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 * 35 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 36 * All rights reserved. 37 * 38 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 39 * 40 * Permission to use, copy, modify and distribute this software and 41 * its documentation is hereby granted, provided that both the copyright 42 * notice and this permission notice appear in all copies of the 43 * software, derivative works or modified versions, and any portions 44 * thereof, and that both notices appear in supporting documentation. 45 * 46 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 47 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 48 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 49 * 50 * Carnegie Mellon requests users of this software to return to 51 * 52 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 53 * School of Computer Science 54 * Carnegie Mellon University 55 * Pittsburgh PA 15213-3890 56 * 57 * any improvements or extensions that they make and grant Carnegie the 58 * rights to redistribute these changes. 59 */ 60 61 /* 62 * Virtual memory mapping module. 63 */ 64 65 #include <sys/param.h> 66 #include <sys/systm.h> 67 #include <sys/elf.h> 68 #include <sys/kernel.h> 69 #include <sys/ktr.h> 70 #include <sys/lock.h> 71 #include <sys/mutex.h> 72 #include <sys/proc.h> 73 #include <sys/vmmeter.h> 74 #include <sys/mman.h> 75 #include <sys/vnode.h> 76 #include <sys/racct.h> 77 #include <sys/resourcevar.h> 78 #include <sys/rwlock.h> 79 #include <sys/file.h> 80 #include <sys/sysctl.h> 81 #include <sys/sysent.h> 82 #include <sys/shm.h> 83 84 #include <vm/vm.h> 85 #include <vm/vm_param.h> 86 #include <vm/pmap.h> 87 #include <vm/vm_map.h> 88 #include <vm/vm_page.h> 89 #include <vm/vm_pageout.h> 90 #include <vm/vm_object.h> 91 #include <vm/vm_pager.h> 92 #include <vm/vm_kern.h> 93 #include <vm/vm_extern.h> 94 #include <vm/vnode_pager.h> 95 #include <vm/swap_pager.h> 96 #include <vm/uma.h> 97 98 /* 99 * Virtual memory maps provide for the mapping, protection, 100 * and sharing of virtual memory objects. In addition, 101 * this module provides for an efficient virtual copy of 102 * memory from one map to another. 103 * 104 * Synchronization is required prior to most operations. 105 * 106 * Maps consist of an ordered doubly-linked list of simple 107 * entries; a self-adjusting binary search tree of these 108 * entries is used to speed up lookups. 109 * 110 * Since portions of maps are specified by start/end addresses, 111 * which may not align with existing map entries, all 112 * routines merely "clip" entries to these start/end values. 113 * [That is, an entry is split into two, bordering at a 114 * start or end value.] Note that these clippings may not 115 * always be necessary (as the two resulting entries are then 116 * not changed); however, the clipping is done for convenience. 117 * 118 * As mentioned above, virtual copy operations are performed 119 * by copying VM object references from one map to 120 * another, and then marking both regions as copy-on-write. 121 */ 122 123 static struct mtx map_sleep_mtx; 124 static uma_zone_t mapentzone; 125 static uma_zone_t kmapentzone; 126 static uma_zone_t vmspace_zone; 127 static int vmspace_zinit(void *mem, int size, int flags); 128 static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, 129 vm_offset_t max); 130 static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map); 131 static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry); 132 static void vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry); 133 static int vm_map_growstack(vm_map_t map, vm_offset_t addr, 134 vm_map_entry_t gap_entry); 135 static void vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 136 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags); 137 #ifdef INVARIANTS 138 static void vmspace_zdtor(void *mem, int size, void *arg); 139 #endif 140 static int vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, 141 vm_size_t max_ssize, vm_size_t growsize, vm_prot_t prot, vm_prot_t max, 142 int cow); 143 static void vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 144 vm_offset_t failed_addr); 145 146 #define CONTAINS_BITS(set, bits) ((~(set) & (bits)) == 0) 147 148 #define ENTRY_CHARGED(e) ((e)->cred != NULL || \ 149 ((e)->object.vm_object != NULL && (e)->object.vm_object->cred != NULL && \ 150 !((e)->eflags & MAP_ENTRY_NEEDS_COPY))) 151 152 /* 153 * PROC_VMSPACE_{UN,}LOCK() can be a noop as long as vmspaces are type 154 * stable. 155 */ 156 #define PROC_VMSPACE_LOCK(p) do { } while (0) 157 #define PROC_VMSPACE_UNLOCK(p) do { } while (0) 158 159 /* 160 * VM_MAP_RANGE_CHECK: [ internal use only ] 161 * 162 * Asserts that the starting and ending region 163 * addresses fall within the valid range of the map. 164 */ 165 #define VM_MAP_RANGE_CHECK(map, start, end) \ 166 { \ 167 if (start < vm_map_min(map)) \ 168 start = vm_map_min(map); \ 169 if (end > vm_map_max(map)) \ 170 end = vm_map_max(map); \ 171 if (start > end) \ 172 start = end; \ 173 } 174 175 #ifndef UMA_USE_DMAP 176 177 /* 178 * Allocate a new slab for kernel map entries. The kernel map may be locked or 179 * unlocked, depending on whether the request is coming from the kernel map or a 180 * submap. This function allocates a virtual address range directly from the 181 * kernel map instead of the kmem_* layer to avoid recursion on the kernel map 182 * lock and also to avoid triggering allocator recursion in the vmem boundary 183 * tag allocator. 184 */ 185 static void * 186 kmapent_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 187 int wait) 188 { 189 vm_offset_t addr; 190 int error, locked; 191 192 *pflag = UMA_SLAB_PRIV; 193 194 if (!(locked = vm_map_locked(kernel_map))) 195 vm_map_lock(kernel_map); 196 addr = vm_map_findspace(kernel_map, vm_map_min(kernel_map), bytes); 197 if (addr + bytes < addr || addr + bytes > vm_map_max(kernel_map)) 198 panic("%s: kernel map is exhausted", __func__); 199 error = vm_map_insert(kernel_map, NULL, 0, addr, addr + bytes, 200 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 201 if (error != KERN_SUCCESS) 202 panic("%s: vm_map_insert() failed: %d", __func__, error); 203 if (!locked) 204 vm_map_unlock(kernel_map); 205 error = kmem_back_domain(domain, kernel_object, addr, bytes, M_NOWAIT | 206 M_USE_RESERVE | (wait & M_ZERO)); 207 if (error == KERN_SUCCESS) { 208 return ((void *)addr); 209 } else { 210 if (!locked) 211 vm_map_lock(kernel_map); 212 vm_map_delete(kernel_map, addr, bytes); 213 if (!locked) 214 vm_map_unlock(kernel_map); 215 return (NULL); 216 } 217 } 218 219 static void 220 kmapent_free(void *item, vm_size_t size, uint8_t pflag) 221 { 222 vm_offset_t addr; 223 int error __diagused; 224 225 if ((pflag & UMA_SLAB_PRIV) == 0) 226 /* XXX leaked */ 227 return; 228 229 addr = (vm_offset_t)item; 230 kmem_unback(kernel_object, addr, size); 231 error = vm_map_remove(kernel_map, addr, addr + size); 232 KASSERT(error == KERN_SUCCESS, 233 ("%s: vm_map_remove failed: %d", __func__, error)); 234 } 235 236 /* 237 * The worst-case upper bound on the number of kernel map entries that may be 238 * created before the zone must be replenished in _vm_map_unlock(). 239 */ 240 #define KMAPENT_RESERVE 1 241 242 #endif /* !UMD_MD_SMALL_ALLOC */ 243 244 /* 245 * vm_map_startup: 246 * 247 * Initialize the vm_map module. Must be called before any other vm_map 248 * routines. 249 * 250 * User map and entry structures are allocated from the general purpose 251 * memory pool. Kernel maps are statically defined. Kernel map entries 252 * require special handling to avoid recursion; see the comments above 253 * kmapent_alloc() and in vm_map_entry_create(). 254 */ 255 void 256 vm_map_startup(void) 257 { 258 mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF); 259 260 /* 261 * Disable the use of per-CPU buckets: map entry allocation is 262 * serialized by the kernel map lock. 263 */ 264 kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry), 265 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 266 UMA_ZONE_VM | UMA_ZONE_NOBUCKET); 267 #ifndef UMA_USE_DMAP 268 /* Reserve an extra map entry for use when replenishing the reserve. */ 269 uma_zone_reserve(kmapentzone, KMAPENT_RESERVE + 1); 270 uma_prealloc(kmapentzone, KMAPENT_RESERVE + 1); 271 uma_zone_set_allocf(kmapentzone, kmapent_alloc); 272 uma_zone_set_freef(kmapentzone, kmapent_free); 273 #endif 274 275 mapentzone = uma_zcreate("MAP ENTRY", sizeof(struct vm_map_entry), 276 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 277 vmspace_zone = uma_zcreate("VMSPACE", sizeof(struct vmspace), NULL, 278 #ifdef INVARIANTS 279 vmspace_zdtor, 280 #else 281 NULL, 282 #endif 283 vmspace_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 284 } 285 286 static int 287 vmspace_zinit(void *mem, int size, int flags) 288 { 289 struct vmspace *vm; 290 vm_map_t map; 291 292 vm = (struct vmspace *)mem; 293 map = &vm->vm_map; 294 295 memset(map, 0, sizeof(*map)); /* set MAP_SYSTEM_MAP to false */ 296 sx_init(&map->lock, "vm map (user)"); 297 PMAP_LOCK_INIT(vmspace_pmap(vm)); 298 return (0); 299 } 300 301 #ifdef INVARIANTS 302 static void 303 vmspace_zdtor(void *mem, int size, void *arg) 304 { 305 struct vmspace *vm; 306 307 vm = (struct vmspace *)mem; 308 KASSERT(vm->vm_map.nentries == 0, 309 ("vmspace %p nentries == %d on free", vm, vm->vm_map.nentries)); 310 KASSERT(vm->vm_map.size == 0, 311 ("vmspace %p size == %ju on free", vm, (uintmax_t)vm->vm_map.size)); 312 } 313 #endif /* INVARIANTS */ 314 315 /* 316 * Allocate a vmspace structure, including a vm_map and pmap, 317 * and initialize those structures. The refcnt is set to 1. 318 */ 319 struct vmspace * 320 vmspace_alloc(vm_offset_t min, vm_offset_t max, pmap_pinit_t pinit) 321 { 322 struct vmspace *vm; 323 324 vm = uma_zalloc(vmspace_zone, M_WAITOK); 325 KASSERT(vm->vm_map.pmap == NULL, ("vm_map.pmap must be NULL")); 326 if (!pinit(vmspace_pmap(vm))) { 327 uma_zfree(vmspace_zone, vm); 328 return (NULL); 329 } 330 CTR1(KTR_VM, "vmspace_alloc: %p", vm); 331 _vm_map_init(&vm->vm_map, vmspace_pmap(vm), min, max); 332 refcount_init(&vm->vm_refcnt, 1); 333 vm->vm_shm = NULL; 334 vm->vm_swrss = 0; 335 vm->vm_tsize = 0; 336 vm->vm_dsize = 0; 337 vm->vm_ssize = 0; 338 vm->vm_taddr = 0; 339 vm->vm_daddr = 0; 340 vm->vm_maxsaddr = 0; 341 return (vm); 342 } 343 344 #ifdef RACCT 345 static void 346 vmspace_container_reset(struct proc *p) 347 { 348 349 PROC_LOCK(p); 350 racct_set(p, RACCT_DATA, 0); 351 racct_set(p, RACCT_STACK, 0); 352 racct_set(p, RACCT_RSS, 0); 353 racct_set(p, RACCT_MEMLOCK, 0); 354 racct_set(p, RACCT_VMEM, 0); 355 PROC_UNLOCK(p); 356 } 357 #endif 358 359 static inline void 360 vmspace_dofree(struct vmspace *vm) 361 { 362 363 CTR1(KTR_VM, "vmspace_free: %p", vm); 364 365 /* 366 * Make sure any SysV shm is freed, it might not have been in 367 * exit1(). 368 */ 369 shmexit(vm); 370 371 /* 372 * Lock the map, to wait out all other references to it. 373 * Delete all of the mappings and pages they hold, then call 374 * the pmap module to reclaim anything left. 375 */ 376 (void)vm_map_remove(&vm->vm_map, vm_map_min(&vm->vm_map), 377 vm_map_max(&vm->vm_map)); 378 379 pmap_release(vmspace_pmap(vm)); 380 vm->vm_map.pmap = NULL; 381 uma_zfree(vmspace_zone, vm); 382 } 383 384 void 385 vmspace_free(struct vmspace *vm) 386 { 387 388 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 389 "vmspace_free() called"); 390 391 if (refcount_release(&vm->vm_refcnt)) 392 vmspace_dofree(vm); 393 } 394 395 void 396 vmspace_exitfree(struct proc *p) 397 { 398 struct vmspace *vm; 399 400 PROC_VMSPACE_LOCK(p); 401 vm = p->p_vmspace; 402 p->p_vmspace = NULL; 403 PROC_VMSPACE_UNLOCK(p); 404 KASSERT(vm == &vmspace0, ("vmspace_exitfree: wrong vmspace")); 405 vmspace_free(vm); 406 } 407 408 void 409 vmspace_exit(struct thread *td) 410 { 411 struct vmspace *vm; 412 struct proc *p; 413 bool released; 414 415 p = td->td_proc; 416 vm = p->p_vmspace; 417 418 /* 419 * Prepare to release the vmspace reference. The thread that releases 420 * the last reference is responsible for tearing down the vmspace. 421 * However, threads not releasing the final reference must switch to the 422 * kernel's vmspace0 before the decrement so that the subsequent pmap 423 * deactivation does not modify a freed vmspace. 424 */ 425 refcount_acquire(&vmspace0.vm_refcnt); 426 if (!(released = refcount_release_if_last(&vm->vm_refcnt))) { 427 if (p->p_vmspace != &vmspace0) { 428 PROC_VMSPACE_LOCK(p); 429 p->p_vmspace = &vmspace0; 430 PROC_VMSPACE_UNLOCK(p); 431 pmap_activate(td); 432 } 433 released = refcount_release(&vm->vm_refcnt); 434 } 435 if (released) { 436 /* 437 * pmap_remove_pages() expects the pmap to be active, so switch 438 * back first if necessary. 439 */ 440 if (p->p_vmspace != vm) { 441 PROC_VMSPACE_LOCK(p); 442 p->p_vmspace = vm; 443 PROC_VMSPACE_UNLOCK(p); 444 pmap_activate(td); 445 } 446 pmap_remove_pages(vmspace_pmap(vm)); 447 PROC_VMSPACE_LOCK(p); 448 p->p_vmspace = &vmspace0; 449 PROC_VMSPACE_UNLOCK(p); 450 pmap_activate(td); 451 vmspace_dofree(vm); 452 } 453 #ifdef RACCT 454 if (racct_enable) 455 vmspace_container_reset(p); 456 #endif 457 } 458 459 /* Acquire reference to vmspace owned by another process. */ 460 461 struct vmspace * 462 vmspace_acquire_ref(struct proc *p) 463 { 464 struct vmspace *vm; 465 466 PROC_VMSPACE_LOCK(p); 467 vm = p->p_vmspace; 468 if (vm == NULL || !refcount_acquire_if_not_zero(&vm->vm_refcnt)) { 469 PROC_VMSPACE_UNLOCK(p); 470 return (NULL); 471 } 472 if (vm != p->p_vmspace) { 473 PROC_VMSPACE_UNLOCK(p); 474 vmspace_free(vm); 475 return (NULL); 476 } 477 PROC_VMSPACE_UNLOCK(p); 478 return (vm); 479 } 480 481 /* 482 * Switch between vmspaces in an AIO kernel process. 483 * 484 * The new vmspace is either the vmspace of a user process obtained 485 * from an active AIO request or the initial vmspace of the AIO kernel 486 * process (when it is idling). Because user processes will block to 487 * drain any active AIO requests before proceeding in exit() or 488 * execve(), the reference count for vmspaces from AIO requests can 489 * never be 0. Similarly, AIO kernel processes hold an extra 490 * reference on their initial vmspace for the life of the process. As 491 * a result, the 'newvm' vmspace always has a non-zero reference 492 * count. This permits an additional reference on 'newvm' to be 493 * acquired via a simple atomic increment rather than the loop in 494 * vmspace_acquire_ref() above. 495 */ 496 void 497 vmspace_switch_aio(struct vmspace *newvm) 498 { 499 struct vmspace *oldvm; 500 501 /* XXX: Need some way to assert that this is an aio daemon. */ 502 503 KASSERT(refcount_load(&newvm->vm_refcnt) > 0, 504 ("vmspace_switch_aio: newvm unreferenced")); 505 506 oldvm = curproc->p_vmspace; 507 if (oldvm == newvm) 508 return; 509 510 /* 511 * Point to the new address space and refer to it. 512 */ 513 curproc->p_vmspace = newvm; 514 refcount_acquire(&newvm->vm_refcnt); 515 516 /* Activate the new mapping. */ 517 pmap_activate(curthread); 518 519 vmspace_free(oldvm); 520 } 521 522 void 523 _vm_map_lock(vm_map_t map, const char *file, int line) 524 { 525 526 if (vm_map_is_system(map)) 527 mtx_lock_flags_(&map->system_mtx, 0, file, line); 528 else 529 sx_xlock_(&map->lock, file, line); 530 map->timestamp++; 531 } 532 533 void 534 vm_map_entry_set_vnode_text(vm_map_entry_t entry, bool add) 535 { 536 vm_object_t object; 537 struct vnode *vp; 538 bool vp_held; 539 540 if ((entry->eflags & MAP_ENTRY_VN_EXEC) == 0) 541 return; 542 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 543 ("Submap with execs")); 544 object = entry->object.vm_object; 545 KASSERT(object != NULL, ("No object for text, entry %p", entry)); 546 if ((object->flags & OBJ_ANON) != 0) 547 object = object->handle; 548 else 549 KASSERT(object->backing_object == NULL, 550 ("non-anon object %p shadows", object)); 551 KASSERT(object != NULL, ("No content object for text, entry %p obj %p", 552 entry, entry->object.vm_object)); 553 554 /* 555 * Mostly, we do not lock the backing object. It is 556 * referenced by the entry we are processing, so it cannot go 557 * away. 558 */ 559 vm_pager_getvp(object, &vp, &vp_held); 560 if (vp != NULL) { 561 if (add) { 562 VOP_SET_TEXT_CHECKED(vp); 563 } else { 564 vn_lock(vp, LK_SHARED | LK_RETRY); 565 VOP_UNSET_TEXT_CHECKED(vp); 566 VOP_UNLOCK(vp); 567 } 568 if (vp_held) 569 vdrop(vp); 570 } 571 } 572 573 /* 574 * Use a different name for this vm_map_entry field when it's use 575 * is not consistent with its use as part of an ordered search tree. 576 */ 577 #define defer_next right 578 579 static void 580 vm_map_process_deferred(void) 581 { 582 struct thread *td; 583 vm_map_entry_t entry, next; 584 vm_object_t object; 585 586 td = curthread; 587 entry = td->td_map_def_user; 588 td->td_map_def_user = NULL; 589 while (entry != NULL) { 590 next = entry->defer_next; 591 MPASS((entry->eflags & (MAP_ENTRY_WRITECNT | 592 MAP_ENTRY_VN_EXEC)) != (MAP_ENTRY_WRITECNT | 593 MAP_ENTRY_VN_EXEC)); 594 if ((entry->eflags & MAP_ENTRY_WRITECNT) != 0) { 595 /* 596 * Decrement the object's writemappings and 597 * possibly the vnode's v_writecount. 598 */ 599 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 600 ("Submap with writecount")); 601 object = entry->object.vm_object; 602 KASSERT(object != NULL, ("No object for writecount")); 603 vm_pager_release_writecount(object, entry->start, 604 entry->end); 605 } 606 vm_map_entry_set_vnode_text(entry, false); 607 vm_map_entry_deallocate(entry, FALSE); 608 entry = next; 609 } 610 } 611 612 #ifdef INVARIANTS 613 static void 614 _vm_map_assert_locked(vm_map_t map, const char *file, int line) 615 { 616 617 if (vm_map_is_system(map)) 618 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 619 else 620 sx_assert_(&map->lock, SA_XLOCKED, file, line); 621 } 622 623 #define VM_MAP_ASSERT_LOCKED(map) \ 624 _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE) 625 626 enum { VMMAP_CHECK_NONE, VMMAP_CHECK_UNLOCK, VMMAP_CHECK_ALL }; 627 #ifdef DIAGNOSTIC 628 static int enable_vmmap_check = VMMAP_CHECK_UNLOCK; 629 #else 630 static int enable_vmmap_check = VMMAP_CHECK_NONE; 631 #endif 632 SYSCTL_INT(_debug, OID_AUTO, vmmap_check, CTLFLAG_RWTUN, 633 &enable_vmmap_check, 0, "Enable vm map consistency checking"); 634 635 static void _vm_map_assert_consistent(vm_map_t map, int check); 636 637 #define VM_MAP_ASSERT_CONSISTENT(map) \ 638 _vm_map_assert_consistent(map, VMMAP_CHECK_ALL) 639 #ifdef DIAGNOSTIC 640 #define VM_MAP_UNLOCK_CONSISTENT(map) do { \ 641 if (map->nupdates > map->nentries) { \ 642 _vm_map_assert_consistent(map, VMMAP_CHECK_UNLOCK); \ 643 map->nupdates = 0; \ 644 } \ 645 } while (0) 646 #else 647 #define VM_MAP_UNLOCK_CONSISTENT(map) 648 #endif 649 #else 650 #define VM_MAP_ASSERT_LOCKED(map) 651 #define VM_MAP_ASSERT_CONSISTENT(map) 652 #define VM_MAP_UNLOCK_CONSISTENT(map) 653 #endif /* INVARIANTS */ 654 655 void 656 _vm_map_unlock(vm_map_t map, const char *file, int line) 657 { 658 659 VM_MAP_UNLOCK_CONSISTENT(map); 660 if (vm_map_is_system(map)) { 661 #ifndef UMA_USE_DMAP 662 if (map == kernel_map && (map->flags & MAP_REPLENISH) != 0) { 663 uma_prealloc(kmapentzone, 1); 664 map->flags &= ~MAP_REPLENISH; 665 } 666 #endif 667 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 668 } else { 669 sx_xunlock_(&map->lock, file, line); 670 vm_map_process_deferred(); 671 } 672 } 673 674 void 675 _vm_map_lock_read(vm_map_t map, const char *file, int line) 676 { 677 678 if (vm_map_is_system(map)) 679 mtx_lock_flags_(&map->system_mtx, 0, file, line); 680 else 681 sx_slock_(&map->lock, file, line); 682 } 683 684 void 685 _vm_map_unlock_read(vm_map_t map, const char *file, int line) 686 { 687 688 if (vm_map_is_system(map)) { 689 KASSERT((map->flags & MAP_REPLENISH) == 0, 690 ("%s: MAP_REPLENISH leaked", __func__)); 691 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 692 } else { 693 sx_sunlock_(&map->lock, file, line); 694 vm_map_process_deferred(); 695 } 696 } 697 698 int 699 _vm_map_trylock(vm_map_t map, const char *file, int line) 700 { 701 int error; 702 703 error = vm_map_is_system(map) ? 704 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 705 !sx_try_xlock_(&map->lock, file, line); 706 if (error == 0) 707 map->timestamp++; 708 return (error == 0); 709 } 710 711 int 712 _vm_map_trylock_read(vm_map_t map, const char *file, int line) 713 { 714 int error; 715 716 error = vm_map_is_system(map) ? 717 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 718 !sx_try_slock_(&map->lock, file, line); 719 return (error == 0); 720 } 721 722 /* 723 * _vm_map_lock_upgrade: [ internal use only ] 724 * 725 * Tries to upgrade a read (shared) lock on the specified map to a write 726 * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a 727 * non-zero value if the upgrade fails. If the upgrade fails, the map is 728 * returned without a read or write lock held. 729 * 730 * Requires that the map be read locked. 731 */ 732 int 733 _vm_map_lock_upgrade(vm_map_t map, const char *file, int line) 734 { 735 unsigned int last_timestamp; 736 737 if (vm_map_is_system(map)) { 738 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 739 } else { 740 if (!sx_try_upgrade_(&map->lock, file, line)) { 741 last_timestamp = map->timestamp; 742 sx_sunlock_(&map->lock, file, line); 743 vm_map_process_deferred(); 744 /* 745 * If the map's timestamp does not change while the 746 * map is unlocked, then the upgrade succeeds. 747 */ 748 sx_xlock_(&map->lock, file, line); 749 if (last_timestamp != map->timestamp) { 750 sx_xunlock_(&map->lock, file, line); 751 return (1); 752 } 753 } 754 } 755 map->timestamp++; 756 return (0); 757 } 758 759 void 760 _vm_map_lock_downgrade(vm_map_t map, const char *file, int line) 761 { 762 763 if (vm_map_is_system(map)) { 764 KASSERT((map->flags & MAP_REPLENISH) == 0, 765 ("%s: MAP_REPLENISH leaked", __func__)); 766 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 767 } else { 768 VM_MAP_UNLOCK_CONSISTENT(map); 769 sx_downgrade_(&map->lock, file, line); 770 } 771 } 772 773 /* 774 * vm_map_locked: 775 * 776 * Returns a non-zero value if the caller holds a write (exclusive) lock 777 * on the specified map and the value "0" otherwise. 778 */ 779 int 780 vm_map_locked(vm_map_t map) 781 { 782 783 if (vm_map_is_system(map)) 784 return (mtx_owned(&map->system_mtx)); 785 return (sx_xlocked(&map->lock)); 786 } 787 788 /* 789 * _vm_map_unlock_and_wait: 790 * 791 * Atomically releases the lock on the specified map and puts the calling 792 * thread to sleep. The calling thread will remain asleep until either 793 * vm_map_wakeup() is performed on the map or the specified timeout is 794 * exceeded. 795 * 796 * WARNING! This function does not perform deferred deallocations of 797 * objects and map entries. Therefore, the calling thread is expected to 798 * reacquire the map lock after reawakening and later perform an ordinary 799 * unlock operation, such as vm_map_unlock(), before completing its 800 * operation on the map. 801 */ 802 int 803 _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line) 804 { 805 806 VM_MAP_UNLOCK_CONSISTENT(map); 807 mtx_lock(&map_sleep_mtx); 808 if (vm_map_is_system(map)) { 809 KASSERT((map->flags & MAP_REPLENISH) == 0, 810 ("%s: MAP_REPLENISH leaked", __func__)); 811 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 812 } else { 813 sx_xunlock_(&map->lock, file, line); 814 } 815 return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps", 816 timo)); 817 } 818 819 /* 820 * vm_map_wakeup: 821 * 822 * Awaken any threads that have slept on the map using 823 * vm_map_unlock_and_wait(). 824 */ 825 void 826 vm_map_wakeup(vm_map_t map) 827 { 828 829 /* 830 * Acquire and release map_sleep_mtx to prevent a wakeup() 831 * from being performed (and lost) between the map unlock 832 * and the msleep() in _vm_map_unlock_and_wait(). 833 */ 834 mtx_lock(&map_sleep_mtx); 835 mtx_unlock(&map_sleep_mtx); 836 wakeup(&map->root); 837 } 838 839 void 840 vm_map_busy(vm_map_t map) 841 { 842 843 VM_MAP_ASSERT_LOCKED(map); 844 map->busy++; 845 } 846 847 void 848 vm_map_unbusy(vm_map_t map) 849 { 850 851 VM_MAP_ASSERT_LOCKED(map); 852 KASSERT(map->busy, ("vm_map_unbusy: not busy")); 853 if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) { 854 vm_map_modflags(map, 0, MAP_BUSY_WAKEUP); 855 wakeup(&map->busy); 856 } 857 } 858 859 void 860 vm_map_wait_busy(vm_map_t map) 861 { 862 863 VM_MAP_ASSERT_LOCKED(map); 864 while (map->busy) { 865 vm_map_modflags(map, MAP_BUSY_WAKEUP, 0); 866 if (vm_map_is_system(map)) 867 msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0); 868 else 869 sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0); 870 } 871 map->timestamp++; 872 } 873 874 long 875 vmspace_resident_count(struct vmspace *vmspace) 876 { 877 return pmap_resident_count(vmspace_pmap(vmspace)); 878 } 879 880 /* 881 * Initialize an existing vm_map structure 882 * such as that in the vmspace structure. 883 */ 884 static void 885 _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 886 { 887 888 map->header.eflags = MAP_ENTRY_HEADER; 889 map->pmap = pmap; 890 map->header.end = min; 891 map->header.start = max; 892 map->flags = 0; 893 map->header.left = map->header.right = &map->header; 894 map->root = NULL; 895 map->timestamp = 0; 896 map->busy = 0; 897 map->anon_loc = 0; 898 #ifdef DIAGNOSTIC 899 map->nupdates = 0; 900 #endif 901 } 902 903 void 904 vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 905 { 906 _vm_map_init(map, pmap, min, max); 907 sx_init(&map->lock, "vm map (user)"); 908 } 909 910 void 911 vm_map_init_system(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 912 { 913 _vm_map_init(map, pmap, min, max); 914 vm_map_modflags(map, MAP_SYSTEM_MAP, 0); 915 mtx_init(&map->system_mtx, "vm map (system)", NULL, MTX_DEF | 916 MTX_DUPOK); 917 } 918 919 /* 920 * vm_map_entry_dispose: [ internal use only ] 921 * 922 * Inverse of vm_map_entry_create. 923 */ 924 static void 925 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry) 926 { 927 uma_zfree(vm_map_is_system(map) ? kmapentzone : mapentzone, entry); 928 } 929 930 /* 931 * vm_map_entry_create: [ internal use only ] 932 * 933 * Allocates a VM map entry for insertion. 934 * No entry fields are filled in. 935 */ 936 static vm_map_entry_t 937 vm_map_entry_create(vm_map_t map) 938 { 939 vm_map_entry_t new_entry; 940 941 #ifndef UMA_USE_DMAP 942 if (map == kernel_map) { 943 VM_MAP_ASSERT_LOCKED(map); 944 945 /* 946 * A new slab of kernel map entries cannot be allocated at this 947 * point because the kernel map has not yet been updated to 948 * reflect the caller's request. Therefore, we allocate a new 949 * map entry, dipping into the reserve if necessary, and set a 950 * flag indicating that the reserve must be replenished before 951 * the map is unlocked. 952 */ 953 new_entry = uma_zalloc(kmapentzone, M_NOWAIT | M_NOVM); 954 if (new_entry == NULL) { 955 new_entry = uma_zalloc(kmapentzone, 956 M_NOWAIT | M_NOVM | M_USE_RESERVE); 957 kernel_map->flags |= MAP_REPLENISH; 958 } 959 } else 960 #endif 961 if (vm_map_is_system(map)) { 962 new_entry = uma_zalloc(kmapentzone, M_NOWAIT); 963 } else { 964 new_entry = uma_zalloc(mapentzone, M_WAITOK); 965 } 966 KASSERT(new_entry != NULL, 967 ("vm_map_entry_create: kernel resources exhausted")); 968 return (new_entry); 969 } 970 971 /* 972 * vm_map_entry_set_behavior: 973 * 974 * Set the expected access behavior, either normal, random, or 975 * sequential. 976 */ 977 static inline void 978 vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior) 979 { 980 entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) | 981 (behavior & MAP_ENTRY_BEHAV_MASK); 982 } 983 984 /* 985 * vm_map_entry_max_free_{left,right}: 986 * 987 * Compute the size of the largest free gap between two entries, 988 * one the root of a tree and the other the ancestor of that root 989 * that is the least or greatest ancestor found on the search path. 990 */ 991 static inline vm_size_t 992 vm_map_entry_max_free_left(vm_map_entry_t root, vm_map_entry_t left_ancestor) 993 { 994 995 return (root->left != left_ancestor ? 996 root->left->max_free : root->start - left_ancestor->end); 997 } 998 999 static inline vm_size_t 1000 vm_map_entry_max_free_right(vm_map_entry_t root, vm_map_entry_t right_ancestor) 1001 { 1002 1003 return (root->right != right_ancestor ? 1004 root->right->max_free : right_ancestor->start - root->end); 1005 } 1006 1007 /* 1008 * vm_map_entry_{pred,succ}: 1009 * 1010 * Find the {predecessor, successor} of the entry by taking one step 1011 * in the appropriate direction and backtracking as much as necessary. 1012 * vm_map_entry_succ is defined in vm_map.h. 1013 */ 1014 static inline vm_map_entry_t 1015 vm_map_entry_pred(vm_map_entry_t entry) 1016 { 1017 vm_map_entry_t prior; 1018 1019 prior = entry->left; 1020 if (prior->right->start < entry->start) { 1021 do 1022 prior = prior->right; 1023 while (prior->right != entry); 1024 } 1025 return (prior); 1026 } 1027 1028 static inline vm_size_t 1029 vm_size_max(vm_size_t a, vm_size_t b) 1030 { 1031 1032 return (a > b ? a : b); 1033 } 1034 1035 #define SPLAY_LEFT_STEP(root, y, llist, rlist, test) do { \ 1036 vm_map_entry_t z; \ 1037 vm_size_t max_free; \ 1038 \ 1039 /* \ 1040 * Infer root->right->max_free == root->max_free when \ 1041 * y->max_free < root->max_free || root->max_free == 0. \ 1042 * Otherwise, look right to find it. \ 1043 */ \ 1044 y = root->left; \ 1045 max_free = root->max_free; \ 1046 KASSERT(max_free == vm_size_max( \ 1047 vm_map_entry_max_free_left(root, llist), \ 1048 vm_map_entry_max_free_right(root, rlist)), \ 1049 ("%s: max_free invariant fails", __func__)); \ 1050 if (max_free - 1 < vm_map_entry_max_free_left(root, llist)) \ 1051 max_free = vm_map_entry_max_free_right(root, rlist); \ 1052 if (y != llist && (test)) { \ 1053 /* Rotate right and make y root. */ \ 1054 z = y->right; \ 1055 if (z != root) { \ 1056 root->left = z; \ 1057 y->right = root; \ 1058 if (max_free < y->max_free) \ 1059 root->max_free = max_free = \ 1060 vm_size_max(max_free, z->max_free); \ 1061 } else if (max_free < y->max_free) \ 1062 root->max_free = max_free = \ 1063 vm_size_max(max_free, root->start - y->end);\ 1064 root = y; \ 1065 y = root->left; \ 1066 } \ 1067 /* Copy right->max_free. Put root on rlist. */ \ 1068 root->max_free = max_free; \ 1069 KASSERT(max_free == vm_map_entry_max_free_right(root, rlist), \ 1070 ("%s: max_free not copied from right", __func__)); \ 1071 root->left = rlist; \ 1072 rlist = root; \ 1073 root = y != llist ? y : NULL; \ 1074 } while (0) 1075 1076 #define SPLAY_RIGHT_STEP(root, y, llist, rlist, test) do { \ 1077 vm_map_entry_t z; \ 1078 vm_size_t max_free; \ 1079 \ 1080 /* \ 1081 * Infer root->left->max_free == root->max_free when \ 1082 * y->max_free < root->max_free || root->max_free == 0. \ 1083 * Otherwise, look left to find it. \ 1084 */ \ 1085 y = root->right; \ 1086 max_free = root->max_free; \ 1087 KASSERT(max_free == vm_size_max( \ 1088 vm_map_entry_max_free_left(root, llist), \ 1089 vm_map_entry_max_free_right(root, rlist)), \ 1090 ("%s: max_free invariant fails", __func__)); \ 1091 if (max_free - 1 < vm_map_entry_max_free_right(root, rlist)) \ 1092 max_free = vm_map_entry_max_free_left(root, llist); \ 1093 if (y != rlist && (test)) { \ 1094 /* Rotate left and make y root. */ \ 1095 z = y->left; \ 1096 if (z != root) { \ 1097 root->right = z; \ 1098 y->left = root; \ 1099 if (max_free < y->max_free) \ 1100 root->max_free = max_free = \ 1101 vm_size_max(max_free, z->max_free); \ 1102 } else if (max_free < y->max_free) \ 1103 root->max_free = max_free = \ 1104 vm_size_max(max_free, y->start - root->end);\ 1105 root = y; \ 1106 y = root->right; \ 1107 } \ 1108 /* Copy left->max_free. Put root on llist. */ \ 1109 root->max_free = max_free; \ 1110 KASSERT(max_free == vm_map_entry_max_free_left(root, llist), \ 1111 ("%s: max_free not copied from left", __func__)); \ 1112 root->right = llist; \ 1113 llist = root; \ 1114 root = y != rlist ? y : NULL; \ 1115 } while (0) 1116 1117 /* 1118 * Walk down the tree until we find addr or a gap where addr would go, breaking 1119 * off left and right subtrees of nodes less than, or greater than addr. Treat 1120 * subtrees with root->max_free < length as empty trees. llist and rlist are 1121 * the two sides in reverse order (bottom-up), with llist linked by the right 1122 * pointer and rlist linked by the left pointer in the vm_map_entry, and both 1123 * lists terminated by &map->header. This function, and the subsequent call to 1124 * vm_map_splay_merge_{left,right,pred,succ}, rely on the start and end address 1125 * values in &map->header. 1126 */ 1127 static __always_inline vm_map_entry_t 1128 vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length, 1129 vm_map_entry_t *llist, vm_map_entry_t *rlist) 1130 { 1131 vm_map_entry_t left, right, root, y; 1132 1133 left = right = &map->header; 1134 root = map->root; 1135 while (root != NULL && root->max_free >= length) { 1136 KASSERT(left->end <= root->start && 1137 root->end <= right->start, 1138 ("%s: root not within tree bounds", __func__)); 1139 if (addr < root->start) { 1140 SPLAY_LEFT_STEP(root, y, left, right, 1141 y->max_free >= length && addr < y->start); 1142 } else if (addr >= root->end) { 1143 SPLAY_RIGHT_STEP(root, y, left, right, 1144 y->max_free >= length && addr >= y->end); 1145 } else 1146 break; 1147 } 1148 *llist = left; 1149 *rlist = right; 1150 return (root); 1151 } 1152 1153 static __always_inline void 1154 vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *rlist) 1155 { 1156 vm_map_entry_t hi, right, y; 1157 1158 right = *rlist; 1159 hi = root->right == right ? NULL : root->right; 1160 if (hi == NULL) 1161 return; 1162 do 1163 SPLAY_LEFT_STEP(hi, y, root, right, true); 1164 while (hi != NULL); 1165 *rlist = right; 1166 } 1167 1168 static __always_inline void 1169 vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *llist) 1170 { 1171 vm_map_entry_t left, lo, y; 1172 1173 left = *llist; 1174 lo = root->left == left ? NULL : root->left; 1175 if (lo == NULL) 1176 return; 1177 do 1178 SPLAY_RIGHT_STEP(lo, y, left, root, true); 1179 while (lo != NULL); 1180 *llist = left; 1181 } 1182 1183 static inline void 1184 vm_map_entry_swap(vm_map_entry_t *a, vm_map_entry_t *b) 1185 { 1186 vm_map_entry_t tmp; 1187 1188 tmp = *b; 1189 *b = *a; 1190 *a = tmp; 1191 } 1192 1193 /* 1194 * Walk back up the two spines, flip the pointers and set max_free. The 1195 * subtrees of the root go at the bottom of llist and rlist. 1196 */ 1197 static vm_size_t 1198 vm_map_splay_merge_left_walk(vm_map_entry_t header, vm_map_entry_t root, 1199 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t llist) 1200 { 1201 do { 1202 /* 1203 * The max_free values of the children of llist are in 1204 * llist->max_free and max_free. Update with the 1205 * max value. 1206 */ 1207 llist->max_free = max_free = 1208 vm_size_max(llist->max_free, max_free); 1209 vm_map_entry_swap(&llist->right, &tail); 1210 vm_map_entry_swap(&tail, &llist); 1211 } while (llist != header); 1212 root->left = tail; 1213 return (max_free); 1214 } 1215 1216 /* 1217 * When llist is known to be the predecessor of root. 1218 */ 1219 static inline vm_size_t 1220 vm_map_splay_merge_pred(vm_map_entry_t header, vm_map_entry_t root, 1221 vm_map_entry_t llist) 1222 { 1223 vm_size_t max_free; 1224 1225 max_free = root->start - llist->end; 1226 if (llist != header) { 1227 max_free = vm_map_splay_merge_left_walk(header, root, 1228 root, max_free, llist); 1229 } else { 1230 root->left = header; 1231 header->right = root; 1232 } 1233 return (max_free); 1234 } 1235 1236 /* 1237 * When llist may or may not be the predecessor of root. 1238 */ 1239 static inline vm_size_t 1240 vm_map_splay_merge_left(vm_map_entry_t header, vm_map_entry_t root, 1241 vm_map_entry_t llist) 1242 { 1243 vm_size_t max_free; 1244 1245 max_free = vm_map_entry_max_free_left(root, llist); 1246 if (llist != header) { 1247 max_free = vm_map_splay_merge_left_walk(header, root, 1248 root->left == llist ? root : root->left, 1249 max_free, llist); 1250 } 1251 return (max_free); 1252 } 1253 1254 static vm_size_t 1255 vm_map_splay_merge_right_walk(vm_map_entry_t header, vm_map_entry_t root, 1256 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t rlist) 1257 { 1258 do { 1259 /* 1260 * The max_free values of the children of rlist are in 1261 * rlist->max_free and max_free. Update with the 1262 * max value. 1263 */ 1264 rlist->max_free = max_free = 1265 vm_size_max(rlist->max_free, max_free); 1266 vm_map_entry_swap(&rlist->left, &tail); 1267 vm_map_entry_swap(&tail, &rlist); 1268 } while (rlist != header); 1269 root->right = tail; 1270 return (max_free); 1271 } 1272 1273 /* 1274 * When rlist is known to be the succecessor of root. 1275 */ 1276 static inline vm_size_t 1277 vm_map_splay_merge_succ(vm_map_entry_t header, vm_map_entry_t root, 1278 vm_map_entry_t rlist) 1279 { 1280 vm_size_t max_free; 1281 1282 max_free = rlist->start - root->end; 1283 if (rlist != header) { 1284 max_free = vm_map_splay_merge_right_walk(header, root, 1285 root, max_free, rlist); 1286 } else { 1287 root->right = header; 1288 header->left = root; 1289 } 1290 return (max_free); 1291 } 1292 1293 /* 1294 * When rlist may or may not be the succecessor of root. 1295 */ 1296 static inline vm_size_t 1297 vm_map_splay_merge_right(vm_map_entry_t header, vm_map_entry_t root, 1298 vm_map_entry_t rlist) 1299 { 1300 vm_size_t max_free; 1301 1302 max_free = vm_map_entry_max_free_right(root, rlist); 1303 if (rlist != header) { 1304 max_free = vm_map_splay_merge_right_walk(header, root, 1305 root->right == rlist ? root : root->right, 1306 max_free, rlist); 1307 } 1308 return (max_free); 1309 } 1310 1311 /* 1312 * vm_map_splay: 1313 * 1314 * The Sleator and Tarjan top-down splay algorithm with the 1315 * following variation. Max_free must be computed bottom-up, so 1316 * on the downward pass, maintain the left and right spines in 1317 * reverse order. Then, make a second pass up each side to fix 1318 * the pointers and compute max_free. The time bound is O(log n) 1319 * amortized. 1320 * 1321 * The tree is threaded, which means that there are no null pointers. 1322 * When a node has no left child, its left pointer points to its 1323 * predecessor, which the last ancestor on the search path from the root 1324 * where the search branched right. Likewise, when a node has no right 1325 * child, its right pointer points to its successor. The map header node 1326 * is the predecessor of the first map entry, and the successor of the 1327 * last. 1328 * 1329 * The new root is the vm_map_entry containing "addr", or else an 1330 * adjacent entry (lower if possible) if addr is not in the tree. 1331 * 1332 * The map must be locked, and leaves it so. 1333 * 1334 * Returns: the new root. 1335 */ 1336 static vm_map_entry_t 1337 vm_map_splay(vm_map_t map, vm_offset_t addr) 1338 { 1339 vm_map_entry_t header, llist, rlist, root; 1340 vm_size_t max_free_left, max_free_right; 1341 1342 header = &map->header; 1343 root = vm_map_splay_split(map, addr, 0, &llist, &rlist); 1344 if (root != NULL) { 1345 max_free_left = vm_map_splay_merge_left(header, root, llist); 1346 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1347 } else if (llist != header) { 1348 /* 1349 * Recover the greatest node in the left 1350 * subtree and make it the root. 1351 */ 1352 root = llist; 1353 llist = root->right; 1354 max_free_left = vm_map_splay_merge_left(header, root, llist); 1355 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1356 } else if (rlist != header) { 1357 /* 1358 * Recover the least node in the right 1359 * subtree and make it the root. 1360 */ 1361 root = rlist; 1362 rlist = root->left; 1363 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1364 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1365 } else { 1366 /* There is no root. */ 1367 return (NULL); 1368 } 1369 root->max_free = vm_size_max(max_free_left, max_free_right); 1370 map->root = root; 1371 VM_MAP_ASSERT_CONSISTENT(map); 1372 return (root); 1373 } 1374 1375 /* 1376 * vm_map_entry_{un,}link: 1377 * 1378 * Insert/remove entries from maps. On linking, if new entry clips 1379 * existing entry, trim existing entry to avoid overlap, and manage 1380 * offsets. On unlinking, merge disappearing entry with neighbor, if 1381 * called for, and manage offsets. Callers should not modify fields in 1382 * entries already mapped. 1383 */ 1384 static void 1385 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) 1386 { 1387 vm_map_entry_t header, llist, rlist, root; 1388 vm_size_t max_free_left, max_free_right; 1389 1390 CTR3(KTR_VM, 1391 "vm_map_entry_link: map %p, nentries %d, entry %p", map, 1392 map->nentries, entry); 1393 VM_MAP_ASSERT_LOCKED(map); 1394 map->nentries++; 1395 header = &map->header; 1396 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1397 if (root == NULL) { 1398 /* 1399 * The new entry does not overlap any existing entry in the 1400 * map, so it becomes the new root of the map tree. 1401 */ 1402 max_free_left = vm_map_splay_merge_pred(header, entry, llist); 1403 max_free_right = vm_map_splay_merge_succ(header, entry, rlist); 1404 } else if (entry->start == root->start) { 1405 /* 1406 * The new entry is a clone of root, with only the end field 1407 * changed. The root entry will be shrunk to abut the new 1408 * entry, and will be the right child of the new root entry in 1409 * the modified map. 1410 */ 1411 KASSERT(entry->end < root->end, 1412 ("%s: clip_start not within entry", __func__)); 1413 vm_map_splay_findprev(root, &llist); 1414 if ((root->eflags & MAP_ENTRY_STACK_GAP) == 0) 1415 root->offset += entry->end - root->start; 1416 root->start = entry->end; 1417 max_free_left = vm_map_splay_merge_pred(header, entry, llist); 1418 max_free_right = root->max_free = vm_size_max( 1419 vm_map_splay_merge_pred(entry, root, entry), 1420 vm_map_splay_merge_right(header, root, rlist)); 1421 } else { 1422 /* 1423 * The new entry is a clone of root, with only the start field 1424 * changed. The root entry will be shrunk to abut the new 1425 * entry, and will be the left child of the new root entry in 1426 * the modified map. 1427 */ 1428 KASSERT(entry->end == root->end, 1429 ("%s: clip_start not within entry", __func__)); 1430 vm_map_splay_findnext(root, &rlist); 1431 if ((entry->eflags & MAP_ENTRY_STACK_GAP) == 0) 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 MPASS((rlist->eflags & MAP_ENTRY_STACK_GAP) == 0); 1467 rlist->offset = root->offset; 1468 } 1469 if (llist != header) { 1470 root = llist; 1471 llist = root->right; 1472 max_free_left = vm_map_splay_merge_left(header, root, llist); 1473 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1474 } else if (rlist != header) { 1475 root = rlist; 1476 rlist = root->left; 1477 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1478 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1479 } else { 1480 header->left = header->right = header; 1481 root = NULL; 1482 } 1483 if (root != NULL) 1484 root->max_free = vm_size_max(max_free_left, max_free_right); 1485 map->root = root; 1486 VM_MAP_ASSERT_CONSISTENT(map); 1487 map->nentries--; 1488 CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map, 1489 map->nentries, entry); 1490 } 1491 1492 /* 1493 * vm_map_entry_resize: 1494 * 1495 * Resize a vm_map_entry, recompute the amount of free space that 1496 * follows it and propagate that value up the tree. 1497 * 1498 * The map must be locked, and leaves it so. 1499 */ 1500 static void 1501 vm_map_entry_resize(vm_map_t map, vm_map_entry_t entry, vm_size_t grow_amount) 1502 { 1503 vm_map_entry_t header, llist, rlist, root; 1504 1505 VM_MAP_ASSERT_LOCKED(map); 1506 header = &map->header; 1507 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1508 KASSERT(root != NULL, ("%s: resize object not mapped", __func__)); 1509 vm_map_splay_findnext(root, &rlist); 1510 entry->end += grow_amount; 1511 root->max_free = vm_size_max( 1512 vm_map_splay_merge_left(header, root, llist), 1513 vm_map_splay_merge_succ(header, root, rlist)); 1514 map->root = root; 1515 VM_MAP_ASSERT_CONSISTENT(map); 1516 CTR4(KTR_VM, "%s: map %p, nentries %d, entry %p", 1517 __func__, map, map->nentries, entry); 1518 } 1519 1520 /* 1521 * vm_map_lookup_entry: [ internal use only ] 1522 * 1523 * Finds the map entry containing (or 1524 * immediately preceding) the specified address 1525 * in the given map; the entry is returned 1526 * in the "entry" parameter. The boolean 1527 * result indicates whether the address is 1528 * actually contained in the map. 1529 */ 1530 boolean_t 1531 vm_map_lookup_entry( 1532 vm_map_t map, 1533 vm_offset_t address, 1534 vm_map_entry_t *entry) /* OUT */ 1535 { 1536 vm_map_entry_t cur, header, lbound, ubound; 1537 boolean_t locked; 1538 1539 /* 1540 * If the map is empty, then the map entry immediately preceding 1541 * "address" is the map's header. 1542 */ 1543 header = &map->header; 1544 cur = map->root; 1545 if (cur == NULL) { 1546 *entry = header; 1547 return (FALSE); 1548 } 1549 if (address >= cur->start && cur->end > address) { 1550 *entry = cur; 1551 return (TRUE); 1552 } 1553 if ((locked = vm_map_locked(map)) || 1554 sx_try_upgrade(&map->lock)) { 1555 /* 1556 * Splay requires a write lock on the map. However, it only 1557 * restructures the binary search tree; it does not otherwise 1558 * change the map. Thus, the map's timestamp need not change 1559 * on a temporary upgrade. 1560 */ 1561 cur = vm_map_splay(map, address); 1562 if (!locked) { 1563 VM_MAP_UNLOCK_CONSISTENT(map); 1564 sx_downgrade(&map->lock); 1565 } 1566 1567 /* 1568 * If "address" is contained within a map entry, the new root 1569 * is that map entry. Otherwise, the new root is a map entry 1570 * immediately before or after "address". 1571 */ 1572 if (address < cur->start) { 1573 *entry = header; 1574 return (FALSE); 1575 } 1576 *entry = cur; 1577 return (address < cur->end); 1578 } 1579 /* 1580 * Since the map is only locked for read access, perform a 1581 * standard binary search tree lookup for "address". 1582 */ 1583 lbound = ubound = header; 1584 for (;;) { 1585 if (address < cur->start) { 1586 ubound = cur; 1587 cur = cur->left; 1588 if (cur == lbound) 1589 break; 1590 } else if (cur->end <= address) { 1591 lbound = cur; 1592 cur = cur->right; 1593 if (cur == ubound) 1594 break; 1595 } else { 1596 *entry = cur; 1597 return (TRUE); 1598 } 1599 } 1600 *entry = lbound; 1601 return (FALSE); 1602 } 1603 1604 /* 1605 * vm_map_insert1() is identical to vm_map_insert() except that it 1606 * returns the newly inserted map entry in '*res'. In case the new 1607 * entry is coalesced with a neighbor or an existing entry was 1608 * resized, that entry is returned. In any case, the returned entry 1609 * covers the specified address range. 1610 */ 1611 static int 1612 vm_map_insert1(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1613 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow, 1614 vm_map_entry_t *res) 1615 { 1616 vm_map_entry_t new_entry, next_entry, prev_entry; 1617 struct ucred *cred; 1618 vm_eflags_t protoeflags; 1619 vm_inherit_t inheritance; 1620 u_long bdry; 1621 u_int bidx; 1622 1623 VM_MAP_ASSERT_LOCKED(map); 1624 KASSERT(object != kernel_object || 1625 (cow & MAP_COPY_ON_WRITE) == 0, 1626 ("vm_map_insert: kernel object and COW")); 1627 KASSERT(object == NULL || (cow & MAP_NOFAULT) == 0 || 1628 (cow & MAP_SPLIT_BOUNDARY_MASK) != 0, 1629 ("vm_map_insert: paradoxical MAP_NOFAULT request, obj %p cow %#x", 1630 object, cow)); 1631 KASSERT((prot & ~max) == 0, 1632 ("prot %#x is not subset of max_prot %#x", prot, max)); 1633 1634 /* 1635 * Check that the start and end points are not bogus. 1636 */ 1637 if (start == end || !vm_map_range_valid(map, start, end)) 1638 return (KERN_INVALID_ADDRESS); 1639 1640 if ((map->flags & MAP_WXORX) != 0 && (prot & (VM_PROT_WRITE | 1641 VM_PROT_EXECUTE)) == (VM_PROT_WRITE | VM_PROT_EXECUTE)) 1642 return (KERN_PROTECTION_FAILURE); 1643 1644 /* 1645 * Find the entry prior to the proposed starting address; if it's part 1646 * of an existing entry, this range is bogus. 1647 */ 1648 if (vm_map_lookup_entry(map, start, &prev_entry)) 1649 return (KERN_NO_SPACE); 1650 1651 /* 1652 * Assert that the next entry doesn't overlap the end point. 1653 */ 1654 next_entry = vm_map_entry_succ(prev_entry); 1655 if (next_entry->start < end) 1656 return (KERN_NO_SPACE); 1657 1658 if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL || 1659 max != VM_PROT_NONE)) 1660 return (KERN_INVALID_ARGUMENT); 1661 1662 protoeflags = 0; 1663 if (cow & MAP_COPY_ON_WRITE) 1664 protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; 1665 if (cow & MAP_NOFAULT) 1666 protoeflags |= MAP_ENTRY_NOFAULT; 1667 if (cow & MAP_DISABLE_SYNCER) 1668 protoeflags |= MAP_ENTRY_NOSYNC; 1669 if (cow & MAP_DISABLE_COREDUMP) 1670 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1671 if (cow & MAP_STACK_AREA) 1672 protoeflags |= MAP_ENTRY_GROWS_DOWN; 1673 if (cow & MAP_WRITECOUNT) 1674 protoeflags |= MAP_ENTRY_WRITECNT; 1675 if (cow & MAP_VN_EXEC) 1676 protoeflags |= MAP_ENTRY_VN_EXEC; 1677 if ((cow & MAP_CREATE_GUARD) != 0) 1678 protoeflags |= MAP_ENTRY_GUARD; 1679 if ((cow & MAP_CREATE_STACK_GAP) != 0) 1680 protoeflags |= MAP_ENTRY_STACK_GAP; 1681 if (cow & MAP_INHERIT_SHARE) 1682 inheritance = VM_INHERIT_SHARE; 1683 else 1684 inheritance = VM_INHERIT_DEFAULT; 1685 if ((cow & MAP_SPLIT_BOUNDARY_MASK) != 0) { 1686 /* This magically ignores index 0, for usual page size. */ 1687 bidx = (cow & MAP_SPLIT_BOUNDARY_MASK) >> 1688 MAP_SPLIT_BOUNDARY_SHIFT; 1689 if (bidx >= MAXPAGESIZES) 1690 return (KERN_INVALID_ARGUMENT); 1691 bdry = pagesizes[bidx] - 1; 1692 if ((start & bdry) != 0 || (end & bdry) != 0) 1693 return (KERN_INVALID_ARGUMENT); 1694 protoeflags |= bidx << MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 1695 } 1696 1697 cred = NULL; 1698 if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0) 1699 goto charged; 1700 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) && 1701 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) { 1702 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start)) 1703 return (KERN_RESOURCE_SHORTAGE); 1704 KASSERT(object == NULL || 1705 (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 || 1706 object->cred == NULL, 1707 ("overcommit: vm_map_insert o %p", object)); 1708 cred = curthread->td_ucred; 1709 } 1710 1711 charged: 1712 /* Expand the kernel pmap, if necessary. */ 1713 if (map == kernel_map && end > kernel_vm_end) 1714 pmap_growkernel(end); 1715 if (object != NULL) { 1716 /* 1717 * OBJ_ONEMAPPING must be cleared unless this mapping 1718 * is trivially proven to be the only mapping for any 1719 * of the object's pages. (Object granularity 1720 * reference counting is insufficient to recognize 1721 * aliases with precision.) 1722 */ 1723 if ((object->flags & OBJ_ANON) != 0) { 1724 VM_OBJECT_WLOCK(object); 1725 if (object->ref_count > 1 || object->shadow_count != 0) 1726 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1727 VM_OBJECT_WUNLOCK(object); 1728 } 1729 } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) == 1730 protoeflags && 1731 (cow & (MAP_STACK_AREA | MAP_VN_EXEC)) == 0 && 1732 prev_entry->end == start && (prev_entry->cred == cred || 1733 (prev_entry->object.vm_object != NULL && 1734 prev_entry->object.vm_object->cred == cred)) && 1735 vm_object_coalesce(prev_entry->object.vm_object, 1736 prev_entry->offset, 1737 (vm_size_t)(prev_entry->end - prev_entry->start), 1738 (vm_size_t)(end - prev_entry->end), cred != NULL && 1739 (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) { 1740 /* 1741 * We were able to extend the object. Determine if we 1742 * can extend the previous map entry to include the 1743 * new range as well. 1744 */ 1745 if (prev_entry->inheritance == inheritance && 1746 prev_entry->protection == prot && 1747 prev_entry->max_protection == max && 1748 prev_entry->wired_count == 0) { 1749 KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) == 1750 0, ("prev_entry %p has incoherent wiring", 1751 prev_entry)); 1752 if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0) 1753 map->size += end - prev_entry->end; 1754 vm_map_entry_resize(map, prev_entry, 1755 end - prev_entry->end); 1756 *res = vm_map_try_merge_entries(map, prev_entry, 1757 next_entry); 1758 return (KERN_SUCCESS); 1759 } 1760 1761 /* 1762 * If we can extend the object but cannot extend the 1763 * map entry, we have to create a new map entry. We 1764 * must bump the ref count on the extended object to 1765 * account for it. object may be NULL. 1766 */ 1767 object = prev_entry->object.vm_object; 1768 offset = prev_entry->offset + 1769 (prev_entry->end - prev_entry->start); 1770 vm_object_reference(object); 1771 if (cred != NULL && object != NULL && object->cred != NULL && 1772 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 1773 /* Object already accounts for this uid. */ 1774 cred = NULL; 1775 } 1776 } 1777 if (cred != NULL) 1778 crhold(cred); 1779 1780 /* 1781 * Create a new entry 1782 */ 1783 new_entry = vm_map_entry_create(map); 1784 new_entry->start = start; 1785 new_entry->end = end; 1786 new_entry->cred = NULL; 1787 1788 new_entry->eflags = protoeflags; 1789 new_entry->object.vm_object = object; 1790 new_entry->offset = offset; 1791 1792 new_entry->inheritance = inheritance; 1793 new_entry->protection = prot; 1794 new_entry->max_protection = max; 1795 new_entry->wired_count = 0; 1796 new_entry->wiring_thread = NULL; 1797 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT; 1798 new_entry->next_read = start; 1799 1800 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry), 1801 ("overcommit: vm_map_insert leaks vm_map %p", new_entry)); 1802 new_entry->cred = cred; 1803 1804 /* 1805 * Insert the new entry into the list 1806 */ 1807 vm_map_entry_link(map, new_entry); 1808 if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0) 1809 map->size += new_entry->end - new_entry->start; 1810 1811 /* 1812 * Try to coalesce the new entry with both the previous and next 1813 * entries in the list. Previously, we only attempted to coalesce 1814 * with the previous entry when object is NULL. Here, we handle the 1815 * other cases, which are less common. 1816 */ 1817 vm_map_try_merge_entries(map, prev_entry, new_entry); 1818 *res = vm_map_try_merge_entries(map, new_entry, next_entry); 1819 1820 if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) { 1821 vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset), 1822 end - start, cow & MAP_PREFAULT_PARTIAL); 1823 } 1824 1825 return (KERN_SUCCESS); 1826 } 1827 1828 /* 1829 * vm_map_insert: 1830 * 1831 * Inserts the given VM object into the target map at the 1832 * specified address range. 1833 * 1834 * Requires that the map be locked, and leaves it so. 1835 * 1836 * If object is non-NULL, ref count must be bumped by caller 1837 * prior to making call to account for the new entry. 1838 */ 1839 int 1840 vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1841 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow) 1842 { 1843 vm_map_entry_t res; 1844 1845 return (vm_map_insert1(map, object, offset, start, end, prot, max, 1846 cow, &res)); 1847 } 1848 1849 /* 1850 * vm_map_findspace: 1851 * 1852 * Find the first fit (lowest VM address) for "length" free bytes 1853 * beginning at address >= start in the given map. 1854 * 1855 * In a vm_map_entry, "max_free" is the maximum amount of 1856 * contiguous free space between an entry in its subtree and a 1857 * neighbor of that entry. This allows finding a free region in 1858 * one path down the tree, so O(log n) amortized with splay 1859 * trees. 1860 * 1861 * The map must be locked, and leaves it so. 1862 * 1863 * Returns: starting address if sufficient space, 1864 * vm_map_max(map)-length+1 if insufficient space. 1865 */ 1866 vm_offset_t 1867 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length) 1868 { 1869 vm_map_entry_t header, llist, rlist, root, y; 1870 vm_size_t left_length, max_free_left, max_free_right; 1871 vm_offset_t gap_end; 1872 1873 VM_MAP_ASSERT_LOCKED(map); 1874 1875 /* 1876 * Request must fit within min/max VM address and must avoid 1877 * address wrap. 1878 */ 1879 start = MAX(start, vm_map_min(map)); 1880 if (start >= vm_map_max(map) || length > vm_map_max(map) - start) 1881 return (vm_map_max(map) - length + 1); 1882 1883 /* Empty tree means wide open address space. */ 1884 if (map->root == NULL) 1885 return (start); 1886 1887 /* 1888 * After splay_split, if start is within an entry, push it to the start 1889 * of the following gap. If rlist is at the end of the gap containing 1890 * start, save the end of that gap in gap_end to see if the gap is big 1891 * enough; otherwise set gap_end to start skip gap-checking and move 1892 * directly to a search of the right subtree. 1893 */ 1894 header = &map->header; 1895 root = vm_map_splay_split(map, start, length, &llist, &rlist); 1896 gap_end = rlist->start; 1897 if (root != NULL) { 1898 start = root->end; 1899 if (root->right != rlist) 1900 gap_end = start; 1901 max_free_left = vm_map_splay_merge_left(header, root, llist); 1902 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1903 } else if (rlist != header) { 1904 root = rlist; 1905 rlist = root->left; 1906 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1907 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1908 } else { 1909 root = llist; 1910 llist = root->right; 1911 max_free_left = vm_map_splay_merge_left(header, root, llist); 1912 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1913 } 1914 root->max_free = vm_size_max(max_free_left, max_free_right); 1915 map->root = root; 1916 VM_MAP_ASSERT_CONSISTENT(map); 1917 if (length <= gap_end - start) 1918 return (start); 1919 1920 /* With max_free, can immediately tell if no solution. */ 1921 if (root->right == header || length > root->right->max_free) 1922 return (vm_map_max(map) - length + 1); 1923 1924 /* 1925 * Splay for the least large-enough gap in the right subtree. 1926 */ 1927 llist = rlist = header; 1928 for (left_length = 0;; 1929 left_length = vm_map_entry_max_free_left(root, llist)) { 1930 if (length <= left_length) 1931 SPLAY_LEFT_STEP(root, y, llist, rlist, 1932 length <= vm_map_entry_max_free_left(y, llist)); 1933 else 1934 SPLAY_RIGHT_STEP(root, y, llist, rlist, 1935 length > vm_map_entry_max_free_left(y, root)); 1936 if (root == NULL) 1937 break; 1938 } 1939 root = llist; 1940 llist = root->right; 1941 max_free_left = vm_map_splay_merge_left(header, root, llist); 1942 if (rlist == header) { 1943 root->max_free = vm_size_max(max_free_left, 1944 vm_map_splay_merge_succ(header, root, rlist)); 1945 } else { 1946 y = rlist; 1947 rlist = y->left; 1948 y->max_free = vm_size_max( 1949 vm_map_splay_merge_pred(root, y, root), 1950 vm_map_splay_merge_right(header, y, rlist)); 1951 root->max_free = vm_size_max(max_free_left, y->max_free); 1952 } 1953 map->root = root; 1954 VM_MAP_ASSERT_CONSISTENT(map); 1955 return (root->end); 1956 } 1957 1958 int 1959 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1960 vm_offset_t start, vm_size_t length, vm_prot_t prot, 1961 vm_prot_t max, int cow) 1962 { 1963 vm_offset_t end; 1964 int result; 1965 1966 end = start + length; 1967 KASSERT((cow & MAP_STACK_AREA) == 0 || object == NULL, 1968 ("vm_map_fixed: non-NULL backing object for stack")); 1969 vm_map_lock(map); 1970 VM_MAP_RANGE_CHECK(map, start, end); 1971 if ((cow & MAP_CHECK_EXCL) == 0) { 1972 result = vm_map_delete(map, start, end); 1973 if (result != KERN_SUCCESS) 1974 goto out; 1975 } 1976 if ((cow & MAP_STACK_AREA) != 0) { 1977 result = vm_map_stack_locked(map, start, length, sgrowsiz, 1978 prot, max, cow); 1979 } else { 1980 result = vm_map_insert(map, object, offset, start, end, 1981 prot, max, cow); 1982 } 1983 out: 1984 vm_map_unlock(map); 1985 return (result); 1986 } 1987 1988 #if VM_NRESERVLEVEL <= 1 1989 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10}; 1990 static const int aslr_pages_rnd_32[2] = {0x100, 0x4}; 1991 #elif VM_NRESERVLEVEL == 2 1992 static const int aslr_pages_rnd_64[3] = {0x1000, 0x1000, 0x10}; 1993 static const int aslr_pages_rnd_32[3] = {0x100, 0x100, 0x4}; 1994 #else 1995 #error "Unsupported VM_NRESERVLEVEL" 1996 #endif 1997 1998 static int cluster_anon = 1; 1999 SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW, 2000 &cluster_anon, 0, 2001 "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always"); 2002 2003 static bool 2004 clustering_anon_allowed(vm_offset_t addr, int cow) 2005 { 2006 2007 switch (cluster_anon) { 2008 case 0: 2009 return (false); 2010 case 1: 2011 return (addr == 0 || (cow & MAP_NO_HINT) != 0); 2012 case 2: 2013 default: 2014 return (true); 2015 } 2016 } 2017 2018 static long aslr_restarts; 2019 SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD, 2020 &aslr_restarts, 0, 2021 "Number of aslr failures"); 2022 2023 /* 2024 * Searches for the specified amount of free space in the given map with the 2025 * specified alignment. Performs an address-ordered, first-fit search from 2026 * the given address "*addr", with an optional upper bound "max_addr". If the 2027 * parameter "alignment" is zero, then the alignment is computed from the 2028 * given (object, offset) pair so as to enable the greatest possible use of 2029 * superpage mappings. Returns KERN_SUCCESS and the address of the free space 2030 * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE. 2031 * 2032 * The map must be locked. Initially, there must be at least "length" bytes 2033 * of free space at the given address. 2034 */ 2035 static int 2036 vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2037 vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, 2038 vm_offset_t alignment) 2039 { 2040 vm_offset_t aligned_addr, free_addr; 2041 2042 VM_MAP_ASSERT_LOCKED(map); 2043 free_addr = *addr; 2044 KASSERT(free_addr == vm_map_findspace(map, free_addr, length), 2045 ("caller failed to provide space %#jx at address %p", 2046 (uintmax_t)length, (void *)free_addr)); 2047 for (;;) { 2048 /* 2049 * At the start of every iteration, the free space at address 2050 * "*addr" is at least "length" bytes. 2051 */ 2052 if (alignment == 0) 2053 pmap_align_superpage(object, offset, addr, length); 2054 else 2055 *addr = roundup2(*addr, alignment); 2056 aligned_addr = *addr; 2057 if (aligned_addr == free_addr) { 2058 /* 2059 * Alignment did not change "*addr", so "*addr" must 2060 * still provide sufficient free space. 2061 */ 2062 return (KERN_SUCCESS); 2063 } 2064 2065 /* 2066 * Test for address wrap on "*addr". A wrapped "*addr" could 2067 * be a valid address, in which case vm_map_findspace() cannot 2068 * be relied upon to fail. 2069 */ 2070 if (aligned_addr < free_addr) 2071 return (KERN_NO_SPACE); 2072 *addr = vm_map_findspace(map, aligned_addr, length); 2073 if (*addr + length > vm_map_max(map) || 2074 (max_addr != 0 && *addr + length > max_addr)) 2075 return (KERN_NO_SPACE); 2076 free_addr = *addr; 2077 if (free_addr == aligned_addr) { 2078 /* 2079 * If a successful call to vm_map_findspace() did not 2080 * change "*addr", then "*addr" must still be aligned 2081 * and provide sufficient free space. 2082 */ 2083 return (KERN_SUCCESS); 2084 } 2085 } 2086 } 2087 2088 int 2089 vm_map_find_aligned(vm_map_t map, vm_offset_t *addr, vm_size_t length, 2090 vm_offset_t max_addr, vm_offset_t alignment) 2091 { 2092 /* XXXKIB ASLR eh ? */ 2093 *addr = vm_map_findspace(map, *addr, length); 2094 if (*addr + length > vm_map_max(map) || 2095 (max_addr != 0 && *addr + length > max_addr)) 2096 return (KERN_NO_SPACE); 2097 return (vm_map_alignspace(map, NULL, 0, addr, length, max_addr, 2098 alignment)); 2099 } 2100 2101 /* 2102 * vm_map_find finds an unallocated region in the target address 2103 * map with the given length. The search is defined to be 2104 * first-fit from the specified address; the region found is 2105 * returned in the same parameter. 2106 * 2107 * If object is non-NULL, ref count must be bumped by caller 2108 * prior to making call to account for the new entry. 2109 */ 2110 int 2111 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2112 vm_offset_t *addr, /* IN/OUT */ 2113 vm_size_t length, vm_offset_t max_addr, int find_space, 2114 vm_prot_t prot, vm_prot_t max, int cow) 2115 { 2116 int rv; 2117 2118 vm_map_lock(map); 2119 rv = vm_map_find_locked(map, object, offset, addr, length, max_addr, 2120 find_space, prot, max, cow); 2121 vm_map_unlock(map); 2122 return (rv); 2123 } 2124 2125 int 2126 vm_map_find_locked(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2127 vm_offset_t *addr, /* IN/OUT */ 2128 vm_size_t length, vm_offset_t max_addr, int find_space, 2129 vm_prot_t prot, vm_prot_t max, int cow) 2130 { 2131 vm_offset_t alignment, curr_min_addr, min_addr; 2132 int gap, pidx, rv, try; 2133 bool cluster, en_aslr, update_anon; 2134 2135 KASSERT((cow & MAP_STACK_AREA) == 0 || object == NULL, 2136 ("non-NULL backing object for stack")); 2137 MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE && 2138 (cow & MAP_STACK_AREA) == 0)); 2139 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL || 2140 (object->flags & OBJ_COLORED) == 0)) 2141 find_space = VMFS_ANY_SPACE; 2142 if (find_space >> 8 != 0) { 2143 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags")); 2144 alignment = (vm_offset_t)1 << (find_space >> 8); 2145 } else 2146 alignment = 0; 2147 en_aslr = (map->flags & MAP_ASLR) != 0; 2148 update_anon = cluster = clustering_anon_allowed(*addr, cow) && 2149 (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 && 2150 find_space != VMFS_NO_SPACE && object == NULL && 2151 (cow & (MAP_INHERIT_SHARE | MAP_STACK_AREA)) == 0 && 2152 prot != PROT_NONE; 2153 curr_min_addr = min_addr = *addr; 2154 if (en_aslr && min_addr == 0 && !cluster && 2155 find_space != VMFS_NO_SPACE && 2156 (map->flags & MAP_ASLR_IGNSTART) != 0) 2157 curr_min_addr = min_addr = vm_map_min(map); 2158 try = 0; 2159 if (cluster) { 2160 curr_min_addr = map->anon_loc; 2161 if (curr_min_addr == 0) 2162 cluster = false; 2163 } 2164 if (find_space != VMFS_NO_SPACE) { 2165 KASSERT(find_space == VMFS_ANY_SPACE || 2166 find_space == VMFS_OPTIMAL_SPACE || 2167 find_space == VMFS_SUPER_SPACE || 2168 alignment != 0, ("unexpected VMFS flag")); 2169 again: 2170 /* 2171 * When creating an anonymous mapping, try clustering 2172 * with an existing anonymous mapping first. 2173 * 2174 * We make up to two attempts to find address space 2175 * for a given find_space value. The first attempt may 2176 * apply randomization or may cluster with an existing 2177 * anonymous mapping. If this first attempt fails, 2178 * perform a first-fit search of the available address 2179 * space. 2180 * 2181 * If all tries failed, and find_space is 2182 * VMFS_OPTIMAL_SPACE, fallback to VMFS_ANY_SPACE. 2183 * Again enable clustering and randomization. 2184 */ 2185 try++; 2186 MPASS(try <= 2); 2187 2188 if (try == 2) { 2189 /* 2190 * Second try: we failed either to find a 2191 * suitable region for randomizing the 2192 * allocation, or to cluster with an existing 2193 * mapping. Retry with free run. 2194 */ 2195 curr_min_addr = (map->flags & MAP_ASLR_IGNSTART) != 0 ? 2196 vm_map_min(map) : min_addr; 2197 atomic_add_long(&aslr_restarts, 1); 2198 } 2199 2200 if (try == 1 && en_aslr && !cluster) { 2201 /* 2202 * Find space for allocation, including 2203 * gap needed for later randomization. 2204 */ 2205 pidx = 0; 2206 #if VM_NRESERVLEVEL > 0 2207 if ((find_space == VMFS_SUPER_SPACE || 2208 find_space == VMFS_OPTIMAL_SPACE) && 2209 pagesizes[VM_NRESERVLEVEL] != 0) { 2210 /* 2211 * Do not pointlessly increase the space that 2212 * is requested from vm_map_findspace(). 2213 * pmap_align_superpage() will only change a 2214 * mapping's alignment if that mapping is at 2215 * least a superpage in size. 2216 */ 2217 pidx = VM_NRESERVLEVEL; 2218 while (pidx > 0 && length < pagesizes[pidx]) 2219 pidx--; 2220 } 2221 #endif 2222 gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR && 2223 (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ? 2224 aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx]; 2225 *addr = vm_map_findspace(map, curr_min_addr, 2226 length + gap * pagesizes[pidx]); 2227 if (*addr + length + gap * pagesizes[pidx] > 2228 vm_map_max(map)) 2229 goto again; 2230 /* And randomize the start address. */ 2231 *addr += (arc4random() % gap) * pagesizes[pidx]; 2232 if (max_addr != 0 && *addr + length > max_addr) 2233 goto again; 2234 } else { 2235 *addr = vm_map_findspace(map, curr_min_addr, length); 2236 if (*addr + length > vm_map_max(map) || 2237 (max_addr != 0 && *addr + length > max_addr)) { 2238 if (cluster) { 2239 cluster = false; 2240 MPASS(try == 1); 2241 goto again; 2242 } 2243 return (KERN_NO_SPACE); 2244 } 2245 } 2246 2247 if (find_space != VMFS_ANY_SPACE && 2248 (rv = vm_map_alignspace(map, object, offset, addr, length, 2249 max_addr, alignment)) != KERN_SUCCESS) { 2250 if (find_space == VMFS_OPTIMAL_SPACE) { 2251 find_space = VMFS_ANY_SPACE; 2252 curr_min_addr = min_addr; 2253 cluster = update_anon; 2254 try = 0; 2255 goto again; 2256 } 2257 return (rv); 2258 } 2259 } else if ((cow & MAP_REMAP) != 0) { 2260 if (!vm_map_range_valid(map, *addr, *addr + length)) 2261 return (KERN_INVALID_ADDRESS); 2262 rv = vm_map_delete(map, *addr, *addr + length); 2263 if (rv != KERN_SUCCESS) 2264 return (rv); 2265 } 2266 if ((cow & MAP_STACK_AREA) != 0) { 2267 rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot, 2268 max, cow); 2269 } else { 2270 rv = vm_map_insert(map, object, offset, *addr, *addr + length, 2271 prot, max, cow); 2272 } 2273 2274 /* 2275 * Update the starting address for clustered anonymous memory mappings 2276 * if a starting address was not previously defined or an ASLR restart 2277 * placed an anonymous memory mapping at a lower address. 2278 */ 2279 if (update_anon && rv == KERN_SUCCESS && (map->anon_loc == 0 || 2280 *addr < map->anon_loc)) 2281 map->anon_loc = *addr; 2282 return (rv); 2283 } 2284 2285 /* 2286 * vm_map_find_min() is a variant of vm_map_find() that takes an 2287 * additional parameter ("default_addr") and treats the given address 2288 * ("*addr") differently. Specifically, it treats "*addr" as a hint 2289 * and not as the minimum address where the mapping is created. 2290 * 2291 * This function works in two phases. First, it tries to 2292 * allocate above the hint. If that fails and the hint is 2293 * greater than "default_addr", it performs a second pass, replacing 2294 * the hint with "default_addr" as the minimum address for the 2295 * allocation. 2296 */ 2297 int 2298 vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2299 vm_offset_t *addr, vm_size_t length, vm_offset_t default_addr, 2300 vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max, 2301 int cow) 2302 { 2303 vm_offset_t hint; 2304 int rv; 2305 2306 hint = *addr; 2307 if (hint == 0) { 2308 cow |= MAP_NO_HINT; 2309 *addr = hint = default_addr; 2310 } 2311 for (;;) { 2312 rv = vm_map_find(map, object, offset, addr, length, max_addr, 2313 find_space, prot, max, cow); 2314 if (rv == KERN_SUCCESS || default_addr >= hint) 2315 return (rv); 2316 *addr = hint = default_addr; 2317 } 2318 } 2319 2320 /* 2321 * A map entry with any of the following flags set must not be merged with 2322 * another entry. 2323 */ 2324 #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | \ 2325 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC | \ 2326 MAP_ENTRY_STACK_GAP) 2327 2328 static bool 2329 vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry) 2330 { 2331 2332 KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 || 2333 (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0, 2334 ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable", 2335 prev, entry)); 2336 return (prev->end == entry->start && 2337 prev->object.vm_object == entry->object.vm_object && 2338 (prev->object.vm_object == NULL || 2339 prev->offset + (prev->end - prev->start) == entry->offset) && 2340 prev->eflags == entry->eflags && 2341 prev->protection == entry->protection && 2342 prev->max_protection == entry->max_protection && 2343 prev->inheritance == entry->inheritance && 2344 prev->wired_count == entry->wired_count && 2345 prev->cred == entry->cred); 2346 } 2347 2348 static void 2349 vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry) 2350 { 2351 2352 /* 2353 * If the backing object is a vnode object, vm_object_deallocate() 2354 * calls vrele(). However, vrele() does not lock the vnode because 2355 * the vnode has additional references. Thus, the map lock can be 2356 * kept without causing a lock-order reversal with the vnode lock. 2357 * 2358 * Since we count the number of virtual page mappings in 2359 * object->un_pager.vnp.writemappings, the writemappings value 2360 * should not be adjusted when the entry is disposed of. 2361 */ 2362 if (entry->object.vm_object != NULL) 2363 vm_object_deallocate(entry->object.vm_object); 2364 if (entry->cred != NULL) 2365 crfree(entry->cred); 2366 vm_map_entry_dispose(map, entry); 2367 } 2368 2369 /* 2370 * vm_map_try_merge_entries: 2371 * 2372 * Compare two map entries that represent consecutive ranges. If 2373 * the entries can be merged, expand the range of the second to 2374 * cover the range of the first and delete the first. Then return 2375 * the map entry that includes the first range. 2376 * 2377 * The map must be locked. 2378 */ 2379 vm_map_entry_t 2380 vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev_entry, 2381 vm_map_entry_t entry) 2382 { 2383 2384 VM_MAP_ASSERT_LOCKED(map); 2385 if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 && 2386 vm_map_mergeable_neighbors(prev_entry, entry)) { 2387 vm_map_entry_unlink(map, prev_entry, UNLINK_MERGE_NEXT); 2388 vm_map_merged_neighbor_dispose(map, prev_entry); 2389 return (entry); 2390 } 2391 return (prev_entry); 2392 } 2393 2394 /* 2395 * vm_map_entry_back: 2396 * 2397 * Allocate an object to back a map entry. 2398 */ 2399 static inline void 2400 vm_map_entry_back(vm_map_entry_t entry) 2401 { 2402 vm_object_t object; 2403 2404 KASSERT(entry->object.vm_object == NULL, 2405 ("map entry %p has backing object", entry)); 2406 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2407 ("map entry %p is a submap", entry)); 2408 object = vm_object_allocate_anon(atop(entry->end - entry->start), NULL, 2409 entry->cred, entry->end - entry->start); 2410 entry->object.vm_object = object; 2411 entry->offset = 0; 2412 entry->cred = NULL; 2413 } 2414 2415 /* 2416 * vm_map_entry_charge_object 2417 * 2418 * If there is no object backing this entry, create one. Otherwise, if 2419 * the entry has cred, give it to the backing object. 2420 */ 2421 static inline void 2422 vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry) 2423 { 2424 2425 VM_MAP_ASSERT_LOCKED(map); 2426 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2427 ("map entry %p is a submap", entry)); 2428 if (entry->object.vm_object == NULL && !vm_map_is_system(map) && 2429 (entry->eflags & MAP_ENTRY_GUARD) == 0) 2430 vm_map_entry_back(entry); 2431 else if (entry->object.vm_object != NULL && 2432 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) && 2433 entry->cred != NULL) { 2434 VM_OBJECT_WLOCK(entry->object.vm_object); 2435 KASSERT(entry->object.vm_object->cred == NULL, 2436 ("OVERCOMMIT: %s: both cred e %p", __func__, entry)); 2437 entry->object.vm_object->cred = entry->cred; 2438 entry->object.vm_object->charge = entry->end - entry->start; 2439 VM_OBJECT_WUNLOCK(entry->object.vm_object); 2440 entry->cred = NULL; 2441 } 2442 } 2443 2444 /* 2445 * vm_map_entry_clone 2446 * 2447 * Create a duplicate map entry for clipping. 2448 */ 2449 static vm_map_entry_t 2450 vm_map_entry_clone(vm_map_t map, vm_map_entry_t entry) 2451 { 2452 vm_map_entry_t new_entry; 2453 2454 VM_MAP_ASSERT_LOCKED(map); 2455 2456 /* 2457 * Create a backing object now, if none exists, so that more individual 2458 * objects won't be created after the map entry is split. 2459 */ 2460 vm_map_entry_charge_object(map, entry); 2461 2462 /* Clone the entry. */ 2463 new_entry = vm_map_entry_create(map); 2464 *new_entry = *entry; 2465 if (new_entry->cred != NULL) 2466 crhold(entry->cred); 2467 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2468 vm_object_reference(new_entry->object.vm_object); 2469 vm_map_entry_set_vnode_text(new_entry, true); 2470 /* 2471 * The object->un_pager.vnp.writemappings for the object of 2472 * MAP_ENTRY_WRITECNT type entry shall be kept as is here. The 2473 * virtual pages are re-distributed among the clipped entries, 2474 * so the sum is left the same. 2475 */ 2476 } 2477 return (new_entry); 2478 } 2479 2480 /* 2481 * vm_map_clip_start: [ internal use only ] 2482 * 2483 * Asserts that the given entry begins at or after 2484 * the specified address; if necessary, 2485 * it splits the entry into two. 2486 */ 2487 static int 2488 vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t startaddr) 2489 { 2490 vm_map_entry_t new_entry; 2491 int bdry_idx; 2492 2493 if (!vm_map_is_system(map)) 2494 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2495 "%s: map %p entry %p start 0x%jx", __func__, map, entry, 2496 (uintmax_t)startaddr); 2497 2498 if (startaddr <= entry->start) 2499 return (KERN_SUCCESS); 2500 2501 VM_MAP_ASSERT_LOCKED(map); 2502 KASSERT(entry->end > startaddr && entry->start < startaddr, 2503 ("%s: invalid clip of entry %p", __func__, entry)); 2504 2505 bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry); 2506 if (bdry_idx != 0) { 2507 if ((startaddr & (pagesizes[bdry_idx] - 1)) != 0) 2508 return (KERN_INVALID_ARGUMENT); 2509 } 2510 2511 new_entry = vm_map_entry_clone(map, entry); 2512 2513 /* 2514 * Split off the front portion. Insert the new entry BEFORE this one, 2515 * so that this entry has the specified starting address. 2516 */ 2517 new_entry->end = startaddr; 2518 vm_map_entry_link(map, new_entry); 2519 return (KERN_SUCCESS); 2520 } 2521 2522 /* 2523 * vm_map_lookup_clip_start: 2524 * 2525 * Find the entry at or just after 'start', and clip it if 'start' is in 2526 * the interior of the entry. Return entry after 'start', and in 2527 * prev_entry set the entry before 'start'. 2528 */ 2529 static int 2530 vm_map_lookup_clip_start(vm_map_t map, vm_offset_t start, 2531 vm_map_entry_t *res_entry, vm_map_entry_t *prev_entry) 2532 { 2533 vm_map_entry_t entry; 2534 int rv; 2535 2536 if (!vm_map_is_system(map)) 2537 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2538 "%s: map %p start 0x%jx prev %p", __func__, map, 2539 (uintmax_t)start, prev_entry); 2540 2541 if (vm_map_lookup_entry(map, start, prev_entry)) { 2542 entry = *prev_entry; 2543 rv = vm_map_clip_start(map, entry, start); 2544 if (rv != KERN_SUCCESS) 2545 return (rv); 2546 *prev_entry = vm_map_entry_pred(entry); 2547 } else 2548 entry = vm_map_entry_succ(*prev_entry); 2549 *res_entry = entry; 2550 return (KERN_SUCCESS); 2551 } 2552 2553 /* 2554 * vm_map_clip_end: [ internal use only ] 2555 * 2556 * Asserts that the given entry ends at or before 2557 * the specified address; if necessary, 2558 * it splits the entry into two. 2559 */ 2560 static int 2561 vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t endaddr) 2562 { 2563 vm_map_entry_t new_entry; 2564 int bdry_idx; 2565 2566 if (!vm_map_is_system(map)) 2567 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2568 "%s: map %p entry %p end 0x%jx", __func__, map, entry, 2569 (uintmax_t)endaddr); 2570 2571 if (endaddr >= entry->end) 2572 return (KERN_SUCCESS); 2573 2574 VM_MAP_ASSERT_LOCKED(map); 2575 KASSERT(entry->start < endaddr && entry->end > endaddr, 2576 ("%s: invalid clip of entry %p", __func__, entry)); 2577 2578 bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry); 2579 if (bdry_idx != 0) { 2580 if ((endaddr & (pagesizes[bdry_idx] - 1)) != 0) 2581 return (KERN_INVALID_ARGUMENT); 2582 } 2583 2584 new_entry = vm_map_entry_clone(map, entry); 2585 2586 /* 2587 * Split off the back portion. Insert the new entry AFTER this one, 2588 * so that this entry has the specified ending address. 2589 */ 2590 new_entry->start = endaddr; 2591 vm_map_entry_link(map, new_entry); 2592 2593 return (KERN_SUCCESS); 2594 } 2595 2596 /* 2597 * vm_map_submap: [ kernel use only ] 2598 * 2599 * Mark the given range as handled by a subordinate map. 2600 * 2601 * This range must have been created with vm_map_find, 2602 * and no other operations may have been performed on this 2603 * range prior to calling vm_map_submap. 2604 * 2605 * Only a limited number of operations can be performed 2606 * within this rage after calling vm_map_submap: 2607 * vm_fault 2608 * [Don't try vm_map_copy!] 2609 * 2610 * To remove a submapping, one must first remove the 2611 * range from the superior map, and then destroy the 2612 * submap (if desired). [Better yet, don't try it.] 2613 */ 2614 int 2615 vm_map_submap( 2616 vm_map_t map, 2617 vm_offset_t start, 2618 vm_offset_t end, 2619 vm_map_t submap) 2620 { 2621 vm_map_entry_t entry; 2622 int result; 2623 2624 result = KERN_INVALID_ARGUMENT; 2625 2626 vm_map_lock(submap); 2627 submap->flags |= MAP_IS_SUB_MAP; 2628 vm_map_unlock(submap); 2629 2630 vm_map_lock(map); 2631 VM_MAP_RANGE_CHECK(map, start, end); 2632 if (vm_map_lookup_entry(map, start, &entry) && entry->end >= end && 2633 (entry->eflags & MAP_ENTRY_COW) == 0 && 2634 entry->object.vm_object == NULL) { 2635 result = vm_map_clip_start(map, entry, start); 2636 if (result != KERN_SUCCESS) 2637 goto unlock; 2638 result = vm_map_clip_end(map, entry, end); 2639 if (result != KERN_SUCCESS) 2640 goto unlock; 2641 entry->object.sub_map = submap; 2642 entry->eflags |= MAP_ENTRY_IS_SUB_MAP; 2643 result = KERN_SUCCESS; 2644 } 2645 unlock: 2646 vm_map_unlock(map); 2647 2648 if (result != KERN_SUCCESS) { 2649 vm_map_lock(submap); 2650 submap->flags &= ~MAP_IS_SUB_MAP; 2651 vm_map_unlock(submap); 2652 } 2653 return (result); 2654 } 2655 2656 /* 2657 * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified 2658 */ 2659 #define MAX_INIT_PT 96 2660 2661 /* 2662 * vm_map_pmap_enter: 2663 * 2664 * Preload the specified map's pmap with mappings to the specified 2665 * object's memory-resident pages. No further physical pages are 2666 * allocated, and no further virtual pages are retrieved from secondary 2667 * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a 2668 * limited number of page mappings are created at the low-end of the 2669 * specified address range. (For this purpose, a superpage mapping 2670 * counts as one page mapping.) Otherwise, all resident pages within 2671 * the specified address range are mapped. 2672 */ 2673 static void 2674 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 2675 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags) 2676 { 2677 vm_offset_t start; 2678 vm_page_t p, p_start; 2679 vm_pindex_t mask, psize, threshold, tmpidx; 2680 int psind; 2681 2682 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL) 2683 return; 2684 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2685 VM_OBJECT_WLOCK(object); 2686 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2687 pmap_object_init_pt(map->pmap, addr, object, pindex, 2688 size); 2689 VM_OBJECT_WUNLOCK(object); 2690 return; 2691 } 2692 VM_OBJECT_LOCK_DOWNGRADE(object); 2693 } else 2694 VM_OBJECT_RLOCK(object); 2695 2696 psize = atop(size); 2697 if (psize + pindex > object->size) { 2698 if (pindex >= object->size) { 2699 VM_OBJECT_RUNLOCK(object); 2700 return; 2701 } 2702 psize = object->size - pindex; 2703 } 2704 2705 start = 0; 2706 p_start = NULL; 2707 threshold = MAX_INIT_PT; 2708 2709 p = vm_page_find_least(object, pindex); 2710 /* 2711 * Assert: the variable p is either (1) the page with the 2712 * least pindex greater than or equal to the parameter pindex 2713 * or (2) NULL. 2714 */ 2715 for (; 2716 p != NULL && (tmpidx = p->pindex - pindex) < psize; 2717 p = TAILQ_NEXT(p, listq)) { 2718 /* 2719 * don't allow an madvise to blow away our really 2720 * free pages allocating pv entries. 2721 */ 2722 if (((flags & MAP_PREFAULT_MADVISE) != 0 && 2723 vm_page_count_severe()) || 2724 ((flags & MAP_PREFAULT_PARTIAL) != 0 && 2725 tmpidx >= threshold)) { 2726 psize = tmpidx; 2727 break; 2728 } 2729 if (vm_page_all_valid(p)) { 2730 if (p_start == NULL) { 2731 start = addr + ptoa(tmpidx); 2732 p_start = p; 2733 } 2734 /* Jump ahead if a superpage mapping is possible. */ 2735 for (psind = p->psind; psind > 0; psind--) { 2736 if (((addr + ptoa(tmpidx)) & 2737 (pagesizes[psind] - 1)) == 0) { 2738 mask = atop(pagesizes[psind]) - 1; 2739 if (tmpidx + mask < psize && 2740 vm_page_ps_test(p, psind, 2741 PS_ALL_VALID, NULL)) { 2742 p += mask; 2743 threshold += mask; 2744 break; 2745 } 2746 } 2747 } 2748 } else if (p_start != NULL) { 2749 pmap_enter_object(map->pmap, start, addr + 2750 ptoa(tmpidx), p_start, prot); 2751 p_start = NULL; 2752 } 2753 } 2754 if (p_start != NULL) 2755 pmap_enter_object(map->pmap, start, addr + ptoa(psize), 2756 p_start, prot); 2757 VM_OBJECT_RUNLOCK(object); 2758 } 2759 2760 static void 2761 vm_map_protect_guard(vm_map_entry_t entry, vm_prot_t new_prot, 2762 vm_prot_t new_maxprot, int flags) 2763 { 2764 vm_prot_t old_prot; 2765 2766 MPASS((entry->eflags & MAP_ENTRY_GUARD) != 0); 2767 if ((entry->eflags & MAP_ENTRY_STACK_GAP) == 0) 2768 return; 2769 2770 old_prot = PROT_EXTRACT(entry->offset); 2771 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) { 2772 entry->offset = PROT_MAX(new_maxprot) | 2773 (new_maxprot & old_prot); 2774 } 2775 if ((flags & VM_MAP_PROTECT_SET_PROT) != 0) { 2776 entry->offset = new_prot | PROT_MAX( 2777 PROT_MAX_EXTRACT(entry->offset)); 2778 } 2779 } 2780 2781 /* 2782 * vm_map_protect: 2783 * 2784 * Sets the protection and/or the maximum protection of the 2785 * specified address region in the target map. 2786 */ 2787 int 2788 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 2789 vm_prot_t new_prot, vm_prot_t new_maxprot, int flags) 2790 { 2791 vm_map_entry_t entry, first_entry, in_tran, prev_entry; 2792 vm_object_t obj; 2793 struct ucred *cred; 2794 vm_offset_t orig_start; 2795 vm_prot_t check_prot, max_prot, old_prot; 2796 int rv; 2797 2798 if (start == end) 2799 return (KERN_SUCCESS); 2800 2801 if (CONTAINS_BITS(flags, VM_MAP_PROTECT_SET_PROT | 2802 VM_MAP_PROTECT_SET_MAXPROT) && 2803 !CONTAINS_BITS(new_maxprot, new_prot)) 2804 return (KERN_OUT_OF_BOUNDS); 2805 2806 orig_start = start; 2807 again: 2808 in_tran = NULL; 2809 start = orig_start; 2810 vm_map_lock(map); 2811 2812 if ((map->flags & MAP_WXORX) != 0 && 2813 (flags & VM_MAP_PROTECT_SET_PROT) != 0 && 2814 CONTAINS_BITS(new_prot, VM_PROT_WRITE | VM_PROT_EXECUTE)) { 2815 vm_map_unlock(map); 2816 return (KERN_PROTECTION_FAILURE); 2817 } 2818 2819 /* 2820 * Ensure that we are not concurrently wiring pages. vm_map_wire() may 2821 * need to fault pages into the map and will drop the map lock while 2822 * doing so, and the VM object may end up in an inconsistent state if we 2823 * update the protection on the map entry in between faults. 2824 */ 2825 vm_map_wait_busy(map); 2826 2827 VM_MAP_RANGE_CHECK(map, start, end); 2828 2829 if (!vm_map_lookup_entry(map, start, &first_entry)) 2830 first_entry = vm_map_entry_succ(first_entry); 2831 2832 if ((flags & VM_MAP_PROTECT_GROWSDOWN) != 0 && 2833 (first_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0) { 2834 /* 2835 * Handle Linux's PROT_GROWSDOWN flag. 2836 * It means that protection is applied down to the 2837 * whole stack, including the specified range of the 2838 * mapped region, and the grow down region (AKA 2839 * guard). 2840 */ 2841 while (!CONTAINS_BITS(first_entry->eflags, 2842 MAP_ENTRY_GUARD | MAP_ENTRY_STACK_GAP) && 2843 first_entry != vm_map_entry_first(map)) 2844 first_entry = vm_map_entry_pred(first_entry); 2845 start = first_entry->start; 2846 } 2847 2848 /* 2849 * Make a first pass to check for protection violations. 2850 */ 2851 check_prot = 0; 2852 if ((flags & VM_MAP_PROTECT_SET_PROT) != 0) 2853 check_prot |= new_prot; 2854 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) 2855 check_prot |= new_maxprot; 2856 for (entry = first_entry; entry->start < end; 2857 entry = vm_map_entry_succ(entry)) { 2858 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 2859 vm_map_unlock(map); 2860 return (KERN_INVALID_ARGUMENT); 2861 } 2862 if ((entry->eflags & (MAP_ENTRY_GUARD | 2863 MAP_ENTRY_STACK_GAP)) == MAP_ENTRY_GUARD) 2864 continue; 2865 max_prot = (entry->eflags & MAP_ENTRY_STACK_GAP) != 0 ? 2866 PROT_MAX_EXTRACT(entry->offset) : entry->max_protection; 2867 if (!CONTAINS_BITS(max_prot, check_prot)) { 2868 vm_map_unlock(map); 2869 return (KERN_PROTECTION_FAILURE); 2870 } 2871 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0) 2872 in_tran = entry; 2873 } 2874 2875 /* 2876 * Postpone the operation until all in-transition map entries have 2877 * stabilized. An in-transition entry might already have its pages 2878 * wired and wired_count incremented, but not yet have its 2879 * MAP_ENTRY_USER_WIRED flag set. In which case, we would fail to call 2880 * vm_fault_copy_entry() in the final loop below. 2881 */ 2882 if (in_tran != NULL) { 2883 in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2884 vm_map_unlock_and_wait(map, 0); 2885 goto again; 2886 } 2887 2888 /* 2889 * Before changing the protections, try to reserve swap space for any 2890 * private (i.e., copy-on-write) mappings that are transitioning from 2891 * read-only to read/write access. If a reservation fails, break out 2892 * of this loop early and let the next loop simplify the entries, since 2893 * some may now be mergeable. 2894 */ 2895 rv = vm_map_clip_start(map, first_entry, start); 2896 if (rv != KERN_SUCCESS) { 2897 vm_map_unlock(map); 2898 return (rv); 2899 } 2900 for (entry = first_entry; entry->start < end; 2901 entry = vm_map_entry_succ(entry)) { 2902 rv = vm_map_clip_end(map, entry, end); 2903 if (rv != KERN_SUCCESS) { 2904 vm_map_unlock(map); 2905 return (rv); 2906 } 2907 2908 if ((flags & VM_MAP_PROTECT_SET_PROT) == 0 || 2909 ((new_prot & ~entry->protection) & VM_PROT_WRITE) == 0 || 2910 ENTRY_CHARGED(entry) || 2911 (entry->eflags & MAP_ENTRY_GUARD) != 0) 2912 continue; 2913 2914 cred = curthread->td_ucred; 2915 obj = entry->object.vm_object; 2916 2917 if (obj == NULL || 2918 (entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0) { 2919 if (!swap_reserve(entry->end - entry->start)) { 2920 rv = KERN_RESOURCE_SHORTAGE; 2921 end = entry->end; 2922 break; 2923 } 2924 crhold(cred); 2925 entry->cred = cred; 2926 continue; 2927 } 2928 2929 VM_OBJECT_WLOCK(obj); 2930 if ((obj->flags & OBJ_SWAP) == 0) { 2931 VM_OBJECT_WUNLOCK(obj); 2932 continue; 2933 } 2934 2935 /* 2936 * Charge for the whole object allocation now, since 2937 * we cannot distinguish between non-charged and 2938 * charged clipped mapping of the same object later. 2939 */ 2940 KASSERT(obj->charge == 0, 2941 ("vm_map_protect: object %p overcharged (entry %p)", 2942 obj, entry)); 2943 if (!swap_reserve(ptoa(obj->size))) { 2944 VM_OBJECT_WUNLOCK(obj); 2945 rv = KERN_RESOURCE_SHORTAGE; 2946 end = entry->end; 2947 break; 2948 } 2949 2950 crhold(cred); 2951 obj->cred = cred; 2952 obj->charge = ptoa(obj->size); 2953 VM_OBJECT_WUNLOCK(obj); 2954 } 2955 2956 /* 2957 * If enough swap space was available, go back and fix up protections. 2958 * Otherwise, just simplify entries, since some may have been modified. 2959 * [Note that clipping is not necessary the second time.] 2960 */ 2961 for (prev_entry = vm_map_entry_pred(first_entry), entry = first_entry; 2962 entry->start < end; 2963 vm_map_try_merge_entries(map, prev_entry, entry), 2964 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 2965 if (rv != KERN_SUCCESS) 2966 continue; 2967 2968 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { 2969 vm_map_protect_guard(entry, new_prot, new_maxprot, 2970 flags); 2971 continue; 2972 } 2973 2974 old_prot = entry->protection; 2975 2976 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) { 2977 entry->max_protection = new_maxprot; 2978 entry->protection = new_maxprot & old_prot; 2979 } 2980 if ((flags & VM_MAP_PROTECT_SET_PROT) != 0) 2981 entry->protection = new_prot; 2982 2983 /* 2984 * For user wired map entries, the normal lazy evaluation of 2985 * write access upgrades through soft page faults is 2986 * undesirable. Instead, immediately copy any pages that are 2987 * copy-on-write and enable write access in the physical map. 2988 */ 2989 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0 && 2990 (entry->protection & VM_PROT_WRITE) != 0 && 2991 (old_prot & VM_PROT_WRITE) == 0) 2992 vm_fault_copy_entry(map, map, entry, entry, NULL); 2993 2994 /* 2995 * When restricting access, update the physical map. Worry 2996 * about copy-on-write here. 2997 */ 2998 if ((old_prot & ~entry->protection) != 0) { 2999 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 3000 VM_PROT_ALL) 3001 pmap_protect(map->pmap, entry->start, 3002 entry->end, 3003 entry->protection & MASK(entry)); 3004 #undef MASK 3005 } 3006 } 3007 vm_map_try_merge_entries(map, prev_entry, entry); 3008 vm_map_unlock(map); 3009 return (rv); 3010 } 3011 3012 /* 3013 * vm_map_madvise: 3014 * 3015 * This routine traverses a processes map handling the madvise 3016 * system call. Advisories are classified as either those effecting 3017 * the vm_map_entry structure, or those effecting the underlying 3018 * objects. 3019 */ 3020 int 3021 vm_map_madvise( 3022 vm_map_t map, 3023 vm_offset_t start, 3024 vm_offset_t end, 3025 int behav) 3026 { 3027 vm_map_entry_t entry, prev_entry; 3028 int rv; 3029 bool modify_map; 3030 3031 /* 3032 * Some madvise calls directly modify the vm_map_entry, in which case 3033 * we need to use an exclusive lock on the map and we need to perform 3034 * various clipping operations. Otherwise we only need a read-lock 3035 * on the map. 3036 */ 3037 switch(behav) { 3038 case MADV_NORMAL: 3039 case MADV_SEQUENTIAL: 3040 case MADV_RANDOM: 3041 case MADV_NOSYNC: 3042 case MADV_AUTOSYNC: 3043 case MADV_NOCORE: 3044 case MADV_CORE: 3045 if (start == end) 3046 return (0); 3047 modify_map = true; 3048 vm_map_lock(map); 3049 break; 3050 case MADV_WILLNEED: 3051 case MADV_DONTNEED: 3052 case MADV_FREE: 3053 if (start == end) 3054 return (0); 3055 modify_map = false; 3056 vm_map_lock_read(map); 3057 break; 3058 default: 3059 return (EINVAL); 3060 } 3061 3062 /* 3063 * Locate starting entry and clip if necessary. 3064 */ 3065 VM_MAP_RANGE_CHECK(map, start, end); 3066 3067 if (modify_map) { 3068 /* 3069 * madvise behaviors that are implemented in the vm_map_entry. 3070 * 3071 * We clip the vm_map_entry so that behavioral changes are 3072 * limited to the specified address range. 3073 */ 3074 rv = vm_map_lookup_clip_start(map, start, &entry, &prev_entry); 3075 if (rv != KERN_SUCCESS) { 3076 vm_map_unlock(map); 3077 return (vm_mmap_to_errno(rv)); 3078 } 3079 3080 for (; entry->start < end; prev_entry = entry, 3081 entry = vm_map_entry_succ(entry)) { 3082 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 3083 continue; 3084 3085 rv = vm_map_clip_end(map, entry, end); 3086 if (rv != KERN_SUCCESS) { 3087 vm_map_unlock(map); 3088 return (vm_mmap_to_errno(rv)); 3089 } 3090 3091 switch (behav) { 3092 case MADV_NORMAL: 3093 vm_map_entry_set_behavior(entry, 3094 MAP_ENTRY_BEHAV_NORMAL); 3095 break; 3096 case MADV_SEQUENTIAL: 3097 vm_map_entry_set_behavior(entry, 3098 MAP_ENTRY_BEHAV_SEQUENTIAL); 3099 break; 3100 case MADV_RANDOM: 3101 vm_map_entry_set_behavior(entry, 3102 MAP_ENTRY_BEHAV_RANDOM); 3103 break; 3104 case MADV_NOSYNC: 3105 entry->eflags |= MAP_ENTRY_NOSYNC; 3106 break; 3107 case MADV_AUTOSYNC: 3108 entry->eflags &= ~MAP_ENTRY_NOSYNC; 3109 break; 3110 case MADV_NOCORE: 3111 entry->eflags |= MAP_ENTRY_NOCOREDUMP; 3112 break; 3113 case MADV_CORE: 3114 entry->eflags &= ~MAP_ENTRY_NOCOREDUMP; 3115 break; 3116 default: 3117 break; 3118 } 3119 vm_map_try_merge_entries(map, prev_entry, entry); 3120 } 3121 vm_map_try_merge_entries(map, prev_entry, entry); 3122 vm_map_unlock(map); 3123 } else { 3124 vm_pindex_t pstart, pend; 3125 3126 /* 3127 * madvise behaviors that are implemented in the underlying 3128 * vm_object. 3129 * 3130 * Since we don't clip the vm_map_entry, we have to clip 3131 * the vm_object pindex and count. 3132 */ 3133 if (!vm_map_lookup_entry(map, start, &entry)) 3134 entry = vm_map_entry_succ(entry); 3135 for (; entry->start < end; 3136 entry = vm_map_entry_succ(entry)) { 3137 vm_offset_t useEnd, useStart; 3138 3139 if ((entry->eflags & (MAP_ENTRY_IS_SUB_MAP | 3140 MAP_ENTRY_GUARD)) != 0) 3141 continue; 3142 3143 /* 3144 * MADV_FREE would otherwise rewind time to 3145 * the creation of the shadow object. Because 3146 * we hold the VM map read-locked, neither the 3147 * entry's object nor the presence of a 3148 * backing object can change. 3149 */ 3150 if (behav == MADV_FREE && 3151 entry->object.vm_object != NULL && 3152 entry->object.vm_object->backing_object != NULL) 3153 continue; 3154 3155 pstart = OFF_TO_IDX(entry->offset); 3156 pend = pstart + atop(entry->end - entry->start); 3157 useStart = entry->start; 3158 useEnd = entry->end; 3159 3160 if (entry->start < start) { 3161 pstart += atop(start - entry->start); 3162 useStart = start; 3163 } 3164 if (entry->end > end) { 3165 pend -= atop(entry->end - end); 3166 useEnd = end; 3167 } 3168 3169 if (pstart >= pend) 3170 continue; 3171 3172 /* 3173 * Perform the pmap_advise() before clearing 3174 * PGA_REFERENCED in vm_page_advise(). Otherwise, a 3175 * concurrent pmap operation, such as pmap_remove(), 3176 * could clear a reference in the pmap and set 3177 * PGA_REFERENCED on the page before the pmap_advise() 3178 * had completed. Consequently, the page would appear 3179 * referenced based upon an old reference that 3180 * occurred before this pmap_advise() ran. 3181 */ 3182 if (behav == MADV_DONTNEED || behav == MADV_FREE) 3183 pmap_advise(map->pmap, useStart, useEnd, 3184 behav); 3185 3186 vm_object_madvise(entry->object.vm_object, pstart, 3187 pend, behav); 3188 3189 /* 3190 * Pre-populate paging structures in the 3191 * WILLNEED case. For wired entries, the 3192 * paging structures are already populated. 3193 */ 3194 if (behav == MADV_WILLNEED && 3195 entry->wired_count == 0) { 3196 vm_map_pmap_enter(map, 3197 useStart, 3198 entry->protection, 3199 entry->object.vm_object, 3200 pstart, 3201 ptoa(pend - pstart), 3202 MAP_PREFAULT_MADVISE 3203 ); 3204 } 3205 } 3206 vm_map_unlock_read(map); 3207 } 3208 return (0); 3209 } 3210 3211 /* 3212 * vm_map_inherit: 3213 * 3214 * Sets the inheritance of the specified address 3215 * range in the target map. Inheritance 3216 * affects how the map will be shared with 3217 * child maps at the time of vmspace_fork. 3218 */ 3219 int 3220 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 3221 vm_inherit_t new_inheritance) 3222 { 3223 vm_map_entry_t entry, lentry, prev_entry, start_entry; 3224 int rv; 3225 3226 switch (new_inheritance) { 3227 case VM_INHERIT_NONE: 3228 case VM_INHERIT_COPY: 3229 case VM_INHERIT_SHARE: 3230 case VM_INHERIT_ZERO: 3231 break; 3232 default: 3233 return (KERN_INVALID_ARGUMENT); 3234 } 3235 if (start == end) 3236 return (KERN_SUCCESS); 3237 vm_map_lock(map); 3238 VM_MAP_RANGE_CHECK(map, start, end); 3239 rv = vm_map_lookup_clip_start(map, start, &start_entry, &prev_entry); 3240 if (rv != KERN_SUCCESS) 3241 goto unlock; 3242 if (vm_map_lookup_entry(map, end - 1, &lentry)) { 3243 rv = vm_map_clip_end(map, lentry, end); 3244 if (rv != KERN_SUCCESS) 3245 goto unlock; 3246 } 3247 if (new_inheritance == VM_INHERIT_COPY) { 3248 for (entry = start_entry; entry->start < end; 3249 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3250 if ((entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) 3251 != 0) { 3252 rv = KERN_INVALID_ARGUMENT; 3253 goto unlock; 3254 } 3255 } 3256 } 3257 for (entry = start_entry; entry->start < end; prev_entry = entry, 3258 entry = vm_map_entry_succ(entry)) { 3259 KASSERT(entry->end <= end, ("non-clipped entry %p end %jx %jx", 3260 entry, (uintmax_t)entry->end, (uintmax_t)end)); 3261 if ((entry->eflags & MAP_ENTRY_GUARD) == 0 || 3262 new_inheritance != VM_INHERIT_ZERO) 3263 entry->inheritance = new_inheritance; 3264 vm_map_try_merge_entries(map, prev_entry, entry); 3265 } 3266 vm_map_try_merge_entries(map, prev_entry, entry); 3267 unlock: 3268 vm_map_unlock(map); 3269 return (rv); 3270 } 3271 3272 /* 3273 * vm_map_entry_in_transition: 3274 * 3275 * Release the map lock, and sleep until the entry is no longer in 3276 * transition. Awake and acquire the map lock. If the map changed while 3277 * another held the lock, lookup a possibly-changed entry at or after the 3278 * 'start' position of the old entry. 3279 */ 3280 static vm_map_entry_t 3281 vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start, 3282 vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry) 3283 { 3284 vm_map_entry_t entry; 3285 vm_offset_t start; 3286 u_int last_timestamp; 3287 3288 VM_MAP_ASSERT_LOCKED(map); 3289 KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3290 ("not in-tranition map entry %p", in_entry)); 3291 /* 3292 * We have not yet clipped the entry. 3293 */ 3294 start = MAX(in_start, in_entry->start); 3295 in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3296 last_timestamp = map->timestamp; 3297 if (vm_map_unlock_and_wait(map, 0)) { 3298 /* 3299 * Allow interruption of user wiring/unwiring? 3300 */ 3301 } 3302 vm_map_lock(map); 3303 if (last_timestamp + 1 == map->timestamp) 3304 return (in_entry); 3305 3306 /* 3307 * Look again for the entry because the map was modified while it was 3308 * unlocked. Specifically, the entry may have been clipped, merged, or 3309 * deleted. 3310 */ 3311 if (!vm_map_lookup_entry(map, start, &entry)) { 3312 if (!holes_ok) { 3313 *io_end = start; 3314 return (NULL); 3315 } 3316 entry = vm_map_entry_succ(entry); 3317 } 3318 return (entry); 3319 } 3320 3321 /* 3322 * vm_map_unwire: 3323 * 3324 * Implements both kernel and user unwiring. 3325 */ 3326 int 3327 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 3328 int flags) 3329 { 3330 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3331 int rv; 3332 bool holes_ok, need_wakeup, user_unwire; 3333 3334 if (start == end) 3335 return (KERN_SUCCESS); 3336 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3337 user_unwire = (flags & VM_MAP_WIRE_USER) != 0; 3338 vm_map_lock(map); 3339 VM_MAP_RANGE_CHECK(map, start, end); 3340 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3341 if (holes_ok) 3342 first_entry = vm_map_entry_succ(first_entry); 3343 else { 3344 vm_map_unlock(map); 3345 return (KERN_INVALID_ADDRESS); 3346 } 3347 } 3348 rv = KERN_SUCCESS; 3349 for (entry = first_entry; entry->start < end; entry = next_entry) { 3350 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3351 /* 3352 * We have not yet clipped the entry. 3353 */ 3354 next_entry = vm_map_entry_in_transition(map, start, 3355 &end, holes_ok, entry); 3356 if (next_entry == NULL) { 3357 if (entry == first_entry) { 3358 vm_map_unlock(map); 3359 return (KERN_INVALID_ADDRESS); 3360 } 3361 rv = KERN_INVALID_ADDRESS; 3362 break; 3363 } 3364 first_entry = (entry == first_entry) ? 3365 next_entry : NULL; 3366 continue; 3367 } 3368 rv = vm_map_clip_start(map, entry, start); 3369 if (rv != KERN_SUCCESS) 3370 break; 3371 rv = vm_map_clip_end(map, entry, end); 3372 if (rv != KERN_SUCCESS) 3373 break; 3374 3375 /* 3376 * Mark the entry in case the map lock is released. (See 3377 * above.) 3378 */ 3379 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3380 entry->wiring_thread == NULL, 3381 ("owned map entry %p", entry)); 3382 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3383 entry->wiring_thread = curthread; 3384 next_entry = vm_map_entry_succ(entry); 3385 /* 3386 * Check the map for holes in the specified region. 3387 * If holes_ok, skip this check. 3388 */ 3389 if (!holes_ok && 3390 entry->end < end && next_entry->start > entry->end) { 3391 end = entry->end; 3392 rv = KERN_INVALID_ADDRESS; 3393 break; 3394 } 3395 /* 3396 * If system unwiring, require that the entry is system wired. 3397 */ 3398 if (!user_unwire && 3399 vm_map_entry_system_wired_count(entry) == 0) { 3400 end = entry->end; 3401 rv = KERN_INVALID_ARGUMENT; 3402 break; 3403 } 3404 } 3405 need_wakeup = false; 3406 if (first_entry == NULL && 3407 !vm_map_lookup_entry(map, start, &first_entry)) { 3408 KASSERT(holes_ok, ("vm_map_unwire: lookup failed")); 3409 prev_entry = first_entry; 3410 entry = vm_map_entry_succ(first_entry); 3411 } else { 3412 prev_entry = vm_map_entry_pred(first_entry); 3413 entry = first_entry; 3414 } 3415 for (; entry->start < end; 3416 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3417 /* 3418 * If holes_ok was specified, an empty 3419 * space in the unwired region could have been mapped 3420 * while the map lock was dropped for draining 3421 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread 3422 * could be simultaneously wiring this new mapping 3423 * entry. Detect these cases and skip any entries 3424 * marked as in transition by us. 3425 */ 3426 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3427 entry->wiring_thread != curthread) { 3428 KASSERT(holes_ok, 3429 ("vm_map_unwire: !HOLESOK and new/changed entry")); 3430 continue; 3431 } 3432 3433 if (rv == KERN_SUCCESS && (!user_unwire || 3434 (entry->eflags & MAP_ENTRY_USER_WIRED))) { 3435 if (entry->wired_count == 1) 3436 vm_map_entry_unwire(map, entry); 3437 else 3438 entry->wired_count--; 3439 if (user_unwire) 3440 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3441 } 3442 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3443 ("vm_map_unwire: in-transition flag missing %p", entry)); 3444 KASSERT(entry->wiring_thread == curthread, 3445 ("vm_map_unwire: alien wire %p", entry)); 3446 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 3447 entry->wiring_thread = NULL; 3448 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3449 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3450 need_wakeup = true; 3451 } 3452 vm_map_try_merge_entries(map, prev_entry, entry); 3453 } 3454 vm_map_try_merge_entries(map, prev_entry, entry); 3455 vm_map_unlock(map); 3456 if (need_wakeup) 3457 vm_map_wakeup(map); 3458 return (rv); 3459 } 3460 3461 static void 3462 vm_map_wire_user_count_sub(u_long npages) 3463 { 3464 3465 atomic_subtract_long(&vm_user_wire_count, npages); 3466 } 3467 3468 static bool 3469 vm_map_wire_user_count_add(u_long npages) 3470 { 3471 u_long wired; 3472 3473 wired = vm_user_wire_count; 3474 do { 3475 if (npages + wired > vm_page_max_user_wired) 3476 return (false); 3477 } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired, 3478 npages + wired)); 3479 3480 return (true); 3481 } 3482 3483 /* 3484 * vm_map_wire_entry_failure: 3485 * 3486 * Handle a wiring failure on the given entry. 3487 * 3488 * The map should be locked. 3489 */ 3490 static void 3491 vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 3492 vm_offset_t failed_addr) 3493 { 3494 3495 VM_MAP_ASSERT_LOCKED(map); 3496 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 && 3497 entry->wired_count == 1, 3498 ("vm_map_wire_entry_failure: entry %p isn't being wired", entry)); 3499 KASSERT(failed_addr < entry->end, 3500 ("vm_map_wire_entry_failure: entry %p was fully wired", entry)); 3501 3502 /* 3503 * If any pages at the start of this entry were successfully wired, 3504 * then unwire them. 3505 */ 3506 if (failed_addr > entry->start) { 3507 pmap_unwire(map->pmap, entry->start, failed_addr); 3508 vm_object_unwire(entry->object.vm_object, entry->offset, 3509 failed_addr - entry->start, PQ_ACTIVE); 3510 } 3511 3512 /* 3513 * Assign an out-of-range value to represent the failure to wire this 3514 * entry. 3515 */ 3516 entry->wired_count = -1; 3517 } 3518 3519 int 3520 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3521 { 3522 int rv; 3523 3524 vm_map_lock(map); 3525 rv = vm_map_wire_locked(map, start, end, flags); 3526 vm_map_unlock(map); 3527 return (rv); 3528 } 3529 3530 /* 3531 * vm_map_wire_locked: 3532 * 3533 * Implements both kernel and user wiring. Returns with the map locked, 3534 * the map lock may be dropped. 3535 */ 3536 int 3537 vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3538 { 3539 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3540 vm_offset_t faddr, saved_end, saved_start; 3541 u_long incr, npages; 3542 u_int bidx, last_timestamp; 3543 int rv; 3544 bool holes_ok, need_wakeup, user_wire; 3545 vm_prot_t prot; 3546 3547 VM_MAP_ASSERT_LOCKED(map); 3548 3549 if (start == end) 3550 return (KERN_SUCCESS); 3551 prot = 0; 3552 if (flags & VM_MAP_WIRE_WRITE) 3553 prot |= VM_PROT_WRITE; 3554 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3555 user_wire = (flags & VM_MAP_WIRE_USER) != 0; 3556 VM_MAP_RANGE_CHECK(map, start, end); 3557 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3558 if (holes_ok) 3559 first_entry = vm_map_entry_succ(first_entry); 3560 else 3561 return (KERN_INVALID_ADDRESS); 3562 } 3563 for (entry = first_entry; entry->start < end; entry = next_entry) { 3564 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3565 /* 3566 * We have not yet clipped the entry. 3567 */ 3568 next_entry = vm_map_entry_in_transition(map, start, 3569 &end, holes_ok, entry); 3570 if (next_entry == NULL) { 3571 if (entry == first_entry) 3572 return (KERN_INVALID_ADDRESS); 3573 rv = KERN_INVALID_ADDRESS; 3574 goto done; 3575 } 3576 first_entry = (entry == first_entry) ? 3577 next_entry : NULL; 3578 continue; 3579 } 3580 rv = vm_map_clip_start(map, entry, start); 3581 if (rv != KERN_SUCCESS) 3582 goto done; 3583 rv = vm_map_clip_end(map, entry, end); 3584 if (rv != KERN_SUCCESS) 3585 goto done; 3586 3587 /* 3588 * Mark the entry in case the map lock is released. (See 3589 * above.) 3590 */ 3591 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3592 entry->wiring_thread == NULL, 3593 ("owned map entry %p", entry)); 3594 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3595 entry->wiring_thread = curthread; 3596 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 3597 || (entry->protection & prot) != prot) { 3598 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED; 3599 if (!holes_ok) { 3600 end = entry->end; 3601 rv = KERN_INVALID_ADDRESS; 3602 goto done; 3603 } 3604 } else if (entry->wired_count == 0) { 3605 entry->wired_count++; 3606 3607 npages = atop(entry->end - entry->start); 3608 if (user_wire && !vm_map_wire_user_count_add(npages)) { 3609 vm_map_wire_entry_failure(map, entry, 3610 entry->start); 3611 end = entry->end; 3612 rv = KERN_RESOURCE_SHORTAGE; 3613 goto done; 3614 } 3615 3616 /* 3617 * Release the map lock, relying on the in-transition 3618 * mark. Mark the map busy for fork. 3619 */ 3620 saved_start = entry->start; 3621 saved_end = entry->end; 3622 last_timestamp = map->timestamp; 3623 bidx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry); 3624 incr = pagesizes[bidx]; 3625 vm_map_busy(map); 3626 vm_map_unlock(map); 3627 3628 for (faddr = saved_start; faddr < saved_end; 3629 faddr += incr) { 3630 /* 3631 * Simulate a fault to get the page and enter 3632 * it into the physical map. 3633 */ 3634 rv = vm_fault(map, faddr, VM_PROT_NONE, 3635 VM_FAULT_WIRE, NULL); 3636 if (rv != KERN_SUCCESS) 3637 break; 3638 } 3639 vm_map_lock(map); 3640 vm_map_unbusy(map); 3641 if (last_timestamp + 1 != map->timestamp) { 3642 /* 3643 * Look again for the entry because the map was 3644 * modified while it was unlocked. The entry 3645 * may have been clipped, but NOT merged or 3646 * deleted. 3647 */ 3648 if (!vm_map_lookup_entry(map, saved_start, 3649 &next_entry)) 3650 KASSERT(false, 3651 ("vm_map_wire: lookup failed")); 3652 first_entry = (entry == first_entry) ? 3653 next_entry : NULL; 3654 for (entry = next_entry; entry->end < saved_end; 3655 entry = vm_map_entry_succ(entry)) { 3656 /* 3657 * In case of failure, handle entries 3658 * that were not fully wired here; 3659 * fully wired entries are handled 3660 * later. 3661 */ 3662 if (rv != KERN_SUCCESS && 3663 faddr < entry->end) 3664 vm_map_wire_entry_failure(map, 3665 entry, faddr); 3666 } 3667 } 3668 if (rv != KERN_SUCCESS) { 3669 vm_map_wire_entry_failure(map, entry, faddr); 3670 if (user_wire) 3671 vm_map_wire_user_count_sub(npages); 3672 end = entry->end; 3673 goto done; 3674 } 3675 } else if (!user_wire || 3676 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3677 entry->wired_count++; 3678 } 3679 /* 3680 * Check the map for holes in the specified region. 3681 * If holes_ok was specified, skip this check. 3682 */ 3683 next_entry = vm_map_entry_succ(entry); 3684 if (!holes_ok && 3685 entry->end < end && next_entry->start > entry->end) { 3686 end = entry->end; 3687 rv = KERN_INVALID_ADDRESS; 3688 goto done; 3689 } 3690 } 3691 rv = KERN_SUCCESS; 3692 done: 3693 need_wakeup = false; 3694 if (first_entry == NULL && 3695 !vm_map_lookup_entry(map, start, &first_entry)) { 3696 KASSERT(holes_ok, ("vm_map_wire: lookup failed")); 3697 prev_entry = first_entry; 3698 entry = vm_map_entry_succ(first_entry); 3699 } else { 3700 prev_entry = vm_map_entry_pred(first_entry); 3701 entry = first_entry; 3702 } 3703 for (; entry->start < end; 3704 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3705 /* 3706 * If holes_ok was specified, an empty 3707 * space in the unwired region could have been mapped 3708 * while the map lock was dropped for faulting in the 3709 * pages or draining MAP_ENTRY_IN_TRANSITION. 3710 * Moreover, another thread could be simultaneously 3711 * wiring this new mapping entry. Detect these cases 3712 * and skip any entries marked as in transition not by us. 3713 * 3714 * Another way to get an entry not marked with 3715 * MAP_ENTRY_IN_TRANSITION is after failed clipping, 3716 * which set rv to KERN_INVALID_ARGUMENT. 3717 */ 3718 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3719 entry->wiring_thread != curthread) { 3720 KASSERT(holes_ok || rv == KERN_INVALID_ARGUMENT, 3721 ("vm_map_wire: !HOLESOK and new/changed entry")); 3722 continue; 3723 } 3724 3725 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) { 3726 /* do nothing */ 3727 } else if (rv == KERN_SUCCESS) { 3728 if (user_wire) 3729 entry->eflags |= MAP_ENTRY_USER_WIRED; 3730 } else if (entry->wired_count == -1) { 3731 /* 3732 * Wiring failed on this entry. Thus, unwiring is 3733 * unnecessary. 3734 */ 3735 entry->wired_count = 0; 3736 } else if (!user_wire || 3737 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3738 /* 3739 * Undo the wiring. Wiring succeeded on this entry 3740 * but failed on a later entry. 3741 */ 3742 if (entry->wired_count == 1) { 3743 vm_map_entry_unwire(map, entry); 3744 if (user_wire) 3745 vm_map_wire_user_count_sub( 3746 atop(entry->end - entry->start)); 3747 } else 3748 entry->wired_count--; 3749 } 3750 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3751 ("vm_map_wire: in-transition flag missing %p", entry)); 3752 KASSERT(entry->wiring_thread == curthread, 3753 ("vm_map_wire: alien wire %p", entry)); 3754 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION | 3755 MAP_ENTRY_WIRE_SKIPPED); 3756 entry->wiring_thread = NULL; 3757 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3758 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3759 need_wakeup = true; 3760 } 3761 vm_map_try_merge_entries(map, prev_entry, entry); 3762 } 3763 vm_map_try_merge_entries(map, prev_entry, entry); 3764 if (need_wakeup) 3765 vm_map_wakeup(map); 3766 return (rv); 3767 } 3768 3769 /* 3770 * vm_map_sync 3771 * 3772 * Push any dirty cached pages in the address range to their pager. 3773 * If syncio is TRUE, dirty pages are written synchronously. 3774 * If invalidate is TRUE, any cached pages are freed as well. 3775 * 3776 * If the size of the region from start to end is zero, we are 3777 * supposed to flush all modified pages within the region containing 3778 * start. Unfortunately, a region can be split or coalesced with 3779 * neighboring regions, making it difficult to determine what the 3780 * original region was. Therefore, we approximate this requirement by 3781 * flushing the current region containing start. 3782 * 3783 * Returns an error if any part of the specified range is not mapped. 3784 */ 3785 int 3786 vm_map_sync( 3787 vm_map_t map, 3788 vm_offset_t start, 3789 vm_offset_t end, 3790 boolean_t syncio, 3791 boolean_t invalidate) 3792 { 3793 vm_map_entry_t entry, first_entry, next_entry; 3794 vm_size_t size; 3795 vm_object_t object; 3796 vm_ooffset_t offset; 3797 unsigned int last_timestamp; 3798 int bdry_idx; 3799 boolean_t failed; 3800 3801 vm_map_lock_read(map); 3802 VM_MAP_RANGE_CHECK(map, start, end); 3803 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3804 vm_map_unlock_read(map); 3805 return (KERN_INVALID_ADDRESS); 3806 } else if (start == end) { 3807 start = first_entry->start; 3808 end = first_entry->end; 3809 } 3810 3811 /* 3812 * Make a first pass to check for user-wired memory, holes, 3813 * and partial invalidation of largepage mappings. 3814 */ 3815 for (entry = first_entry; entry->start < end; entry = next_entry) { 3816 if (invalidate) { 3817 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) { 3818 vm_map_unlock_read(map); 3819 return (KERN_INVALID_ARGUMENT); 3820 } 3821 bdry_idx = MAP_ENTRY_SPLIT_BOUNDARY_INDEX(entry); 3822 if (bdry_idx != 0 && 3823 ((start & (pagesizes[bdry_idx] - 1)) != 0 || 3824 (end & (pagesizes[bdry_idx] - 1)) != 0)) { 3825 vm_map_unlock_read(map); 3826 return (KERN_INVALID_ARGUMENT); 3827 } 3828 } 3829 next_entry = vm_map_entry_succ(entry); 3830 if (end > entry->end && 3831 entry->end != next_entry->start) { 3832 vm_map_unlock_read(map); 3833 return (KERN_INVALID_ADDRESS); 3834 } 3835 } 3836 3837 if (invalidate) 3838 pmap_remove(map->pmap, start, end); 3839 failed = FALSE; 3840 3841 /* 3842 * Make a second pass, cleaning/uncaching pages from the indicated 3843 * objects as we go. 3844 */ 3845 for (entry = first_entry; entry->start < end;) { 3846 offset = entry->offset + (start - entry->start); 3847 size = (end <= entry->end ? end : entry->end) - start; 3848 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 3849 vm_map_t smap; 3850 vm_map_entry_t tentry; 3851 vm_size_t tsize; 3852 3853 smap = entry->object.sub_map; 3854 vm_map_lock_read(smap); 3855 (void) vm_map_lookup_entry(smap, offset, &tentry); 3856 tsize = tentry->end - offset; 3857 if (tsize < size) 3858 size = tsize; 3859 object = tentry->object.vm_object; 3860 offset = tentry->offset + (offset - tentry->start); 3861 vm_map_unlock_read(smap); 3862 } else { 3863 object = entry->object.vm_object; 3864 } 3865 vm_object_reference(object); 3866 last_timestamp = map->timestamp; 3867 vm_map_unlock_read(map); 3868 if (!vm_object_sync(object, offset, size, syncio, invalidate)) 3869 failed = TRUE; 3870 start += size; 3871 vm_object_deallocate(object); 3872 vm_map_lock_read(map); 3873 if (last_timestamp == map->timestamp || 3874 !vm_map_lookup_entry(map, start, &entry)) 3875 entry = vm_map_entry_succ(entry); 3876 } 3877 3878 vm_map_unlock_read(map); 3879 return (failed ? KERN_FAILURE : KERN_SUCCESS); 3880 } 3881 3882 /* 3883 * vm_map_entry_unwire: [ internal use only ] 3884 * 3885 * Make the region specified by this entry pageable. 3886 * 3887 * The map in question should be locked. 3888 * [This is the reason for this routine's existence.] 3889 */ 3890 static void 3891 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 3892 { 3893 vm_size_t size; 3894 3895 VM_MAP_ASSERT_LOCKED(map); 3896 KASSERT(entry->wired_count > 0, 3897 ("vm_map_entry_unwire: entry %p isn't wired", entry)); 3898 3899 size = entry->end - entry->start; 3900 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) 3901 vm_map_wire_user_count_sub(atop(size)); 3902 pmap_unwire(map->pmap, entry->start, entry->end); 3903 vm_object_unwire(entry->object.vm_object, entry->offset, size, 3904 PQ_ACTIVE); 3905 entry->wired_count = 0; 3906 } 3907 3908 static void 3909 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map) 3910 { 3911 3912 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) 3913 vm_object_deallocate(entry->object.vm_object); 3914 uma_zfree(system_map ? kmapentzone : mapentzone, entry); 3915 } 3916 3917 /* 3918 * vm_map_entry_delete: [ internal use only ] 3919 * 3920 * Deallocate the given entry from the target map. 3921 */ 3922 static void 3923 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry) 3924 { 3925 vm_object_t object; 3926 vm_pindex_t offidxstart, offidxend, size1; 3927 vm_size_t size; 3928 3929 vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE); 3930 object = entry->object.vm_object; 3931 3932 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { 3933 MPASS(entry->cred == NULL); 3934 MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0); 3935 MPASS(object == NULL); 3936 vm_map_entry_deallocate(entry, vm_map_is_system(map)); 3937 return; 3938 } 3939 3940 size = entry->end - entry->start; 3941 map->size -= size; 3942 3943 if (entry->cred != NULL) { 3944 swap_release_by_cred(size, entry->cred); 3945 crfree(entry->cred); 3946 } 3947 3948 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || object == NULL) { 3949 entry->object.vm_object = NULL; 3950 } else if ((object->flags & OBJ_ANON) != 0 || 3951 object == kernel_object) { 3952 KASSERT(entry->cred == NULL || object->cred == NULL || 3953 (entry->eflags & MAP_ENTRY_NEEDS_COPY), 3954 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry)); 3955 offidxstart = OFF_TO_IDX(entry->offset); 3956 offidxend = offidxstart + atop(size); 3957 VM_OBJECT_WLOCK(object); 3958 if (object->ref_count != 1 && 3959 ((object->flags & OBJ_ONEMAPPING) != 0 || 3960 object == kernel_object)) { 3961 vm_object_collapse(object); 3962 3963 /* 3964 * The option OBJPR_NOTMAPPED can be passed here 3965 * because vm_map_delete() already performed 3966 * pmap_remove() on the only mapping to this range 3967 * of pages. 3968 */ 3969 vm_object_page_remove(object, offidxstart, offidxend, 3970 OBJPR_NOTMAPPED); 3971 if (offidxend >= object->size && 3972 offidxstart < object->size) { 3973 size1 = object->size; 3974 object->size = offidxstart; 3975 if (object->cred != NULL) { 3976 size1 -= object->size; 3977 KASSERT(object->charge >= ptoa(size1), 3978 ("object %p charge < 0", object)); 3979 swap_release_by_cred(ptoa(size1), 3980 object->cred); 3981 object->charge -= ptoa(size1); 3982 } 3983 } 3984 } 3985 VM_OBJECT_WUNLOCK(object); 3986 } 3987 if (vm_map_is_system(map)) 3988 vm_map_entry_deallocate(entry, TRUE); 3989 else { 3990 entry->defer_next = curthread->td_map_def_user; 3991 curthread->td_map_def_user = entry; 3992 } 3993 } 3994 3995 /* 3996 * vm_map_delete: [ internal use only ] 3997 * 3998 * Deallocates the given address range from the target 3999 * map. 4000 */ 4001 int 4002 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end) 4003 { 4004 vm_map_entry_t entry, next_entry, scratch_entry; 4005 int rv; 4006 4007 VM_MAP_ASSERT_LOCKED(map); 4008 4009 if (start == end) 4010 return (KERN_SUCCESS); 4011 4012 /* 4013 * Find the start of the region, and clip it. 4014 * Step through all entries in this region. 4015 */ 4016 rv = vm_map_lookup_clip_start(map, start, &entry, &scratch_entry); 4017 if (rv != KERN_SUCCESS) 4018 return (rv); 4019 for (; entry->start < end; entry = next_entry) { 4020 /* 4021 * Wait for wiring or unwiring of an entry to complete. 4022 * Also wait for any system wirings to disappear on 4023 * user maps. 4024 */ 4025 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 || 4026 (vm_map_pmap(map) != kernel_pmap && 4027 vm_map_entry_system_wired_count(entry) != 0)) { 4028 unsigned int last_timestamp; 4029 vm_offset_t saved_start; 4030 4031 saved_start = entry->start; 4032 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 4033 last_timestamp = map->timestamp; 4034 (void) vm_map_unlock_and_wait(map, 0); 4035 vm_map_lock(map); 4036 if (last_timestamp + 1 != map->timestamp) { 4037 /* 4038 * Look again for the entry because the map was 4039 * modified while it was unlocked. 4040 * Specifically, the entry may have been 4041 * clipped, merged, or deleted. 4042 */ 4043 rv = vm_map_lookup_clip_start(map, saved_start, 4044 &next_entry, &scratch_entry); 4045 if (rv != KERN_SUCCESS) 4046 break; 4047 } else 4048 next_entry = entry; 4049 continue; 4050 } 4051 4052 /* XXXKIB or delete to the upper superpage boundary ? */ 4053 rv = vm_map_clip_end(map, entry, end); 4054 if (rv != KERN_SUCCESS) 4055 break; 4056 next_entry = vm_map_entry_succ(entry); 4057 4058 /* 4059 * Unwire before removing addresses from the pmap; otherwise, 4060 * unwiring will put the entries back in the pmap. 4061 */ 4062 if (entry->wired_count != 0) 4063 vm_map_entry_unwire(map, entry); 4064 4065 /* 4066 * Remove mappings for the pages, but only if the 4067 * mappings could exist. For instance, it does not 4068 * make sense to call pmap_remove() for guard entries. 4069 */ 4070 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 4071 entry->object.vm_object != NULL) 4072 pmap_map_delete(map->pmap, entry->start, entry->end); 4073 4074 /* 4075 * Delete the entry only after removing all pmap 4076 * entries pointing to its pages. (Otherwise, its 4077 * page frames may be reallocated, and any modify bits 4078 * will be set in the wrong object!) 4079 */ 4080 vm_map_entry_delete(map, entry); 4081 } 4082 return (rv); 4083 } 4084 4085 /* 4086 * vm_map_remove: 4087 * 4088 * Remove the given address range from the target map. 4089 * This is the exported form of vm_map_delete. 4090 */ 4091 int 4092 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 4093 { 4094 int result; 4095 4096 vm_map_lock(map); 4097 VM_MAP_RANGE_CHECK(map, start, end); 4098 result = vm_map_delete(map, start, end); 4099 vm_map_unlock(map); 4100 return (result); 4101 } 4102 4103 /* 4104 * vm_map_check_protection: 4105 * 4106 * Assert that the target map allows the specified privilege on the 4107 * entire address region given. The entire region must be allocated. 4108 * 4109 * WARNING! This code does not and should not check whether the 4110 * contents of the region is accessible. For example a smaller file 4111 * might be mapped into a larger address space. 4112 * 4113 * NOTE! This code is also called by munmap(). 4114 * 4115 * The map must be locked. A read lock is sufficient. 4116 */ 4117 boolean_t 4118 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 4119 vm_prot_t protection) 4120 { 4121 vm_map_entry_t entry; 4122 vm_map_entry_t tmp_entry; 4123 4124 if (!vm_map_lookup_entry(map, start, &tmp_entry)) 4125 return (FALSE); 4126 entry = tmp_entry; 4127 4128 while (start < end) { 4129 /* 4130 * No holes allowed! 4131 */ 4132 if (start < entry->start) 4133 return (FALSE); 4134 /* 4135 * Check protection associated with entry. 4136 */ 4137 if ((entry->protection & protection) != protection) 4138 return (FALSE); 4139 /* go to next entry */ 4140 start = entry->end; 4141 entry = vm_map_entry_succ(entry); 4142 } 4143 return (TRUE); 4144 } 4145 4146 /* 4147 * 4148 * vm_map_copy_swap_object: 4149 * 4150 * Copies a swap-backed object from an existing map entry to a 4151 * new one. Carries forward the swap charge. May change the 4152 * src object on return. 4153 */ 4154 static void 4155 vm_map_copy_swap_object(vm_map_entry_t src_entry, vm_map_entry_t dst_entry, 4156 vm_offset_t size, vm_ooffset_t *fork_charge) 4157 { 4158 vm_object_t src_object; 4159 struct ucred *cred; 4160 int charged; 4161 4162 src_object = src_entry->object.vm_object; 4163 charged = ENTRY_CHARGED(src_entry); 4164 if ((src_object->flags & OBJ_ANON) != 0) { 4165 VM_OBJECT_WLOCK(src_object); 4166 vm_object_collapse(src_object); 4167 if ((src_object->flags & OBJ_ONEMAPPING) != 0) { 4168 vm_object_split(src_entry); 4169 src_object = src_entry->object.vm_object; 4170 } 4171 vm_object_reference_locked(src_object); 4172 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 4173 VM_OBJECT_WUNLOCK(src_object); 4174 } else 4175 vm_object_reference(src_object); 4176 if (src_entry->cred != NULL && 4177 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 4178 KASSERT(src_object->cred == NULL, 4179 ("OVERCOMMIT: vm_map_copy_anon_entry: cred %p", 4180 src_object)); 4181 src_object->cred = src_entry->cred; 4182 src_object->charge = size; 4183 } 4184 dst_entry->object.vm_object = src_object; 4185 if (charged) { 4186 cred = curthread->td_ucred; 4187 crhold(cred); 4188 dst_entry->cred = cred; 4189 *fork_charge += size; 4190 if (!(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 4191 crhold(cred); 4192 src_entry->cred = cred; 4193 *fork_charge += size; 4194 } 4195 } 4196 } 4197 4198 /* 4199 * vm_map_copy_entry: 4200 * 4201 * Copies the contents of the source entry to the destination 4202 * entry. The entries *must* be aligned properly. 4203 */ 4204 static void 4205 vm_map_copy_entry( 4206 vm_map_t src_map, 4207 vm_map_t dst_map, 4208 vm_map_entry_t src_entry, 4209 vm_map_entry_t dst_entry, 4210 vm_ooffset_t *fork_charge) 4211 { 4212 vm_object_t src_object; 4213 vm_map_entry_t fake_entry; 4214 vm_offset_t size; 4215 4216 VM_MAP_ASSERT_LOCKED(dst_map); 4217 4218 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) 4219 return; 4220 4221 if (src_entry->wired_count == 0 || 4222 (src_entry->protection & VM_PROT_WRITE) == 0) { 4223 /* 4224 * If the source entry is marked needs_copy, it is already 4225 * write-protected. 4226 */ 4227 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && 4228 (src_entry->protection & VM_PROT_WRITE) != 0) { 4229 pmap_protect(src_map->pmap, 4230 src_entry->start, 4231 src_entry->end, 4232 src_entry->protection & ~VM_PROT_WRITE); 4233 } 4234 4235 /* 4236 * Make a copy of the object. 4237 */ 4238 size = src_entry->end - src_entry->start; 4239 if ((src_object = src_entry->object.vm_object) != NULL) { 4240 if ((src_object->flags & OBJ_SWAP) != 0) { 4241 vm_map_copy_swap_object(src_entry, dst_entry, 4242 size, fork_charge); 4243 /* May have split/collapsed, reload obj. */ 4244 src_object = src_entry->object.vm_object; 4245 } else { 4246 vm_object_reference(src_object); 4247 dst_entry->object.vm_object = src_object; 4248 } 4249 src_entry->eflags |= MAP_ENTRY_COW | 4250 MAP_ENTRY_NEEDS_COPY; 4251 dst_entry->eflags |= MAP_ENTRY_COW | 4252 MAP_ENTRY_NEEDS_COPY; 4253 dst_entry->offset = src_entry->offset; 4254 if (src_entry->eflags & MAP_ENTRY_WRITECNT) { 4255 /* 4256 * MAP_ENTRY_WRITECNT cannot 4257 * indicate write reference from 4258 * src_entry, since the entry is 4259 * marked as needs copy. Allocate a 4260 * fake entry that is used to 4261 * decrement object->un_pager writecount 4262 * at the appropriate time. Attach 4263 * fake_entry to the deferred list. 4264 */ 4265 fake_entry = vm_map_entry_create(dst_map); 4266 fake_entry->eflags = MAP_ENTRY_WRITECNT; 4267 src_entry->eflags &= ~MAP_ENTRY_WRITECNT; 4268 vm_object_reference(src_object); 4269 fake_entry->object.vm_object = src_object; 4270 fake_entry->start = src_entry->start; 4271 fake_entry->end = src_entry->end; 4272 fake_entry->defer_next = 4273 curthread->td_map_def_user; 4274 curthread->td_map_def_user = fake_entry; 4275 } 4276 4277 pmap_copy(dst_map->pmap, src_map->pmap, 4278 dst_entry->start, dst_entry->end - dst_entry->start, 4279 src_entry->start); 4280 } else { 4281 dst_entry->object.vm_object = NULL; 4282 if ((dst_entry->eflags & MAP_ENTRY_GUARD) == 0) 4283 dst_entry->offset = 0; 4284 if (src_entry->cred != NULL) { 4285 dst_entry->cred = curthread->td_ucred; 4286 crhold(dst_entry->cred); 4287 *fork_charge += size; 4288 } 4289 } 4290 } else { 4291 /* 4292 * We don't want to make writeable wired pages copy-on-write. 4293 * Immediately copy these pages into the new map by simulating 4294 * page faults. The new pages are pageable. 4295 */ 4296 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry, 4297 fork_charge); 4298 } 4299 } 4300 4301 /* 4302 * vmspace_map_entry_forked: 4303 * Update the newly-forked vmspace each time a map entry is inherited 4304 * or copied. The values for vm_dsize and vm_tsize are approximate 4305 * (and mostly-obsolete ideas in the face of mmap(2) et al.) 4306 */ 4307 static void 4308 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2, 4309 vm_map_entry_t entry) 4310 { 4311 vm_size_t entrysize; 4312 vm_offset_t newend; 4313 4314 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 4315 return; 4316 entrysize = entry->end - entry->start; 4317 vm2->vm_map.size += entrysize; 4318 if ((entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0) { 4319 vm2->vm_ssize += btoc(entrysize); 4320 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr && 4321 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) { 4322 newend = MIN(entry->end, 4323 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)); 4324 vm2->vm_dsize += btoc(newend - entry->start); 4325 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr && 4326 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) { 4327 newend = MIN(entry->end, 4328 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)); 4329 vm2->vm_tsize += btoc(newend - entry->start); 4330 } 4331 } 4332 4333 /* 4334 * vmspace_fork: 4335 * Create a new process vmspace structure and vm_map 4336 * based on those of an existing process. The new map 4337 * is based on the old map, according to the inheritance 4338 * values on the regions in that map. 4339 * 4340 * XXX It might be worth coalescing the entries added to the new vmspace. 4341 * 4342 * The source map must not be locked. 4343 */ 4344 struct vmspace * 4345 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge) 4346 { 4347 struct vmspace *vm2; 4348 vm_map_t new_map, old_map; 4349 vm_map_entry_t new_entry, old_entry; 4350 vm_object_t object; 4351 int error, locked __diagused; 4352 vm_inherit_t inh; 4353 4354 old_map = &vm1->vm_map; 4355 /* Copy immutable fields of vm1 to vm2. */ 4356 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map), 4357 pmap_pinit); 4358 if (vm2 == NULL) 4359 return (NULL); 4360 4361 vm2->vm_taddr = vm1->vm_taddr; 4362 vm2->vm_daddr = vm1->vm_daddr; 4363 vm2->vm_maxsaddr = vm1->vm_maxsaddr; 4364 vm2->vm_stacktop = vm1->vm_stacktop; 4365 vm2->vm_shp_base = vm1->vm_shp_base; 4366 vm_map_lock(old_map); 4367 if (old_map->busy) 4368 vm_map_wait_busy(old_map); 4369 new_map = &vm2->vm_map; 4370 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */ 4371 KASSERT(locked, ("vmspace_fork: lock failed")); 4372 4373 error = pmap_vmspace_copy(new_map->pmap, old_map->pmap); 4374 if (error != 0) { 4375 sx_xunlock(&old_map->lock); 4376 sx_xunlock(&new_map->lock); 4377 vm_map_process_deferred(); 4378 vmspace_free(vm2); 4379 return (NULL); 4380 } 4381 4382 new_map->anon_loc = old_map->anon_loc; 4383 new_map->flags |= old_map->flags & (MAP_ASLR | MAP_ASLR_IGNSTART | 4384 MAP_ASLR_STACK | MAP_WXORX); 4385 4386 VM_MAP_ENTRY_FOREACH(old_entry, old_map) { 4387 if ((old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 4388 panic("vm_map_fork: encountered a submap"); 4389 4390 inh = old_entry->inheritance; 4391 if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 && 4392 inh != VM_INHERIT_NONE) 4393 inh = VM_INHERIT_COPY; 4394 4395 switch (inh) { 4396 case VM_INHERIT_NONE: 4397 break; 4398 4399 case VM_INHERIT_SHARE: 4400 /* 4401 * Clone the entry, creating the shared object if 4402 * necessary. 4403 */ 4404 object = old_entry->object.vm_object; 4405 if (object == NULL) { 4406 vm_map_entry_back(old_entry); 4407 object = old_entry->object.vm_object; 4408 } 4409 4410 /* 4411 * Add the reference before calling vm_object_shadow 4412 * to insure that a shadow object is created. 4413 */ 4414 vm_object_reference(object); 4415 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4416 vm_object_shadow(&old_entry->object.vm_object, 4417 &old_entry->offset, 4418 old_entry->end - old_entry->start, 4419 old_entry->cred, 4420 /* Transfer the second reference too. */ 4421 true); 4422 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 4423 old_entry->cred = NULL; 4424 4425 /* 4426 * As in vm_map_merged_neighbor_dispose(), 4427 * the vnode lock will not be acquired in 4428 * this call to vm_object_deallocate(). 4429 */ 4430 vm_object_deallocate(object); 4431 object = old_entry->object.vm_object; 4432 } else { 4433 VM_OBJECT_WLOCK(object); 4434 vm_object_clear_flag(object, OBJ_ONEMAPPING); 4435 if (old_entry->cred != NULL) { 4436 KASSERT(object->cred == NULL, 4437 ("vmspace_fork both cred")); 4438 object->cred = old_entry->cred; 4439 object->charge = old_entry->end - 4440 old_entry->start; 4441 old_entry->cred = NULL; 4442 } 4443 4444 /* 4445 * Assert the correct state of the vnode 4446 * v_writecount while the object is locked, to 4447 * not relock it later for the assertion 4448 * correctness. 4449 */ 4450 if (old_entry->eflags & MAP_ENTRY_WRITECNT && 4451 object->type == OBJT_VNODE) { 4452 KASSERT(((struct vnode *)object-> 4453 handle)->v_writecount > 0, 4454 ("vmspace_fork: v_writecount %p", 4455 object)); 4456 KASSERT(object->un_pager.vnp. 4457 writemappings > 0, 4458 ("vmspace_fork: vnp.writecount %p", 4459 object)); 4460 } 4461 VM_OBJECT_WUNLOCK(object); 4462 } 4463 4464 /* 4465 * Clone the entry, referencing the shared object. 4466 */ 4467 new_entry = vm_map_entry_create(new_map); 4468 *new_entry = *old_entry; 4469 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4470 MAP_ENTRY_IN_TRANSITION); 4471 new_entry->wiring_thread = NULL; 4472 new_entry->wired_count = 0; 4473 if (new_entry->eflags & MAP_ENTRY_WRITECNT) { 4474 vm_pager_update_writecount(object, 4475 new_entry->start, new_entry->end); 4476 } 4477 vm_map_entry_set_vnode_text(new_entry, true); 4478 4479 /* 4480 * Insert the entry into the new map -- we know we're 4481 * inserting at the end of the new map. 4482 */ 4483 vm_map_entry_link(new_map, new_entry); 4484 vmspace_map_entry_forked(vm1, vm2, new_entry); 4485 4486 /* 4487 * Update the physical map 4488 */ 4489 pmap_copy(new_map->pmap, old_map->pmap, 4490 new_entry->start, 4491 (old_entry->end - old_entry->start), 4492 old_entry->start); 4493 break; 4494 4495 case VM_INHERIT_COPY: 4496 /* 4497 * Clone the entry and link into the map. 4498 */ 4499 new_entry = vm_map_entry_create(new_map); 4500 *new_entry = *old_entry; 4501 /* 4502 * Copied entry is COW over the old object. 4503 */ 4504 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4505 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT); 4506 new_entry->wiring_thread = NULL; 4507 new_entry->wired_count = 0; 4508 new_entry->object.vm_object = NULL; 4509 new_entry->cred = NULL; 4510 vm_map_entry_link(new_map, new_entry); 4511 vmspace_map_entry_forked(vm1, vm2, new_entry); 4512 vm_map_copy_entry(old_map, new_map, old_entry, 4513 new_entry, fork_charge); 4514 vm_map_entry_set_vnode_text(new_entry, true); 4515 break; 4516 4517 case VM_INHERIT_ZERO: 4518 /* 4519 * Create a new anonymous mapping entry modelled from 4520 * the old one. 4521 */ 4522 new_entry = vm_map_entry_create(new_map); 4523 memset(new_entry, 0, sizeof(*new_entry)); 4524 4525 new_entry->start = old_entry->start; 4526 new_entry->end = old_entry->end; 4527 new_entry->eflags = old_entry->eflags & 4528 ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION | 4529 MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC | 4530 MAP_ENTRY_SPLIT_BOUNDARY_MASK); 4531 new_entry->protection = old_entry->protection; 4532 new_entry->max_protection = old_entry->max_protection; 4533 new_entry->inheritance = VM_INHERIT_ZERO; 4534 4535 vm_map_entry_link(new_map, new_entry); 4536 vmspace_map_entry_forked(vm1, vm2, new_entry); 4537 4538 new_entry->cred = curthread->td_ucred; 4539 crhold(new_entry->cred); 4540 *fork_charge += (new_entry->end - new_entry->start); 4541 4542 break; 4543 } 4544 } 4545 /* 4546 * Use inlined vm_map_unlock() to postpone handling the deferred 4547 * map entries, which cannot be done until both old_map and 4548 * new_map locks are released. 4549 */ 4550 sx_xunlock(&old_map->lock); 4551 sx_xunlock(&new_map->lock); 4552 vm_map_process_deferred(); 4553 4554 return (vm2); 4555 } 4556 4557 /* 4558 * Create a process's stack for exec_new_vmspace(). This function is never 4559 * asked to wire the newly created stack. 4560 */ 4561 int 4562 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4563 vm_prot_t prot, vm_prot_t max, int cow) 4564 { 4565 vm_size_t growsize, init_ssize; 4566 rlim_t vmemlim; 4567 int rv; 4568 4569 MPASS((map->flags & MAP_WIREFUTURE) == 0); 4570 growsize = sgrowsiz; 4571 init_ssize = (max_ssize < growsize) ? max_ssize : growsize; 4572 vm_map_lock(map); 4573 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4574 /* If we would blow our VMEM resource limit, no go */ 4575 if (map->size + init_ssize > vmemlim) { 4576 rv = KERN_NO_SPACE; 4577 goto out; 4578 } 4579 rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot, 4580 max, cow); 4581 out: 4582 vm_map_unlock(map); 4583 return (rv); 4584 } 4585 4586 static int stack_guard_page = 1; 4587 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN, 4588 &stack_guard_page, 0, 4589 "Specifies the number of guard pages for a stack that grows"); 4590 4591 static int 4592 vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4593 vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow) 4594 { 4595 vm_map_entry_t gap_entry, new_entry, prev_entry; 4596 vm_offset_t bot, gap_bot, gap_top, top; 4597 vm_size_t init_ssize, sgp; 4598 int rv; 4599 4600 KASSERT((cow & MAP_STACK_AREA) != 0, 4601 ("New mapping is not a stack")); 4602 4603 if (max_ssize == 0 || 4604 !vm_map_range_valid(map, addrbos, addrbos + max_ssize)) 4605 return (KERN_INVALID_ADDRESS); 4606 sgp = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4607 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4608 (vm_size_t)stack_guard_page * PAGE_SIZE; 4609 if (sgp >= max_ssize) 4610 return (KERN_INVALID_ARGUMENT); 4611 4612 init_ssize = growsize; 4613 if (max_ssize < init_ssize + sgp) 4614 init_ssize = max_ssize - sgp; 4615 4616 /* If addr is already mapped, no go */ 4617 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) 4618 return (KERN_NO_SPACE); 4619 4620 /* 4621 * If we can't accommodate max_ssize in the current mapping, no go. 4622 */ 4623 if (vm_map_entry_succ(prev_entry)->start < addrbos + max_ssize) 4624 return (KERN_NO_SPACE); 4625 4626 /* 4627 * We initially map a stack of only init_ssize, at the top of 4628 * the range. We will grow as needed later. 4629 * 4630 * Note: we would normally expect prot and max to be VM_PROT_ALL, 4631 * and cow to be 0. Possibly we should eliminate these as input 4632 * parameters, and just pass these values here in the insert call. 4633 */ 4634 bot = addrbos + max_ssize - init_ssize; 4635 top = bot + init_ssize; 4636 gap_bot = addrbos; 4637 gap_top = bot; 4638 rv = vm_map_insert1(map, NULL, 0, bot, top, prot, max, cow, 4639 &new_entry); 4640 if (rv != KERN_SUCCESS) 4641 return (rv); 4642 KASSERT(new_entry->end == top || new_entry->start == bot, 4643 ("Bad entry start/end for new stack entry")); 4644 KASSERT((new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0, 4645 ("new entry lacks MAP_ENTRY_GROWS_DOWN")); 4646 if (gap_bot == gap_top) 4647 return (KERN_SUCCESS); 4648 rv = vm_map_insert1(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE, 4649 VM_PROT_NONE, MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP, 4650 &gap_entry); 4651 if (rv == KERN_SUCCESS) { 4652 KASSERT((gap_entry->eflags & MAP_ENTRY_GUARD) != 0, 4653 ("entry %p not gap %#x", gap_entry, gap_entry->eflags)); 4654 KASSERT((gap_entry->eflags & MAP_ENTRY_STACK_GAP) != 0, 4655 ("entry %p not stack gap %#x", gap_entry, 4656 gap_entry->eflags)); 4657 4658 /* 4659 * Gap can never successfully handle a fault, so 4660 * read-ahead logic is never used for it. Re-use 4661 * next_read of the gap entry to store 4662 * stack_guard_page for vm_map_growstack(). 4663 * Similarly, since a gap cannot have a backing object, 4664 * store the original stack protections in the 4665 * object offset. 4666 */ 4667 gap_entry->next_read = sgp; 4668 gap_entry->offset = prot | PROT_MAX(max); 4669 } else { 4670 (void)vm_map_delete(map, bot, top); 4671 } 4672 return (rv); 4673 } 4674 4675 /* 4676 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we 4677 * successfully grow the stack. 4678 */ 4679 static int 4680 vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry) 4681 { 4682 vm_map_entry_t stack_entry; 4683 struct proc *p; 4684 struct vmspace *vm; 4685 vm_offset_t gap_end, gap_start, grow_start; 4686 vm_size_t grow_amount, guard, max_grow, sgp; 4687 vm_prot_t prot, max; 4688 rlim_t lmemlim, stacklim, vmemlim; 4689 int rv, rv1 __diagused; 4690 bool gap_deleted, is_procstack; 4691 #ifdef notyet 4692 uint64_t limit; 4693 #endif 4694 #ifdef RACCT 4695 int error __diagused; 4696 #endif 4697 4698 p = curproc; 4699 vm = p->p_vmspace; 4700 4701 /* 4702 * Disallow stack growth when the access is performed by a 4703 * debugger or AIO daemon. The reason is that the wrong 4704 * resource limits are applied. 4705 */ 4706 if (p != initproc && (map != &p->p_vmspace->vm_map || 4707 p->p_textvp == NULL)) 4708 return (KERN_FAILURE); 4709 4710 MPASS(!vm_map_is_system(map)); 4711 4712 lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK); 4713 stacklim = lim_cur(curthread, RLIMIT_STACK); 4714 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4715 retry: 4716 /* If addr is not in a hole for a stack grow area, no need to grow. */ 4717 if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry)) 4718 return (KERN_FAILURE); 4719 if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0) 4720 return (KERN_SUCCESS); 4721 if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP) != 0) { 4722 stack_entry = vm_map_entry_succ(gap_entry); 4723 if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 || 4724 stack_entry->start != gap_entry->end) 4725 return (KERN_FAILURE); 4726 grow_amount = round_page(stack_entry->start - addr); 4727 } else { 4728 return (KERN_FAILURE); 4729 } 4730 guard = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4731 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4732 gap_entry->next_read; 4733 max_grow = gap_entry->end - gap_entry->start; 4734 if (guard > max_grow) 4735 return (KERN_NO_SPACE); 4736 max_grow -= guard; 4737 if (grow_amount > max_grow) 4738 return (KERN_NO_SPACE); 4739 4740 /* 4741 * If this is the main process stack, see if we're over the stack 4742 * limit. 4743 */ 4744 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr && 4745 addr < (vm_offset_t)vm->vm_stacktop; 4746 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) 4747 return (KERN_NO_SPACE); 4748 4749 #ifdef RACCT 4750 if (racct_enable) { 4751 PROC_LOCK(p); 4752 if (is_procstack && racct_set(p, RACCT_STACK, 4753 ctob(vm->vm_ssize) + grow_amount)) { 4754 PROC_UNLOCK(p); 4755 return (KERN_NO_SPACE); 4756 } 4757 PROC_UNLOCK(p); 4758 } 4759 #endif 4760 4761 grow_amount = roundup(grow_amount, sgrowsiz); 4762 if (grow_amount > max_grow) 4763 grow_amount = max_grow; 4764 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) { 4765 grow_amount = trunc_page((vm_size_t)stacklim) - 4766 ctob(vm->vm_ssize); 4767 } 4768 4769 #ifdef notyet 4770 PROC_LOCK(p); 4771 limit = racct_get_available(p, RACCT_STACK); 4772 PROC_UNLOCK(p); 4773 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit)) 4774 grow_amount = limit - ctob(vm->vm_ssize); 4775 #endif 4776 4777 if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) { 4778 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) { 4779 rv = KERN_NO_SPACE; 4780 goto out; 4781 } 4782 #ifdef RACCT 4783 if (racct_enable) { 4784 PROC_LOCK(p); 4785 if (racct_set(p, RACCT_MEMLOCK, 4786 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) { 4787 PROC_UNLOCK(p); 4788 rv = KERN_NO_SPACE; 4789 goto out; 4790 } 4791 PROC_UNLOCK(p); 4792 } 4793 #endif 4794 } 4795 4796 /* If we would blow our VMEM resource limit, no go */ 4797 if (map->size + grow_amount > vmemlim) { 4798 rv = KERN_NO_SPACE; 4799 goto out; 4800 } 4801 #ifdef RACCT 4802 if (racct_enable) { 4803 PROC_LOCK(p); 4804 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) { 4805 PROC_UNLOCK(p); 4806 rv = KERN_NO_SPACE; 4807 goto out; 4808 } 4809 PROC_UNLOCK(p); 4810 } 4811 #endif 4812 4813 if (vm_map_lock_upgrade(map)) { 4814 gap_entry = NULL; 4815 vm_map_lock_read(map); 4816 goto retry; 4817 } 4818 4819 /* 4820 * The gap_entry "offset" field is overloaded. See 4821 * vm_map_stack_locked(). 4822 */ 4823 prot = PROT_EXTRACT(gap_entry->offset); 4824 max = PROT_MAX_EXTRACT(gap_entry->offset); 4825 sgp = gap_entry->next_read; 4826 4827 grow_start = gap_entry->end - grow_amount; 4828 if (gap_entry->start + grow_amount == gap_entry->end) { 4829 gap_start = gap_entry->start; 4830 gap_end = gap_entry->end; 4831 vm_map_entry_delete(map, gap_entry); 4832 gap_deleted = true; 4833 } else { 4834 MPASS(gap_entry->start < gap_entry->end - grow_amount); 4835 vm_map_entry_resize(map, gap_entry, -grow_amount); 4836 gap_deleted = false; 4837 } 4838 rv = vm_map_insert(map, NULL, 0, grow_start, 4839 grow_start + grow_amount, prot, max, MAP_STACK_AREA); 4840 if (rv != KERN_SUCCESS) { 4841 if (gap_deleted) { 4842 rv1 = vm_map_insert1(map, NULL, 0, gap_start, 4843 gap_end, VM_PROT_NONE, VM_PROT_NONE, 4844 MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP, 4845 &gap_entry); 4846 MPASS(rv1 == KERN_SUCCESS); 4847 gap_entry->next_read = sgp; 4848 gap_entry->offset = prot | PROT_MAX(max); 4849 } else { 4850 vm_map_entry_resize(map, gap_entry, 4851 grow_amount); 4852 } 4853 } 4854 if (rv == KERN_SUCCESS && is_procstack) 4855 vm->vm_ssize += btoc(grow_amount); 4856 4857 /* 4858 * Heed the MAP_WIREFUTURE flag if it was set for this process. 4859 */ 4860 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) { 4861 rv = vm_map_wire_locked(map, grow_start, 4862 grow_start + grow_amount, 4863 VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); 4864 } 4865 vm_map_lock_downgrade(map); 4866 4867 out: 4868 #ifdef RACCT 4869 if (racct_enable && rv != KERN_SUCCESS) { 4870 PROC_LOCK(p); 4871 error = racct_set(p, RACCT_VMEM, map->size); 4872 KASSERT(error == 0, ("decreasing RACCT_VMEM failed")); 4873 if (!old_mlock) { 4874 error = racct_set(p, RACCT_MEMLOCK, 4875 ptoa(pmap_wired_count(map->pmap))); 4876 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed")); 4877 } 4878 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize)); 4879 KASSERT(error == 0, ("decreasing RACCT_STACK failed")); 4880 PROC_UNLOCK(p); 4881 } 4882 #endif 4883 4884 return (rv); 4885 } 4886 4887 /* 4888 * Unshare the specified VM space for exec. If other processes are 4889 * mapped to it, then create a new one. The new vmspace is null. 4890 */ 4891 int 4892 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser) 4893 { 4894 struct vmspace *oldvmspace = p->p_vmspace; 4895 struct vmspace *newvmspace; 4896 4897 KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0, 4898 ("vmspace_exec recursed")); 4899 newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit); 4900 if (newvmspace == NULL) 4901 return (ENOMEM); 4902 newvmspace->vm_swrss = oldvmspace->vm_swrss; 4903 /* 4904 * This code is written like this for prototype purposes. The 4905 * goal is to avoid running down the vmspace here, but let the 4906 * other process's that are still using the vmspace to finally 4907 * run it down. Even though there is little or no chance of blocking 4908 * here, it is a good idea to keep this form for future mods. 4909 */ 4910 PROC_VMSPACE_LOCK(p); 4911 p->p_vmspace = newvmspace; 4912 PROC_VMSPACE_UNLOCK(p); 4913 if (p == curthread->td_proc) 4914 pmap_activate(curthread); 4915 curthread->td_pflags |= TDP_EXECVMSPC; 4916 return (0); 4917 } 4918 4919 /* 4920 * Unshare the specified VM space for forcing COW. This 4921 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 4922 */ 4923 int 4924 vmspace_unshare(struct proc *p) 4925 { 4926 struct vmspace *oldvmspace = p->p_vmspace; 4927 struct vmspace *newvmspace; 4928 vm_ooffset_t fork_charge; 4929 4930 /* 4931 * The caller is responsible for ensuring that the reference count 4932 * cannot concurrently transition 1 -> 2. 4933 */ 4934 if (refcount_load(&oldvmspace->vm_refcnt) == 1) 4935 return (0); 4936 fork_charge = 0; 4937 newvmspace = vmspace_fork(oldvmspace, &fork_charge); 4938 if (newvmspace == NULL) 4939 return (ENOMEM); 4940 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) { 4941 vmspace_free(newvmspace); 4942 return (ENOMEM); 4943 } 4944 PROC_VMSPACE_LOCK(p); 4945 p->p_vmspace = newvmspace; 4946 PROC_VMSPACE_UNLOCK(p); 4947 if (p == curthread->td_proc) 4948 pmap_activate(curthread); 4949 vmspace_free(oldvmspace); 4950 return (0); 4951 } 4952 4953 /* 4954 * vm_map_lookup: 4955 * 4956 * Finds the VM object, offset, and 4957 * protection for a given virtual address in the 4958 * specified map, assuming a page fault of the 4959 * type specified. 4960 * 4961 * Leaves the map in question locked for read; return 4962 * values are guaranteed until a vm_map_lookup_done 4963 * call is performed. Note that the map argument 4964 * is in/out; the returned map must be used in 4965 * the call to vm_map_lookup_done. 4966 * 4967 * A handle (out_entry) is returned for use in 4968 * vm_map_lookup_done, to make that fast. 4969 * 4970 * If a lookup is requested with "write protection" 4971 * specified, the map may be changed to perform virtual 4972 * copying operations, although the data referenced will 4973 * remain the same. 4974 */ 4975 int 4976 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4977 vm_offset_t vaddr, 4978 vm_prot_t fault_typea, 4979 vm_map_entry_t *out_entry, /* OUT */ 4980 vm_object_t *object, /* OUT */ 4981 vm_pindex_t *pindex, /* OUT */ 4982 vm_prot_t *out_prot, /* OUT */ 4983 boolean_t *wired) /* OUT */ 4984 { 4985 vm_map_entry_t entry; 4986 vm_map_t map = *var_map; 4987 vm_prot_t prot; 4988 vm_prot_t fault_type; 4989 vm_object_t eobject; 4990 vm_size_t size; 4991 struct ucred *cred; 4992 4993 RetryLookup: 4994 4995 vm_map_lock_read(map); 4996 4997 RetryLookupLocked: 4998 /* 4999 * Lookup the faulting address. 5000 */ 5001 if (!vm_map_lookup_entry(map, vaddr, out_entry)) { 5002 vm_map_unlock_read(map); 5003 return (KERN_INVALID_ADDRESS); 5004 } 5005 5006 entry = *out_entry; 5007 5008 /* 5009 * Handle submaps. 5010 */ 5011 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 5012 vm_map_t old_map = map; 5013 5014 *var_map = map = entry->object.sub_map; 5015 vm_map_unlock_read(old_map); 5016 goto RetryLookup; 5017 } 5018 5019 /* 5020 * Check whether this task is allowed to have this page. 5021 */ 5022 prot = entry->protection; 5023 if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) { 5024 fault_typea &= ~VM_PROT_FAULT_LOOKUP; 5025 if (prot == VM_PROT_NONE && map != kernel_map && 5026 (entry->eflags & MAP_ENTRY_GUARD) != 0 && 5027 (entry->eflags & MAP_ENTRY_STACK_GAP) != 0 && 5028 vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS) 5029 goto RetryLookupLocked; 5030 } 5031 fault_type = fault_typea & VM_PROT_ALL; 5032 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) { 5033 vm_map_unlock_read(map); 5034 return (KERN_PROTECTION_FAILURE); 5035 } 5036 KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags & 5037 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) != 5038 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY), 5039 ("entry %p flags %x", entry, entry->eflags)); 5040 if ((fault_typea & VM_PROT_COPY) != 0 && 5041 (entry->max_protection & VM_PROT_WRITE) == 0 && 5042 (entry->eflags & MAP_ENTRY_COW) == 0) { 5043 vm_map_unlock_read(map); 5044 return (KERN_PROTECTION_FAILURE); 5045 } 5046 5047 /* 5048 * If this page is not pageable, we have to get it for all possible 5049 * accesses. 5050 */ 5051 *wired = (entry->wired_count != 0); 5052 if (*wired) 5053 fault_type = entry->protection; 5054 size = entry->end - entry->start; 5055 5056 /* 5057 * If the entry was copy-on-write, we either ... 5058 */ 5059 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 5060 /* 5061 * If we want to write the page, we may as well handle that 5062 * now since we've got the map locked. 5063 * 5064 * If we don't need to write the page, we just demote the 5065 * permissions allowed. 5066 */ 5067 if ((fault_type & VM_PROT_WRITE) != 0 || 5068 (fault_typea & VM_PROT_COPY) != 0) { 5069 /* 5070 * Make a new object, and place it in the object 5071 * chain. Note that no new references have appeared 5072 * -- one just moved from the map to the new 5073 * object. 5074 */ 5075 if (vm_map_lock_upgrade(map)) 5076 goto RetryLookup; 5077 5078 if (entry->cred == NULL) { 5079 /* 5080 * The debugger owner is charged for 5081 * the memory. 5082 */ 5083 cred = curthread->td_ucred; 5084 crhold(cred); 5085 if (!swap_reserve_by_cred(size, cred)) { 5086 crfree(cred); 5087 vm_map_unlock(map); 5088 return (KERN_RESOURCE_SHORTAGE); 5089 } 5090 entry->cred = cred; 5091 } 5092 eobject = entry->object.vm_object; 5093 vm_object_shadow(&entry->object.vm_object, 5094 &entry->offset, size, entry->cred, false); 5095 if (eobject == entry->object.vm_object) { 5096 /* 5097 * The object was not shadowed. 5098 */ 5099 swap_release_by_cred(size, entry->cred); 5100 crfree(entry->cred); 5101 } 5102 entry->cred = NULL; 5103 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 5104 5105 vm_map_lock_downgrade(map); 5106 } else { 5107 /* 5108 * We're attempting to read a copy-on-write page -- 5109 * don't allow writes. 5110 */ 5111 prot &= ~VM_PROT_WRITE; 5112 } 5113 } 5114 5115 /* 5116 * Create an object if necessary. 5117 */ 5118 if (entry->object.vm_object == NULL && !vm_map_is_system(map)) { 5119 if (vm_map_lock_upgrade(map)) 5120 goto RetryLookup; 5121 entry->object.vm_object = vm_object_allocate_anon(atop(size), 5122 NULL, entry->cred, size); 5123 entry->offset = 0; 5124 entry->cred = NULL; 5125 vm_map_lock_downgrade(map); 5126 } 5127 5128 /* 5129 * Return the object/offset from this entry. If the entry was 5130 * copy-on-write or empty, it has been fixed up. 5131 */ 5132 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 5133 *object = entry->object.vm_object; 5134 5135 *out_prot = prot; 5136 return (KERN_SUCCESS); 5137 } 5138 5139 /* 5140 * vm_map_lookup_locked: 5141 * 5142 * Lookup the faulting address. A version of vm_map_lookup that returns 5143 * KERN_FAILURE instead of blocking on map lock or memory allocation. 5144 */ 5145 int 5146 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */ 5147 vm_offset_t vaddr, 5148 vm_prot_t fault_typea, 5149 vm_map_entry_t *out_entry, /* OUT */ 5150 vm_object_t *object, /* OUT */ 5151 vm_pindex_t *pindex, /* OUT */ 5152 vm_prot_t *out_prot, /* OUT */ 5153 boolean_t *wired) /* OUT */ 5154 { 5155 vm_map_entry_t entry; 5156 vm_map_t map = *var_map; 5157 vm_prot_t prot; 5158 vm_prot_t fault_type = fault_typea; 5159 5160 /* 5161 * Lookup the faulting address. 5162 */ 5163 if (!vm_map_lookup_entry(map, vaddr, out_entry)) 5164 return (KERN_INVALID_ADDRESS); 5165 5166 entry = *out_entry; 5167 5168 /* 5169 * Fail if the entry refers to a submap. 5170 */ 5171 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) 5172 return (KERN_FAILURE); 5173 5174 /* 5175 * Check whether this task is allowed to have this page. 5176 */ 5177 prot = entry->protection; 5178 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 5179 if ((fault_type & prot) != fault_type) 5180 return (KERN_PROTECTION_FAILURE); 5181 5182 /* 5183 * If this page is not pageable, we have to get it for all possible 5184 * accesses. 5185 */ 5186 *wired = (entry->wired_count != 0); 5187 if (*wired) 5188 fault_type = entry->protection; 5189 5190 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 5191 /* 5192 * Fail if the entry was copy-on-write for a write fault. 5193 */ 5194 if (fault_type & VM_PROT_WRITE) 5195 return (KERN_FAILURE); 5196 /* 5197 * We're attempting to read a copy-on-write page -- 5198 * don't allow writes. 5199 */ 5200 prot &= ~VM_PROT_WRITE; 5201 } 5202 5203 /* 5204 * Fail if an object should be created. 5205 */ 5206 if (entry->object.vm_object == NULL && !vm_map_is_system(map)) 5207 return (KERN_FAILURE); 5208 5209 /* 5210 * Return the object/offset from this entry. If the entry was 5211 * copy-on-write or empty, it has been fixed up. 5212 */ 5213 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 5214 *object = entry->object.vm_object; 5215 5216 *out_prot = prot; 5217 return (KERN_SUCCESS); 5218 } 5219 5220 /* 5221 * vm_map_lookup_done: 5222 * 5223 * Releases locks acquired by a vm_map_lookup 5224 * (according to the handle returned by that lookup). 5225 */ 5226 void 5227 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry) 5228 { 5229 /* 5230 * Unlock the main-level map 5231 */ 5232 vm_map_unlock_read(map); 5233 } 5234 5235 vm_offset_t 5236 vm_map_max_KBI(const struct vm_map *map) 5237 { 5238 5239 return (vm_map_max(map)); 5240 } 5241 5242 vm_offset_t 5243 vm_map_min_KBI(const struct vm_map *map) 5244 { 5245 5246 return (vm_map_min(map)); 5247 } 5248 5249 pmap_t 5250 vm_map_pmap_KBI(vm_map_t map) 5251 { 5252 5253 return (map->pmap); 5254 } 5255 5256 bool 5257 vm_map_range_valid_KBI(vm_map_t map, vm_offset_t start, vm_offset_t end) 5258 { 5259 5260 return (vm_map_range_valid(map, start, end)); 5261 } 5262 5263 #ifdef INVARIANTS 5264 static void 5265 _vm_map_assert_consistent(vm_map_t map, int check) 5266 { 5267 vm_map_entry_t entry, prev; 5268 vm_map_entry_t cur, header, lbound, ubound; 5269 vm_size_t max_left, max_right; 5270 5271 #ifdef DIAGNOSTIC 5272 ++map->nupdates; 5273 #endif 5274 if (enable_vmmap_check != check) 5275 return; 5276 5277 header = prev = &map->header; 5278 VM_MAP_ENTRY_FOREACH(entry, map) { 5279 KASSERT(prev->end <= entry->start, 5280 ("map %p prev->end = %jx, start = %jx", map, 5281 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5282 KASSERT(entry->start < entry->end, 5283 ("map %p start = %jx, end = %jx", map, 5284 (uintmax_t)entry->start, (uintmax_t)entry->end)); 5285 KASSERT(entry->left == header || 5286 entry->left->start < entry->start, 5287 ("map %p left->start = %jx, start = %jx", map, 5288 (uintmax_t)entry->left->start, (uintmax_t)entry->start)); 5289 KASSERT(entry->right == header || 5290 entry->start < entry->right->start, 5291 ("map %p start = %jx, right->start = %jx", map, 5292 (uintmax_t)entry->start, (uintmax_t)entry->right->start)); 5293 cur = map->root; 5294 lbound = ubound = header; 5295 for (;;) { 5296 if (entry->start < cur->start) { 5297 ubound = cur; 5298 cur = cur->left; 5299 KASSERT(cur != lbound, 5300 ("map %p cannot find %jx", 5301 map, (uintmax_t)entry->start)); 5302 } else if (cur->end <= entry->start) { 5303 lbound = cur; 5304 cur = cur->right; 5305 KASSERT(cur != ubound, 5306 ("map %p cannot find %jx", 5307 map, (uintmax_t)entry->start)); 5308 } else { 5309 KASSERT(cur == entry, 5310 ("map %p cannot find %jx", 5311 map, (uintmax_t)entry->start)); 5312 break; 5313 } 5314 } 5315 max_left = vm_map_entry_max_free_left(entry, lbound); 5316 max_right = vm_map_entry_max_free_right(entry, ubound); 5317 KASSERT(entry->max_free == vm_size_max(max_left, max_right), 5318 ("map %p max = %jx, max_left = %jx, max_right = %jx", map, 5319 (uintmax_t)entry->max_free, 5320 (uintmax_t)max_left, (uintmax_t)max_right)); 5321 prev = entry; 5322 } 5323 KASSERT(prev->end <= entry->start, 5324 ("map %p prev->end = %jx, start = %jx", map, 5325 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5326 } 5327 #endif 5328 5329 #include "opt_ddb.h" 5330 #ifdef DDB 5331 #include <sys/kernel.h> 5332 5333 #include <ddb/ddb.h> 5334 5335 static void 5336 vm_map_print(vm_map_t map) 5337 { 5338 vm_map_entry_t entry, prev; 5339 5340 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 5341 (void *)map, 5342 (void *)map->pmap, map->nentries, map->timestamp); 5343 5344 db_indent += 2; 5345 prev = &map->header; 5346 VM_MAP_ENTRY_FOREACH(entry, map) { 5347 db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n", 5348 (void *)entry, (void *)entry->start, (void *)entry->end, 5349 entry->eflags); 5350 { 5351 static const char * const inheritance_name[4] = 5352 {"share", "copy", "none", "donate_copy"}; 5353 5354 db_iprintf(" prot=%x/%x/%s", 5355 entry->protection, 5356 entry->max_protection, 5357 inheritance_name[(int)(unsigned char) 5358 entry->inheritance]); 5359 if (entry->wired_count != 0) 5360 db_printf(", wired"); 5361 } 5362 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 5363 db_printf(", share=%p, offset=0x%jx\n", 5364 (void *)entry->object.sub_map, 5365 (uintmax_t)entry->offset); 5366 if (prev == &map->header || 5367 prev->object.sub_map != 5368 entry->object.sub_map) { 5369 db_indent += 2; 5370 vm_map_print((vm_map_t)entry->object.sub_map); 5371 db_indent -= 2; 5372 } 5373 } else { 5374 if (entry->cred != NULL) 5375 db_printf(", ruid %d", entry->cred->cr_ruid); 5376 db_printf(", object=%p, offset=0x%jx", 5377 (void *)entry->object.vm_object, 5378 (uintmax_t)entry->offset); 5379 if (entry->object.vm_object && entry->object.vm_object->cred) 5380 db_printf(", obj ruid %d charge %jx", 5381 entry->object.vm_object->cred->cr_ruid, 5382 (uintmax_t)entry->object.vm_object->charge); 5383 if (entry->eflags & MAP_ENTRY_COW) 5384 db_printf(", copy (%s)", 5385 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 5386 db_printf("\n"); 5387 5388 if (prev == &map->header || 5389 prev->object.vm_object != 5390 entry->object.vm_object) { 5391 db_indent += 2; 5392 vm_object_print((db_expr_t)(intptr_t) 5393 entry->object.vm_object, 5394 0, 0, (char *)0); 5395 db_indent -= 2; 5396 } 5397 } 5398 prev = entry; 5399 } 5400 db_indent -= 2; 5401 } 5402 5403 DB_SHOW_COMMAND(map, map) 5404 { 5405 5406 if (!have_addr) { 5407 db_printf("usage: show map <addr>\n"); 5408 return; 5409 } 5410 vm_map_print((vm_map_t)addr); 5411 } 5412 5413 DB_SHOW_COMMAND(procvm, procvm) 5414 { 5415 struct proc *p; 5416 5417 if (have_addr) { 5418 p = db_lookup_proc(addr); 5419 } else { 5420 p = curproc; 5421 } 5422 5423 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 5424 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 5425 (void *)vmspace_pmap(p->p_vmspace)); 5426 5427 vm_map_print((vm_map_t)&p->p_vmspace->vm_map); 5428 } 5429 5430 #endif /* DDB */ 5431