset = 0; 4189 if (src_entry->cred != NULL) { 4190 dst_entry->cred = curthread->td_ucred; 4191 crhold(dst_entry->cred); 4192 *fork_charge += size; 4193 } 4194 } 4195 } else { 4196 /* 4197 * We don't want to make writeable wired pages copy-on-write. 4198 * Immediately copy these pages into the new map by simulating 4199 * page faults. The new pages are pageable. 4200 */ 4201 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry, 4202 fork_charge); 4203 } 4204 } 4205 4206 /* 4207 * vmspace_map_entry_forked: 4208 * Update the newly-forked vmspace each time a map entry is inherited 4209 * or copied. The values for vm_dsize and vm_tsize are approximate 4210 * (and mostly-obsolete ideas in the face of mmap(2) et al.) 4211 */ 4212 static void 4213 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2, 4214 vm_map_entry_t entry) 4215 { 4216 vm_size_t entrysize; 4217 vm_offset_t newend; 4218 4219 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 4220 return; 4221 entrysize = entry->end - entry->start; 4222 vm2->vm_map.size += entrysize; 4223 if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) { 4224 vm2->vm_ssize += btoc(entrysize); 4225 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr && 4226 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) { 4227 newend = MIN(entry->end, 4228 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)); 4229 vm2->vm_dsize += btoc(newend - entry->start); 4230 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr && 4231 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) { 4232 newend = MIN(entry->end, 4233 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)); 4234 vm2->vm_tsize += btoc(newend - entry->start); 4235 } 4236 } 4237 4238 /* 4239 * vmspace_fork: 4240 * Create a new process vmspace structure and vm_map 4241 * based on those of an existing process. The new map 4242 * is based on the old map, according to the inheritance 4243 * values on the regions in that map. 4244 * 4245 * XXX It might be worth coalescing the entries added to the new vmspace. 4246 * 4247 * The source map must not be locked. 4248 */ 4249 struct vmspace * 4250 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge) 4251 { 4252 struct vmspace *vm2; 4253 vm_map_t new_map, old_map; 4254 vm_map_entry_t new_entry, old_entry; 4255 vm_object_t object; 4256 int error, locked __diagused; 4257 vm_inherit_t inh; 4258 4259 old_map = &vm1->vm_map; 4260 /* Copy immutable fields of vm1 to vm2. */ 4261 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map), 4262 pmap_pinit); 4263 if (vm2 == NULL) 4264 return (NULL); 4265 4266 vm2->vm_taddr = vm1->vm_taddr; 4267 vm2->vm_daddr = vm1->vm_daddr; 4268 vm2->vm_maxsaddr = vm1->vm_maxsaddr; 4269 vm2->vm_stacktop = vm1->vm_stacktop; 4270 vm_map_lock(old_map); 4271 if (old_map->busy) 4272 vm_map_wait_busy(old_map); 4273 new_map = &vm2->vm_map; 4274 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */ 4275 KASSERT(locked, ("vmspace_fork: lock failed")); 4276 4277 error = pmap_vmspace_copy(new_map->pmap, old_map->pmap); 4278 if (error != 0) { 4279 sx_xunlock(&old_map->lock); 4280 sx_xunlock(&new_map->lock); 4281 vm_map_process_deferred(); 4282 vmspace_free(vm2); 4283 return (NULL); 4284 } 4285 4286 new_map->anon_loc = old_map->anon_loc; 4287 new_map->flags |= old_map->flags & (MAP_ASLR | MAP_ASLR_IGNSTART | 4288 MAP_ASLR_STACK | MAP_WXORX); 4289 4290 VM_MAP_ENTRY_FOREACH(old_entry, old_map) { 4291 if ((old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 4292 panic("vm_map_fork: encountered a submap"); 4293 4294 inh = old_entry->inheritance; 4295 if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 && 4296 inh != VM_INHERIT_NONE) 4297 inh = VM_INHERIT_COPY; 4298 4299 switch (inh) { 4300 case VM_INHERIT_NONE: 4301 break; 4302 4303 case VM_INHERIT_SHARE: 4304 /* 4305 * Clone the entry, creating the shared object if 4306 * necessary. 4307 */ 4308 object = old_entry->object.vm_object; 4309 if (object == NULL) { 4310 vm_map_entry_back(old_entry); 4311 object = old_entry->object.vm_object; 4312 } 4313 4314 /* 4315 * Add the reference before calling vm_object_shadow 4316 * to insure that a shadow object is created. 4317 */ 4318 vm_object_reference(object); 4319 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4320 vm_object_shadow(&old_entry->object.vm_object, 4321 &old_entry->offset, 4322 old_entry->end - old_entry->start, 4323 old_entry->cred, 4324 /* Transfer the second reference too. */ 4325 true); 4326 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 4327 old_entry->cred = NULL; 4328 4329 /* 4330 * As in vm_map_merged_neighbor_dispose(), 4331 * the vnode lock will not be acquired in 4332 * this call to vm_object_deallocate(). 4333 */ 4334 vm_object_deallocate(object); 4335 object = old_entry->object.vm_object; 4336 } else { 4337 VM_OBJECT_WLOCK(object); 4338 vm_object_clear_flag(object, OBJ_ONEMAPPING); 4339 if (old_entry->cred != NULL) { 4340 KASSERT(object->cred == NULL, 4341 ("vmspace_fork both cred")); 4342 object->cred = old_entry->cred; 4343 object->charge = old_entry->end - 4344 old_entry->start; 4345 old_entry->cred = NULL; 4346 } 4347 4348 /* 4349 * Assert the correct state of the vnode 4350 * v_writecount while the object is locked, to 4351 * not relock it later for the assertion 4352 * correctness. 4353 */ 4354 if (old_entry->eflags & MAP_ENTRY_WRITECNT && 4355 object->type == OBJT_VNODE) { 4356 KASSERT(((struct vnode *)object-> 4357 handle)->v_writecount > 0, 4358 ("vmspace_fork: v_writecount %p", 4359 object)); 4360 KASSERT(object->un_pager.vnp. 4361 writemappings > 0, 4362 ("vmspace_fork: vnp.writecount %p", 4363 object)); 4364 } 4365 VM_OBJECT_WUNLOCK(object); 4366 } 4367 4368 /* 4369 * Clone the entry, referencing the shared object. 4370 */ 4371 new_entry = vm_map_entry_create(new_map); 4372 *new_entry = *old_entry; 4373 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4374 MAP_ENTRY_IN_TRANSITION); 4375 new_entry->wiring_thread = NULL; 4376 new_entry->wired_count = 0; 4377 if (new_entry->eflags & MAP_ENTRY_WRITECNT) { 4378 vm_pager_update_writecount(object, 4379 new_entry->start, new_entry->end); 4380 } 4381 vm_map_entry_set_vnode_text(new_entry, true); 4382 4383 /* 4384 * Insert the entry into the new map -- we know we're 4385 * inserting at the end of the new map. 4386 */ 4387 vm_map_entry_link(new_map, new_entry); 4388 vmspace_map_entry_forked(vm1, vm2, new_entry); 4389 4390 /* 4391 * Update the physical map 4392 */ 4393 pmap_copy(new_map->pmap, old_map->pmap, 4394 new_entry->start, 4395 (old_entry->end - old_entry->start), 4396 old_entry->start); 4397 break; 4398 4399 case VM_INHERIT_COPY: 4400 /* 4401 * Clone the entry and link into the map. 4402 */ 4403 new_entry = vm_map_entry_create(new_map); 4404 *new_entry = *old_entry; 4405 /* 4406 * Copied entry is COW over the old object. 4407 */ 4408 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4409 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT); 4410 new_entry->wiring_thread = NULL; 4411 new_entry->wired_count = 0; 4412 new_entry->object.vm_object = NULL; 4413 new_entry->cred = NULL; 4414 vm_map_entry_link(new_map, new_entry); 4415 vmspace_map_entry_forked(vm1, vm2, new_entry); 4416 vm_map_copy_entry(old_map, new_map, old_entry, 4417 new_entry, fork_charge); 4418 vm_map_entry_set_vnode_text(new_entry, true); 4419 break; 4420 4421 case VM_INHERIT_ZERO: 4422 /* 4423 * Create a new anonymous mapping entry modelled from 4424 * the old one. 4425 */ 4426 new_entry = vm_map_entry_create(new_map); 4427 memset(new_entry, 0, sizeof(*new_entry)); 4428 4429 new_entry->start = old_entry->start; 4430 new_entry->end = old_entry->end; 4431 new_entry->eflags = old_entry->eflags & 4432 ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION | 4433 MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC | 4434 MAP_ENTRY_SPLIT_BOUNDARY_MASK); 4435 new_entry->protection = old_entry->protection; 4436 new_entry->max_protection = old_entry->max_protection; 4437 new_entry->inheritance = VM_INHERIT_ZERO; 4438 4439 vm_map_entry_link(new_map, new_entry); 4440 vmspace_map_entry_forked(vm1, vm2, new_entry); 4441 4442 new_entry->cred = curthread->td_ucred; 4443 crhold(new_entry->cred); 4444 *fork_charge += (new_entry->end - new_entry->start); 4445 4446 break; 4447 } 4448 } 4449 /* 4450 * Use inlined vm_map_unlock() to postpone handling the deferred 4451 * map entries, which cannot be done until both old_map and 4452 * new_map locks are released. 4453 */ 4454 sx_xunlock(&old_map->lock); 4455 sx_xunlock(&new_map->lock); 4456 vm_map_process_deferred(); 4457 4458 return (vm2); 4459 } 4460 4461 /* 4462 * Create a process's stack for exec_new_vmspace(). This function is never 4463 * asked to wire the newly created stack. 4464 */ 4465 int 4466 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4467 vm_prot_t prot, vm_prot_t max, int cow) 4468 { 4469 vm_size_t growsize, init_ssize; 4470 rlim_t vmemlim; 4471 int rv; 4472 4473 MPASS((map->flags & MAP_WIREFUTURE) == 0); 4474 growsize = sgrowsiz; 4475 init_ssize = (max_ssize < growsize) ? max_ssize : growsize; 4476 vm_map_lock(map); 4477 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4478 /* If we would blow our VMEM resource limit, no go */ 4479 if (map->size + init_ssize > vmemlim) { 4480 rv = KERN_NO_SPACE; 4481 goto out; 4482 } 4483 rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot, 4484 max, cow); 4485 out: 4486 vm_map_unlock(map); 4487 return (rv); 4488 } 4489 4490 static int stack_guard_page = 1; 4491 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN, 4492 &stack_guard_page, 0, 4493 "Specifies the number of guard pages for a stack that grows"); 4494 4495 static int 4496 vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4497 vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow) 4498 { 4499 vm_map_entry_t new_entry, prev_entry; 4500 vm_offset_t bot, gap_bot, gap_top, top; 4501 vm_size_t init_ssize, sgp; 4502 int orient, rv; 4503 4504 /* 4505 * The stack orientation is piggybacked with the cow argument. 4506 * Extract it into orient and mask the cow argument so that we 4507 * don't pass it around further. 4508 */ 4509 orient = cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP); 4510 KASSERT(orient != 0, ("No stack grow direction")); 4511 KASSERT(orient != (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP), 4512 ("bi-dir stack")); 4513 4514 if (max_ssize == 0 || 4515 !vm_map_range_valid(map, addrbos, addrbos + max_ssize)) 4516 return (KERN_INVALID_ADDRESS); 4517 sgp = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4518 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4519 (vm_size_t)stack_guard_page * PAGE_SIZE; 4520 if (sgp >= max_ssize) 4521 return (KERN_INVALID_ARGUMENT); 4522 4523 init_ssize = growsize; 4524 if (max_ssize < init_ssize + sgp) 4525 init_ssize = max_ssize - sgp; 4526 4527 /* If addr is already mapped, no go */ 4528 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) 4529 return (KERN_NO_SPACE); 4530 4531 /* 4532 * If we can't accommodate max_ssize in the current mapping, no go. 4533 */ 4534 if (vm_map_entry_succ(prev_entry)->start < addrbos + max_ssize) 4535 return (KERN_NO_SPACE); 4536 4537 /* 4538 * We initially map a stack of only init_ssize. We will grow as 4539 * needed later. Depending on the orientation of the stack (i.e. 4540 * the grow direction) we either map at the top of the range, the 4541 * bottom of the range or in the middle. 4542 * 4543 * Note: we would normally expect prot and max to be VM_PROT_ALL, 4544 * and cow to be 0. Possibly we should eliminate these as input 4545 * parameters, and just pass these values here in the insert call. 4546 */ 4547 if (orient == MAP_STACK_GROWS_DOWN) { 4548 bot = addrbos + max_ssize - init_ssize; 4549 top = bot + init_ssize; 4550 gap_bot = addrbos; 4551 gap_top = bot; 4552 } else /* if (orient == MAP_STACK_GROWS_UP) */ { 4553 bot = addrbos; 4554 top = bot + init_ssize; 4555 gap_bot = top; 4556 gap_top = addrbos + max_ssize; 4557 } 4558 rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow); 4559 if (rv != KERN_SUCCESS) 4560 return (rv); 4561 new_entry = vm_map_entry_succ(prev_entry); 4562 KASSERT(new_entry->end == top || new_entry->start == bot, 4563 ("Bad entry start/end for new stack entry")); 4564 KASSERT((orient & MAP_STACK_GROWS_DOWN) == 0 || 4565 (new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0, 4566 ("new entry lacks MAP_ENTRY_GROWS_DOWN")); 4567 KASSERT((orient & MAP_STACK_GROWS_UP) == 0 || 4568 (new_entry->eflags & MAP_ENTRY_GROWS_UP) != 0, 4569 ("new entry lacks MAP_ENTRY_GROWS_UP")); 4570 if (gap_bot == gap_top) 4571 return (KERN_SUCCESS); 4572 rv = vm_map_insert(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE, 4573 VM_PROT_NONE, MAP_CREATE_GUARD | (orient == MAP_STACK_GROWS_DOWN ? 4574 MAP_CREATE_STACK_GAP_DN : MAP_CREATE_STACK_GAP_UP)); 4575 if (rv == KERN_SUCCESS) { 4576 /* 4577 * Gap can never successfully handle a fault, so 4578 * read-ahead logic is never used for it. Re-use 4579 * next_read of the gap entry to store 4580 * stack_guard_page for vm_map_growstack(). 4581 */ 4582 if (orient == MAP_STACK_GROWS_DOWN) 4583 vm_map_entry_pred(new_entry)->next_read = sgp; 4584 else 4585 vm_map_entry_succ(new_entry)->next_read = sgp; 4586 } else { 4587 (void)vm_map_delete(map, bot, top); 4588 } 4589 return (rv); 4590 } 4591 4592 /* 4593 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we 4594 * successfully grow the stack. 4595 */ 4596 static int 4597 vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry) 4598 { 4599 vm_map_entry_t stack_entry; 4600 struct proc *p; 4601 struct vmspace *vm; 4602 struct ucred *cred; 4603 vm_offset_t gap_end, gap_start, grow_start; 4604 vm_size_t grow_amount, guard, max_grow; 4605 rlim_t lmemlim, stacklim, vmemlim; 4606 int rv, rv1 __diagused; 4607 bool gap_deleted, grow_down, is_procstack; 4608 #ifdef notyet 4609 uint64_t limit; 4610 #endif 4611 #ifdef RACCT 4612 int error __diagused; 4613 #endif 4614 4615 p = curproc; 4616 vm = p->p_vmspace; 4617 4618 /* 4619 * Disallow stack growth when the access is performed by a 4620 * debugger or AIO daemon. The reason is that the wrong 4621 * resource limits are applied. 4622 */ 4623 if (p != initproc && (map != &p->p_vmspace->vm_map || 4624 p->p_textvp == NULL)) 4625 return (KERN_FAILURE); 4626 4627 MPASS(!map->system_map); 4628 4629 lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK); 4630 stacklim = lim_cur(curthread, RLIMIT_STACK); 4631 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4632 retry: 4633 /* If addr is not in a hole for a stack grow area, no need to grow. */ 4634 if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry)) 4635 return (KERN_FAILURE); 4636 if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0) 4637 return (KERN_SUCCESS); 4638 if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_DN) != 0) { 4639 stack_entry = vm_map_entry_succ(gap_entry); 4640 if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 || 4641 stack_entry->start != gap_entry->end) 4642 return (KERN_FAILURE); 4643 grow_amount = round_page(stack_entry->start - addr); 4644 grow_down = true; 4645 } else if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_UP) != 0) { 4646 stack_entry = vm_map_entry_pred(gap_entry); 4647 if ((stack_entry->eflags & MAP_ENTRY_GROWS_UP) == 0 || 4648 stack_entry->end != gap_entry->start) 4649 return (KERN_FAILURE); 4650 grow_amount = round_page(addr + 1 - stack_entry->end); 4651 grow_down = false; 4652 } else { 4653 return (KERN_FAILURE); 4654 } 4655 guard = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4656 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4657 gap_entry->next_read; 4658 max_grow = gap_entry->end - gap_entry->start; 4659 if (guard > max_grow) 4660 return (KERN_NO_SPACE); 4661 max_grow -= guard; 4662 if (grow_amount > max_grow) 4663 return (KERN_NO_SPACE); 4664 4665 /* 4666 * If this is the main process stack, see if we're over the stack 4667 * limit. 4668 */ 4669 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr && 4670 addr < (vm_offset_t)vm->vm_stacktop; 4671 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) 4672 return (KERN_NO_SPACE); 4673 4674 #ifdef RACCT 4675 if (racct_enable) { 4676 PROC_LOCK(p); 4677 if (is_procstack && racct_set(p, RACCT_STACK, 4678 ctob(vm->vm_ssize) + grow_amount)) { 4679 PROC_UNLOCK(p); 4680 return (KERN_NO_SPACE); 4681 } 4682 PROC_UNLOCK(p); 4683 } 4684 #endif 4685 4686 grow_amount = roundup(grow_amount, sgrowsiz); 4687 if (grow_amount > max_grow) 4688 grow_amount = max_grow; 4689 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) { 4690 grow_amount = trunc_page((vm_size_t)stacklim) - 4691 ctob(vm->vm_ssize); 4692 } 4693 4694 #ifdef notyet 4695 PROC_LOCK(p); 4696 limit = racct_get_available(p, RACCT_STACK); 4697 PROC_UNLOCK(p); 4698 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit)) 4699 grow_amount = limit - ctob(vm->vm_ssize); 4700 #endif 4701 4702 if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) { 4703 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) { 4704 rv = KERN_NO_SPACE; 4705 goto out; 4706 } 4707 #ifdef RACCT 4708 if (racct_enable) { 4709 PROC_LOCK(p); 4710 if (racct_set(p, RACCT_MEMLOCK, 4711 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) { 4712 PROC_UNLOCK(p); 4713 rv = KERN_NO_SPACE; 4714 goto out; 4715 } 4716 PROC_UNLOCK(p); 4717 } 4718 #endif 4719 } 4720 4721 /* If we would blow our VMEM resource limit, no go */ 4722 if (map->size + grow_amount > vmemlim) { 4723 rv = KERN_NO_SPACE; 4724 goto out; 4725 } 4726 #ifdef RACCT 4727 if (racct_enable) { 4728 PROC_LOCK(p); 4729 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) { 4730 PROC_UNLOCK(p); 4731 rv = KERN_NO_SPACE; 4732 goto out; 4733 } 4734 PROC_UNLOCK(p); 4735 } 4736 #endif 4737 4738 if (vm_map_lock_upgrade(map)) { 4739 gap_entry = NULL; 4740 vm_map_lock_read(map); 4741 goto retry; 4742 } 4743 4744 if (grow_down) { 4745 grow_start = gap_entry->end - grow_amount; 4746 if (gap_entry->start + grow_amount == gap_entry->end) { 4747 gap_start = gap_entry->start; 4748 gap_end = gap_entry->end; 4749 vm_map_entry_delete(map, gap_entry); 4750 gap_deleted = true; 4751 } else { 4752 MPASS(gap_entry->start < gap_entry->end - grow_amount); 4753 vm_map_entry_resize(map, gap_entry, -grow_amount); 4754 gap_deleted = false; 4755 } 4756 rv = vm_map_insert(map, NULL, 0, grow_start, 4757 grow_start + grow_amount, 4758 stack_entry->protection, stack_entry->max_protection, 4759 MAP_STACK_GROWS_DOWN); 4760 if (rv != KERN_SUCCESS) { 4761 if (gap_deleted) { 4762 rv1 = vm_map_insert(map, NULL, 0, gap_start, 4763 gap_end, VM_PROT_NONE, VM_PROT_NONE, 4764 MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP_DN); 4765 MPASS(rv1 == KERN_SUCCESS); 4766 } else 4767 vm_map_entry_resize(map, gap_entry, 4768 grow_amount); 4769 } 4770 } else { 4771 grow_start = stack_entry->end; 4772 cred = stack_entry->cred; 4773 if (cred == NULL && stack_entry->object.vm_object != NULL) 4774 cred = stack_entry->object.vm_object->cred; 4775 if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred)) 4776 rv = KERN_NO_SPACE; 4777 /* Grow the underlying object if applicable. */ 4778 else if (stack_entry->object.vm_object == NULL || 4779 vm_object_coalesce(stack_entry->object.vm_object, 4780 stack_entry->offset, 4781 (vm_size_t)(stack_entry->end - stack_entry->start), 4782 grow_amount, cred != NULL)) { 4783 if (gap_entry->start + grow_amount == gap_entry->end) { 4784 vm_map_entry_delete(map, gap_entry); 4785 vm_map_entry_resize(map, stack_entry, 4786 grow_amount); 4787 } else { 4788 gap_entry->start += grow_amount; 4789 stack_entry->end += grow_amount; 4790 } 4791 map->size += grow_amount; 4792 rv = KERN_SUCCESS; 4793 } else 4794 rv = KERN_FAILURE; 4795 } 4796 if (rv == KERN_SUCCESS && is_procstack) 4797 vm->vm_ssize += btoc(grow_amount); 4798 4799 /* 4800 * Heed the MAP_WIREFUTURE flag if it was set for this process. 4801 */ 4802 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) { 4803 rv = vm_map_wire_locked(map, grow_start, 4804 grow_start + grow_amount, 4805 VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); 4806 } 4807 vm_map_lock_downgrade(map); 4808 4809 out: 4810 #ifdef RACCT 4811 if (racct_enable && rv != KERN_SUCCESS) { 4812 PROC_LOCK(p); 4813 error = racct_set(p, RACCT_VMEM, map->size); 4814 KASSERT(error == 0, ("decreasing RACCT_VMEM failed")); 4815 if (!old_mlock) { 4816 error = racct_set(p, RACCT_MEMLOCK, 4817 ptoa(pmap_wired_count(map->pmap))); 4818 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed")); 4819 } 4820 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize)); 4821 KASSERT(error == 0, ("decreasing RACCT_STACK failed")); 4822 PROC_UNLOCK(p); 4823 } 4824 #endif 4825 4826 return (rv); 4827 } 4828 4829 /* 4830 * Unshare the specified VM space for exec. If other processes are 4831 * mapped to it, then create a new one. The new vmspace is null. 4832 */ 4833 int 4834 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser) 4835 { 4836 struct vmspace *oldvmspace = p->p_vmspace; 4837 struct vmspace *newvmspace; 4838 4839 KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0, 4840 ("vmspace_exec recursed")); 4841 newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit); 4842 if (newvmspace == NULL) 4843 return (ENOMEM); 4844 newvmspace->vm_swrss = oldvmspace->vm_swrss; 4845 /* 4846 * This code is written like this for prototype purposes. The 4847 * goal is to avoid running down the vmspace here, but let the 4848 * other process's that are still using the vmspace to finally 4849 * run it down. Even though there is little or no chance of blocking 4850 * here, it is a good idea to keep this form for future mods. 4851 */ 4852 PROC_VMSPACE_LOCK(p); 4853 p->p_vmspace = newvmspace; 4854 PROC_VMSPACE_UNLOCK(p); 4855 if (p == curthread->td_proc) 4856 pmap_activate(curthread); 4857 curthread->td_pflags |= TDP_EXECVMSPC; 4858 return (0); 4859 } 4860 4861 /* 4862 * Unshare the specified VM space for forcing COW. This 4863 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 4864 */ 4865 int 4866 vmspace_unshare(struct proc *p) 4867 { 4868 struct vmspace *oldvmspace = p->p_vmspace; 4869 struct vmspace *newvmspace; 4870 vm_ooffset_t fork_charge; 4871 4872 /* 4873 * The caller is responsible for ensuring that the reference count 4874 * cannot concurrently transition 1 -> 2. 4875 */ 4876 if (refcount_load(&oldvmspace->vm_refcnt) == 1) 4877 return (0); 4878 fork_charge = 0; 4879 newvmspace = vmspace_fork(oldvmspace, &fork_charge); 4880 if (newvmspace == NULL) 4881 return (ENOMEM); 4882 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) { 4883 vmspace_free(newvmspace); 4884 return (ENOMEM); 4885 } 4886 PROC_VMSPACE_LOCK(p); 4887 p->p_vmspace = newvmspace; 4888 PROC_VMSPACE_UNLOCK(p); 4889 if (p == curthread->td_proc) 4890 pmap_activate(curthread); 4891 vmspace_free(oldvmspace); 4892 return (0); 4893 } 4894 4895 /* 4896 * vm_map_lookup: 4897 * 4898 * Finds the VM object, offset, and 4899 * protection for a given virtual address in the 4900 * specified map, assuming a page fault of the 4901 * type specified. 4902 * 4903 * Leaves the map in question locked for read; return 4904 * values are guaranteed until a vm_map_lookup_done 4905 * call is performed. Note that the map argument 4906 * is in/out; the returned map must be used in 4907 * the call to vm_map_lookup_done. 4908 * 4909 * A handle (out_entry) is returned for use in 4910 * vm_map_lookup_done, to make that fast. 4911 * 4912 * If a lookup is requested with "write protection" 4913 * specified, the map may be changed to perform virtual 4914 * copying operations, although the data referenced will 4915 * remain the same. 4916 */ 4917 int 4918 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4919 vm_offset_t vaddr, 4920 vm_prot_t fault_typea, 4921 vm_map_entry_t *out_entry, /* OUT */ 4922 vm_object_t *object, /* OUT */ 4923 vm_pindex_t *pindex, /* OUT */ 4924 vm_prot_t *out_prot, /* OUT */ 4925 boolean_t *wired) /* OUT */ 4926 { 4927 vm_map_entry_t entry; 4928 vm_map_t map = *var_map; 4929 vm_prot_t prot; 4930 vm_prot_t fault_type; 4931 vm_object_t eobject; 4932 vm_size_t size; 4933 struct ucred *cred; 4934 4935 RetryLookup: 4936 4937 vm_map_lock_read(map); 4938 4939 RetryLookupLocked: 4940 /* 4941 * Lookup the faulting address. 4942 */ 4943 if (!vm_map_lookup_entry(map, vaddr, out_entry)) { 4944 vm_map_unlock_read(map); 4945 return (KERN_INVALID_ADDRESS); 4946 } 4947 4948 entry = *out_entry; 4949 4950 /* 4951 * Handle submaps. 4952 */ 4953 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 4954 vm_map_t old_map = map; 4955 4956 *var_map = map = entry->object.sub_map; 4957 vm_map_unlock_read(old_map); 4958 goto RetryLookup; 4959 } 4960 4961 /* 4962 * Check whether this task is allowed to have this page. 4963 */ 4964 prot = entry->protection; 4965 if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) { 4966 fault_typea &= ~VM_PROT_FAULT_LOOKUP; 4967 if (prot == VM_PROT_NONE && map != kernel_map && 4968 (entry->eflags & MAP_ENTRY_GUARD) != 0 && 4969 (entry->eflags & (MAP_ENTRY_STACK_GAP_DN | 4970 MAP_ENTRY_STACK_GAP_UP)) != 0 && 4971 vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS) 4972 goto RetryLookupLocked; 4973 } 4974 fault_type = fault_typea & VM_PROT_ALL; 4975 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) { 4976 vm_map_unlock_read(map); 4977 return (KERN_PROTECTION_FAILURE); 4978 } 4979 KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags & 4980 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) != 4981 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY), 4982 ("entry %p flags %x", entry, entry->eflags)); 4983 if ((fault_typea & VM_PROT_COPY) != 0 && 4984 (entry->max_protection & VM_PROT_WRITE) == 0 && 4985 (entry->eflags & MAP_ENTRY_COW) == 0) { 4986 vm_map_unlock_read(map); 4987 return (KERN_PROTECTION_FAILURE); 4988 } 4989 4990 /* 4991 * If this page is not pageable, we have to get it for all possible 4992 * accesses. 4993 */ 4994 *wired = (entry->wired_count != 0); 4995 if (*wired) 4996 fault_type = entry->protection; 4997 size = entry->end - entry->start; 4998 4999 /* 5000 * If the entry was copy-on-write, we either ... 5001 */ 5002 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 5003 /* 5004 * If we want to write the page, we may as well handle that 5005 * now since we've got the map locked. 5006 * 5007 * If we don't need to write the page, we just demote the 5008 * permissions allowed. 5009 */ 5010 if ((fault_type & VM_PROT_WRITE) != 0 || 5011 (fault_typea & VM_PROT_COPY) != 0) { 5012 /* 5013 * Make a new object, and place it in the object 5014 * chain. Note that no new references have appeared 5015 * -- one just moved from the map to the new 5016 * object. 5017 */ 5018 if (vm_map_lock_upgrade(map)) 5019 goto RetryLookup; 5020 5021 if (entry->cred == NULL) { 5022 /* 5023 * The debugger owner is charged for 5024 * the memory. 5025 */ 5026 cred = curthread->td_ucred; 5027 crhold(cred); 5028 if (!swap_reserve_by_cred(size, cred)) { 5029 crfree(cred); 5030 vm_map_unlock(map); 5031 return (KERN_RESOURCE_SHORTAGE); 5032 } 5033 entry->cred = cred; 5034 } 5035 eobject = entry->object.vm_object; 5036 vm_object_shadow(&entry->object.vm_object, 5037 &entry->offset, size, entry->cred, false); 5038 if (eobject == entry->object.vm_object) { 5039 /* 5040 * The object was not shadowed. 5041 */ 5042 swap_release_by_cred(size, entry->cred); 5043 crfree(entry->cred); 5044 } 5045 entry->cred = NULL; 5046 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 5047 5048 vm_map_lock_downgrade(map); 5049 } else { 5050 /* 5051 * We're attempting to read a copy-on-write page -- 5052 * don't allow writes. 5053 */ 5054 prot &= ~VM_PROT_WRITE; 5055 } 5056 } 5057 5058 /* 5059 * Create an object if necessary. 5060 */ 5061 if (entry->object.vm_object == NULL && !map->system_map) { 5062 if (vm_map_lock_upgrade(map)) 5063 goto RetryLookup; 5064 entry->object.vm_object = vm_object_allocate_anon(atop(size), 5065 NULL, entry->cred, entry->cred != NULL ? size : 0); 5066 entry->offset = 0; 5067 entry->cred = NULL; 5068 vm_map_lock_downgrade(map); 5069 } 5070 5071 /* 5072 * Return the object/offset from this entry. If the entry was 5073 * copy-on-write or empty, it has been fixed up. 5074 */ 5075 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 5076 *object = entry->object.vm_object; 5077 5078 *out_prot = prot; 5079 return (KERN_SUCCESS); 5080 } 5081 5082 /* 5083 * vm_map_lookup_locked: 5084 * 5085 * Lookup the faulting address. A version of vm_map_lookup that returns 5086 * KERN_FAILURE instead of blocking on map lock or memory allocation. 5087 */ 5088 int 5089 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */ 5090 vm_offset_t vaddr, 5091 vm_prot_t fault_typea, 5092 vm_map_entry_t *out_entry, /* OUT */ 5093 vm_object_t *object, /* OUT */ 5094 vm_pindex_t *pindex, /* OUT */ 5095 vm_prot_t *out_prot, /* OUT */ 5096 boolean_t *wired) /* OUT */ 5097 { 5098 vm_map_entry_t entry; 5099 vm_map_t map = *var_map; 5100 vm_prot_t prot; 5101 vm_prot_t fault_type = fault_typea; 5102 5103 /* 5104 * Lookup the faulting address. 5105 */ 5106 if (!vm_map_lookup_entry(map, vaddr, out_entry)) 5107 return (KERN_INVALID_ADDRESS); 5108 5109 entry = *out_entry; 5110 5111 /* 5112 * Fail if the entry refers to a submap. 5113 */ 5114 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) 5115 return (KERN_FAILURE); 5116 5117 /* 5118 * Check whether this task is allowed to have this page. 5119 */ 5120 prot = entry->protection; 5121 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 5122 if ((fault_type & prot) != fault_type) 5123 return (KERN_PROTECTION_FAILURE); 5124 5125 /* 5126 * If this page is not pageable, we have to get it for all possible 5127 * accesses. 5128 */ 5129 *wired = (entry->wired_count != 0); 5130 if (*wired) 5131 fault_type = entry->protection; 5132 5133 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 5134 /* 5135 * Fail if the entry was copy-on-write for a write fault. 5136 */ 5137 if (fault_type & VM_PROT_WRITE) 5138 return (KERN_FAILURE); 5139 /* 5140 * We're attempting to read a copy-on-write page -- 5141 * don't allow writes. 5142 */ 5143 prot &= ~VM_PROT_WRITE; 5144 } 5145 5146 /* 5147 * Fail if an object should be created. 5148 */ 5149 if (entry->object.vm_object == NULL && !map->system_map) 5150 return (KERN_FAILURE); 5151 5152 /* 5153 * Return the object/offset from this entry. If the entry was 5154 * copy-on-write or empty, it has been fixed up. 5155 */ 5156 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 5157 *object = entry->object.vm_object; 5158 5159 *out_prot = prot; 5160 return (KERN_SUCCESS); 5161 } 5162 5163 /* 5164 * vm_map_lookup_done: 5165 * 5166 * Releases locks acquired by a vm_map_lookup 5167 * (according to the handle returned by that lookup). 5168 */ 5169 void 5170 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry) 5171 { 5172 /* 5173 * Unlock the main-level map 5174 */ 5175 vm_map_unlock_read(map); 5176 } 5177 5178 vm_offset_t 5179 vm_map_max_KBI(const struct vm_map *map) 5180 { 5181 5182 return (vm_map_max(map)); 5183 } 5184 5185 vm_offset_t 5186 vm_map_min_KBI(const struct vm_map *map) 5187 { 5188 5189 return (vm_map_min(map)); 5190 } 5191 5192 pmap_t 5193 vm_map_pmap_KBI(vm_map_t map) 5194 { 5195 5196 return (map->pmap); 5197 } 5198 5199 bool 5200 vm_map_range_valid_KBI(vm_map_t map, vm_offset_t start, vm_offset_t end) 5201 { 5202 5203 return (vm_map_range_valid(map, start, end)); 5204 } 5205 5206 #ifdef INVARIANTS 5207 static void 5208 _vm_map_assert_consistent(vm_map_t map, int check) 5209 { 5210 vm_map_entry_t entry, prev; 5211 vm_map_entry_t cur, header, lbound, ubound; 5212 vm_size_t max_left, max_right; 5213 5214 #ifdef DIAGNOSTIC 5215 ++map->nupdates; 5216 #endif 5217 if (enable_vmmap_check != check) 5218 return; 5219 5220 header = prev = &map->header; 5221 VM_MAP_ENTRY_FOREACH(entry, map) { 5222 KASSERT(prev->end <= entry->start, 5223 ("map %p prev->end = %jx, start = %jx", map, 5224 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5225 KASSERT(entry->start < entry->end, 5226 ("map %p start = %jx, end = %jx", map, 5227 (uintmax_t)entry->start, (uintmax_t)entry->end)); 5228 KASSERT(entry->left == header || 5229 entry->left->start < entry->start, 5230 ("map %p left->start = %jx, start = %jx", map, 5231 (uintmax_t)entry->left->start, (uintmax_t)entry->start)); 5232 KASSERT(entry->right == header || 5233 entry->start < entry->right->start, 5234 ("map %p start = %jx, right->start = %jx", map, 5235 (uintmax_t)entry->start, (uintmax_t)entry->right->start)); 5236 cur = map->root; 5237 lbound = ubound = header; 5238 for (;;) { 5239 if (entry->start < cur->start) { 5240 ubound = cur; 5241 cur = cur->left; 5242 KASSERT(cur != lbound, 5243 ("map %p cannot find %jx", 5244 map, (uintmax_t)entry->start)); 5245 } else if (cur->end <= entry->start) { 5246 lbound = cur; 5247 cur = cur->right; 5248 KASSERT(cur != ubound, 5249 ("map %p cannot find %jx", 5250 map, (uintmax_t)entry->start)); 5251 } else { 5252 KASSERT(cur == entry, 5253 ("map %p cannot find %jx", 5254 map, (uintmax_t)entry->start)); 5255 break; 5256 } 5257 } 5258 max_left = vm_map_entry_max_free_left(entry, lbound); 5259 max_right = vm_map_entry_max_free_right(entry, ubound); 5260 KASSERT(entry->max_free == vm_size_max(max_left, max_right), 5261 ("map %p max = %jx, max_left = %jx, max_right = %jx", map, 5262 (uintmax_t)entry->max_free, 5263 (uintmax_t)max_left, (uintmax_t)max_right)); 5264 prev = entry; 5265 } 5266 KASSERT(prev->end <= entry->start, 5267 ("map %p prev->end = %jx, start = %jx", map, 5268 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5269 } 5270 #endif 5271 5272 #include "opt_ddb.h" 5273 #ifdef DDB 5274 #include <sys/kernel.h> 5275 5276 #include <ddb/ddb.h> 5277 5278 static void 5279 vm_map_print(vm_map_t map) 5280 { 5281 vm_map_entry_t entry, prev; 5282 5283 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 5284 (void *)map, 5285 (void *)map->pmap, map->nentries, map->timestamp); 5286 5287 db_indent += 2; 5288 prev = &map->header; 5289 VM_MAP_ENTRY_FOREACH(entry, map) { 5290 db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n", 5291 (void *)entry, (void *)entry->start, (void *)entry->end, 5292 entry->eflags); 5293 { 5294 static const char * const inheritance_name[4] = 5295 {"share", "copy", "none", "donate_copy"}; 5296 5297 db_iprintf(" prot=%x/%x/%s", 5298 entry->protection, 5299 entry->max_protection, 5300 inheritance_name[(int)(unsigned char) 5301 entry->inheritance]); 5302 if (entry->wired_count != 0) 5303 db_printf(", wired"); 5304 } 5305 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 5306 db_printf(", share=%p, offset=0x%jx\n", 5307 (void *)entry->object.sub_map, 5308 (uintmax_t)entry->offset); 5309 if (prev == &map->header || 5310 prev->object.sub_map != 5311 entry->object.sub_map) { 5312 db_indent += 2; 5313 vm_map_print((vm_map_t)entry->object.sub_map); 5314 db_indent -= 2; 5315 } 5316 } else { 5317 if (entry->cred != NULL) 5318 db_printf(", ruid %d", entry->cred->cr_ruid); 5319 db_printf(", object=%p, offset=0x%jx", 5320 (void *)entry->object.vm_object, 5321 (uintmax_t)entry->offset); 5322 if (entry->object.vm_object && entry->object.vm_object->cred) 5323 db_printf(", obj ruid %d charge %jx", 5324 entry->object.vm_object->cred->cr_ruid, 5325 (uintmax_t)entry->object.vm_object->charge); 5326 if (entry->eflags & MAP_ENTRY_COW) 5327 db_printf(", copy (%s)", 5328 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 5329 db_printf("\n"); 5330 5331 if (prev == &map->header || 5332 prev->object.vm_object != 5333 entry->object.vm_object) { 5334 db_indent += 2; 5335 vm_object_print((db_expr_t)(intptr_t) 5336 entry->object.vm_object, 5337 0, 0, (char *)0); 5338 db_indent -= 2; 5339 } 5340 } 5341 prev = entry; 5342 } 5343 db_indent -= 2; 5344 } 5345 5346 DB_SHOW_COMMAND(map, map) 5347 { 5348 5349 if (!have_addr) { 5350 db_printf("usage: show map <addr>\n"); 5351 return; 5352 } 5353 vm_map_print((vm_map_t)addr); 5354 } 5355 5356 DB_SHOW_COMMAND(procvm, procvm) 5357 { 5358 struct proc *p; 5359 5360 if (have_addr) { 5361 p = db_lookup_proc(addr); 5362 } else { 5363 p = curproc; 5364 } 5365 5366 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 5367 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 5368 (void *)vmspace_pmap(p->p_vmspace)); 5369 5370 vm_map_print((vm_map_t)&p->p_vmspace->vm_map); 5371 } 5372 5373 #endif /* DDB */ 5374