1 /*- 2 * SPDX-License-Identifier: (BSD-3-Clause AND MIT-CMU) 3 * 4 * Copyright (c) 1991, 1993 5 * The Regents of the University of California. All rights reserved. 6 * 7 * This code is derived from software contributed to Berkeley by 8 * The Mach Operating System project at Carnegie-Mellon University. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94 35 * 36 * 37 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 38 * All rights reserved. 39 * 40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 41 * 42 * Permission to use, copy, modify and distribute this software and 43 * its documentation is hereby granted, provided that both the copyright 44 * notice and this permission notice appear in all copies of the 45 * software, derivative works or modified versions, and any portions 46 * thereof, and that both notices appear in supporting documentation. 47 * 48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 51 * 52 * Carnegie Mellon requests users of this software to return to 53 * 54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 55 * School of Computer Science 56 * Carnegie Mellon University 57 * Pittsburgh PA 15213-3890 58 * 59 * any improvements or extensions that they make and grant Carnegie the 60 * rights to redistribute these changes. 61 */ 62 63 /* 64 * Virtual memory mapping module. 65 */ 66 67 #include <sys/cdefs.h> 68 __FBSDID("$FreeBSD$"); 69 70 #include <sys/param.h> 71 #include <sys/systm.h> 72 #include <sys/elf.h> 73 #include <sys/kernel.h> 74 #include <sys/ktr.h> 75 #include <sys/lock.h> 76 #include <sys/mutex.h> 77 #include <sys/proc.h> 78 #include <sys/vmmeter.h> 79 #include <sys/mman.h> 80 #include <sys/vnode.h> 81 #include <sys/racct.h> 82 #include <sys/resourcevar.h> 83 #include <sys/rwlock.h> 84 #include <sys/file.h> 85 #include <sys/sysctl.h> 86 #include <sys/sysent.h> 87 #include <sys/shm.h> 88 89 #include <vm/vm.h> 90 #include <vm/vm_param.h> 91 #include <vm/pmap.h> 92 #include <vm/vm_map.h> 93 #include <vm/vm_page.h> 94 #include <vm/vm_pageout.h> 95 #include <vm/vm_object.h> 96 #include <vm/vm_pager.h> 97 #include <vm/vm_kern.h> 98 #include <vm/vm_extern.h> 99 #include <vm/vnode_pager.h> 100 #include <vm/swap_pager.h> 101 #include <vm/uma.h> 102 103 /* 104 * Virtual memory maps provide for the mapping, protection, 105 * and sharing of virtual memory objects. In addition, 106 * this module provides for an efficient virtual copy of 107 * memory from one map to another. 108 * 109 * Synchronization is required prior to most operations. 110 * 111 * Maps consist of an ordered doubly-linked list of simple 112 * entries; a self-adjusting binary search tree of these 113 * entries is used to speed up lookups. 114 * 115 * Since portions of maps are specified by start/end addresses, 116 * which may not align with existing map entries, all 117 * routines merely "clip" entries to these start/end values. 118 * [That is, an entry is split into two, bordering at a 119 * start or end value.] Note that these clippings may not 120 * always be necessary (as the two resulting entries are then 121 * not changed); however, the clipping is done for convenience. 122 * 123 * As mentioned above, virtual copy operations are performed 124 * by copying VM object references from one map to 125 * another, and then marking both regions as copy-on-write. 126 */ 127 128 static struct mtx map_sleep_mtx; 129 static uma_zone_t mapentzone; 130 static uma_zone_t kmapentzone; 131 static uma_zone_t vmspace_zone; 132 static int vmspace_zinit(void *mem, int size, int flags); 133 static void _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, 134 vm_offset_t max); 135 static void vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map); 136 static void vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry); 137 static void vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry); 138 static int vm_map_growstack(vm_map_t map, vm_offset_t addr, 139 vm_map_entry_t gap_entry); 140 static void vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 141 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags); 142 #ifdef INVARIANTS 143 static void vmspace_zdtor(void *mem, int size, void *arg); 144 #endif 145 static int vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, 146 vm_size_t max_ssize, vm_size_t growsize, vm_prot_t prot, vm_prot_t max, 147 int cow); 148 static void vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 149 vm_offset_t failed_addr); 150 151 #define ENTRY_CHARGED(e) ((e)->cred != NULL || \ 152 ((e)->object.vm_object != NULL && (e)->object.vm_object->cred != NULL && \ 153 !((e)->eflags & MAP_ENTRY_NEEDS_COPY))) 154 155 /* 156 * PROC_VMSPACE_{UN,}LOCK() can be a noop as long as vmspaces are type 157 * stable. 158 */ 159 #define PROC_VMSPACE_LOCK(p) do { } while (0) 160 #define PROC_VMSPACE_UNLOCK(p) do { } while (0) 161 162 /* 163 * VM_MAP_RANGE_CHECK: [ internal use only ] 164 * 165 * Asserts that the starting and ending region 166 * addresses fall within the valid range of the map. 167 */ 168 #define VM_MAP_RANGE_CHECK(map, start, end) \ 169 { \ 170 if (start < vm_map_min(map)) \ 171 start = vm_map_min(map); \ 172 if (end > vm_map_max(map)) \ 173 end = vm_map_max(map); \ 174 if (start > end) \ 175 start = end; \ 176 } 177 178 #ifndef UMA_MD_SMALL_ALLOC 179 180 /* 181 * Allocate a new slab for kernel map entries. The kernel map may be locked or 182 * unlocked, depending on whether the request is coming from the kernel map or a 183 * submap. This function allocates a virtual address range directly from the 184 * kernel map instead of the kmem_* layer to avoid recursion on the kernel map 185 * lock and also to avoid triggering allocator recursion in the vmem boundary 186 * tag allocator. 187 */ 188 static void * 189 kmapent_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag, 190 int wait) 191 { 192 vm_offset_t addr; 193 int error, locked; 194 195 *pflag = UMA_SLAB_PRIV; 196 197 if (!(locked = vm_map_locked(kernel_map))) 198 vm_map_lock(kernel_map); 199 addr = vm_map_findspace(kernel_map, vm_map_min(kernel_map), bytes); 200 if (addr + bytes < addr || addr + bytes > vm_map_max(kernel_map)) 201 panic("%s: kernel map is exhausted", __func__); 202 error = vm_map_insert(kernel_map, NULL, 0, addr, addr + bytes, 203 VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); 204 if (error != KERN_SUCCESS) 205 panic("%s: vm_map_insert() failed: %d", __func__, error); 206 if (!locked) 207 vm_map_unlock(kernel_map); 208 error = kmem_back_domain(domain, kernel_object, addr, bytes, M_NOWAIT | 209 M_USE_RESERVE | (wait & M_ZERO)); 210 if (error == KERN_SUCCESS) { 211 return ((void *)addr); 212 } else { 213 if (!locked) 214 vm_map_lock(kernel_map); 215 vm_map_delete(kernel_map, addr, bytes); 216 if (!locked) 217 vm_map_unlock(kernel_map); 218 return (NULL); 219 } 220 } 221 222 static void 223 kmapent_free(void *item, vm_size_t size, uint8_t pflag) 224 { 225 vm_offset_t addr; 226 int error; 227 228 if ((pflag & UMA_SLAB_PRIV) == 0) 229 /* XXX leaked */ 230 return; 231 232 addr = (vm_offset_t)item; 233 kmem_unback(kernel_object, addr, size); 234 error = vm_map_remove(kernel_map, addr, addr + size); 235 KASSERT(error == KERN_SUCCESS, 236 ("%s: vm_map_remove failed: %d", __func__, error)); 237 } 238 239 /* 240 * The worst-case upper bound on the number of kernel map entries that may be 241 * created before the zone must be replenished in _vm_map_unlock(). 242 */ 243 #define KMAPENT_RESERVE 1 244 245 #endif /* !UMD_MD_SMALL_ALLOC */ 246 247 /* 248 * vm_map_startup: 249 * 250 * Initialize the vm_map module. Must be called before any other vm_map 251 * routines. 252 * 253 * User map and entry structures are allocated from the general purpose 254 * memory pool. Kernel maps are statically defined. Kernel map entries 255 * require special handling to avoid recursion; see the comments above 256 * kmapent_alloc() and in vm_map_entry_create(). 257 */ 258 void 259 vm_map_startup(void) 260 { 261 mtx_init(&map_sleep_mtx, "vm map sleep mutex", NULL, MTX_DEF); 262 263 /* 264 * Disable the use of per-CPU buckets: map entry allocation is 265 * serialized by the kernel map lock. 266 */ 267 kmapentzone = uma_zcreate("KMAP ENTRY", sizeof(struct vm_map_entry), 268 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 269 UMA_ZONE_VM | UMA_ZONE_NOBUCKET); 270 #ifndef UMA_MD_SMALL_ALLOC 271 /* Reserve an extra map entry for use when replenishing the reserve. */ 272 uma_zone_reserve(kmapentzone, KMAPENT_RESERVE + 1); 273 uma_prealloc(kmapentzone, KMAPENT_RESERVE + 1); 274 uma_zone_set_allocf(kmapentzone, kmapent_alloc); 275 uma_zone_set_freef(kmapentzone, kmapent_free); 276 #endif 277 278 mapentzone = uma_zcreate("MAP ENTRY", sizeof(struct vm_map_entry), 279 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); 280 vmspace_zone = uma_zcreate("VMSPACE", sizeof(struct vmspace), NULL, 281 #ifdef INVARIANTS 282 vmspace_zdtor, 283 #else 284 NULL, 285 #endif 286 vmspace_zinit, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE); 287 } 288 289 static int 290 vmspace_zinit(void *mem, int size, int flags) 291 { 292 struct vmspace *vm; 293 vm_map_t map; 294 295 vm = (struct vmspace *)mem; 296 map = &vm->vm_map; 297 298 memset(map, 0, sizeof(*map)); 299 mtx_init(&map->system_mtx, "vm map (system)", NULL, 300 MTX_DEF | MTX_DUPOK); 301 sx_init(&map->lock, "vm map (user)"); 302 PMAP_LOCK_INIT(vmspace_pmap(vm)); 303 return (0); 304 } 305 306 #ifdef INVARIANTS 307 static void 308 vmspace_zdtor(void *mem, int size, void *arg) 309 { 310 struct vmspace *vm; 311 312 vm = (struct vmspace *)mem; 313 KASSERT(vm->vm_map.nentries == 0, 314 ("vmspace %p nentries == %d on free", vm, vm->vm_map.nentries)); 315 KASSERT(vm->vm_map.size == 0, 316 ("vmspace %p size == %ju on free", vm, (uintmax_t)vm->vm_map.size)); 317 } 318 #endif /* INVARIANTS */ 319 320 /* 321 * Allocate a vmspace structure, including a vm_map and pmap, 322 * and initialize those structures. The refcnt is set to 1. 323 */ 324 struct vmspace * 325 vmspace_alloc(vm_offset_t min, vm_offset_t max, pmap_pinit_t pinit) 326 { 327 struct vmspace *vm; 328 329 vm = uma_zalloc(vmspace_zone, M_WAITOK); 330 KASSERT(vm->vm_map.pmap == NULL, ("vm_map.pmap must be NULL")); 331 if (!pinit(vmspace_pmap(vm))) { 332 uma_zfree(vmspace_zone, vm); 333 return (NULL); 334 } 335 CTR1(KTR_VM, "vmspace_alloc: %p", vm); 336 _vm_map_init(&vm->vm_map, vmspace_pmap(vm), min, max); 337 refcount_init(&vm->vm_refcnt, 1); 338 vm->vm_shm = NULL; 339 vm->vm_swrss = 0; 340 vm->vm_tsize = 0; 341 vm->vm_dsize = 0; 342 vm->vm_ssize = 0; 343 vm->vm_taddr = 0; 344 vm->vm_daddr = 0; 345 vm->vm_maxsaddr = 0; 346 return (vm); 347 } 348 349 #ifdef RACCT 350 static void 351 vmspace_container_reset(struct proc *p) 352 { 353 354 PROC_LOCK(p); 355 racct_set(p, RACCT_DATA, 0); 356 racct_set(p, RACCT_STACK, 0); 357 racct_set(p, RACCT_RSS, 0); 358 racct_set(p, RACCT_MEMLOCK, 0); 359 racct_set(p, RACCT_VMEM, 0); 360 PROC_UNLOCK(p); 361 } 362 #endif 363 364 static inline void 365 vmspace_dofree(struct vmspace *vm) 366 { 367 368 CTR1(KTR_VM, "vmspace_free: %p", vm); 369 370 /* 371 * Make sure any SysV shm is freed, it might not have been in 372 * exit1(). 373 */ 374 shmexit(vm); 375 376 /* 377 * Lock the map, to wait out all other references to it. 378 * Delete all of the mappings and pages they hold, then call 379 * the pmap module to reclaim anything left. 380 */ 381 (void)vm_map_remove(&vm->vm_map, vm_map_min(&vm->vm_map), 382 vm_map_max(&vm->vm_map)); 383 384 pmap_release(vmspace_pmap(vm)); 385 vm->vm_map.pmap = NULL; 386 uma_zfree(vmspace_zone, vm); 387 } 388 389 void 390 vmspace_free(struct vmspace *vm) 391 { 392 393 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 394 "vmspace_free() called"); 395 396 if (refcount_release(&vm->vm_refcnt)) 397 vmspace_dofree(vm); 398 } 399 400 void 401 vmspace_exitfree(struct proc *p) 402 { 403 struct vmspace *vm; 404 405 PROC_VMSPACE_LOCK(p); 406 vm = p->p_vmspace; 407 p->p_vmspace = NULL; 408 PROC_VMSPACE_UNLOCK(p); 409 KASSERT(vm == &vmspace0, ("vmspace_exitfree: wrong vmspace")); 410 vmspace_free(vm); 411 } 412 413 void 414 vmspace_exit(struct thread *td) 415 { 416 struct vmspace *vm; 417 struct proc *p; 418 bool released; 419 420 p = td->td_proc; 421 vm = p->p_vmspace; 422 423 /* 424 * Prepare to release the vmspace reference. The thread that releases 425 * the last reference is responsible for tearing down the vmspace. 426 * However, threads not releasing the final reference must switch to the 427 * kernel's vmspace0 before the decrement so that the subsequent pmap 428 * deactivation does not modify a freed vmspace. 429 */ 430 refcount_acquire(&vmspace0.vm_refcnt); 431 if (!(released = refcount_release_if_last(&vm->vm_refcnt))) { 432 if (p->p_vmspace != &vmspace0) { 433 PROC_VMSPACE_LOCK(p); 434 p->p_vmspace = &vmspace0; 435 PROC_VMSPACE_UNLOCK(p); 436 pmap_activate(td); 437 } 438 released = refcount_release(&vm->vm_refcnt); 439 } 440 if (released) { 441 /* 442 * pmap_remove_pages() expects the pmap to be active, so switch 443 * back first if necessary. 444 */ 445 if (p->p_vmspace != vm) { 446 PROC_VMSPACE_LOCK(p); 447 p->p_vmspace = vm; 448 PROC_VMSPACE_UNLOCK(p); 449 pmap_activate(td); 450 } 451 pmap_remove_pages(vmspace_pmap(vm)); 452 PROC_VMSPACE_LOCK(p); 453 p->p_vmspace = &vmspace0; 454 PROC_VMSPACE_UNLOCK(p); 455 pmap_activate(td); 456 vmspace_dofree(vm); 457 } 458 #ifdef RACCT 459 if (racct_enable) 460 vmspace_container_reset(p); 461 #endif 462 } 463 464 /* Acquire reference to vmspace owned by another process. */ 465 466 struct vmspace * 467 vmspace_acquire_ref(struct proc *p) 468 { 469 struct vmspace *vm; 470 471 PROC_VMSPACE_LOCK(p); 472 vm = p->p_vmspace; 473 if (vm == NULL || !refcount_acquire_if_not_zero(&vm->vm_refcnt)) { 474 PROC_VMSPACE_UNLOCK(p); 475 return (NULL); 476 } 477 if (vm != p->p_vmspace) { 478 PROC_VMSPACE_UNLOCK(p); 479 vmspace_free(vm); 480 return (NULL); 481 } 482 PROC_VMSPACE_UNLOCK(p); 483 return (vm); 484 } 485 486 /* 487 * Switch between vmspaces in an AIO kernel process. 488 * 489 * The new vmspace is either the vmspace of a user process obtained 490 * from an active AIO request or the initial vmspace of the AIO kernel 491 * process (when it is idling). Because user processes will block to 492 * drain any active AIO requests before proceeding in exit() or 493 * execve(), the reference count for vmspaces from AIO requests can 494 * never be 0. Similarly, AIO kernel processes hold an extra 495 * reference on their initial vmspace for the life of the process. As 496 * a result, the 'newvm' vmspace always has a non-zero reference 497 * count. This permits an additional reference on 'newvm' to be 498 * acquired via a simple atomic increment rather than the loop in 499 * vmspace_acquire_ref() above. 500 */ 501 void 502 vmspace_switch_aio(struct vmspace *newvm) 503 { 504 struct vmspace *oldvm; 505 506 /* XXX: Need some way to assert that this is an aio daemon. */ 507 508 KASSERT(refcount_load(&newvm->vm_refcnt) > 0, 509 ("vmspace_switch_aio: newvm unreferenced")); 510 511 oldvm = curproc->p_vmspace; 512 if (oldvm == newvm) 513 return; 514 515 /* 516 * Point to the new address space and refer to it. 517 */ 518 curproc->p_vmspace = newvm; 519 refcount_acquire(&newvm->vm_refcnt); 520 521 /* Activate the new mapping. */ 522 pmap_activate(curthread); 523 524 vmspace_free(oldvm); 525 } 526 527 void 528 _vm_map_lock(vm_map_t map, const char *file, int line) 529 { 530 531 if (map->system_map) 532 mtx_lock_flags_(&map->system_mtx, 0, file, line); 533 else 534 sx_xlock_(&map->lock, file, line); 535 map->timestamp++; 536 } 537 538 void 539 vm_map_entry_set_vnode_text(vm_map_entry_t entry, bool add) 540 { 541 vm_object_t object; 542 struct vnode *vp; 543 bool vp_held; 544 545 if ((entry->eflags & MAP_ENTRY_VN_EXEC) == 0) 546 return; 547 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 548 ("Submap with execs")); 549 object = entry->object.vm_object; 550 KASSERT(object != NULL, ("No object for text, entry %p", entry)); 551 if ((object->flags & OBJ_ANON) != 0) 552 object = object->handle; 553 else 554 KASSERT(object->backing_object == NULL, 555 ("non-anon object %p shadows", object)); 556 KASSERT(object != NULL, ("No content object for text, entry %p obj %p", 557 entry, entry->object.vm_object)); 558 559 /* 560 * Mostly, we do not lock the backing object. It is 561 * referenced by the entry we are processing, so it cannot go 562 * away. 563 */ 564 vp = NULL; 565 vp_held = false; 566 if (object->type == OBJT_DEAD) { 567 /* 568 * For OBJT_DEAD objects, v_writecount was handled in 569 * vnode_pager_dealloc(). 570 */ 571 } else if (object->type == OBJT_VNODE) { 572 vp = object->handle; 573 } else if (object->type == OBJT_SWAP) { 574 KASSERT((object->flags & OBJ_TMPFS_NODE) != 0, 575 ("vm_map_entry_set_vnode_text: swap and !TMPFS " 576 "entry %p, object %p, add %d", entry, object, add)); 577 /* 578 * Tmpfs VREG node, which was reclaimed, has 579 * OBJ_TMPFS_NODE flag set, but not OBJ_TMPFS. In 580 * this case there is no v_writecount to adjust. 581 */ 582 VM_OBJECT_RLOCK(object); 583 if ((object->flags & OBJ_TMPFS) != 0) { 584 vp = object->un_pager.swp.swp_tmpfs; 585 if (vp != NULL) { 586 vhold(vp); 587 vp_held = true; 588 } 589 } 590 VM_OBJECT_RUNLOCK(object); 591 } else { 592 KASSERT(0, 593 ("vm_map_entry_set_vnode_text: wrong object type, " 594 "entry %p, object %p, add %d", entry, object, add)); 595 } 596 if (vp != NULL) { 597 if (add) { 598 VOP_SET_TEXT_CHECKED(vp); 599 } else { 600 vn_lock(vp, LK_SHARED | LK_RETRY); 601 VOP_UNSET_TEXT_CHECKED(vp); 602 VOP_UNLOCK(vp); 603 } 604 if (vp_held) 605 vdrop(vp); 606 } 607 } 608 609 /* 610 * Use a different name for this vm_map_entry field when it's use 611 * is not consistent with its use as part of an ordered search tree. 612 */ 613 #define defer_next right 614 615 static void 616 vm_map_process_deferred(void) 617 { 618 struct thread *td; 619 vm_map_entry_t entry, next; 620 vm_object_t object; 621 622 td = curthread; 623 entry = td->td_map_def_user; 624 td->td_map_def_user = NULL; 625 while (entry != NULL) { 626 next = entry->defer_next; 627 MPASS((entry->eflags & (MAP_ENTRY_WRITECNT | 628 MAP_ENTRY_VN_EXEC)) != (MAP_ENTRY_WRITECNT | 629 MAP_ENTRY_VN_EXEC)); 630 if ((entry->eflags & MAP_ENTRY_WRITECNT) != 0) { 631 /* 632 * Decrement the object's writemappings and 633 * possibly the vnode's v_writecount. 634 */ 635 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 636 ("Submap with writecount")); 637 object = entry->object.vm_object; 638 KASSERT(object != NULL, ("No object for writecount")); 639 vm_pager_release_writecount(object, entry->start, 640 entry->end); 641 } 642 vm_map_entry_set_vnode_text(entry, false); 643 vm_map_entry_deallocate(entry, FALSE); 644 entry = next; 645 } 646 } 647 648 #ifdef INVARIANTS 649 static void 650 _vm_map_assert_locked(vm_map_t map, const char *file, int line) 651 { 652 653 if (map->system_map) 654 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 655 else 656 sx_assert_(&map->lock, SA_XLOCKED, file, line); 657 } 658 659 #define VM_MAP_ASSERT_LOCKED(map) \ 660 _vm_map_assert_locked(map, LOCK_FILE, LOCK_LINE) 661 662 enum { VMMAP_CHECK_NONE, VMMAP_CHECK_UNLOCK, VMMAP_CHECK_ALL }; 663 #ifdef DIAGNOSTIC 664 static int enable_vmmap_check = VMMAP_CHECK_UNLOCK; 665 #else 666 static int enable_vmmap_check = VMMAP_CHECK_NONE; 667 #endif 668 SYSCTL_INT(_debug, OID_AUTO, vmmap_check, CTLFLAG_RWTUN, 669 &enable_vmmap_check, 0, "Enable vm map consistency checking"); 670 671 static void _vm_map_assert_consistent(vm_map_t map, int check); 672 673 #define VM_MAP_ASSERT_CONSISTENT(map) \ 674 _vm_map_assert_consistent(map, VMMAP_CHECK_ALL) 675 #ifdef DIAGNOSTIC 676 #define VM_MAP_UNLOCK_CONSISTENT(map) do { \ 677 if (map->nupdates > map->nentries) { \ 678 _vm_map_assert_consistent(map, VMMAP_CHECK_UNLOCK); \ 679 map->nupdates = 0; \ 680 } \ 681 } while (0) 682 #else 683 #define VM_MAP_UNLOCK_CONSISTENT(map) 684 #endif 685 #else 686 #define VM_MAP_ASSERT_LOCKED(map) 687 #define VM_MAP_ASSERT_CONSISTENT(map) 688 #define VM_MAP_UNLOCK_CONSISTENT(map) 689 #endif /* INVARIANTS */ 690 691 void 692 _vm_map_unlock(vm_map_t map, const char *file, int line) 693 { 694 695 VM_MAP_UNLOCK_CONSISTENT(map); 696 if (map->system_map) { 697 #ifndef UMA_MD_SMALL_ALLOC 698 if (map == kernel_map && (map->flags & MAP_REPLENISH) != 0) { 699 uma_prealloc(kmapentzone, 1); 700 map->flags &= ~MAP_REPLENISH; 701 } 702 #endif 703 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 704 } else { 705 sx_xunlock_(&map->lock, file, line); 706 vm_map_process_deferred(); 707 } 708 } 709 710 void 711 _vm_map_lock_read(vm_map_t map, const char *file, int line) 712 { 713 714 if (map->system_map) 715 mtx_lock_flags_(&map->system_mtx, 0, file, line); 716 else 717 sx_slock_(&map->lock, file, line); 718 } 719 720 void 721 _vm_map_unlock_read(vm_map_t map, const char *file, int line) 722 { 723 724 if (map->system_map) { 725 KASSERT((map->flags & MAP_REPLENISH) == 0, 726 ("%s: MAP_REPLENISH leaked", __func__)); 727 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 728 } else { 729 sx_sunlock_(&map->lock, file, line); 730 vm_map_process_deferred(); 731 } 732 } 733 734 int 735 _vm_map_trylock(vm_map_t map, const char *file, int line) 736 { 737 int error; 738 739 error = map->system_map ? 740 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 741 !sx_try_xlock_(&map->lock, file, line); 742 if (error == 0) 743 map->timestamp++; 744 return (error == 0); 745 } 746 747 int 748 _vm_map_trylock_read(vm_map_t map, const char *file, int line) 749 { 750 int error; 751 752 error = map->system_map ? 753 !mtx_trylock_flags_(&map->system_mtx, 0, file, line) : 754 !sx_try_slock_(&map->lock, file, line); 755 return (error == 0); 756 } 757 758 /* 759 * _vm_map_lock_upgrade: [ internal use only ] 760 * 761 * Tries to upgrade a read (shared) lock on the specified map to a write 762 * (exclusive) lock. Returns the value "0" if the upgrade succeeds and a 763 * non-zero value if the upgrade fails. If the upgrade fails, the map is 764 * returned without a read or write lock held. 765 * 766 * Requires that the map be read locked. 767 */ 768 int 769 _vm_map_lock_upgrade(vm_map_t map, const char *file, int line) 770 { 771 unsigned int last_timestamp; 772 773 if (map->system_map) { 774 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 775 } else { 776 if (!sx_try_upgrade_(&map->lock, file, line)) { 777 last_timestamp = map->timestamp; 778 sx_sunlock_(&map->lock, file, line); 779 vm_map_process_deferred(); 780 /* 781 * If the map's timestamp does not change while the 782 * map is unlocked, then the upgrade succeeds. 783 */ 784 sx_xlock_(&map->lock, file, line); 785 if (last_timestamp != map->timestamp) { 786 sx_xunlock_(&map->lock, file, line); 787 return (1); 788 } 789 } 790 } 791 map->timestamp++; 792 return (0); 793 } 794 795 void 796 _vm_map_lock_downgrade(vm_map_t map, const char *file, int line) 797 { 798 799 if (map->system_map) { 800 KASSERT((map->flags & MAP_REPLENISH) == 0, 801 ("%s: MAP_REPLENISH leaked", __func__)); 802 mtx_assert_(&map->system_mtx, MA_OWNED, file, line); 803 } else { 804 VM_MAP_UNLOCK_CONSISTENT(map); 805 sx_downgrade_(&map->lock, file, line); 806 } 807 } 808 809 /* 810 * vm_map_locked: 811 * 812 * Returns a non-zero value if the caller holds a write (exclusive) lock 813 * on the specified map and the value "0" otherwise. 814 */ 815 int 816 vm_map_locked(vm_map_t map) 817 { 818 819 if (map->system_map) 820 return (mtx_owned(&map->system_mtx)); 821 else 822 return (sx_xlocked(&map->lock)); 823 } 824 825 /* 826 * _vm_map_unlock_and_wait: 827 * 828 * Atomically releases the lock on the specified map and puts the calling 829 * thread to sleep. The calling thread will remain asleep until either 830 * vm_map_wakeup() is performed on the map or the specified timeout is 831 * exceeded. 832 * 833 * WARNING! This function does not perform deferred deallocations of 834 * objects and map entries. Therefore, the calling thread is expected to 835 * reacquire the map lock after reawakening and later perform an ordinary 836 * unlock operation, such as vm_map_unlock(), before completing its 837 * operation on the map. 838 */ 839 int 840 _vm_map_unlock_and_wait(vm_map_t map, int timo, const char *file, int line) 841 { 842 843 VM_MAP_UNLOCK_CONSISTENT(map); 844 mtx_lock(&map_sleep_mtx); 845 if (map->system_map) { 846 KASSERT((map->flags & MAP_REPLENISH) == 0, 847 ("%s: MAP_REPLENISH leaked", __func__)); 848 mtx_unlock_flags_(&map->system_mtx, 0, file, line); 849 } else { 850 sx_xunlock_(&map->lock, file, line); 851 } 852 return (msleep(&map->root, &map_sleep_mtx, PDROP | PVM, "vmmaps", 853 timo)); 854 } 855 856 /* 857 * vm_map_wakeup: 858 * 859 * Awaken any threads that have slept on the map using 860 * vm_map_unlock_and_wait(). 861 */ 862 void 863 vm_map_wakeup(vm_map_t map) 864 { 865 866 /* 867 * Acquire and release map_sleep_mtx to prevent a wakeup() 868 * from being performed (and lost) between the map unlock 869 * and the msleep() in _vm_map_unlock_and_wait(). 870 */ 871 mtx_lock(&map_sleep_mtx); 872 mtx_unlock(&map_sleep_mtx); 873 wakeup(&map->root); 874 } 875 876 void 877 vm_map_busy(vm_map_t map) 878 { 879 880 VM_MAP_ASSERT_LOCKED(map); 881 map->busy++; 882 } 883 884 void 885 vm_map_unbusy(vm_map_t map) 886 { 887 888 VM_MAP_ASSERT_LOCKED(map); 889 KASSERT(map->busy, ("vm_map_unbusy: not busy")); 890 if (--map->busy == 0 && (map->flags & MAP_BUSY_WAKEUP)) { 891 vm_map_modflags(map, 0, MAP_BUSY_WAKEUP); 892 wakeup(&map->busy); 893 } 894 } 895 896 void 897 vm_map_wait_busy(vm_map_t map) 898 { 899 900 VM_MAP_ASSERT_LOCKED(map); 901 while (map->busy) { 902 vm_map_modflags(map, MAP_BUSY_WAKEUP, 0); 903 if (map->system_map) 904 msleep(&map->busy, &map->system_mtx, 0, "mbusy", 0); 905 else 906 sx_sleep(&map->busy, &map->lock, 0, "mbusy", 0); 907 } 908 map->timestamp++; 909 } 910 911 long 912 vmspace_resident_count(struct vmspace *vmspace) 913 { 914 return pmap_resident_count(vmspace_pmap(vmspace)); 915 } 916 917 /* 918 * Initialize an existing vm_map structure 919 * such as that in the vmspace structure. 920 */ 921 static void 922 _vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 923 { 924 925 map->header.eflags = MAP_ENTRY_HEADER; 926 map->needs_wakeup = FALSE; 927 map->system_map = 0; 928 map->pmap = pmap; 929 map->header.end = min; 930 map->header.start = max; 931 map->flags = 0; 932 map->header.left = map->header.right = &map->header; 933 map->root = NULL; 934 map->timestamp = 0; 935 map->busy = 0; 936 map->anon_loc = 0; 937 #ifdef DIAGNOSTIC 938 map->nupdates = 0; 939 #endif 940 } 941 942 void 943 vm_map_init(vm_map_t map, pmap_t pmap, vm_offset_t min, vm_offset_t max) 944 { 945 946 _vm_map_init(map, pmap, min, max); 947 mtx_init(&map->system_mtx, "vm map (system)", NULL, 948 MTX_DEF | MTX_DUPOK); 949 sx_init(&map->lock, "vm map (user)"); 950 } 951 952 /* 953 * vm_map_entry_dispose: [ internal use only ] 954 * 955 * Inverse of vm_map_entry_create. 956 */ 957 static void 958 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry) 959 { 960 uma_zfree(map->system_map ? kmapentzone : mapentzone, entry); 961 } 962 963 /* 964 * vm_map_entry_create: [ internal use only ] 965 * 966 * Allocates a VM map entry for insertion. 967 * No entry fields are filled in. 968 */ 969 static vm_map_entry_t 970 vm_map_entry_create(vm_map_t map) 971 { 972 vm_map_entry_t new_entry; 973 974 #ifndef UMA_MD_SMALL_ALLOC 975 if (map == kernel_map) { 976 VM_MAP_ASSERT_LOCKED(map); 977 978 /* 979 * A new slab of kernel map entries cannot be allocated at this 980 * point because the kernel map has not yet been updated to 981 * reflect the caller's request. Therefore, we allocate a new 982 * map entry, dipping into the reserve if necessary, and set a 983 * flag indicating that the reserve must be replenished before 984 * the map is unlocked. 985 */ 986 new_entry = uma_zalloc(kmapentzone, M_NOWAIT | M_NOVM); 987 if (new_entry == NULL) { 988 new_entry = uma_zalloc(kmapentzone, 989 M_NOWAIT | M_NOVM | M_USE_RESERVE); 990 kernel_map->flags |= MAP_REPLENISH; 991 } 992 } else 993 #endif 994 if (map->system_map) { 995 new_entry = uma_zalloc(kmapentzone, M_NOWAIT); 996 } else { 997 new_entry = uma_zalloc(mapentzone, M_WAITOK); 998 } 999 KASSERT(new_entry != NULL, 1000 ("vm_map_entry_create: kernel resources exhausted")); 1001 return (new_entry); 1002 } 1003 1004 /* 1005 * vm_map_entry_set_behavior: 1006 * 1007 * Set the expected access behavior, either normal, random, or 1008 * sequential. 1009 */ 1010 static inline void 1011 vm_map_entry_set_behavior(vm_map_entry_t entry, u_char behavior) 1012 { 1013 entry->eflags = (entry->eflags & ~MAP_ENTRY_BEHAV_MASK) | 1014 (behavior & MAP_ENTRY_BEHAV_MASK); 1015 } 1016 1017 /* 1018 * vm_map_entry_max_free_{left,right}: 1019 * 1020 * Compute the size of the largest free gap between two entries, 1021 * one the root of a tree and the other the ancestor of that root 1022 * that is the least or greatest ancestor found on the search path. 1023 */ 1024 static inline vm_size_t 1025 vm_map_entry_max_free_left(vm_map_entry_t root, vm_map_entry_t left_ancestor) 1026 { 1027 1028 return (root->left != left_ancestor ? 1029 root->left->max_free : root->start - left_ancestor->end); 1030 } 1031 1032 static inline vm_size_t 1033 vm_map_entry_max_free_right(vm_map_entry_t root, vm_map_entry_t right_ancestor) 1034 { 1035 1036 return (root->right != right_ancestor ? 1037 root->right->max_free : right_ancestor->start - root->end); 1038 } 1039 1040 /* 1041 * vm_map_entry_{pred,succ}: 1042 * 1043 * Find the {predecessor, successor} of the entry by taking one step 1044 * in the appropriate direction and backtracking as much as necessary. 1045 * vm_map_entry_succ is defined in vm_map.h. 1046 */ 1047 static inline vm_map_entry_t 1048 vm_map_entry_pred(vm_map_entry_t entry) 1049 { 1050 vm_map_entry_t prior; 1051 1052 prior = entry->left; 1053 if (prior->right->start < entry->start) { 1054 do 1055 prior = prior->right; 1056 while (prior->right != entry); 1057 } 1058 return (prior); 1059 } 1060 1061 static inline vm_size_t 1062 vm_size_max(vm_size_t a, vm_size_t b) 1063 { 1064 1065 return (a > b ? a : b); 1066 } 1067 1068 #define SPLAY_LEFT_STEP(root, y, llist, rlist, test) do { \ 1069 vm_map_entry_t z; \ 1070 vm_size_t max_free; \ 1071 \ 1072 /* \ 1073 * Infer root->right->max_free == root->max_free when \ 1074 * y->max_free < root->max_free || root->max_free == 0. \ 1075 * Otherwise, look right to find it. \ 1076 */ \ 1077 y = root->left; \ 1078 max_free = root->max_free; \ 1079 KASSERT(max_free == vm_size_max( \ 1080 vm_map_entry_max_free_left(root, llist), \ 1081 vm_map_entry_max_free_right(root, rlist)), \ 1082 ("%s: max_free invariant fails", __func__)); \ 1083 if (max_free - 1 < vm_map_entry_max_free_left(root, llist)) \ 1084 max_free = vm_map_entry_max_free_right(root, rlist); \ 1085 if (y != llist && (test)) { \ 1086 /* Rotate right and make y root. */ \ 1087 z = y->right; \ 1088 if (z != root) { \ 1089 root->left = z; \ 1090 y->right = root; \ 1091 if (max_free < y->max_free) \ 1092 root->max_free = max_free = \ 1093 vm_size_max(max_free, z->max_free); \ 1094 } else if (max_free < y->max_free) \ 1095 root->max_free = max_free = \ 1096 vm_size_max(max_free, root->start - y->end);\ 1097 root = y; \ 1098 y = root->left; \ 1099 } \ 1100 /* Copy right->max_free. Put root on rlist. */ \ 1101 root->max_free = max_free; \ 1102 KASSERT(max_free == vm_map_entry_max_free_right(root, rlist), \ 1103 ("%s: max_free not copied from right", __func__)); \ 1104 root->left = rlist; \ 1105 rlist = root; \ 1106 root = y != llist ? y : NULL; \ 1107 } while (0) 1108 1109 #define SPLAY_RIGHT_STEP(root, y, llist, rlist, test) do { \ 1110 vm_map_entry_t z; \ 1111 vm_size_t max_free; \ 1112 \ 1113 /* \ 1114 * Infer root->left->max_free == root->max_free when \ 1115 * y->max_free < root->max_free || root->max_free == 0. \ 1116 * Otherwise, look left to find it. \ 1117 */ \ 1118 y = root->right; \ 1119 max_free = root->max_free; \ 1120 KASSERT(max_free == vm_size_max( \ 1121 vm_map_entry_max_free_left(root, llist), \ 1122 vm_map_entry_max_free_right(root, rlist)), \ 1123 ("%s: max_free invariant fails", __func__)); \ 1124 if (max_free - 1 < vm_map_entry_max_free_right(root, rlist)) \ 1125 max_free = vm_map_entry_max_free_left(root, llist); \ 1126 if (y != rlist && (test)) { \ 1127 /* Rotate left and make y root. */ \ 1128 z = y->left; \ 1129 if (z != root) { \ 1130 root->right = z; \ 1131 y->left = root; \ 1132 if (max_free < y->max_free) \ 1133 root->max_free = max_free = \ 1134 vm_size_max(max_free, z->max_free); \ 1135 } else if (max_free < y->max_free) \ 1136 root->max_free = max_free = \ 1137 vm_size_max(max_free, y->start - root->end);\ 1138 root = y; \ 1139 y = root->right; \ 1140 } \ 1141 /* Copy left->max_free. Put root on llist. */ \ 1142 root->max_free = max_free; \ 1143 KASSERT(max_free == vm_map_entry_max_free_left(root, llist), \ 1144 ("%s: max_free not copied from left", __func__)); \ 1145 root->right = llist; \ 1146 llist = root; \ 1147 root = y != rlist ? y : NULL; \ 1148 } while (0) 1149 1150 /* 1151 * Walk down the tree until we find addr or a gap where addr would go, breaking 1152 * off left and right subtrees of nodes less than, or greater than addr. Treat 1153 * subtrees with root->max_free < length as empty trees. llist and rlist are 1154 * the two sides in reverse order (bottom-up), with llist linked by the right 1155 * pointer and rlist linked by the left pointer in the vm_map_entry, and both 1156 * lists terminated by &map->header. This function, and the subsequent call to 1157 * vm_map_splay_merge_{left,right,pred,succ}, rely on the start and end address 1158 * values in &map->header. 1159 */ 1160 static __always_inline vm_map_entry_t 1161 vm_map_splay_split(vm_map_t map, vm_offset_t addr, vm_size_t length, 1162 vm_map_entry_t *llist, vm_map_entry_t *rlist) 1163 { 1164 vm_map_entry_t left, right, root, y; 1165 1166 left = right = &map->header; 1167 root = map->root; 1168 while (root != NULL && root->max_free >= length) { 1169 KASSERT(left->end <= root->start && 1170 root->end <= right->start, 1171 ("%s: root not within tree bounds", __func__)); 1172 if (addr < root->start) { 1173 SPLAY_LEFT_STEP(root, y, left, right, 1174 y->max_free >= length && addr < y->start); 1175 } else if (addr >= root->end) { 1176 SPLAY_RIGHT_STEP(root, y, left, right, 1177 y->max_free >= length && addr >= y->end); 1178 } else 1179 break; 1180 } 1181 *llist = left; 1182 *rlist = right; 1183 return (root); 1184 } 1185 1186 static __always_inline void 1187 vm_map_splay_findnext(vm_map_entry_t root, vm_map_entry_t *rlist) 1188 { 1189 vm_map_entry_t hi, right, y; 1190 1191 right = *rlist; 1192 hi = root->right == right ? NULL : root->right; 1193 if (hi == NULL) 1194 return; 1195 do 1196 SPLAY_LEFT_STEP(hi, y, root, right, true); 1197 while (hi != NULL); 1198 *rlist = right; 1199 } 1200 1201 static __always_inline void 1202 vm_map_splay_findprev(vm_map_entry_t root, vm_map_entry_t *llist) 1203 { 1204 vm_map_entry_t left, lo, y; 1205 1206 left = *llist; 1207 lo = root->left == left ? NULL : root->left; 1208 if (lo == NULL) 1209 return; 1210 do 1211 SPLAY_RIGHT_STEP(lo, y, left, root, true); 1212 while (lo != NULL); 1213 *llist = left; 1214 } 1215 1216 static inline void 1217 vm_map_entry_swap(vm_map_entry_t *a, vm_map_entry_t *b) 1218 { 1219 vm_map_entry_t tmp; 1220 1221 tmp = *b; 1222 *b = *a; 1223 *a = tmp; 1224 } 1225 1226 /* 1227 * Walk back up the two spines, flip the pointers and set max_free. The 1228 * subtrees of the root go at the bottom of llist and rlist. 1229 */ 1230 static vm_size_t 1231 vm_map_splay_merge_left_walk(vm_map_entry_t header, vm_map_entry_t root, 1232 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t llist) 1233 { 1234 do { 1235 /* 1236 * The max_free values of the children of llist are in 1237 * llist->max_free and max_free. Update with the 1238 * max value. 1239 */ 1240 llist->max_free = max_free = 1241 vm_size_max(llist->max_free, max_free); 1242 vm_map_entry_swap(&llist->right, &tail); 1243 vm_map_entry_swap(&tail, &llist); 1244 } while (llist != header); 1245 root->left = tail; 1246 return (max_free); 1247 } 1248 1249 /* 1250 * When llist is known to be the predecessor of root. 1251 */ 1252 static inline vm_size_t 1253 vm_map_splay_merge_pred(vm_map_entry_t header, vm_map_entry_t root, 1254 vm_map_entry_t llist) 1255 { 1256 vm_size_t max_free; 1257 1258 max_free = root->start - llist->end; 1259 if (llist != header) { 1260 max_free = vm_map_splay_merge_left_walk(header, root, 1261 root, max_free, llist); 1262 } else { 1263 root->left = header; 1264 header->right = root; 1265 } 1266 return (max_free); 1267 } 1268 1269 /* 1270 * When llist may or may not be the predecessor of root. 1271 */ 1272 static inline vm_size_t 1273 vm_map_splay_merge_left(vm_map_entry_t header, vm_map_entry_t root, 1274 vm_map_entry_t llist) 1275 { 1276 vm_size_t max_free; 1277 1278 max_free = vm_map_entry_max_free_left(root, llist); 1279 if (llist != header) { 1280 max_free = vm_map_splay_merge_left_walk(header, root, 1281 root->left == llist ? root : root->left, 1282 max_free, llist); 1283 } 1284 return (max_free); 1285 } 1286 1287 static vm_size_t 1288 vm_map_splay_merge_right_walk(vm_map_entry_t header, vm_map_entry_t root, 1289 vm_map_entry_t tail, vm_size_t max_free, vm_map_entry_t rlist) 1290 { 1291 do { 1292 /* 1293 * The max_free values of the children of rlist are in 1294 * rlist->max_free and max_free. Update with the 1295 * max value. 1296 */ 1297 rlist->max_free = max_free = 1298 vm_size_max(rlist->max_free, max_free); 1299 vm_map_entry_swap(&rlist->left, &tail); 1300 vm_map_entry_swap(&tail, &rlist); 1301 } while (rlist != header); 1302 root->right = tail; 1303 return (max_free); 1304 } 1305 1306 /* 1307 * When rlist is known to be the succecessor of root. 1308 */ 1309 static inline vm_size_t 1310 vm_map_splay_merge_succ(vm_map_entry_t header, vm_map_entry_t root, 1311 vm_map_entry_t rlist) 1312 { 1313 vm_size_t max_free; 1314 1315 max_free = rlist->start - root->end; 1316 if (rlist != header) { 1317 max_free = vm_map_splay_merge_right_walk(header, root, 1318 root, max_free, rlist); 1319 } else { 1320 root->right = header; 1321 header->left = root; 1322 } 1323 return (max_free); 1324 } 1325 1326 /* 1327 * When rlist may or may not be the succecessor of root. 1328 */ 1329 static inline vm_size_t 1330 vm_map_splay_merge_right(vm_map_entry_t header, vm_map_entry_t root, 1331 vm_map_entry_t rlist) 1332 { 1333 vm_size_t max_free; 1334 1335 max_free = vm_map_entry_max_free_right(root, rlist); 1336 if (rlist != header) { 1337 max_free = vm_map_splay_merge_right_walk(header, root, 1338 root->right == rlist ? root : root->right, 1339 max_free, rlist); 1340 } 1341 return (max_free); 1342 } 1343 1344 /* 1345 * vm_map_splay: 1346 * 1347 * The Sleator and Tarjan top-down splay algorithm with the 1348 * following variation. Max_free must be computed bottom-up, so 1349 * on the downward pass, maintain the left and right spines in 1350 * reverse order. Then, make a second pass up each side to fix 1351 * the pointers and compute max_free. The time bound is O(log n) 1352 * amortized. 1353 * 1354 * The tree is threaded, which means that there are no null pointers. 1355 * When a node has no left child, its left pointer points to its 1356 * predecessor, which the last ancestor on the search path from the root 1357 * where the search branched right. Likewise, when a node has no right 1358 * child, its right pointer points to its successor. The map header node 1359 * is the predecessor of the first map entry, and the successor of the 1360 * last. 1361 * 1362 * The new root is the vm_map_entry containing "addr", or else an 1363 * adjacent entry (lower if possible) if addr is not in the tree. 1364 * 1365 * The map must be locked, and leaves it so. 1366 * 1367 * Returns: the new root. 1368 */ 1369 static vm_map_entry_t 1370 vm_map_splay(vm_map_t map, vm_offset_t addr) 1371 { 1372 vm_map_entry_t header, llist, rlist, root; 1373 vm_size_t max_free_left, max_free_right; 1374 1375 header = &map->header; 1376 root = vm_map_splay_split(map, addr, 0, &llist, &rlist); 1377 if (root != NULL) { 1378 max_free_left = vm_map_splay_merge_left(header, root, llist); 1379 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1380 } else if (llist != header) { 1381 /* 1382 * Recover the greatest node in the left 1383 * subtree and make it the root. 1384 */ 1385 root = llist; 1386 llist = root->right; 1387 max_free_left = vm_map_splay_merge_left(header, root, llist); 1388 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1389 } else if (rlist != header) { 1390 /* 1391 * Recover the least node in the right 1392 * subtree and make it the root. 1393 */ 1394 root = rlist; 1395 rlist = root->left; 1396 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1397 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1398 } else { 1399 /* There is no root. */ 1400 return (NULL); 1401 } 1402 root->max_free = vm_size_max(max_free_left, max_free_right); 1403 map->root = root; 1404 VM_MAP_ASSERT_CONSISTENT(map); 1405 return (root); 1406 } 1407 1408 /* 1409 * vm_map_entry_{un,}link: 1410 * 1411 * Insert/remove entries from maps. On linking, if new entry clips 1412 * existing entry, trim existing entry to avoid overlap, and manage 1413 * offsets. On unlinking, merge disappearing entry with neighbor, if 1414 * called for, and manage offsets. Callers should not modify fields in 1415 * entries already mapped. 1416 */ 1417 static void 1418 vm_map_entry_link(vm_map_t map, vm_map_entry_t entry) 1419 { 1420 vm_map_entry_t header, llist, rlist, root; 1421 vm_size_t max_free_left, max_free_right; 1422 1423 CTR3(KTR_VM, 1424 "vm_map_entry_link: map %p, nentries %d, entry %p", map, 1425 map->nentries, entry); 1426 VM_MAP_ASSERT_LOCKED(map); 1427 map->nentries++; 1428 header = &map->header; 1429 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1430 if (root == NULL) { 1431 /* 1432 * The new entry does not overlap any existing entry in the 1433 * map, so it becomes the new root of the map tree. 1434 */ 1435 max_free_left = vm_map_splay_merge_pred(header, entry, llist); 1436 max_free_right = vm_map_splay_merge_succ(header, entry, rlist); 1437 } else if (entry->start == root->start) { 1438 /* 1439 * The new entry is a clone of root, with only the end field 1440 * changed. The root entry will be shrunk to abut the new 1441 * entry, and will be the right child of the new root entry in 1442 * the modified map. 1443 */ 1444 KASSERT(entry->end < root->end, 1445 ("%s: clip_start not within entry", __func__)); 1446 vm_map_splay_findprev(root, &llist); 1447 root->offset += entry->end - root->start; 1448 root->start = entry->end; 1449 max_free_left = vm_map_splay_merge_pred(header, entry, llist); 1450 max_free_right = root->max_free = vm_size_max( 1451 vm_map_splay_merge_pred(entry, root, entry), 1452 vm_map_splay_merge_right(header, root, rlist)); 1453 } else { 1454 /* 1455 * The new entry is a clone of root, with only the start field 1456 * changed. The root entry will be shrunk to abut the new 1457 * entry, and will be the left child of the new root entry in 1458 * the modified map. 1459 */ 1460 KASSERT(entry->end == root->end, 1461 ("%s: clip_start not within entry", __func__)); 1462 vm_map_splay_findnext(root, &rlist); 1463 entry->offset += entry->start - root->start; 1464 root->end = entry->start; 1465 max_free_left = root->max_free = vm_size_max( 1466 vm_map_splay_merge_left(header, root, llist), 1467 vm_map_splay_merge_succ(entry, root, entry)); 1468 max_free_right = vm_map_splay_merge_succ(header, entry, rlist); 1469 } 1470 entry->max_free = vm_size_max(max_free_left, max_free_right); 1471 map->root = entry; 1472 VM_MAP_ASSERT_CONSISTENT(map); 1473 } 1474 1475 enum unlink_merge_type { 1476 UNLINK_MERGE_NONE, 1477 UNLINK_MERGE_NEXT 1478 }; 1479 1480 static void 1481 vm_map_entry_unlink(vm_map_t map, vm_map_entry_t entry, 1482 enum unlink_merge_type op) 1483 { 1484 vm_map_entry_t header, llist, rlist, root; 1485 vm_size_t max_free_left, max_free_right; 1486 1487 VM_MAP_ASSERT_LOCKED(map); 1488 header = &map->header; 1489 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1490 KASSERT(root != NULL, 1491 ("vm_map_entry_unlink: unlink object not mapped")); 1492 1493 vm_map_splay_findprev(root, &llist); 1494 vm_map_splay_findnext(root, &rlist); 1495 if (op == UNLINK_MERGE_NEXT) { 1496 rlist->start = root->start; 1497 rlist->offset = root->offset; 1498 } 1499 if (llist != header) { 1500 root = llist; 1501 llist = root->right; 1502 max_free_left = vm_map_splay_merge_left(header, root, llist); 1503 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1504 } else if (rlist != header) { 1505 root = rlist; 1506 rlist = root->left; 1507 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1508 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1509 } else { 1510 header->left = header->right = header; 1511 root = NULL; 1512 } 1513 if (root != NULL) 1514 root->max_free = vm_size_max(max_free_left, max_free_right); 1515 map->root = root; 1516 VM_MAP_ASSERT_CONSISTENT(map); 1517 map->nentries--; 1518 CTR3(KTR_VM, "vm_map_entry_unlink: map %p, nentries %d, entry %p", map, 1519 map->nentries, entry); 1520 } 1521 1522 /* 1523 * vm_map_entry_resize: 1524 * 1525 * Resize a vm_map_entry, recompute the amount of free space that 1526 * follows it and propagate that value up the tree. 1527 * 1528 * The map must be locked, and leaves it so. 1529 */ 1530 static void 1531 vm_map_entry_resize(vm_map_t map, vm_map_entry_t entry, vm_size_t grow_amount) 1532 { 1533 vm_map_entry_t header, llist, rlist, root; 1534 1535 VM_MAP_ASSERT_LOCKED(map); 1536 header = &map->header; 1537 root = vm_map_splay_split(map, entry->start, 0, &llist, &rlist); 1538 KASSERT(root != NULL, ("%s: resize object not mapped", __func__)); 1539 vm_map_splay_findnext(root, &rlist); 1540 entry->end += grow_amount; 1541 root->max_free = vm_size_max( 1542 vm_map_splay_merge_left(header, root, llist), 1543 vm_map_splay_merge_succ(header, root, rlist)); 1544 map->root = root; 1545 VM_MAP_ASSERT_CONSISTENT(map); 1546 CTR4(KTR_VM, "%s: map %p, nentries %d, entry %p", 1547 __func__, map, map->nentries, entry); 1548 } 1549 1550 /* 1551 * vm_map_lookup_entry: [ internal use only ] 1552 * 1553 * Finds the map entry containing (or 1554 * immediately preceding) the specified address 1555 * in the given map; the entry is returned 1556 * in the "entry" parameter. The boolean 1557 * result indicates whether the address is 1558 * actually contained in the map. 1559 */ 1560 boolean_t 1561 vm_map_lookup_entry( 1562 vm_map_t map, 1563 vm_offset_t address, 1564 vm_map_entry_t *entry) /* OUT */ 1565 { 1566 vm_map_entry_t cur, header, lbound, ubound; 1567 boolean_t locked; 1568 1569 /* 1570 * If the map is empty, then the map entry immediately preceding 1571 * "address" is the map's header. 1572 */ 1573 header = &map->header; 1574 cur = map->root; 1575 if (cur == NULL) { 1576 *entry = header; 1577 return (FALSE); 1578 } 1579 if (address >= cur->start && cur->end > address) { 1580 *entry = cur; 1581 return (TRUE); 1582 } 1583 if ((locked = vm_map_locked(map)) || 1584 sx_try_upgrade(&map->lock)) { 1585 /* 1586 * Splay requires a write lock on the map. However, it only 1587 * restructures the binary search tree; it does not otherwise 1588 * change the map. Thus, the map's timestamp need not change 1589 * on a temporary upgrade. 1590 */ 1591 cur = vm_map_splay(map, address); 1592 if (!locked) { 1593 VM_MAP_UNLOCK_CONSISTENT(map); 1594 sx_downgrade(&map->lock); 1595 } 1596 1597 /* 1598 * If "address" is contained within a map entry, the new root 1599 * is that map entry. Otherwise, the new root is a map entry 1600 * immediately before or after "address". 1601 */ 1602 if (address < cur->start) { 1603 *entry = header; 1604 return (FALSE); 1605 } 1606 *entry = cur; 1607 return (address < cur->end); 1608 } 1609 /* 1610 * Since the map is only locked for read access, perform a 1611 * standard binary search tree lookup for "address". 1612 */ 1613 lbound = ubound = header; 1614 for (;;) { 1615 if (address < cur->start) { 1616 ubound = cur; 1617 cur = cur->left; 1618 if (cur == lbound) 1619 break; 1620 } else if (cur->end <= address) { 1621 lbound = cur; 1622 cur = cur->right; 1623 if (cur == ubound) 1624 break; 1625 } else { 1626 *entry = cur; 1627 return (TRUE); 1628 } 1629 } 1630 *entry = lbound; 1631 return (FALSE); 1632 } 1633 1634 /* 1635 * vm_map_insert: 1636 * 1637 * Inserts the given whole VM object into the target 1638 * map at the specified address range. The object's 1639 * size should match that of the address range. 1640 * 1641 * Requires that the map be locked, and leaves it so. 1642 * 1643 * If object is non-NULL, ref count must be bumped by caller 1644 * prior to making call to account for the new entry. 1645 */ 1646 int 1647 vm_map_insert(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1648 vm_offset_t start, vm_offset_t end, vm_prot_t prot, vm_prot_t max, int cow) 1649 { 1650 vm_map_entry_t new_entry, next_entry, prev_entry; 1651 struct ucred *cred; 1652 vm_eflags_t protoeflags; 1653 vm_inherit_t inheritance; 1654 u_long bdry; 1655 u_int bidx; 1656 1657 VM_MAP_ASSERT_LOCKED(map); 1658 KASSERT(object != kernel_object || 1659 (cow & MAP_COPY_ON_WRITE) == 0, 1660 ("vm_map_insert: kernel object and COW")); 1661 KASSERT(object == NULL || (cow & MAP_NOFAULT) == 0 || 1662 (cow & MAP_SPLIT_BOUNDARY_MASK) != 0, 1663 ("vm_map_insert: paradoxical MAP_NOFAULT request, obj %p cow %#x", 1664 object, cow)); 1665 KASSERT((prot & ~max) == 0, 1666 ("prot %#x is not subset of max_prot %#x", prot, max)); 1667 1668 /* 1669 * Check that the start and end points are not bogus. 1670 */ 1671 if (start == end || !vm_map_range_valid(map, start, end)) 1672 return (KERN_INVALID_ADDRESS); 1673 1674 if ((map->flags & MAP_WXORX) != 0 && (prot & (VM_PROT_WRITE | 1675 VM_PROT_EXECUTE)) == (VM_PROT_WRITE | VM_PROT_EXECUTE)) 1676 return (KERN_PROTECTION_FAILURE); 1677 1678 /* 1679 * Find the entry prior to the proposed starting address; if it's part 1680 * of an existing entry, this range is bogus. 1681 */ 1682 if (vm_map_lookup_entry(map, start, &prev_entry)) 1683 return (KERN_NO_SPACE); 1684 1685 /* 1686 * Assert that the next entry doesn't overlap the end point. 1687 */ 1688 next_entry = vm_map_entry_succ(prev_entry); 1689 if (next_entry->start < end) 1690 return (KERN_NO_SPACE); 1691 1692 if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL || 1693 max != VM_PROT_NONE)) 1694 return (KERN_INVALID_ARGUMENT); 1695 1696 protoeflags = 0; 1697 if (cow & MAP_COPY_ON_WRITE) 1698 protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; 1699 if (cow & MAP_NOFAULT) 1700 protoeflags |= MAP_ENTRY_NOFAULT; 1701 if (cow & MAP_DISABLE_SYNCER) 1702 protoeflags |= MAP_ENTRY_NOSYNC; 1703 if (cow & MAP_DISABLE_COREDUMP) 1704 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1705 if (cow & MAP_STACK_GROWS_DOWN) 1706 protoeflags |= MAP_ENTRY_GROWS_DOWN; 1707 if (cow & MAP_STACK_GROWS_UP) 1708 protoeflags |= MAP_ENTRY_GROWS_UP; 1709 if (cow & MAP_WRITECOUNT) 1710 protoeflags |= MAP_ENTRY_WRITECNT; 1711 if (cow & MAP_VN_EXEC) 1712 protoeflags |= MAP_ENTRY_VN_EXEC; 1713 if ((cow & MAP_CREATE_GUARD) != 0) 1714 protoeflags |= MAP_ENTRY_GUARD; 1715 if ((cow & MAP_CREATE_STACK_GAP_DN) != 0) 1716 protoeflags |= MAP_ENTRY_STACK_GAP_DN; 1717 if ((cow & MAP_CREATE_STACK_GAP_UP) != 0) 1718 protoeflags |= MAP_ENTRY_STACK_GAP_UP; 1719 if (cow & MAP_INHERIT_SHARE) 1720 inheritance = VM_INHERIT_SHARE; 1721 else 1722 inheritance = VM_INHERIT_DEFAULT; 1723 if ((cow & MAP_SPLIT_BOUNDARY_MASK) != 0) { 1724 /* This magically ignores index 0, for usual page size. */ 1725 bidx = (cow & MAP_SPLIT_BOUNDARY_MASK) >> 1726 MAP_SPLIT_BOUNDARY_SHIFT; 1727 if (bidx >= MAXPAGESIZES) 1728 return (KERN_INVALID_ARGUMENT); 1729 bdry = pagesizes[bidx] - 1; 1730 if ((start & bdry) != 0 || (end & bdry) != 0) 1731 return (KERN_INVALID_ARGUMENT); 1732 protoeflags |= bidx << MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 1733 } 1734 1735 cred = NULL; 1736 if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0) 1737 goto charged; 1738 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) && 1739 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) { 1740 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start)) 1741 return (KERN_RESOURCE_SHORTAGE); 1742 KASSERT(object == NULL || 1743 (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 || 1744 object->cred == NULL, 1745 ("overcommit: vm_map_insert o %p", object)); 1746 cred = curthread->td_ucred; 1747 } 1748 1749 charged: 1750 /* Expand the kernel pmap, if necessary. */ 1751 if (map == kernel_map && end > kernel_vm_end) 1752 pmap_growkernel(end); 1753 if (object != NULL) { 1754 /* 1755 * OBJ_ONEMAPPING must be cleared unless this mapping 1756 * is trivially proven to be the only mapping for any 1757 * of the object's pages. (Object granularity 1758 * reference counting is insufficient to recognize 1759 * aliases with precision.) 1760 */ 1761 if ((object->flags & OBJ_ANON) != 0) { 1762 VM_OBJECT_WLOCK(object); 1763 if (object->ref_count > 1 || object->shadow_count != 0) 1764 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1765 VM_OBJECT_WUNLOCK(object); 1766 } 1767 } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) == 1768 protoeflags && 1769 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP | 1770 MAP_VN_EXEC)) == 0 && 1771 prev_entry->end == start && (prev_entry->cred == cred || 1772 (prev_entry->object.vm_object != NULL && 1773 prev_entry->object.vm_object->cred == cred)) && 1774 vm_object_coalesce(prev_entry->object.vm_object, 1775 prev_entry->offset, 1776 (vm_size_t)(prev_entry->end - prev_entry->start), 1777 (vm_size_t)(end - prev_entry->end), cred != NULL && 1778 (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) { 1779 /* 1780 * We were able to extend the object. Determine if we 1781 * can extend the previous map entry to include the 1782 * new range as well. 1783 */ 1784 if (prev_entry->inheritance == inheritance && 1785 prev_entry->protection == prot && 1786 prev_entry->max_protection == max && 1787 prev_entry->wired_count == 0) { 1788 KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) == 1789 0, ("prev_entry %p has incoherent wiring", 1790 prev_entry)); 1791 if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0) 1792 map->size += end - prev_entry->end; 1793 vm_map_entry_resize(map, prev_entry, 1794 end - prev_entry->end); 1795 vm_map_try_merge_entries(map, prev_entry, next_entry); 1796 return (KERN_SUCCESS); 1797 } 1798 1799 /* 1800 * If we can extend the object but cannot extend the 1801 * map entry, we have to create a new map entry. We 1802 * must bump the ref count on the extended object to 1803 * account for it. object may be NULL. 1804 */ 1805 object = prev_entry->object.vm_object; 1806 offset = prev_entry->offset + 1807 (prev_entry->end - prev_entry->start); 1808 vm_object_reference(object); 1809 if (cred != NULL && object != NULL && object->cred != NULL && 1810 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 1811 /* Object already accounts for this uid. */ 1812 cred = NULL; 1813 } 1814 } 1815 if (cred != NULL) 1816 crhold(cred); 1817 1818 /* 1819 * Create a new entry 1820 */ 1821 new_entry = vm_map_entry_create(map); 1822 new_entry->start = start; 1823 new_entry->end = end; 1824 new_entry->cred = NULL; 1825 1826 new_entry->eflags = protoeflags; 1827 new_entry->object.vm_object = object; 1828 new_entry->offset = offset; 1829 1830 new_entry->inheritance = inheritance; 1831 new_entry->protection = prot; 1832 new_entry->max_protection = max; 1833 new_entry->wired_count = 0; 1834 new_entry->wiring_thread = NULL; 1835 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT; 1836 new_entry->next_read = start; 1837 1838 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry), 1839 ("overcommit: vm_map_insert leaks vm_map %p", new_entry)); 1840 new_entry->cred = cred; 1841 1842 /* 1843 * Insert the new entry into the list 1844 */ 1845 vm_map_entry_link(map, new_entry); 1846 if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0) 1847 map->size += new_entry->end - new_entry->start; 1848 1849 /* 1850 * Try to coalesce the new entry with both the previous and next 1851 * entries in the list. Previously, we only attempted to coalesce 1852 * with the previous entry when object is NULL. Here, we handle the 1853 * other cases, which are less common. 1854 */ 1855 vm_map_try_merge_entries(map, prev_entry, new_entry); 1856 vm_map_try_merge_entries(map, new_entry, next_entry); 1857 1858 if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) { 1859 vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset), 1860 end - start, cow & MAP_PREFAULT_PARTIAL); 1861 } 1862 1863 return (KERN_SUCCESS); 1864 } 1865 1866 /* 1867 * vm_map_findspace: 1868 * 1869 * Find the first fit (lowest VM address) for "length" free bytes 1870 * beginning at address >= start in the given map. 1871 * 1872 * In a vm_map_entry, "max_free" is the maximum amount of 1873 * contiguous free space between an entry in its subtree and a 1874 * neighbor of that entry. This allows finding a free region in 1875 * one path down the tree, so O(log n) amortized with splay 1876 * trees. 1877 * 1878 * The map must be locked, and leaves it so. 1879 * 1880 * Returns: starting address if sufficient space, 1881 * vm_map_max(map)-length+1 if insufficient space. 1882 */ 1883 vm_offset_t 1884 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length) 1885 { 1886 vm_map_entry_t header, llist, rlist, root, y; 1887 vm_size_t left_length, max_free_left, max_free_right; 1888 vm_offset_t gap_end; 1889 1890 VM_MAP_ASSERT_LOCKED(map); 1891 1892 /* 1893 * Request must fit within min/max VM address and must avoid 1894 * address wrap. 1895 */ 1896 start = MAX(start, vm_map_min(map)); 1897 if (start >= vm_map_max(map) || length > vm_map_max(map) - start) 1898 return (vm_map_max(map) - length + 1); 1899 1900 /* Empty tree means wide open address space. */ 1901 if (map->root == NULL) 1902 return (start); 1903 1904 /* 1905 * After splay_split, if start is within an entry, push it to the start 1906 * of the following gap. If rlist is at the end of the gap containing 1907 * start, save the end of that gap in gap_end to see if the gap is big 1908 * enough; otherwise set gap_end to start skip gap-checking and move 1909 * directly to a search of the right subtree. 1910 */ 1911 header = &map->header; 1912 root = vm_map_splay_split(map, start, length, &llist, &rlist); 1913 gap_end = rlist->start; 1914 if (root != NULL) { 1915 start = root->end; 1916 if (root->right != rlist) 1917 gap_end = start; 1918 max_free_left = vm_map_splay_merge_left(header, root, llist); 1919 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1920 } else if (rlist != header) { 1921 root = rlist; 1922 rlist = root->left; 1923 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1924 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1925 } else { 1926 root = llist; 1927 llist = root->right; 1928 max_free_left = vm_map_splay_merge_left(header, root, llist); 1929 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1930 } 1931 root->max_free = vm_size_max(max_free_left, max_free_right); 1932 map->root = root; 1933 VM_MAP_ASSERT_CONSISTENT(map); 1934 if (length <= gap_end - start) 1935 return (start); 1936 1937 /* With max_free, can immediately tell if no solution. */ 1938 if (root->right == header || length > root->right->max_free) 1939 return (vm_map_max(map) - length + 1); 1940 1941 /* 1942 * Splay for the least large-enough gap in the right subtree. 1943 */ 1944 llist = rlist = header; 1945 for (left_length = 0;; 1946 left_length = vm_map_entry_max_free_left(root, llist)) { 1947 if (length <= left_length) 1948 SPLAY_LEFT_STEP(root, y, llist, rlist, 1949 length <= vm_map_entry_max_free_left(y, llist)); 1950 else 1951 SPLAY_RIGHT_STEP(root, y, llist, rlist, 1952 length > vm_map_entry_max_free_left(y, root)); 1953 if (root == NULL) 1954 break; 1955 } 1956 root = llist; 1957 llist = root->right; 1958 max_free_left = vm_map_splay_merge_left(header, root, llist); 1959 if (rlist == header) { 1960 root->max_free = vm_size_max(max_free_left, 1961 vm_map_splay_merge_succ(header, root, rlist)); 1962 } else { 1963 y = rlist; 1964 rlist = y->left; 1965 y->max_free = vm_size_max( 1966 vm_map_splay_merge_pred(root, y, root), 1967 vm_map_splay_merge_right(header, y, rlist)); 1968 root->max_free = vm_size_max(max_free_left, y->max_free); 1969 } 1970 map->root = root; 1971 VM_MAP_ASSERT_CONSISTENT(map); 1972 return (root->end); 1973 } 1974 1975 int 1976 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1977 vm_offset_t start, vm_size_t length, vm_prot_t prot, 1978 vm_prot_t max, int cow) 1979 { 1980 vm_offset_t end; 1981 int result; 1982 1983 end = start + length; 1984 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 1985 object == NULL, 1986 ("vm_map_fixed: non-NULL backing object for stack")); 1987 vm_map_lock(map); 1988 VM_MAP_RANGE_CHECK(map, start, end); 1989 if ((cow & MAP_CHECK_EXCL) == 0) { 1990 result = vm_map_delete(map, start, end); 1991 if (result != KERN_SUCCESS) 1992 goto out; 1993 } 1994 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 1995 result = vm_map_stack_locked(map, start, length, sgrowsiz, 1996 prot, max, cow); 1997 } else { 1998 result = vm_map_insert(map, object, offset, start, end, 1999 prot, max, cow); 2000 } 2001 out: 2002 vm_map_unlock(map); 2003 return (result); 2004 } 2005 2006 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10}; 2007 static const int aslr_pages_rnd_32[2] = {0x100, 0x4}; 2008 2009 static int cluster_anon = 1; 2010 SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW, 2011 &cluster_anon, 0, 2012 "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always"); 2013 2014 static bool 2015 clustering_anon_allowed(vm_offset_t addr) 2016 { 2017 2018 switch (cluster_anon) { 2019 case 0: 2020 return (false); 2021 case 1: 2022 return (addr == 0); 2023 case 2: 2024 default: 2025 return (true); 2026 } 2027 } 2028 2029 static long aslr_restarts; 2030 SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD, 2031 &aslr_restarts, 0, 2032 "Number of aslr failures"); 2033 2034 /* 2035 * Searches for the specified amount of free space in the given map with the 2036 * specified alignment. Performs an address-ordered, first-fit search from 2037 * the given address "*addr", with an optional upper bound "max_addr". If the 2038 * parameter "alignment" is zero, then the alignment is computed from the 2039 * given (object, offset) pair so as to enable the greatest possible use of 2040 * superpage mappings. Returns KERN_SUCCESS and the address of the free space 2041 * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE. 2042 * 2043 * The map must be locked. Initially, there must be at least "length" bytes 2044 * of free space at the given address. 2045 */ 2046 static int 2047 vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2048 vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, 2049 vm_offset_t alignment) 2050 { 2051 vm_offset_t aligned_addr, free_addr; 2052 2053 VM_MAP_ASSERT_LOCKED(map); 2054 free_addr = *addr; 2055 KASSERT(free_addr == vm_map_findspace(map, free_addr, length), 2056 ("caller failed to provide space %#jx at address %p", 2057 (uintmax_t)length, (void *)free_addr)); 2058 for (;;) { 2059 /* 2060 * At the start of every iteration, the free space at address 2061 * "*addr" is at least "length" bytes. 2062 */ 2063 if (alignment == 0) 2064 pmap_align_superpage(object, offset, addr, length); 2065 else if ((*addr & (alignment - 1)) != 0) { 2066 *addr &= ~(alignment - 1); 2067 *addr += alignment; 2068 } 2069 aligned_addr = *addr; 2070 if (aligned_addr == free_addr) { 2071 /* 2072 * Alignment did not change "*addr", so "*addr" must 2073 * still provide sufficient free space. 2074 */ 2075 return (KERN_SUCCESS); 2076 } 2077 2078 /* 2079 * Test for address wrap on "*addr". A wrapped "*addr" could 2080 * be a valid address, in which case vm_map_findspace() cannot 2081 * be relied upon to fail. 2082 */ 2083 if (aligned_addr < free_addr) 2084 return (KERN_NO_SPACE); 2085 *addr = vm_map_findspace(map, aligned_addr, length); 2086 if (*addr + length > vm_map_max(map) || 2087 (max_addr != 0 && *addr + length > max_addr)) 2088 return (KERN_NO_SPACE); 2089 free_addr = *addr; 2090 if (free_addr == aligned_addr) { 2091 /* 2092 * If a successful call to vm_map_findspace() did not 2093 * change "*addr", then "*addr" must still be aligned 2094 * and provide sufficient free space. 2095 */ 2096 return (KERN_SUCCESS); 2097 } 2098 } 2099 } 2100 2101 int 2102 vm_map_find_aligned(vm_map_t map, vm_offset_t *addr, vm_size_t length, 2103 vm_offset_t max_addr, vm_offset_t alignment) 2104 { 2105 /* XXXKIB ASLR eh ? */ 2106 *addr = vm_map_findspace(map, *addr, length); 2107 if (*addr + length > vm_map_max(map) || 2108 (max_addr != 0 && *addr + length > max_addr)) 2109 return (KERN_NO_SPACE); 2110 return (vm_map_alignspace(map, NULL, 0, addr, length, max_addr, 2111 alignment)); 2112 } 2113 2114 /* 2115 * vm_map_find finds an unallocated region in the target address 2116 * map with the given length. The search is defined to be 2117 * first-fit from the specified address; the region found is 2118 * returned in the same parameter. 2119 * 2120 * If object is non-NULL, ref count must be bumped by caller 2121 * prior to making call to account for the new entry. 2122 */ 2123 int 2124 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2125 vm_offset_t *addr, /* IN/OUT */ 2126 vm_size_t length, vm_offset_t max_addr, int find_space, 2127 vm_prot_t prot, vm_prot_t max, int cow) 2128 { 2129 vm_offset_t alignment, curr_min_addr, min_addr; 2130 int gap, pidx, rv, try; 2131 bool cluster, en_aslr, update_anon; 2132 2133 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 2134 object == NULL, 2135 ("vm_map_find: non-NULL backing object for stack")); 2136 MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE && 2137 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0)); 2138 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL || 2139 (object->flags & OBJ_COLORED) == 0)) 2140 find_space = VMFS_ANY_SPACE; 2141 if (find_space >> 8 != 0) { 2142 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags")); 2143 alignment = (vm_offset_t)1 << (find_space >> 8); 2144 } else 2145 alignment = 0; 2146 en_aslr = (map->flags & MAP_ASLR) != 0; 2147 update_anon = cluster = clustering_anon_allowed(*addr) && 2148 (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 && 2149 find_space != VMFS_NO_SPACE && object == NULL && 2150 (cow & (MAP_INHERIT_SHARE | MAP_STACK_GROWS_UP | 2151 MAP_STACK_GROWS_DOWN)) == 0 && prot != PROT_NONE; 2152 curr_min_addr = min_addr = *addr; 2153 if (en_aslr && min_addr == 0 && !cluster && 2154 find_space != VMFS_NO_SPACE && 2155 (map->flags & MAP_ASLR_IGNSTART) != 0) 2156 curr_min_addr = min_addr = vm_map_min(map); 2157 try = 0; 2158 vm_map_lock(map); 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 = MAXPAGESIZES > 1 && pagesizes[1] != 0 && 2206 (find_space == VMFS_SUPER_SPACE || find_space == 2207 VMFS_OPTIMAL_SPACE) ? 1 : 0; 2208 gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR && 2209 (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ? 2210 aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx]; 2211 *addr = vm_map_findspace(map, curr_min_addr, 2212 length + gap * pagesizes[pidx]); 2213 if (*addr + length + gap * pagesizes[pidx] > 2214 vm_map_max(map)) 2215 goto again; 2216 /* And randomize the start address. */ 2217 *addr += (arc4random() % gap) * pagesizes[pidx]; 2218 if (max_addr != 0 && *addr + length > max_addr) 2219 goto again; 2220 } else { 2221 *addr = vm_map_findspace(map, curr_min_addr, length); 2222 if (*addr + length > vm_map_max(map) || 2223 (max_addr != 0 && *addr + length > max_addr)) { 2224 if (cluster) { 2225 cluster = false; 2226 MPASS(try == 1); 2227 goto again; 2228 } 2229 rv = KERN_NO_SPACE; 2230 goto done; 2231 } 2232 } 2233 2234 if (find_space != VMFS_ANY_SPACE && 2235 (rv = vm_map_alignspace(map, object, offset, addr, length, 2236 max_addr, alignment)) != KERN_SUCCESS) { 2237 if (find_space == VMFS_OPTIMAL_SPACE) { 2238 find_space = VMFS_ANY_SPACE; 2239 curr_min_addr = min_addr; 2240 cluster = update_anon; 2241 try = 0; 2242 goto again; 2243 } 2244 goto done; 2245 } 2246 } else if ((cow & MAP_REMAP) != 0) { 2247 if (!vm_map_range_valid(map, *addr, *addr + length)) { 2248 rv = KERN_INVALID_ADDRESS; 2249 goto done; 2250 } 2251 rv = vm_map_delete(map, *addr, *addr + length); 2252 if (rv != KERN_SUCCESS) 2253 goto done; 2254 } 2255 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 2256 rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot, 2257 max, cow); 2258 } else { 2259 rv = vm_map_insert(map, object, offset, *addr, *addr + length, 2260 prot, max, cow); 2261 } 2262 if (rv == KERN_SUCCESS && update_anon) 2263 map->anon_loc = *addr + length; 2264 done: 2265 vm_map_unlock(map); 2266 return (rv); 2267 } 2268 2269 /* 2270 * vm_map_find_min() is a variant of vm_map_find() that takes an 2271 * additional parameter (min_addr) and treats the given address 2272 * (*addr) differently. Specifically, it treats *addr as a hint 2273 * and not as the minimum address where the mapping is created. 2274 * 2275 * This function works in two phases. First, it tries to 2276 * allocate above the hint. If that fails and the hint is 2277 * greater than min_addr, it performs a second pass, replacing 2278 * the hint with min_addr as the minimum address for the 2279 * allocation. 2280 */ 2281 int 2282 vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2283 vm_offset_t *addr, vm_size_t length, vm_offset_t min_addr, 2284 vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max, 2285 int cow) 2286 { 2287 vm_offset_t hint; 2288 int rv; 2289 2290 hint = *addr; 2291 for (;;) { 2292 rv = vm_map_find(map, object, offset, addr, length, max_addr, 2293 find_space, prot, max, cow); 2294 if (rv == KERN_SUCCESS || min_addr >= hint) 2295 return (rv); 2296 *addr = hint = min_addr; 2297 } 2298 } 2299 2300 /* 2301 * A map entry with any of the following flags set must not be merged with 2302 * another entry. 2303 */ 2304 #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP | \ 2305 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC) 2306 2307 static bool 2308 vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry) 2309 { 2310 2311 KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 || 2312 (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0, 2313 ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable", 2314 prev, entry)); 2315 return (prev->end == entry->start && 2316 prev->object.vm_object == entry->object.vm_object && 2317 (prev->object.vm_object == NULL || 2318 prev->offset + (prev->end - prev->start) == entry->offset) && 2319 prev->eflags == entry->eflags && 2320 prev->protection == entry->protection && 2321 prev->max_protection == entry->max_protection && 2322 prev->inheritance == entry->inheritance && 2323 prev->wired_count == entry->wired_count && 2324 prev->cred == entry->cred); 2325 } 2326 2327 static void 2328 vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry) 2329 { 2330 2331 /* 2332 * If the backing object is a vnode object, vm_object_deallocate() 2333 * calls vrele(). However, vrele() does not lock the vnode because 2334 * the vnode has additional references. Thus, the map lock can be 2335 * kept without causing a lock-order reversal with the vnode lock. 2336 * 2337 * Since we count the number of virtual page mappings in 2338 * object->un_pager.vnp.writemappings, the writemappings value 2339 * should not be adjusted when the entry is disposed of. 2340 */ 2341 if (entry->object.vm_object != NULL) 2342 vm_object_deallocate(entry->object.vm_object); 2343 if (entry->cred != NULL) 2344 crfree(entry->cred); 2345 vm_map_entry_dispose(map, entry); 2346 } 2347 2348 /* 2349 * vm_map_try_merge_entries: 2350 * 2351 * Compare the given map entry to its predecessor, and merge its precessor 2352 * into it if possible. The entry remains valid, and may be extended. 2353 * The predecessor may be deleted. 2354 * 2355 * The map must be locked. 2356 */ 2357 void 2358 vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev_entry, 2359 vm_map_entry_t entry) 2360 { 2361 2362 VM_MAP_ASSERT_LOCKED(map); 2363 if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 && 2364 vm_map_mergeable_neighbors(prev_entry, entry)) { 2365 vm_map_entry_unlink(map, prev_entry, UNLINK_MERGE_NEXT); 2366 vm_map_merged_neighbor_dispose(map, prev_entry); 2367 } 2368 } 2369 2370 /* 2371 * vm_map_entry_back: 2372 * 2373 * Allocate an object to back a map entry. 2374 */ 2375 static inline void 2376 vm_map_entry_back(vm_map_entry_t entry) 2377 { 2378 vm_object_t object; 2379 2380 KASSERT(entry->object.vm_object == NULL, 2381 ("map entry %p has backing object", entry)); 2382 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2383 ("map entry %p is a submap", entry)); 2384 object = vm_object_allocate_anon(atop(entry->end - entry->start), NULL, 2385 entry->cred, entry->end - entry->start); 2386 entry->object.vm_object = object; 2387 entry->offset = 0; 2388 entry->cred = NULL; 2389 } 2390 2391 /* 2392 * vm_map_entry_charge_object 2393 * 2394 * If there is no object backing this entry, create one. Otherwise, if 2395 * the entry has cred, give it to the backing object. 2396 */ 2397 static inline void 2398 vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry) 2399 { 2400 2401 VM_MAP_ASSERT_LOCKED(map); 2402 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2403 ("map entry %p is a submap", entry)); 2404 if (entry->object.vm_object == NULL && !map->system_map && 2405 (entry->eflags & MAP_ENTRY_GUARD) == 0) 2406 vm_map_entry_back(entry); 2407 else if (entry->object.vm_object != NULL && 2408 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) && 2409 entry->cred != NULL) { 2410 VM_OBJECT_WLOCK(entry->object.vm_object); 2411 KASSERT(entry->object.vm_object->cred == NULL, 2412 ("OVERCOMMIT: %s: both cred e %p", __func__, entry)); 2413 entry->object.vm_object->cred = entry->cred; 2414 entry->object.vm_object->charge = entry->end - entry->start; 2415 VM_OBJECT_WUNLOCK(entry->object.vm_object); 2416 entry->cred = NULL; 2417 } 2418 } 2419 2420 /* 2421 * vm_map_entry_clone 2422 * 2423 * Create a duplicate map entry for clipping. 2424 */ 2425 static vm_map_entry_t 2426 vm_map_entry_clone(vm_map_t map, vm_map_entry_t entry) 2427 { 2428 vm_map_entry_t new_entry; 2429 2430 VM_MAP_ASSERT_LOCKED(map); 2431 2432 /* 2433 * Create a backing object now, if none exists, so that more individual 2434 * objects won't be created after the map entry is split. 2435 */ 2436 vm_map_entry_charge_object(map, entry); 2437 2438 /* Clone the entry. */ 2439 new_entry = vm_map_entry_create(map); 2440 *new_entry = *entry; 2441 if (new_entry->cred != NULL) 2442 crhold(entry->cred); 2443 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2444 vm_object_reference(new_entry->object.vm_object); 2445 vm_map_entry_set_vnode_text(new_entry, true); 2446 /* 2447 * The object->un_pager.vnp.writemappings for the object of 2448 * MAP_ENTRY_WRITECNT type entry shall be kept as is here. The 2449 * virtual pages are re-distributed among the clipped entries, 2450 * so the sum is left the same. 2451 */ 2452 } 2453 return (new_entry); 2454 } 2455 2456 /* 2457 * vm_map_clip_start: [ internal use only ] 2458 * 2459 * Asserts that the given entry begins at or after 2460 * the specified address; if necessary, 2461 * it splits the entry into two. 2462 */ 2463 static int 2464 vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t startaddr) 2465 { 2466 vm_map_entry_t new_entry; 2467 int bdry_idx; 2468 2469 if (!map->system_map) 2470 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2471 "%s: map %p entry %p start 0x%jx", __func__, map, entry, 2472 (uintmax_t)startaddr); 2473 2474 if (startaddr <= entry->start) 2475 return (KERN_SUCCESS); 2476 2477 VM_MAP_ASSERT_LOCKED(map); 2478 KASSERT(entry->end > startaddr && entry->start < startaddr, 2479 ("%s: invalid clip of entry %p", __func__, entry)); 2480 2481 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 2482 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 2483 if (bdry_idx != 0) { 2484 if ((startaddr & (pagesizes[bdry_idx] - 1)) != 0) 2485 return (KERN_INVALID_ARGUMENT); 2486 } 2487 2488 new_entry = vm_map_entry_clone(map, entry); 2489 2490 /* 2491 * Split off the front portion. Insert the new entry BEFORE this one, 2492 * so that this entry has the specified starting address. 2493 */ 2494 new_entry->end = startaddr; 2495 vm_map_entry_link(map, new_entry); 2496 return (KERN_SUCCESS); 2497 } 2498 2499 /* 2500 * vm_map_lookup_clip_start: 2501 * 2502 * Find the entry at or just after 'start', and clip it if 'start' is in 2503 * the interior of the entry. Return entry after 'start', and in 2504 * prev_entry set the entry before 'start'. 2505 */ 2506 static int 2507 vm_map_lookup_clip_start(vm_map_t map, vm_offset_t start, 2508 vm_map_entry_t *res_entry, vm_map_entry_t *prev_entry) 2509 { 2510 vm_map_entry_t entry; 2511 int rv; 2512 2513 if (!map->system_map) 2514 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2515 "%s: map %p start 0x%jx prev %p", __func__, map, 2516 (uintmax_t)start, prev_entry); 2517 2518 if (vm_map_lookup_entry(map, start, prev_entry)) { 2519 entry = *prev_entry; 2520 rv = vm_map_clip_start(map, entry, start); 2521 if (rv != KERN_SUCCESS) 2522 return (rv); 2523 *prev_entry = vm_map_entry_pred(entry); 2524 } else 2525 entry = vm_map_entry_succ(*prev_entry); 2526 *res_entry = entry; 2527 return (KERN_SUCCESS); 2528 } 2529 2530 /* 2531 * vm_map_clip_end: [ internal use only ] 2532 * 2533 * Asserts that the given entry ends at or before 2534 * the specified address; if necessary, 2535 * it splits the entry into two. 2536 */ 2537 static int 2538 vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t endaddr) 2539 { 2540 vm_map_entry_t new_entry; 2541 int bdry_idx; 2542 2543 if (!map->system_map) 2544 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2545 "%s: map %p entry %p end 0x%jx", __func__, map, entry, 2546 (uintmax_t)endaddr); 2547 2548 if (endaddr >= entry->end) 2549 return (KERN_SUCCESS); 2550 2551 VM_MAP_ASSERT_LOCKED(map); 2552 KASSERT(entry->start < endaddr && entry->end > endaddr, 2553 ("%s: invalid clip of entry %p", __func__, entry)); 2554 2555 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 2556 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 2557 if (bdry_idx != 0) { 2558 if ((endaddr & (pagesizes[bdry_idx] - 1)) != 0) 2559 return (KERN_INVALID_ARGUMENT); 2560 } 2561 2562 new_entry = vm_map_entry_clone(map, entry); 2563 2564 /* 2565 * Split off the back portion. Insert the new entry AFTER this one, 2566 * so that this entry has the specified ending address. 2567 */ 2568 new_entry->start = endaddr; 2569 vm_map_entry_link(map, new_entry); 2570 2571 return (KERN_SUCCESS); 2572 } 2573 2574 /* 2575 * vm_map_submap: [ kernel use only ] 2576 * 2577 * Mark the given range as handled by a subordinate map. 2578 * 2579 * This range must have been created with vm_map_find, 2580 * and no other operations may have been performed on this 2581 * range prior to calling vm_map_submap. 2582 * 2583 * Only a limited number of operations can be performed 2584 * within this rage after calling vm_map_submap: 2585 * vm_fault 2586 * [Don't try vm_map_copy!] 2587 * 2588 * To remove a submapping, one must first remove the 2589 * range from the superior map, and then destroy the 2590 * submap (if desired). [Better yet, don't try it.] 2591 */ 2592 int 2593 vm_map_submap( 2594 vm_map_t map, 2595 vm_offset_t start, 2596 vm_offset_t end, 2597 vm_map_t submap) 2598 { 2599 vm_map_entry_t entry; 2600 int result; 2601 2602 result = KERN_INVALID_ARGUMENT; 2603 2604 vm_map_lock(submap); 2605 submap->flags |= MAP_IS_SUB_MAP; 2606 vm_map_unlock(submap); 2607 2608 vm_map_lock(map); 2609 VM_MAP_RANGE_CHECK(map, start, end); 2610 if (vm_map_lookup_entry(map, start, &entry) && entry->end >= end && 2611 (entry->eflags & MAP_ENTRY_COW) == 0 && 2612 entry->object.vm_object == NULL) { 2613 result = vm_map_clip_start(map, entry, start); 2614 if (result != KERN_SUCCESS) 2615 goto unlock; 2616 result = vm_map_clip_end(map, entry, end); 2617 if (result != KERN_SUCCESS) 2618 goto unlock; 2619 entry->object.sub_map = submap; 2620 entry->eflags |= MAP_ENTRY_IS_SUB_MAP; 2621 result = KERN_SUCCESS; 2622 } 2623 unlock: 2624 vm_map_unlock(map); 2625 2626 if (result != KERN_SUCCESS) { 2627 vm_map_lock(submap); 2628 submap->flags &= ~MAP_IS_SUB_MAP; 2629 vm_map_unlock(submap); 2630 } 2631 return (result); 2632 } 2633 2634 /* 2635 * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified 2636 */ 2637 #define MAX_INIT_PT 96 2638 2639 /* 2640 * vm_map_pmap_enter: 2641 * 2642 * Preload the specified map's pmap with mappings to the specified 2643 * object's memory-resident pages. No further physical pages are 2644 * allocated, and no further virtual pages are retrieved from secondary 2645 * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a 2646 * limited number of page mappings are created at the low-end of the 2647 * specified address range. (For this purpose, a superpage mapping 2648 * counts as one page mapping.) Otherwise, all resident pages within 2649 * the specified address range are mapped. 2650 */ 2651 static void 2652 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 2653 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags) 2654 { 2655 vm_offset_t start; 2656 vm_page_t p, p_start; 2657 vm_pindex_t mask, psize, threshold, tmpidx; 2658 2659 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL) 2660 return; 2661 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2662 VM_OBJECT_WLOCK(object); 2663 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2664 pmap_object_init_pt(map->pmap, addr, object, pindex, 2665 size); 2666 VM_OBJECT_WUNLOCK(object); 2667 return; 2668 } 2669 VM_OBJECT_LOCK_DOWNGRADE(object); 2670 } else 2671 VM_OBJECT_RLOCK(object); 2672 2673 psize = atop(size); 2674 if (psize + pindex > object->size) { 2675 if (pindex >= object->size) { 2676 VM_OBJECT_RUNLOCK(object); 2677 return; 2678 } 2679 psize = object->size - pindex; 2680 } 2681 2682 start = 0; 2683 p_start = NULL; 2684 threshold = MAX_INIT_PT; 2685 2686 p = vm_page_find_least(object, pindex); 2687 /* 2688 * Assert: the variable p is either (1) the page with the 2689 * least pindex greater than or equal to the parameter pindex 2690 * or (2) NULL. 2691 */ 2692 for (; 2693 p != NULL && (tmpidx = p->pindex - pindex) < psize; 2694 p = TAILQ_NEXT(p, listq)) { 2695 /* 2696 * don't allow an madvise to blow away our really 2697 * free pages allocating pv entries. 2698 */ 2699 if (((flags & MAP_PREFAULT_MADVISE) != 0 && 2700 vm_page_count_severe()) || 2701 ((flags & MAP_PREFAULT_PARTIAL) != 0 && 2702 tmpidx >= threshold)) { 2703 psize = tmpidx; 2704 break; 2705 } 2706 if (vm_page_all_valid(p)) { 2707 if (p_start == NULL) { 2708 start = addr + ptoa(tmpidx); 2709 p_start = p; 2710 } 2711 /* Jump ahead if a superpage mapping is possible. */ 2712 if (p->psind > 0 && ((addr + ptoa(tmpidx)) & 2713 (pagesizes[p->psind] - 1)) == 0) { 2714 mask = atop(pagesizes[p->psind]) - 1; 2715 if (tmpidx + mask < psize && 2716 vm_page_ps_test(p, PS_ALL_VALID, NULL)) { 2717 p += mask; 2718 threshold += mask; 2719 } 2720 } 2721 } else if (p_start != NULL) { 2722 pmap_enter_object(map->pmap, start, addr + 2723 ptoa(tmpidx), p_start, prot); 2724 p_start = NULL; 2725 } 2726 } 2727 if (p_start != NULL) 2728 pmap_enter_object(map->pmap, start, addr + ptoa(psize), 2729 p_start, prot); 2730 VM_OBJECT_RUNLOCK(object); 2731 } 2732 2733 /* 2734 * vm_map_protect: 2735 * 2736 * Sets the protection and/or the maximum protection of the 2737 * specified address region in the target map. 2738 */ 2739 int 2740 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 2741 vm_prot_t new_prot, vm_prot_t new_maxprot, int flags) 2742 { 2743 vm_map_entry_t entry, first_entry, in_tran, prev_entry; 2744 vm_object_t obj; 2745 struct ucred *cred; 2746 vm_prot_t old_prot; 2747 int rv; 2748 2749 if (start == end) 2750 return (KERN_SUCCESS); 2751 2752 if ((flags & (VM_MAP_PROTECT_SET_PROT | VM_MAP_PROTECT_SET_MAXPROT)) == 2753 (VM_MAP_PROTECT_SET_PROT | VM_MAP_PROTECT_SET_MAXPROT) && 2754 (new_prot & new_maxprot) != new_prot) 2755 return (KERN_OUT_OF_BOUNDS); 2756 2757 again: 2758 in_tran = NULL; 2759 vm_map_lock(map); 2760 2761 if ((map->flags & MAP_WXORX) != 0 && 2762 (flags & VM_MAP_PROTECT_SET_PROT) != 0 && 2763 (new_prot & (VM_PROT_WRITE | VM_PROT_EXECUTE)) == (VM_PROT_WRITE | 2764 VM_PROT_EXECUTE)) { 2765 vm_map_unlock(map); 2766 return (KERN_PROTECTION_FAILURE); 2767 } 2768 2769 /* 2770 * Ensure that we are not concurrently wiring pages. vm_map_wire() may 2771 * need to fault pages into the map and will drop the map lock while 2772 * doing so, and the VM object may end up in an inconsistent state if we 2773 * update the protection on the map entry in between faults. 2774 */ 2775 vm_map_wait_busy(map); 2776 2777 VM_MAP_RANGE_CHECK(map, start, end); 2778 2779 if (!vm_map_lookup_entry(map, start, &first_entry)) 2780 first_entry = vm_map_entry_succ(first_entry); 2781 2782 /* 2783 * Make a first pass to check for protection violations. 2784 */ 2785 for (entry = first_entry; entry->start < end; 2786 entry = vm_map_entry_succ(entry)) { 2787 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 2788 continue; 2789 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 2790 vm_map_unlock(map); 2791 return (KERN_INVALID_ARGUMENT); 2792 } 2793 if ((flags & VM_MAP_PROTECT_SET_PROT) == 0) 2794 new_prot = entry->protection; 2795 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) == 0) 2796 new_maxprot = entry->max_protection; 2797 if ((new_prot & entry->max_protection) != new_prot || 2798 (new_maxprot & entry->max_protection) != new_maxprot) { 2799 vm_map_unlock(map); 2800 return (KERN_PROTECTION_FAILURE); 2801 } 2802 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0) 2803 in_tran = entry; 2804 } 2805 2806 /* 2807 * Postpone the operation until all in-transition map entries have 2808 * stabilized. An in-transition entry might already have its pages 2809 * wired and wired_count incremented, but not yet have its 2810 * MAP_ENTRY_USER_WIRED flag set. In which case, we would fail to call 2811 * vm_fault_copy_entry() in the final loop below. 2812 */ 2813 if (in_tran != NULL) { 2814 in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2815 vm_map_unlock_and_wait(map, 0); 2816 goto again; 2817 } 2818 2819 /* 2820 * Before changing the protections, try to reserve swap space for any 2821 * private (i.e., copy-on-write) mappings that are transitioning from 2822 * read-only to read/write access. If a reservation fails, break out 2823 * of this loop early and let the next loop simplify the entries, since 2824 * some may now be mergeable. 2825 */ 2826 rv = vm_map_clip_start(map, first_entry, start); 2827 if (rv != KERN_SUCCESS) { 2828 vm_map_unlock(map); 2829 return (rv); 2830 } 2831 for (entry = first_entry; entry->start < end; 2832 entry = vm_map_entry_succ(entry)) { 2833 rv = vm_map_clip_end(map, entry, end); 2834 if (rv != KERN_SUCCESS) { 2835 vm_map_unlock(map); 2836 return (rv); 2837 } 2838 2839 if ((flags & VM_MAP_PROTECT_SET_PROT) == 0) 2840 new_prot = entry->protection; 2841 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) == 0) 2842 new_maxprot = entry->max_protection; 2843 2844 if ((flags & VM_MAP_PROTECT_SET_PROT) == 0 || 2845 ((new_prot & ~entry->protection) & VM_PROT_WRITE) == 0 || 2846 ENTRY_CHARGED(entry) || 2847 (entry->eflags & MAP_ENTRY_GUARD) != 0) 2848 continue; 2849 2850 cred = curthread->td_ucred; 2851 obj = entry->object.vm_object; 2852 2853 if (obj == NULL || 2854 (entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0) { 2855 if (!swap_reserve(entry->end - entry->start)) { 2856 rv = KERN_RESOURCE_SHORTAGE; 2857 end = entry->end; 2858 break; 2859 } 2860 crhold(cred); 2861 entry->cred = cred; 2862 continue; 2863 } 2864 2865 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) 2866 continue; 2867 VM_OBJECT_WLOCK(obj); 2868 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) { 2869 VM_OBJECT_WUNLOCK(obj); 2870 continue; 2871 } 2872 2873 /* 2874 * Charge for the whole object allocation now, since 2875 * we cannot distinguish between non-charged and 2876 * charged clipped mapping of the same object later. 2877 */ 2878 KASSERT(obj->charge == 0, 2879 ("vm_map_protect: object %p overcharged (entry %p)", 2880 obj, entry)); 2881 if (!swap_reserve(ptoa(obj->size))) { 2882 VM_OBJECT_WUNLOCK(obj); 2883 rv = KERN_RESOURCE_SHORTAGE; 2884 end = entry->end; 2885 break; 2886 } 2887 2888 crhold(cred); 2889 obj->cred = cred; 2890 obj->charge = ptoa(obj->size); 2891 VM_OBJECT_WUNLOCK(obj); 2892 } 2893 2894 /* 2895 * If enough swap space was available, go back and fix up protections. 2896 * Otherwise, just simplify entries, since some may have been modified. 2897 * [Note that clipping is not necessary the second time.] 2898 */ 2899 for (prev_entry = vm_map_entry_pred(first_entry), entry = first_entry; 2900 entry->start < end; 2901 vm_map_try_merge_entries(map, prev_entry, entry), 2902 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 2903 if (rv != KERN_SUCCESS || 2904 (entry->eflags & MAP_ENTRY_GUARD) != 0) 2905 continue; 2906 2907 old_prot = entry->protection; 2908 2909 if ((flags & VM_MAP_PROTECT_SET_MAXPROT) != 0) { 2910 entry->max_protection = new_maxprot; 2911 entry->protection = new_maxprot & old_prot; 2912 } 2913 if ((flags & VM_MAP_PROTECT_SET_PROT) != 0) 2914 entry->protection = new_prot; 2915 2916 /* 2917 * For user wired map entries, the normal lazy evaluation of 2918 * write access upgrades through soft page faults is 2919 * undesirable. Instead, immediately copy any pages that are 2920 * copy-on-write and enable write access in the physical map. 2921 */ 2922 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0 && 2923 (entry->protection & VM_PROT_WRITE) != 0 && 2924 (old_prot & VM_PROT_WRITE) == 0) 2925 vm_fault_copy_entry(map, map, entry, entry, NULL); 2926 2927 /* 2928 * When restricting access, update the physical map. Worry 2929 * about copy-on-write here. 2930 */ 2931 if ((old_prot & ~entry->protection) != 0) { 2932 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 2933 VM_PROT_ALL) 2934 pmap_protect(map->pmap, entry->start, 2935 entry->end, 2936 entry->protection & MASK(entry)); 2937 #undef MASK 2938 } 2939 } 2940 vm_map_try_merge_entries(map, prev_entry, entry); 2941 vm_map_unlock(map); 2942 return (rv); 2943 } 2944 2945 /* 2946 * vm_map_madvise: 2947 * 2948 * This routine traverses a processes map handling the madvise 2949 * system call. Advisories are classified as either those effecting 2950 * the vm_map_entry structure, or those effecting the underlying 2951 * objects. 2952 */ 2953 int 2954 vm_map_madvise( 2955 vm_map_t map, 2956 vm_offset_t start, 2957 vm_offset_t end, 2958 int behav) 2959 { 2960 vm_map_entry_t entry, prev_entry; 2961 int rv; 2962 bool modify_map; 2963 2964 /* 2965 * Some madvise calls directly modify the vm_map_entry, in which case 2966 * we need to use an exclusive lock on the map and we need to perform 2967 * various clipping operations. Otherwise we only need a read-lock 2968 * on the map. 2969 */ 2970 switch(behav) { 2971 case MADV_NORMAL: 2972 case MADV_SEQUENTIAL: 2973 case MADV_RANDOM: 2974 case MADV_NOSYNC: 2975 case MADV_AUTOSYNC: 2976 case MADV_NOCORE: 2977 case MADV_CORE: 2978 if (start == end) 2979 return (0); 2980 modify_map = true; 2981 vm_map_lock(map); 2982 break; 2983 case MADV_WILLNEED: 2984 case MADV_DONTNEED: 2985 case MADV_FREE: 2986 if (start == end) 2987 return (0); 2988 modify_map = false; 2989 vm_map_lock_read(map); 2990 break; 2991 default: 2992 return (EINVAL); 2993 } 2994 2995 /* 2996 * Locate starting entry and clip if necessary. 2997 */ 2998 VM_MAP_RANGE_CHECK(map, start, end); 2999 3000 if (modify_map) { 3001 /* 3002 * madvise behaviors that are implemented in the vm_map_entry. 3003 * 3004 * We clip the vm_map_entry so that behavioral changes are 3005 * limited to the specified address range. 3006 */ 3007 rv = vm_map_lookup_clip_start(map, start, &entry, &prev_entry); 3008 if (rv != KERN_SUCCESS) { 3009 vm_map_unlock(map); 3010 return (vm_mmap_to_errno(rv)); 3011 } 3012 3013 for (; entry->start < end; prev_entry = entry, 3014 entry = vm_map_entry_succ(entry)) { 3015 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 3016 continue; 3017 3018 rv = vm_map_clip_end(map, entry, end); 3019 if (rv != KERN_SUCCESS) { 3020 vm_map_unlock(map); 3021 return (vm_mmap_to_errno(rv)); 3022 } 3023 3024 switch (behav) { 3025 case MADV_NORMAL: 3026 vm_map_entry_set_behavior(entry, 3027 MAP_ENTRY_BEHAV_NORMAL); 3028 break; 3029 case MADV_SEQUENTIAL: 3030 vm_map_entry_set_behavior(entry, 3031 MAP_ENTRY_BEHAV_SEQUENTIAL); 3032 break; 3033 case MADV_RANDOM: 3034 vm_map_entry_set_behavior(entry, 3035 MAP_ENTRY_BEHAV_RANDOM); 3036 break; 3037 case MADV_NOSYNC: 3038 entry->eflags |= MAP_ENTRY_NOSYNC; 3039 break; 3040 case MADV_AUTOSYNC: 3041 entry->eflags &= ~MAP_ENTRY_NOSYNC; 3042 break; 3043 case MADV_NOCORE: 3044 entry->eflags |= MAP_ENTRY_NOCOREDUMP; 3045 break; 3046 case MADV_CORE: 3047 entry->eflags &= ~MAP_ENTRY_NOCOREDUMP; 3048 break; 3049 default: 3050 break; 3051 } 3052 vm_map_try_merge_entries(map, prev_entry, entry); 3053 } 3054 vm_map_try_merge_entries(map, prev_entry, entry); 3055 vm_map_unlock(map); 3056 } else { 3057 vm_pindex_t pstart, pend; 3058 3059 /* 3060 * madvise behaviors that are implemented in the underlying 3061 * vm_object. 3062 * 3063 * Since we don't clip the vm_map_entry, we have to clip 3064 * the vm_object pindex and count. 3065 */ 3066 if (!vm_map_lookup_entry(map, start, &entry)) 3067 entry = vm_map_entry_succ(entry); 3068 for (; entry->start < end; 3069 entry = vm_map_entry_succ(entry)) { 3070 vm_offset_t useEnd, useStart; 3071 3072 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 3073 continue; 3074 3075 /* 3076 * MADV_FREE would otherwise rewind time to 3077 * the creation of the shadow object. Because 3078 * we hold the VM map read-locked, neither the 3079 * entry's object nor the presence of a 3080 * backing object can change. 3081 */ 3082 if (behav == MADV_FREE && 3083 entry->object.vm_object != NULL && 3084 entry->object.vm_object->backing_object != NULL) 3085 continue; 3086 3087 pstart = OFF_TO_IDX(entry->offset); 3088 pend = pstart + atop(entry->end - entry->start); 3089 useStart = entry->start; 3090 useEnd = entry->end; 3091 3092 if (entry->start < start) { 3093 pstart += atop(start - entry->start); 3094 useStart = start; 3095 } 3096 if (entry->end > end) { 3097 pend -= atop(entry->end - end); 3098 useEnd = end; 3099 } 3100 3101 if (pstart >= pend) 3102 continue; 3103 3104 /* 3105 * Perform the pmap_advise() before clearing 3106 * PGA_REFERENCED in vm_page_advise(). Otherwise, a 3107 * concurrent pmap operation, such as pmap_remove(), 3108 * could clear a reference in the pmap and set 3109 * PGA_REFERENCED on the page before the pmap_advise() 3110 * had completed. Consequently, the page would appear 3111 * referenced based upon an old reference that 3112 * occurred before this pmap_advise() ran. 3113 */ 3114 if (behav == MADV_DONTNEED || behav == MADV_FREE) 3115 pmap_advise(map->pmap, useStart, useEnd, 3116 behav); 3117 3118 vm_object_madvise(entry->object.vm_object, pstart, 3119 pend, behav); 3120 3121 /* 3122 * Pre-populate paging structures in the 3123 * WILLNEED case. For wired entries, the 3124 * paging structures are already populated. 3125 */ 3126 if (behav == MADV_WILLNEED && 3127 entry->wired_count == 0) { 3128 vm_map_pmap_enter(map, 3129 useStart, 3130 entry->protection, 3131 entry->object.vm_object, 3132 pstart, 3133 ptoa(pend - pstart), 3134 MAP_PREFAULT_MADVISE 3135 ); 3136 } 3137 } 3138 vm_map_unlock_read(map); 3139 } 3140 return (0); 3141 } 3142 3143 /* 3144 * vm_map_inherit: 3145 * 3146 * Sets the inheritance of the specified address 3147 * range in the target map. Inheritance 3148 * affects how the map will be shared with 3149 * child maps at the time of vmspace_fork. 3150 */ 3151 int 3152 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 3153 vm_inherit_t new_inheritance) 3154 { 3155 vm_map_entry_t entry, lentry, prev_entry, start_entry; 3156 int rv; 3157 3158 switch (new_inheritance) { 3159 case VM_INHERIT_NONE: 3160 case VM_INHERIT_COPY: 3161 case VM_INHERIT_SHARE: 3162 case VM_INHERIT_ZERO: 3163 break; 3164 default: 3165 return (KERN_INVALID_ARGUMENT); 3166 } 3167 if (start == end) 3168 return (KERN_SUCCESS); 3169 vm_map_lock(map); 3170 VM_MAP_RANGE_CHECK(map, start, end); 3171 rv = vm_map_lookup_clip_start(map, start, &start_entry, &prev_entry); 3172 if (rv != KERN_SUCCESS) 3173 goto unlock; 3174 if (vm_map_lookup_entry(map, end - 1, &lentry)) { 3175 rv = vm_map_clip_end(map, lentry, end); 3176 if (rv != KERN_SUCCESS) 3177 goto unlock; 3178 } 3179 if (new_inheritance == VM_INHERIT_COPY) { 3180 for (entry = start_entry; entry->start < end; 3181 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3182 if ((entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) 3183 != 0) { 3184 rv = KERN_INVALID_ARGUMENT; 3185 goto unlock; 3186 } 3187 } 3188 } 3189 for (entry = start_entry; entry->start < end; prev_entry = entry, 3190 entry = vm_map_entry_succ(entry)) { 3191 KASSERT(entry->end <= end, ("non-clipped entry %p end %jx %jx", 3192 entry, (uintmax_t)entry->end, (uintmax_t)end)); 3193 if ((entry->eflags & MAP_ENTRY_GUARD) == 0 || 3194 new_inheritance != VM_INHERIT_ZERO) 3195 entry->inheritance = new_inheritance; 3196 vm_map_try_merge_entries(map, prev_entry, entry); 3197 } 3198 vm_map_try_merge_entries(map, prev_entry, entry); 3199 unlock: 3200 vm_map_unlock(map); 3201 return (rv); 3202 } 3203 3204 /* 3205 * vm_map_entry_in_transition: 3206 * 3207 * Release the map lock, and sleep until the entry is no longer in 3208 * transition. Awake and acquire the map lock. If the map changed while 3209 * another held the lock, lookup a possibly-changed entry at or after the 3210 * 'start' position of the old entry. 3211 */ 3212 static vm_map_entry_t 3213 vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start, 3214 vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry) 3215 { 3216 vm_map_entry_t entry; 3217 vm_offset_t start; 3218 u_int last_timestamp; 3219 3220 VM_MAP_ASSERT_LOCKED(map); 3221 KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3222 ("not in-tranition map entry %p", in_entry)); 3223 /* 3224 * We have not yet clipped the entry. 3225 */ 3226 start = MAX(in_start, in_entry->start); 3227 in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3228 last_timestamp = map->timestamp; 3229 if (vm_map_unlock_and_wait(map, 0)) { 3230 /* 3231 * Allow interruption of user wiring/unwiring? 3232 */ 3233 } 3234 vm_map_lock(map); 3235 if (last_timestamp + 1 == map->timestamp) 3236 return (in_entry); 3237 3238 /* 3239 * Look again for the entry because the map was modified while it was 3240 * unlocked. Specifically, the entry may have been clipped, merged, or 3241 * deleted. 3242 */ 3243 if (!vm_map_lookup_entry(map, start, &entry)) { 3244 if (!holes_ok) { 3245 *io_end = start; 3246 return (NULL); 3247 } 3248 entry = vm_map_entry_succ(entry); 3249 } 3250 return (entry); 3251 } 3252 3253 /* 3254 * vm_map_unwire: 3255 * 3256 * Implements both kernel and user unwiring. 3257 */ 3258 int 3259 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 3260 int flags) 3261 { 3262 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3263 int rv; 3264 bool holes_ok, need_wakeup, user_unwire; 3265 3266 if (start == end) 3267 return (KERN_SUCCESS); 3268 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3269 user_unwire = (flags & VM_MAP_WIRE_USER) != 0; 3270 vm_map_lock(map); 3271 VM_MAP_RANGE_CHECK(map, start, end); 3272 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3273 if (holes_ok) 3274 first_entry = vm_map_entry_succ(first_entry); 3275 else { 3276 vm_map_unlock(map); 3277 return (KERN_INVALID_ADDRESS); 3278 } 3279 } 3280 rv = KERN_SUCCESS; 3281 for (entry = first_entry; entry->start < end; entry = next_entry) { 3282 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3283 /* 3284 * We have not yet clipped the entry. 3285 */ 3286 next_entry = vm_map_entry_in_transition(map, start, 3287 &end, holes_ok, entry); 3288 if (next_entry == NULL) { 3289 if (entry == first_entry) { 3290 vm_map_unlock(map); 3291 return (KERN_INVALID_ADDRESS); 3292 } 3293 rv = KERN_INVALID_ADDRESS; 3294 break; 3295 } 3296 first_entry = (entry == first_entry) ? 3297 next_entry : NULL; 3298 continue; 3299 } 3300 rv = vm_map_clip_start(map, entry, start); 3301 if (rv != KERN_SUCCESS) 3302 break; 3303 rv = vm_map_clip_end(map, entry, end); 3304 if (rv != KERN_SUCCESS) 3305 break; 3306 3307 /* 3308 * Mark the entry in case the map lock is released. (See 3309 * above.) 3310 */ 3311 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3312 entry->wiring_thread == NULL, 3313 ("owned map entry %p", entry)); 3314 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3315 entry->wiring_thread = curthread; 3316 next_entry = vm_map_entry_succ(entry); 3317 /* 3318 * Check the map for holes in the specified region. 3319 * If holes_ok, skip this check. 3320 */ 3321 if (!holes_ok && 3322 entry->end < end && next_entry->start > entry->end) { 3323 end = entry->end; 3324 rv = KERN_INVALID_ADDRESS; 3325 break; 3326 } 3327 /* 3328 * If system unwiring, require that the entry is system wired. 3329 */ 3330 if (!user_unwire && 3331 vm_map_entry_system_wired_count(entry) == 0) { 3332 end = entry->end; 3333 rv = KERN_INVALID_ARGUMENT; 3334 break; 3335 } 3336 } 3337 need_wakeup = false; 3338 if (first_entry == NULL && 3339 !vm_map_lookup_entry(map, start, &first_entry)) { 3340 KASSERT(holes_ok, ("vm_map_unwire: lookup failed")); 3341 prev_entry = first_entry; 3342 entry = vm_map_entry_succ(first_entry); 3343 } else { 3344 prev_entry = vm_map_entry_pred(first_entry); 3345 entry = first_entry; 3346 } 3347 for (; entry->start < end; 3348 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3349 /* 3350 * If holes_ok was specified, an empty 3351 * space in the unwired region could have been mapped 3352 * while the map lock was dropped for draining 3353 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread 3354 * could be simultaneously wiring this new mapping 3355 * entry. Detect these cases and skip any entries 3356 * marked as in transition by us. 3357 */ 3358 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3359 entry->wiring_thread != curthread) { 3360 KASSERT(holes_ok, 3361 ("vm_map_unwire: !HOLESOK and new/changed entry")); 3362 continue; 3363 } 3364 3365 if (rv == KERN_SUCCESS && (!user_unwire || 3366 (entry->eflags & MAP_ENTRY_USER_WIRED))) { 3367 if (entry->wired_count == 1) 3368 vm_map_entry_unwire(map, entry); 3369 else 3370 entry->wired_count--; 3371 if (user_unwire) 3372 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3373 } 3374 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3375 ("vm_map_unwire: in-transition flag missing %p", entry)); 3376 KASSERT(entry->wiring_thread == curthread, 3377 ("vm_map_unwire: alien wire %p", entry)); 3378 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 3379 entry->wiring_thread = NULL; 3380 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3381 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3382 need_wakeup = true; 3383 } 3384 vm_map_try_merge_entries(map, prev_entry, entry); 3385 } 3386 vm_map_try_merge_entries(map, prev_entry, entry); 3387 vm_map_unlock(map); 3388 if (need_wakeup) 3389 vm_map_wakeup(map); 3390 return (rv); 3391 } 3392 3393 static void 3394 vm_map_wire_user_count_sub(u_long npages) 3395 { 3396 3397 atomic_subtract_long(&vm_user_wire_count, npages); 3398 } 3399 3400 static bool 3401 vm_map_wire_user_count_add(u_long npages) 3402 { 3403 u_long wired; 3404 3405 wired = vm_user_wire_count; 3406 do { 3407 if (npages + wired > vm_page_max_user_wired) 3408 return (false); 3409 } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired, 3410 npages + wired)); 3411 3412 return (true); 3413 } 3414 3415 /* 3416 * vm_map_wire_entry_failure: 3417 * 3418 * Handle a wiring failure on the given entry. 3419 * 3420 * The map should be locked. 3421 */ 3422 static void 3423 vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 3424 vm_offset_t failed_addr) 3425 { 3426 3427 VM_MAP_ASSERT_LOCKED(map); 3428 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 && 3429 entry->wired_count == 1, 3430 ("vm_map_wire_entry_failure: entry %p isn't being wired", entry)); 3431 KASSERT(failed_addr < entry->end, 3432 ("vm_map_wire_entry_failure: entry %p was fully wired", entry)); 3433 3434 /* 3435 * If any pages at the start of this entry were successfully wired, 3436 * then unwire them. 3437 */ 3438 if (failed_addr > entry->start) { 3439 pmap_unwire(map->pmap, entry->start, failed_addr); 3440 vm_object_unwire(entry->object.vm_object, entry->offset, 3441 failed_addr - entry->start, PQ_ACTIVE); 3442 } 3443 3444 /* 3445 * Assign an out-of-range value to represent the failure to wire this 3446 * entry. 3447 */ 3448 entry->wired_count = -1; 3449 } 3450 3451 int 3452 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3453 { 3454 int rv; 3455 3456 vm_map_lock(map); 3457 rv = vm_map_wire_locked(map, start, end, flags); 3458 vm_map_unlock(map); 3459 return (rv); 3460 } 3461 3462 /* 3463 * vm_map_wire_locked: 3464 * 3465 * Implements both kernel and user wiring. Returns with the map locked, 3466 * the map lock may be dropped. 3467 */ 3468 int 3469 vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3470 { 3471 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3472 vm_offset_t faddr, saved_end, saved_start; 3473 u_long incr, npages; 3474 u_int bidx, last_timestamp; 3475 int rv; 3476 bool holes_ok, need_wakeup, user_wire; 3477 vm_prot_t prot; 3478 3479 VM_MAP_ASSERT_LOCKED(map); 3480 3481 if (start == end) 3482 return (KERN_SUCCESS); 3483 prot = 0; 3484 if (flags & VM_MAP_WIRE_WRITE) 3485 prot |= VM_PROT_WRITE; 3486 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3487 user_wire = (flags & VM_MAP_WIRE_USER) != 0; 3488 VM_MAP_RANGE_CHECK(map, start, end); 3489 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3490 if (holes_ok) 3491 first_entry = vm_map_entry_succ(first_entry); 3492 else 3493 return (KERN_INVALID_ADDRESS); 3494 } 3495 for (entry = first_entry; entry->start < end; entry = next_entry) { 3496 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3497 /* 3498 * We have not yet clipped the entry. 3499 */ 3500 next_entry = vm_map_entry_in_transition(map, start, 3501 &end, holes_ok, entry); 3502 if (next_entry == NULL) { 3503 if (entry == first_entry) 3504 return (KERN_INVALID_ADDRESS); 3505 rv = KERN_INVALID_ADDRESS; 3506 goto done; 3507 } 3508 first_entry = (entry == first_entry) ? 3509 next_entry : NULL; 3510 continue; 3511 } 3512 rv = vm_map_clip_start(map, entry, start); 3513 if (rv != KERN_SUCCESS) 3514 goto done; 3515 rv = vm_map_clip_end(map, entry, end); 3516 if (rv != KERN_SUCCESS) 3517 goto done; 3518 3519 /* 3520 * Mark the entry in case the map lock is released. (See 3521 * above.) 3522 */ 3523 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3524 entry->wiring_thread == NULL, 3525 ("owned map entry %p", entry)); 3526 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3527 entry->wiring_thread = curthread; 3528 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 3529 || (entry->protection & prot) != prot) { 3530 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED; 3531 if (!holes_ok) { 3532 end = entry->end; 3533 rv = KERN_INVALID_ADDRESS; 3534 goto done; 3535 } 3536 } else if (entry->wired_count == 0) { 3537 entry->wired_count++; 3538 3539 npages = atop(entry->end - entry->start); 3540 if (user_wire && !vm_map_wire_user_count_add(npages)) { 3541 vm_map_wire_entry_failure(map, entry, 3542 entry->start); 3543 end = entry->end; 3544 rv = KERN_RESOURCE_SHORTAGE; 3545 goto done; 3546 } 3547 3548 /* 3549 * Release the map lock, relying on the in-transition 3550 * mark. Mark the map busy for fork. 3551 */ 3552 saved_start = entry->start; 3553 saved_end = entry->end; 3554 last_timestamp = map->timestamp; 3555 bidx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) 3556 >> MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 3557 incr = pagesizes[bidx]; 3558 vm_map_busy(map); 3559 vm_map_unlock(map); 3560 3561 for (faddr = saved_start; faddr < saved_end; 3562 faddr += incr) { 3563 /* 3564 * Simulate a fault to get the page and enter 3565 * it into the physical map. 3566 */ 3567 rv = vm_fault(map, faddr, VM_PROT_NONE, 3568 VM_FAULT_WIRE, NULL); 3569 if (rv != KERN_SUCCESS) 3570 break; 3571 } 3572 vm_map_lock(map); 3573 vm_map_unbusy(map); 3574 if (last_timestamp + 1 != map->timestamp) { 3575 /* 3576 * Look again for the entry because the map was 3577 * modified while it was unlocked. The entry 3578 * may have been clipped, but NOT merged or 3579 * deleted. 3580 */ 3581 if (!vm_map_lookup_entry(map, saved_start, 3582 &next_entry)) 3583 KASSERT(false, 3584 ("vm_map_wire: lookup failed")); 3585 first_entry = (entry == first_entry) ? 3586 next_entry : NULL; 3587 for (entry = next_entry; entry->end < saved_end; 3588 entry = vm_map_entry_succ(entry)) { 3589 /* 3590 * In case of failure, handle entries 3591 * that were not fully wired here; 3592 * fully wired entries are handled 3593 * later. 3594 */ 3595 if (rv != KERN_SUCCESS && 3596 faddr < entry->end) 3597 vm_map_wire_entry_failure(map, 3598 entry, faddr); 3599 } 3600 } 3601 if (rv != KERN_SUCCESS) { 3602 vm_map_wire_entry_failure(map, entry, faddr); 3603 if (user_wire) 3604 vm_map_wire_user_count_sub(npages); 3605 end = entry->end; 3606 goto done; 3607 } 3608 } else if (!user_wire || 3609 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3610 entry->wired_count++; 3611 } 3612 /* 3613 * Check the map for holes in the specified region. 3614 * If holes_ok was specified, skip this check. 3615 */ 3616 next_entry = vm_map_entry_succ(entry); 3617 if (!holes_ok && 3618 entry->end < end && next_entry->start > entry->end) { 3619 end = entry->end; 3620 rv = KERN_INVALID_ADDRESS; 3621 goto done; 3622 } 3623 } 3624 rv = KERN_SUCCESS; 3625 done: 3626 need_wakeup = false; 3627 if (first_entry == NULL && 3628 !vm_map_lookup_entry(map, start, &first_entry)) { 3629 KASSERT(holes_ok, ("vm_map_wire: lookup failed")); 3630 prev_entry = first_entry; 3631 entry = vm_map_entry_succ(first_entry); 3632 } else { 3633 prev_entry = vm_map_entry_pred(first_entry); 3634 entry = first_entry; 3635 } 3636 for (; entry->start < end; 3637 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3638 /* 3639 * If holes_ok was specified, an empty 3640 * space in the unwired region could have been mapped 3641 * while the map lock was dropped for faulting in the 3642 * pages or draining MAP_ENTRY_IN_TRANSITION. 3643 * Moreover, another thread could be simultaneously 3644 * wiring this new mapping entry. Detect these cases 3645 * and skip any entries marked as in transition not by us. 3646 * 3647 * Another way to get an entry not marked with 3648 * MAP_ENTRY_IN_TRANSITION is after failed clipping, 3649 * which set rv to KERN_INVALID_ARGUMENT. 3650 */ 3651 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3652 entry->wiring_thread != curthread) { 3653 KASSERT(holes_ok || rv == KERN_INVALID_ARGUMENT, 3654 ("vm_map_wire: !HOLESOK and new/changed entry")); 3655 continue; 3656 } 3657 3658 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) { 3659 /* do nothing */ 3660 } else if (rv == KERN_SUCCESS) { 3661 if (user_wire) 3662 entry->eflags |= MAP_ENTRY_USER_WIRED; 3663 } else if (entry->wired_count == -1) { 3664 /* 3665 * Wiring failed on this entry. Thus, unwiring is 3666 * unnecessary. 3667 */ 3668 entry->wired_count = 0; 3669 } else if (!user_wire || 3670 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3671 /* 3672 * Undo the wiring. Wiring succeeded on this entry 3673 * but failed on a later entry. 3674 */ 3675 if (entry->wired_count == 1) { 3676 vm_map_entry_unwire(map, entry); 3677 if (user_wire) 3678 vm_map_wire_user_count_sub( 3679 atop(entry->end - entry->start)); 3680 } else 3681 entry->wired_count--; 3682 } 3683 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3684 ("vm_map_wire: in-transition flag missing %p", entry)); 3685 KASSERT(entry->wiring_thread == curthread, 3686 ("vm_map_wire: alien wire %p", entry)); 3687 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION | 3688 MAP_ENTRY_WIRE_SKIPPED); 3689 entry->wiring_thread = NULL; 3690 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3691 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3692 need_wakeup = true; 3693 } 3694 vm_map_try_merge_entries(map, prev_entry, entry); 3695 } 3696 vm_map_try_merge_entries(map, prev_entry, entry); 3697 if (need_wakeup) 3698 vm_map_wakeup(map); 3699 return (rv); 3700 } 3701 3702 /* 3703 * vm_map_sync 3704 * 3705 * Push any dirty cached pages in the address range to their pager. 3706 * If syncio is TRUE, dirty pages are written synchronously. 3707 * If invalidate is TRUE, any cached pages are freed as well. 3708 * 3709 * If the size of the region from start to end is zero, we are 3710 * supposed to flush all modified pages within the region containing 3711 * start. Unfortunately, a region can be split or coalesced with 3712 * neighboring regions, making it difficult to determine what the 3713 * original region was. Therefore, we approximate this requirement by 3714 * flushing the current region containing start. 3715 * 3716 * Returns an error if any part of the specified range is not mapped. 3717 */ 3718 int 3719 vm_map_sync( 3720 vm_map_t map, 3721 vm_offset_t start, 3722 vm_offset_t end, 3723 boolean_t syncio, 3724 boolean_t invalidate) 3725 { 3726 vm_map_entry_t entry, first_entry, next_entry; 3727 vm_size_t size; 3728 vm_object_t object; 3729 vm_ooffset_t offset; 3730 unsigned int last_timestamp; 3731 int bdry_idx; 3732 boolean_t failed; 3733 3734 vm_map_lock_read(map); 3735 VM_MAP_RANGE_CHECK(map, start, end); 3736 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3737 vm_map_unlock_read(map); 3738 return (KERN_INVALID_ADDRESS); 3739 } else if (start == end) { 3740 start = first_entry->start; 3741 end = first_entry->end; 3742 } 3743 3744 /* 3745 * Make a first pass to check for user-wired memory, holes, 3746 * and partial invalidation of largepage mappings. 3747 */ 3748 for (entry = first_entry; entry->start < end; entry = next_entry) { 3749 if (invalidate) { 3750 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) { 3751 vm_map_unlock_read(map); 3752 return (KERN_INVALID_ARGUMENT); 3753 } 3754 bdry_idx = (entry->eflags & 3755 MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 3756 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 3757 if (bdry_idx != 0 && 3758 ((start & (pagesizes[bdry_idx] - 1)) != 0 || 3759 (end & (pagesizes[bdry_idx] - 1)) != 0)) { 3760 vm_map_unlock_read(map); 3761 return (KERN_INVALID_ARGUMENT); 3762 } 3763 } 3764 next_entry = vm_map_entry_succ(entry); 3765 if (end > entry->end && 3766 entry->end != next_entry->start) { 3767 vm_map_unlock_read(map); 3768 return (KERN_INVALID_ADDRESS); 3769 } 3770 } 3771 3772 if (invalidate) 3773 pmap_remove(map->pmap, start, end); 3774 failed = FALSE; 3775 3776 /* 3777 * Make a second pass, cleaning/uncaching pages from the indicated 3778 * objects as we go. 3779 */ 3780 for (entry = first_entry; entry->start < end;) { 3781 offset = entry->offset + (start - entry->start); 3782 size = (end <= entry->end ? end : entry->end) - start; 3783 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 3784 vm_map_t smap; 3785 vm_map_entry_t tentry; 3786 vm_size_t tsize; 3787 3788 smap = entry->object.sub_map; 3789 vm_map_lock_read(smap); 3790 (void) vm_map_lookup_entry(smap, offset, &tentry); 3791 tsize = tentry->end - offset; 3792 if (tsize < size) 3793 size = tsize; 3794 object = tentry->object.vm_object; 3795 offset = tentry->offset + (offset - tentry->start); 3796 vm_map_unlock_read(smap); 3797 } else { 3798 object = entry->object.vm_object; 3799 } 3800 vm_object_reference(object); 3801 last_timestamp = map->timestamp; 3802 vm_map_unlock_read(map); 3803 if (!vm_object_sync(object, offset, size, syncio, invalidate)) 3804 failed = TRUE; 3805 start += size; 3806 vm_object_deallocate(object); 3807 vm_map_lock_read(map); 3808 if (last_timestamp == map->timestamp || 3809 !vm_map_lookup_entry(map, start, &entry)) 3810 entry = vm_map_entry_succ(entry); 3811 } 3812 3813 vm_map_unlock_read(map); 3814 return (failed ? KERN_FAILURE : KERN_SUCCESS); 3815 } 3816 3817 /* 3818 * vm_map_entry_unwire: [ internal use only ] 3819 * 3820 * Make the region specified by this entry pageable. 3821 * 3822 * The map in question should be locked. 3823 * [This is the reason for this routine's existence.] 3824 */ 3825 static void 3826 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 3827 { 3828 vm_size_t size; 3829 3830 VM_MAP_ASSERT_LOCKED(map); 3831 KASSERT(entry->wired_count > 0, 3832 ("vm_map_entry_unwire: entry %p isn't wired", entry)); 3833 3834 size = entry->end - entry->start; 3835 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) 3836 vm_map_wire_user_count_sub(atop(size)); 3837 pmap_unwire(map->pmap, entry->start, entry->end); 3838 vm_object_unwire(entry->object.vm_object, entry->offset, size, 3839 PQ_ACTIVE); 3840 entry->wired_count = 0; 3841 } 3842 3843 static void 3844 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map) 3845 { 3846 3847 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) 3848 vm_object_deallocate(entry->object.vm_object); 3849 uma_zfree(system_map ? kmapentzone : mapentzone, entry); 3850 } 3851 3852 /* 3853 * vm_map_entry_delete: [ internal use only ] 3854 * 3855 * Deallocate the given entry from the target map. 3856 */ 3857 static void 3858 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry) 3859 { 3860 vm_object_t object; 3861 vm_pindex_t offidxstart, offidxend, size1; 3862 vm_size_t size; 3863 3864 vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE); 3865 object = entry->object.vm_object; 3866 3867 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { 3868 MPASS(entry->cred == NULL); 3869 MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0); 3870 MPASS(object == NULL); 3871 vm_map_entry_deallocate(entry, map->system_map); 3872 return; 3873 } 3874 3875 size = entry->end - entry->start; 3876 map->size -= size; 3877 3878 if (entry->cred != NULL) { 3879 swap_release_by_cred(size, entry->cred); 3880 crfree(entry->cred); 3881 } 3882 3883 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || object == NULL) { 3884 entry->object.vm_object = NULL; 3885 } else if ((object->flags & OBJ_ANON) != 0 || 3886 object == kernel_object) { 3887 KASSERT(entry->cred == NULL || object->cred == NULL || 3888 (entry->eflags & MAP_ENTRY_NEEDS_COPY), 3889 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry)); 3890 offidxstart = OFF_TO_IDX(entry->offset); 3891 offidxend = offidxstart + atop(size); 3892 VM_OBJECT_WLOCK(object); 3893 if (object->ref_count != 1 && 3894 ((object->flags & OBJ_ONEMAPPING) != 0 || 3895 object == kernel_object)) { 3896 vm_object_collapse(object); 3897 3898 /* 3899 * The option OBJPR_NOTMAPPED can be passed here 3900 * because vm_map_delete() already performed 3901 * pmap_remove() on the only mapping to this range 3902 * of pages. 3903 */ 3904 vm_object_page_remove(object, offidxstart, offidxend, 3905 OBJPR_NOTMAPPED); 3906 if (offidxend >= object->size && 3907 offidxstart < object->size) { 3908 size1 = object->size; 3909 object->size = offidxstart; 3910 if (object->cred != NULL) { 3911 size1 -= object->size; 3912 KASSERT(object->charge >= ptoa(size1), 3913 ("object %p charge < 0", object)); 3914 swap_release_by_cred(ptoa(size1), 3915 object->cred); 3916 object->charge -= ptoa(size1); 3917 } 3918 } 3919 } 3920 VM_OBJECT_WUNLOCK(object); 3921 } 3922 if (map->system_map) 3923 vm_map_entry_deallocate(entry, TRUE); 3924 else { 3925 entry->defer_next = curthread->td_map_def_user; 3926 curthread->td_map_def_user = entry; 3927 } 3928 } 3929 3930 /* 3931 * vm_map_delete: [ internal use only ] 3932 * 3933 * Deallocates the given address range from the target 3934 * map. 3935 */ 3936 int 3937 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end) 3938 { 3939 vm_map_entry_t entry, next_entry, scratch_entry; 3940 int rv; 3941 3942 VM_MAP_ASSERT_LOCKED(map); 3943 3944 if (start == end) 3945 return (KERN_SUCCESS); 3946 3947 /* 3948 * Find the start of the region, and clip it. 3949 * Step through all entries in this region. 3950 */ 3951 rv = vm_map_lookup_clip_start(map, start, &entry, &scratch_entry); 3952 if (rv != KERN_SUCCESS) 3953 return (rv); 3954 for (; entry->start < end; entry = next_entry) { 3955 /* 3956 * Wait for wiring or unwiring of an entry to complete. 3957 * Also wait for any system wirings to disappear on 3958 * user maps. 3959 */ 3960 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 || 3961 (vm_map_pmap(map) != kernel_pmap && 3962 vm_map_entry_system_wired_count(entry) != 0)) { 3963 unsigned int last_timestamp; 3964 vm_offset_t saved_start; 3965 3966 saved_start = entry->start; 3967 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3968 last_timestamp = map->timestamp; 3969 (void) vm_map_unlock_and_wait(map, 0); 3970 vm_map_lock(map); 3971 if (last_timestamp + 1 != map->timestamp) { 3972 /* 3973 * Look again for the entry because the map was 3974 * modified while it was unlocked. 3975 * Specifically, the entry may have been 3976 * clipped, merged, or deleted. 3977 */ 3978 rv = vm_map_lookup_clip_start(map, saved_start, 3979 &next_entry, &scratch_entry); 3980 if (rv != KERN_SUCCESS) 3981 break; 3982 } else 3983 next_entry = entry; 3984 continue; 3985 } 3986 3987 /* XXXKIB or delete to the upper superpage boundary ? */ 3988 rv = vm_map_clip_end(map, entry, end); 3989 if (rv != KERN_SUCCESS) 3990 break; 3991 next_entry = vm_map_entry_succ(entry); 3992 3993 /* 3994 * Unwire before removing addresses from the pmap; otherwise, 3995 * unwiring will put the entries back in the pmap. 3996 */ 3997 if (entry->wired_count != 0) 3998 vm_map_entry_unwire(map, entry); 3999 4000 /* 4001 * Remove mappings for the pages, but only if the 4002 * mappings could exist. For instance, it does not 4003 * make sense to call pmap_remove() for guard entries. 4004 */ 4005 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 4006 entry->object.vm_object != NULL) 4007 pmap_remove(map->pmap, entry->start, entry->end); 4008 4009 if (entry->end == map->anon_loc) 4010 map->anon_loc = entry->start; 4011 4012 /* 4013 * Delete the entry only after removing all pmap 4014 * entries pointing to its pages. (Otherwise, its 4015 * page frames may be reallocated, and any modify bits 4016 * will be set in the wrong object!) 4017 */ 4018 vm_map_entry_delete(map, entry); 4019 } 4020 return (rv); 4021 } 4022 4023 /* 4024 * vm_map_remove: 4025 * 4026 * Remove the given address range from the target map. 4027 * This is the exported form of vm_map_delete. 4028 */ 4029 int 4030 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 4031 { 4032 int result; 4033 4034 vm_map_lock(map); 4035 VM_MAP_RANGE_CHECK(map, start, end); 4036 result = vm_map_delete(map, start, end); 4037 vm_map_unlock(map); 4038 return (result); 4039 } 4040 4041 /* 4042 * vm_map_check_protection: 4043 * 4044 * Assert that the target map allows the specified privilege on the 4045 * entire address region given. The entire region must be allocated. 4046 * 4047 * WARNING! This code does not and should not check whether the 4048 * contents of the region is accessible. For example a smaller file 4049 * might be mapped into a larger address space. 4050 * 4051 * NOTE! This code is also called by munmap(). 4052 * 4053 * The map must be locked. A read lock is sufficient. 4054 */ 4055 boolean_t 4056 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 4057 vm_prot_t protection) 4058 { 4059 vm_map_entry_t entry; 4060 vm_map_entry_t tmp_entry; 4061 4062 if (!vm_map_lookup_entry(map, start, &tmp_entry)) 4063 return (FALSE); 4064 entry = tmp_entry; 4065 4066 while (start < end) { 4067 /* 4068 * No holes allowed! 4069 */ 4070 if (start < entry->start) 4071 return (FALSE); 4072 /* 4073 * Check protection associated with entry. 4074 */ 4075 if ((entry->protection & protection) != protection) 4076 return (FALSE); 4077 /* go to next entry */ 4078 start = entry->end; 4079 entry = vm_map_entry_succ(entry); 4080 } 4081 return (TRUE); 4082 } 4083 4084 /* 4085 * 4086 * vm_map_copy_swap_object: 4087 * 4088 * Copies a swap-backed object from an existing map entry to a 4089 * new one. Carries forward the swap charge. May change the 4090 * src object on return. 4091 */ 4092 static void 4093 vm_map_copy_swap_object(vm_map_entry_t src_entry, vm_map_entry_t dst_entry, 4094 vm_offset_t size, vm_ooffset_t *fork_charge) 4095 { 4096 vm_object_t src_object; 4097 struct ucred *cred; 4098 int charged; 4099 4100 src_object = src_entry->object.vm_object; 4101 charged = ENTRY_CHARGED(src_entry); 4102 if ((src_object->flags & OBJ_ANON) != 0) { 4103 VM_OBJECT_WLOCK(src_object); 4104 vm_object_collapse(src_object); 4105 if ((src_object->flags & OBJ_ONEMAPPING) != 0) { 4106 vm_object_split(src_entry); 4107 src_object = src_entry->object.vm_object; 4108 } 4109 vm_object_reference_locked(src_object); 4110 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 4111 VM_OBJECT_WUNLOCK(src_object); 4112 } else 4113 vm_object_reference(src_object); 4114 if (src_entry->cred != NULL && 4115 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 4116 KASSERT(src_object->cred == NULL, 4117 ("OVERCOMMIT: vm_map_copy_anon_entry: cred %p", 4118 src_object)); 4119 src_object->cred = src_entry->cred; 4120 src_object->charge = size; 4121 } 4122 dst_entry->object.vm_object = src_object; 4123 if (charged) { 4124 cred = curthread->td_ucred; 4125 crhold(cred); 4126 dst_entry->cred = cred; 4127 *fork_charge += size; 4128 if (!(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 4129 crhold(cred); 4130 src_entry->cred = cred; 4131 *fork_charge += size; 4132 } 4133 } 4134 } 4135 4136 /* 4137 * vm_map_copy_entry: 4138 * 4139 * Copies the contents of the source entry to the destination 4140 * entry. The entries *must* be aligned properly. 4141 */ 4142 static void 4143 vm_map_copy_entry( 4144 vm_map_t src_map, 4145 vm_map_t dst_map, 4146 vm_map_entry_t src_entry, 4147 vm_map_entry_t dst_entry, 4148 vm_ooffset_t *fork_charge) 4149 { 4150 vm_object_t src_object; 4151 vm_map_entry_t fake_entry; 4152 vm_offset_t size; 4153 4154 VM_MAP_ASSERT_LOCKED(dst_map); 4155 4156 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) 4157 return; 4158 4159 if (src_entry->wired_count == 0 || 4160 (src_entry->protection & VM_PROT_WRITE) == 0) { 4161 /* 4162 * If the source entry is marked needs_copy, it is already 4163 * write-protected. 4164 */ 4165 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && 4166 (src_entry->protection & VM_PROT_WRITE) != 0) { 4167 pmap_protect(src_map->pmap, 4168 src_entry->start, 4169 src_entry->end, 4170 src_entry->protection & ~VM_PROT_WRITE); 4171 } 4172 4173 /* 4174 * Make a copy of the object. 4175 */ 4176 size = src_entry->end - src_entry->start; 4177 if ((src_object = src_entry->object.vm_object) != NULL) { 4178 if (src_object->type == OBJT_DEFAULT || 4179 src_object->type == OBJT_SWAP) { 4180 vm_map_copy_swap_object(src_entry, dst_entry, 4181 size, fork_charge); 4182 /* May have split/collapsed, reload obj. */ 4183 src_object = src_entry->object.vm_object; 4184 } else { 4185 vm_object_reference(src_object); 4186 dst_entry->object.vm_object = src_object; 4187 } 4188 src_entry->eflags |= MAP_ENTRY_COW | 4189 MAP_ENTRY_NEEDS_COPY; 4190 dst_entry->eflags |= MAP_ENTRY_COW | 4191 MAP_ENTRY_NEEDS_COPY; 4192 dst_entry->offset = src_entry->offset; 4193 if (src_entry->eflags & MAP_ENTRY_WRITECNT) { 4194 /* 4195 * MAP_ENTRY_WRITECNT cannot 4196 * indicate write reference from 4197 * src_entry, since the entry is 4198 * marked as needs copy. Allocate a 4199 * fake entry that is used to 4200 * decrement object->un_pager writecount 4201 * at the appropriate time. Attach 4202 * fake_entry to the deferred list. 4203 */ 4204 fake_entry = vm_map_entry_create(dst_map); 4205 fake_entry->eflags = MAP_ENTRY_WRITECNT; 4206 src_entry->eflags &= ~MAP_ENTRY_WRITECNT; 4207 vm_object_reference(src_object); 4208 fake_entry->object.vm_object = src_object; 4209 fake_entry->start = src_entry->start; 4210 fake_entry->end = src_entry->end; 4211 fake_entry->defer_next = 4212 curthread->td_map_def_user; 4213 curthread->td_map_def_user = fake_entry; 4214 } 4215 4216 pmap_copy(dst_map->pmap, src_map->pmap, 4217 dst_entry->start, dst_entry->end - dst_entry->start, 4218 src_entry->start); 4219 } else { 4220 dst_entry->object.vm_object = NULL; 4221 dst_entry->offset = 0; 4222 if (src_entry->cred != NULL) { 4223 dst_entry->cred = curthread->td_ucred; 4224 crhold(dst_entry->cred); 4225 *fork_charge += size; 4226 } 4227 } 4228 } else { 4229 /* 4230 * We don't want to make writeable wired pages copy-on-write. 4231 * Immediately copy these pages into the new map by simulating 4232 * page faults. The new pages are pageable. 4233 */ 4234 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry, 4235 fork_charge); 4236 } 4237 } 4238 4239 /* 4240 * vmspace_map_entry_forked: 4241 * Update the newly-forked vmspace each time a map entry is inherited 4242 * or copied. The values for vm_dsize and vm_tsize are approximate 4243 * (and mostly-obsolete ideas in the face of mmap(2) et al.) 4244 */ 4245 static void 4246 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2, 4247 vm_map_entry_t entry) 4248 { 4249 vm_size_t entrysize; 4250 vm_offset_t newend; 4251 4252 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 4253 return; 4254 entrysize = entry->end - entry->start; 4255 vm2->vm_map.size += entrysize; 4256 if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) { 4257 vm2->vm_ssize += btoc(entrysize); 4258 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr && 4259 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) { 4260 newend = MIN(entry->end, 4261 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)); 4262 vm2->vm_dsize += btoc(newend - entry->start); 4263 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr && 4264 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) { 4265 newend = MIN(entry->end, 4266 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)); 4267 vm2->vm_tsize += btoc(newend - entry->start); 4268 } 4269 } 4270 4271 /* 4272 * vmspace_fork: 4273 * Create a new process vmspace structure and vm_map 4274 * based on those of an existing process. The new map 4275 * is based on the old map, according to the inheritance 4276 * values on the regions in that map. 4277 * 4278 * XXX It might be worth coalescing the entries added to the new vmspace. 4279 * 4280 * The source map must not be locked. 4281 */ 4282 struct vmspace * 4283 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge) 4284 { 4285 struct vmspace *vm2; 4286 vm_map_t new_map, old_map; 4287 vm_map_entry_t new_entry, old_entry; 4288 vm_object_t object; 4289 int error, locked; 4290 vm_inherit_t inh; 4291 4292 old_map = &vm1->vm_map; 4293 /* Copy immutable fields of vm1 to vm2. */ 4294 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map), 4295 pmap_pinit); 4296 if (vm2 == NULL) 4297 return (NULL); 4298 4299 vm2->vm_taddr = vm1->vm_taddr; 4300 vm2->vm_daddr = vm1->vm_daddr; 4301 vm2->vm_maxsaddr = vm1->vm_maxsaddr; 4302 vm_map_lock(old_map); 4303 if (old_map->busy) 4304 vm_map_wait_busy(old_map); 4305 new_map = &vm2->vm_map; 4306 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */ 4307 KASSERT(locked, ("vmspace_fork: lock failed")); 4308 4309 error = pmap_vmspace_copy(new_map->pmap, old_map->pmap); 4310 if (error != 0) { 4311 sx_xunlock(&old_map->lock); 4312 sx_xunlock(&new_map->lock); 4313 vm_map_process_deferred(); 4314 vmspace_free(vm2); 4315 return (NULL); 4316 } 4317 4318 new_map->anon_loc = old_map->anon_loc; 4319 new_map->flags |= old_map->flags & (MAP_ASLR | MAP_ASLR_IGNSTART | 4320 MAP_WXORX); 4321 4322 VM_MAP_ENTRY_FOREACH(old_entry, old_map) { 4323 if ((old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 4324 panic("vm_map_fork: encountered a submap"); 4325 4326 inh = old_entry->inheritance; 4327 if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 && 4328 inh != VM_INHERIT_NONE) 4329 inh = VM_INHERIT_COPY; 4330 4331 switch (inh) { 4332 case VM_INHERIT_NONE: 4333 break; 4334 4335 case VM_INHERIT_SHARE: 4336 /* 4337 * Clone the entry, creating the shared object if 4338 * necessary. 4339 */ 4340 object = old_entry->object.vm_object; 4341 if (object == NULL) { 4342 vm_map_entry_back(old_entry); 4343 object = old_entry->object.vm_object; 4344 } 4345 4346 /* 4347 * Add the reference before calling vm_object_shadow 4348 * to insure that a shadow object is created. 4349 */ 4350 vm_object_reference(object); 4351 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4352 vm_object_shadow(&old_entry->object.vm_object, 4353 &old_entry->offset, 4354 old_entry->end - old_entry->start, 4355 old_entry->cred, 4356 /* Transfer the second reference too. */ 4357 true); 4358 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 4359 old_entry->cred = NULL; 4360 4361 /* 4362 * As in vm_map_merged_neighbor_dispose(), 4363 * the vnode lock will not be acquired in 4364 * this call to vm_object_deallocate(). 4365 */ 4366 vm_object_deallocate(object); 4367 object = old_entry->object.vm_object; 4368 } else { 4369 VM_OBJECT_WLOCK(object); 4370 vm_object_clear_flag(object, OBJ_ONEMAPPING); 4371 if (old_entry->cred != NULL) { 4372 KASSERT(object->cred == NULL, 4373 ("vmspace_fork both cred")); 4374 object->cred = old_entry->cred; 4375 object->charge = old_entry->end - 4376 old_entry->start; 4377 old_entry->cred = NULL; 4378 } 4379 4380 /* 4381 * Assert the correct state of the vnode 4382 * v_writecount while the object is locked, to 4383 * not relock it later for the assertion 4384 * correctness. 4385 */ 4386 if (old_entry->eflags & MAP_ENTRY_WRITECNT && 4387 object->type == OBJT_VNODE) { 4388 KASSERT(((struct vnode *)object-> 4389 handle)->v_writecount > 0, 4390 ("vmspace_fork: v_writecount %p", 4391 object)); 4392 KASSERT(object->un_pager.vnp. 4393 writemappings > 0, 4394 ("vmspace_fork: vnp.writecount %p", 4395 object)); 4396 } 4397 VM_OBJECT_WUNLOCK(object); 4398 } 4399 4400 /* 4401 * Clone the entry, referencing the shared object. 4402 */ 4403 new_entry = vm_map_entry_create(new_map); 4404 *new_entry = *old_entry; 4405 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4406 MAP_ENTRY_IN_TRANSITION); 4407 new_entry->wiring_thread = NULL; 4408 new_entry->wired_count = 0; 4409 if (new_entry->eflags & MAP_ENTRY_WRITECNT) { 4410 vm_pager_update_writecount(object, 4411 new_entry->start, new_entry->end); 4412 } 4413 vm_map_entry_set_vnode_text(new_entry, true); 4414 4415 /* 4416 * Insert the entry into the new map -- we know we're 4417 * inserting at the end of the new map. 4418 */ 4419 vm_map_entry_link(new_map, new_entry); 4420 vmspace_map_entry_forked(vm1, vm2, new_entry); 4421 4422 /* 4423 * Update the physical map 4424 */ 4425 pmap_copy(new_map->pmap, old_map->pmap, 4426 new_entry->start, 4427 (old_entry->end - old_entry->start), 4428 old_entry->start); 4429 break; 4430 4431 case VM_INHERIT_COPY: 4432 /* 4433 * Clone the entry and link into the map. 4434 */ 4435 new_entry = vm_map_entry_create(new_map); 4436 *new_entry = *old_entry; 4437 /* 4438 * Copied entry is COW over the old object. 4439 */ 4440 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4441 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT); 4442 new_entry->wiring_thread = NULL; 4443 new_entry->wired_count = 0; 4444 new_entry->object.vm_object = NULL; 4445 new_entry->cred = NULL; 4446 vm_map_entry_link(new_map, new_entry); 4447 vmspace_map_entry_forked(vm1, vm2, new_entry); 4448 vm_map_copy_entry(old_map, new_map, old_entry, 4449 new_entry, fork_charge); 4450 vm_map_entry_set_vnode_text(new_entry, true); 4451 break; 4452 4453 case VM_INHERIT_ZERO: 4454 /* 4455 * Create a new anonymous mapping entry modelled from 4456 * the old one. 4457 */ 4458 new_entry = vm_map_entry_create(new_map); 4459 memset(new_entry, 0, sizeof(*new_entry)); 4460 4461 new_entry->start = old_entry->start; 4462 new_entry->end = old_entry->end; 4463 new_entry->eflags = old_entry->eflags & 4464 ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION | 4465 MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC | 4466 MAP_ENTRY_SPLIT_BOUNDARY_MASK); 4467 new_entry->protection = old_entry->protection; 4468 new_entry->max_protection = old_entry->max_protection; 4469 new_entry->inheritance = VM_INHERIT_ZERO; 4470 4471 vm_map_entry_link(new_map, new_entry); 4472 vmspace_map_entry_forked(vm1, vm2, new_entry); 4473 4474 new_entry->cred = curthread->td_ucred; 4475 crhold(new_entry->cred); 4476 *fork_charge += (new_entry->end - new_entry->start); 4477 4478 break; 4479 } 4480 } 4481 /* 4482 * Use inlined vm_map_unlock() to postpone handling the deferred 4483 * map entries, which cannot be done until both old_map and 4484 * new_map locks are released. 4485 */ 4486 sx_xunlock(&old_map->lock); 4487 sx_xunlock(&new_map->lock); 4488 vm_map_process_deferred(); 4489 4490 return (vm2); 4491 } 4492 4493 /* 4494 * Create a process's stack for exec_new_vmspace(). This function is never 4495 * asked to wire the newly created stack. 4496 */ 4497 int 4498 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4499 vm_prot_t prot, vm_prot_t max, int cow) 4500 { 4501 vm_size_t growsize, init_ssize; 4502 rlim_t vmemlim; 4503 int rv; 4504 4505 MPASS((map->flags & MAP_WIREFUTURE) == 0); 4506 growsize = sgrowsiz; 4507 init_ssize = (max_ssize < growsize) ? max_ssize : growsize; 4508 vm_map_lock(map); 4509 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4510 /* If we would blow our VMEM resource limit, no go */ 4511 if (map->size + init_ssize > vmemlim) { 4512 rv = KERN_NO_SPACE; 4513 goto out; 4514 } 4515 rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot, 4516 max, cow); 4517 out: 4518 vm_map_unlock(map); 4519 return (rv); 4520 } 4521 4522 static int stack_guard_page = 1; 4523 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN, 4524 &stack_guard_page, 0, 4525 "Specifies the number of guard pages for a stack that grows"); 4526 4527 static int 4528 vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4529 vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow) 4530 { 4531 vm_map_entry_t new_entry, prev_entry; 4532 vm_offset_t bot, gap_bot, gap_top, top; 4533 vm_size_t init_ssize, sgp; 4534 int orient, rv; 4535 4536 /* 4537 * The stack orientation is piggybacked with the cow argument. 4538 * Extract it into orient and mask the cow argument so that we 4539 * don't pass it around further. 4540 */ 4541 orient = cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP); 4542 KASSERT(orient != 0, ("No stack grow direction")); 4543 KASSERT(orient != (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP), 4544 ("bi-dir stack")); 4545 4546 if (max_ssize == 0 || 4547 !vm_map_range_valid(map, addrbos, addrbos + max_ssize)) 4548 return (KERN_INVALID_ADDRESS); 4549 sgp = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4550 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4551 (vm_size_t)stack_guard_page * PAGE_SIZE; 4552 if (sgp >= max_ssize) 4553 return (KERN_INVALID_ARGUMENT); 4554 4555 init_ssize = growsize; 4556 if (max_ssize < init_ssize + sgp) 4557 init_ssize = max_ssize - sgp; 4558 4559 /* If addr is already mapped, no go */ 4560 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) 4561 return (KERN_NO_SPACE); 4562 4563 /* 4564 * If we can't accommodate max_ssize in the current mapping, no go. 4565 */ 4566 if (vm_map_entry_succ(prev_entry)->start < addrbos + max_ssize) 4567 return (KERN_NO_SPACE); 4568 4569 /* 4570 * We initially map a stack of only init_ssize. We will grow as 4571 * needed later. Depending on the orientation of the stack (i.e. 4572 * the grow direction) we either map at the top of the range, the 4573 * bottom of the range or in the middle. 4574 * 4575 * Note: we would normally expect prot and max to be VM_PROT_ALL, 4576 * and cow to be 0. Possibly we should eliminate these as input 4577 * parameters, and just pass these values here in the insert call. 4578 */ 4579 if (orient == MAP_STACK_GROWS_DOWN) { 4580 bot = addrbos + max_ssize - init_ssize; 4581 top = bot + init_ssize; 4582 gap_bot = addrbos; 4583 gap_top = bot; 4584 } else /* if (orient == MAP_STACK_GROWS_UP) */ { 4585 bot = addrbos; 4586 top = bot + init_ssize; 4587 gap_bot = top; 4588 gap_top = addrbos + max_ssize; 4589 } 4590 rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow); 4591 if (rv != KERN_SUCCESS) 4592 return (rv); 4593 new_entry = vm_map_entry_succ(prev_entry); 4594 KASSERT(new_entry->end == top || new_entry->start == bot, 4595 ("Bad entry start/end for new stack entry")); 4596 KASSERT((orient & MAP_STACK_GROWS_DOWN) == 0 || 4597 (new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0, 4598 ("new entry lacks MAP_ENTRY_GROWS_DOWN")); 4599 KASSERT((orient & MAP_STACK_GROWS_UP) == 0 || 4600 (new_entry->eflags & MAP_ENTRY_GROWS_UP) != 0, 4601 ("new entry lacks MAP_ENTRY_GROWS_UP")); 4602 if (gap_bot == gap_top) 4603 return (KERN_SUCCESS); 4604 rv = vm_map_insert(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE, 4605 VM_PROT_NONE, MAP_CREATE_GUARD | (orient == MAP_STACK_GROWS_DOWN ? 4606 MAP_CREATE_STACK_GAP_DN : MAP_CREATE_STACK_GAP_UP)); 4607 if (rv == KERN_SUCCESS) { 4608 /* 4609 * Gap can never successfully handle a fault, so 4610 * read-ahead logic is never used for it. Re-use 4611 * next_read of the gap entry to store 4612 * stack_guard_page for vm_map_growstack(). 4613 */ 4614 if (orient == MAP_STACK_GROWS_DOWN) 4615 vm_map_entry_pred(new_entry)->next_read = sgp; 4616 else 4617 vm_map_entry_succ(new_entry)->next_read = sgp; 4618 } else { 4619 (void)vm_map_delete(map, bot, top); 4620 } 4621 return (rv); 4622 } 4623 4624 /* 4625 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we 4626 * successfully grow the stack. 4627 */ 4628 static int 4629 vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry) 4630 { 4631 vm_map_entry_t stack_entry; 4632 struct proc *p; 4633 struct vmspace *vm; 4634 struct ucred *cred; 4635 vm_offset_t gap_end, gap_start, grow_start; 4636 vm_size_t grow_amount, guard, max_grow; 4637 rlim_t lmemlim, stacklim, vmemlim; 4638 int rv, rv1; 4639 bool gap_deleted, grow_down, is_procstack; 4640 #ifdef notyet 4641 uint64_t limit; 4642 #endif 4643 #ifdef RACCT 4644 int error; 4645 #endif 4646 4647 p = curproc; 4648 vm = p->p_vmspace; 4649 4650 /* 4651 * Disallow stack growth when the access is performed by a 4652 * debugger or AIO daemon. The reason is that the wrong 4653 * resource limits are applied. 4654 */ 4655 if (p != initproc && (map != &p->p_vmspace->vm_map || 4656 p->p_textvp == NULL)) 4657 return (KERN_FAILURE); 4658 4659 MPASS(!map->system_map); 4660 4661 lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK); 4662 stacklim = lim_cur(curthread, RLIMIT_STACK); 4663 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4664 retry: 4665 /* If addr is not in a hole for a stack grow area, no need to grow. */ 4666 if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry)) 4667 return (KERN_FAILURE); 4668 if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0) 4669 return (KERN_SUCCESS); 4670 if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_DN) != 0) { 4671 stack_entry = vm_map_entry_succ(gap_entry); 4672 if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 || 4673 stack_entry->start != gap_entry->end) 4674 return (KERN_FAILURE); 4675 grow_amount = round_page(stack_entry->start - addr); 4676 grow_down = true; 4677 } else if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_UP) != 0) { 4678 stack_entry = vm_map_entry_pred(gap_entry); 4679 if ((stack_entry->eflags & MAP_ENTRY_GROWS_UP) == 0 || 4680 stack_entry->end != gap_entry->start) 4681 return (KERN_FAILURE); 4682 grow_amount = round_page(addr + 1 - stack_entry->end); 4683 grow_down = false; 4684 } else { 4685 return (KERN_FAILURE); 4686 } 4687 guard = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4688 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4689 gap_entry->next_read; 4690 max_grow = gap_entry->end - gap_entry->start; 4691 if (guard > max_grow) 4692 return (KERN_NO_SPACE); 4693 max_grow -= guard; 4694 if (grow_amount > max_grow) 4695 return (KERN_NO_SPACE); 4696 4697 /* 4698 * If this is the main process stack, see if we're over the stack 4699 * limit. 4700 */ 4701 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr && 4702 addr < (vm_offset_t)p->p_sysent->sv_usrstack; 4703 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) 4704 return (KERN_NO_SPACE); 4705 4706 #ifdef RACCT 4707 if (racct_enable) { 4708 PROC_LOCK(p); 4709 if (is_procstack && racct_set(p, RACCT_STACK, 4710 ctob(vm->vm_ssize) + grow_amount)) { 4711 PROC_UNLOCK(p); 4712 return (KERN_NO_SPACE); 4713 } 4714 PROC_UNLOCK(p); 4715 } 4716 #endif 4717 4718 grow_amount = roundup(grow_amount, sgrowsiz); 4719 if (grow_amount > max_grow) 4720 grow_amount = max_grow; 4721 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) { 4722 grow_amount = trunc_page((vm_size_t)stacklim) - 4723 ctob(vm->vm_ssize); 4724 } 4725 4726 #ifdef notyet 4727 PROC_LOCK(p); 4728 limit = racct_get_available(p, RACCT_STACK); 4729 PROC_UNLOCK(p); 4730 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit)) 4731 grow_amount = limit - ctob(vm->vm_ssize); 4732 #endif 4733 4734 if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) { 4735 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) { 4736 rv = KERN_NO_SPACE; 4737 goto out; 4738 } 4739 #ifdef RACCT 4740 if (racct_enable) { 4741 PROC_LOCK(p); 4742 if (racct_set(p, RACCT_MEMLOCK, 4743 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) { 4744 PROC_UNLOCK(p); 4745 rv = KERN_NO_SPACE; 4746 goto out; 4747 } 4748 PROC_UNLOCK(p); 4749 } 4750 #endif 4751 } 4752 4753 /* If we would blow our VMEM resource limit, no go */ 4754 if (map->size + grow_amount > vmemlim) { 4755 rv = KERN_NO_SPACE; 4756 goto out; 4757 } 4758 #ifdef RACCT 4759 if (racct_enable) { 4760 PROC_LOCK(p); 4761 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) { 4762 PROC_UNLOCK(p); 4763 rv = KERN_NO_SPACE; 4764 goto out; 4765 } 4766 PROC_UNLOCK(p); 4767 } 4768 #endif 4769 4770 if (vm_map_lock_upgrade(map)) { 4771 gap_entry = NULL; 4772 vm_map_lock_read(map); 4773 goto retry; 4774 } 4775 4776 if (grow_down) { 4777 grow_start = gap_entry->end - grow_amount; 4778 if (gap_entry->start + grow_amount == gap_entry->end) { 4779 gap_start = gap_entry->start; 4780 gap_end = gap_entry->end; 4781 vm_map_entry_delete(map, gap_entry); 4782 gap_deleted = true; 4783 } else { 4784 MPASS(gap_entry->start < gap_entry->end - grow_amount); 4785 vm_map_entry_resize(map, gap_entry, -grow_amount); 4786 gap_deleted = false; 4787 } 4788 rv = vm_map_insert(map, NULL, 0, grow_start, 4789 grow_start + grow_amount, 4790 stack_entry->protection, stack_entry->max_protection, 4791 MAP_STACK_GROWS_DOWN); 4792 if (rv != KERN_SUCCESS) { 4793 if (gap_deleted) { 4794 rv1 = vm_map_insert(map, NULL, 0, gap_start, 4795 gap_end, VM_PROT_NONE, VM_PROT_NONE, 4796 MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP_DN); 4797 MPASS(rv1 == KERN_SUCCESS); 4798 } else 4799 vm_map_entry_resize(map, gap_entry, 4800 grow_amount); 4801 } 4802 } else { 4803 grow_start = stack_entry->end; 4804 cred = stack_entry->cred; 4805 if (cred == NULL && stack_entry->object.vm_object != NULL) 4806 cred = stack_entry->object.vm_object->cred; 4807 if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred)) 4808 rv = KERN_NO_SPACE; 4809 /* Grow the underlying object if applicable. */ 4810 else if (stack_entry->object.vm_object == NULL || 4811 vm_object_coalesce(stack_entry->object.vm_object, 4812 stack_entry->offset, 4813 (vm_size_t)(stack_entry->end - stack_entry->start), 4814 grow_amount, cred != NULL)) { 4815 if (gap_entry->start + grow_amount == gap_entry->end) { 4816 vm_map_entry_delete(map, gap_entry); 4817 vm_map_entry_resize(map, stack_entry, 4818 grow_amount); 4819 } else { 4820 gap_entry->start += grow_amount; 4821 stack_entry->end += grow_amount; 4822 } 4823 map->size += grow_amount; 4824 rv = KERN_SUCCESS; 4825 } else 4826 rv = KERN_FAILURE; 4827 } 4828 if (rv == KERN_SUCCESS && is_procstack) 4829 vm->vm_ssize += btoc(grow_amount); 4830 4831 /* 4832 * Heed the MAP_WIREFUTURE flag if it was set for this process. 4833 */ 4834 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) { 4835 rv = vm_map_wire_locked(map, grow_start, 4836 grow_start + grow_amount, 4837 VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); 4838 } 4839 vm_map_lock_downgrade(map); 4840 4841 out: 4842 #ifdef RACCT 4843 if (racct_enable && rv != KERN_SUCCESS) { 4844 PROC_LOCK(p); 4845 error = racct_set(p, RACCT_VMEM, map->size); 4846 KASSERT(error == 0, ("decreasing RACCT_VMEM failed")); 4847 if (!old_mlock) { 4848 error = racct_set(p, RACCT_MEMLOCK, 4849 ptoa(pmap_wired_count(map->pmap))); 4850 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed")); 4851 } 4852 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize)); 4853 KASSERT(error == 0, ("decreasing RACCT_STACK failed")); 4854 PROC_UNLOCK(p); 4855 } 4856 #endif 4857 4858 return (rv); 4859 } 4860 4861 /* 4862 * Unshare the specified VM space for exec. If other processes are 4863 * mapped to it, then create a new one. The new vmspace is null. 4864 */ 4865 int 4866 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser) 4867 { 4868 struct vmspace *oldvmspace = p->p_vmspace; 4869 struct vmspace *newvmspace; 4870 4871 KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0, 4872 ("vmspace_exec recursed")); 4873 newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit); 4874 if (newvmspace == NULL) 4875 return (ENOMEM); 4876 newvmspace->vm_swrss = oldvmspace->vm_swrss; 4877 /* 4878 * This code is written like this for prototype purposes. The 4879 * goal is to avoid running down the vmspace here, but let the 4880 * other process's that are still using the vmspace to finally 4881 * run it down. Even though there is little or no chance of blocking 4882 * here, it is a good idea to keep this form for future mods. 4883 */ 4884 PROC_VMSPACE_LOCK(p); 4885 p->p_vmspace = newvmspace; 4886 PROC_VMSPACE_UNLOCK(p); 4887 if (p == curthread->td_proc) 4888 pmap_activate(curthread); 4889 curthread->td_pflags |= TDP_EXECVMSPC; 4890 return (0); 4891 } 4892 4893 /* 4894 * Unshare the specified VM space for forcing COW. This 4895 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 4896 */ 4897 int 4898 vmspace_unshare(struct proc *p) 4899 { 4900 struct vmspace *oldvmspace = p->p_vmspace; 4901 struct vmspace *newvmspace; 4902 vm_ooffset_t fork_charge; 4903 4904 if (refcount_load(&oldvmspace->vm_refcnt) == 1) 4905 return (0); 4906 fork_charge = 0; 4907 newvmspace = vmspace_fork(oldvmspace, &fork_charge); 4908 if (newvmspace == NULL) 4909 return (ENOMEM); 4910 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) { 4911 vmspace_free(newvmspace); 4912 return (ENOMEM); 4913 } 4914 PROC_VMSPACE_LOCK(p); 4915 p->p_vmspace = newvmspace; 4916 PROC_VMSPACE_UNLOCK(p); 4917 if (p == curthread->td_proc) 4918 pmap_activate(curthread); 4919 vmspace_free(oldvmspace); 4920 return (0); 4921 } 4922 4923 /* 4924 * vm_map_lookup: 4925 * 4926 * Finds the VM object, offset, and 4927 * protection for a given virtual address in the 4928 * specified map, assuming a page fault of the 4929 * type specified. 4930 * 4931 * Leaves the map in question locked for read; return 4932 * values are guaranteed until a vm_map_lookup_done 4933 * call is performed. Note that the map argument 4934 * is in/out; the returned map must be used in 4935 * the call to vm_map_lookup_done. 4936 * 4937 * A handle (out_entry) is returned for use in 4938 * vm_map_lookup_done, to make that fast. 4939 * 4940 * If a lookup is requested with "write protection" 4941 * specified, the map may be changed to perform virtual 4942 * copying operations, although the data referenced will 4943 * remain the same. 4944 */ 4945 int 4946 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4947 vm_offset_t vaddr, 4948 vm_prot_t fault_typea, 4949 vm_map_entry_t *out_entry, /* OUT */ 4950 vm_object_t *object, /* OUT */ 4951 vm_pindex_t *pindex, /* OUT */ 4952 vm_prot_t *out_prot, /* OUT */ 4953 boolean_t *wired) /* OUT */ 4954 { 4955 vm_map_entry_t entry; 4956 vm_map_t map = *var_map; 4957 vm_prot_t prot; 4958 vm_prot_t fault_type; 4959 vm_object_t eobject; 4960 vm_size_t size; 4961 struct ucred *cred; 4962 4963 RetryLookup: 4964 4965 vm_map_lock_read(map); 4966 4967 RetryLookupLocked: 4968 /* 4969 * Lookup the faulting address. 4970 */ 4971 if (!vm_map_lookup_entry(map, vaddr, out_entry)) { 4972 vm_map_unlock_read(map); 4973 return (KERN_INVALID_ADDRESS); 4974 } 4975 4976 entry = *out_entry; 4977 4978 /* 4979 * Handle submaps. 4980 */ 4981 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 4982 vm_map_t old_map = map; 4983 4984 *var_map = map = entry->object.sub_map; 4985 vm_map_unlock_read(old_map); 4986 goto RetryLookup; 4987 } 4988 4989 /* 4990 * Check whether this task is allowed to have this page. 4991 */ 4992 prot = entry->protection; 4993 if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) { 4994 fault_typea &= ~VM_PROT_FAULT_LOOKUP; 4995 if (prot == VM_PROT_NONE && map != kernel_map && 4996 (entry->eflags & MAP_ENTRY_GUARD) != 0 && 4997 (entry->eflags & (MAP_ENTRY_STACK_GAP_DN | 4998 MAP_ENTRY_STACK_GAP_UP)) != 0 && 4999 vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS) 5000 goto RetryLookupLocked; 5001 } 5002 fault_type = fault_typea & VM_PROT_ALL; 5003 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) { 5004 vm_map_unlock_read(map); 5005 return (KERN_PROTECTION_FAILURE); 5006 } 5007 KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags & 5008 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) != 5009 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY), 5010 ("entry %p flags %x", entry, entry->eflags)); 5011 if ((fault_typea & VM_PROT_COPY) != 0 && 5012 (entry->max_protection & VM_PROT_WRITE) == 0 && 5013 (entry->eflags & MAP_ENTRY_COW) == 0) { 5014 vm_map_unlock_read(map); 5015 return (KERN_PROTECTION_FAILURE); 5016 } 5017 5018 /* 5019 * If this page is not pageable, we have to get it for all possible 5020 * accesses. 5021 */ 5022 *wired = (entry->wired_count != 0); 5023 if (*wired) 5024 fault_type = entry->protection; 5025 size = entry->end - entry->start; 5026 5027 /* 5028 * If the entry was copy-on-write, we either ... 5029 */ 5030 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 5031 /* 5032 * If we want to write the page, we may as well handle that 5033 * now since we've got the map locked. 5034 * 5035 * If we don't need to write the page, we just demote the 5036 * permissions allowed. 5037 */ 5038 if ((fault_type & VM_PROT_WRITE) != 0 || 5039 (fault_typea & VM_PROT_COPY) != 0) { 5040 /* 5041 * Make a new object, and place it in the object 5042 * chain. Note that no new references have appeared 5043 * -- one just moved from the map to the new 5044 * object. 5045 */ 5046 if (vm_map_lock_upgrade(map)) 5047 goto RetryLookup; 5048 5049 if (entry->cred == NULL) { 5050 /* 5051 * The debugger owner is charged for 5052 * the memory. 5053 */ 5054 cred = curthread->td_ucred; 5055 crhold(cred); 5056 if (!swap_reserve_by_cred(size, cred)) { 5057 crfree(cred); 5058 vm_map_unlock(map); 5059 return (KERN_RESOURCE_SHORTAGE); 5060 } 5061 entry->cred = cred; 5062 } 5063 eobject = entry->object.vm_object; 5064 vm_object_shadow(&entry->object.vm_object, 5065 &entry->offset, size, entry->cred, false); 5066 if (eobject == entry->object.vm_object) { 5067 /* 5068 * The object was not shadowed. 5069 */ 5070 swap_release_by_cred(size, entry->cred); 5071 crfree(entry->cred); 5072 } 5073 entry->cred = NULL; 5074 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 5075 5076 vm_map_lock_downgrade(map); 5077 } else { 5078 /* 5079 * We're attempting to read a copy-on-write page -- 5080 * don't allow writes. 5081 */ 5082 prot &= ~VM_PROT_WRITE; 5083 } 5084 } 5085 5086 /* 5087 * Create an object if necessary. 5088 */ 5089 if (entry->object.vm_object == NULL && !map->system_map) { 5090 if (vm_map_lock_upgrade(map)) 5091 goto RetryLookup; 5092 entry->object.vm_object = vm_object_allocate_anon(atop(size), 5093 NULL, entry->cred, entry->cred != NULL ? size : 0); 5094 entry->offset = 0; 5095 entry->cred = NULL; 5096 vm_map_lock_downgrade(map); 5097 } 5098 5099 /* 5100 * Return the object/offset from this entry. If the entry was 5101 * copy-on-write or empty, it has been fixed up. 5102 */ 5103 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 5104 *object = entry->object.vm_object; 5105 5106 *out_prot = prot; 5107 return (KERN_SUCCESS); 5108 } 5109 5110 /* 5111 * vm_map_lookup_locked: 5112 * 5113 * Lookup the faulting address. A version of vm_map_lookup that returns 5114 * KERN_FAILURE instead of blocking on map lock or memory allocation. 5115 */ 5116 int 5117 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */ 5118 vm_offset_t vaddr, 5119 vm_prot_t fault_typea, 5120 vm_map_entry_t *out_entry, /* OUT */ 5121 vm_object_t *object, /* OUT */ 5122 vm_pindex_t *pindex, /* OUT */ 5123 vm_prot_t *out_prot, /* OUT */ 5124 boolean_t *wired) /* OUT */ 5125 { 5126 vm_map_entry_t entry; 5127 vm_map_t map = *var_map; 5128 vm_prot_t prot; 5129 vm_prot_t fault_type = fault_typea; 5130 5131 /* 5132 * Lookup the faulting address. 5133 */ 5134 if (!vm_map_lookup_entry(map, vaddr, out_entry)) 5135 return (KERN_INVALID_ADDRESS); 5136 5137 entry = *out_entry; 5138 5139 /* 5140 * Fail if the entry refers to a submap. 5141 */ 5142 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) 5143 return (KERN_FAILURE); 5144 5145 /* 5146 * Check whether this task is allowed to have this page. 5147 */ 5148 prot = entry->protection; 5149 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 5150 if ((fault_type & prot) != fault_type) 5151 return (KERN_PROTECTION_FAILURE); 5152 5153 /* 5154 * If this page is not pageable, we have to get it for all possible 5155 * accesses. 5156 */ 5157 *wired = (entry->wired_count != 0); 5158 if (*wired) 5159 fault_type = entry->protection; 5160 5161 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 5162 /* 5163 * Fail if the entry was copy-on-write for a write fault. 5164 */ 5165 if (fault_type & VM_PROT_WRITE) 5166 return (KERN_FAILURE); 5167 /* 5168 * We're attempting to read a copy-on-write page -- 5169 * don't allow writes. 5170 */ 5171 prot &= ~VM_PROT_WRITE; 5172 } 5173 5174 /* 5175 * Fail if an object should be created. 5176 */ 5177 if (entry->object.vm_object == NULL && !map->system_map) 5178 return (KERN_FAILURE); 5179 5180 /* 5181 * Return the object/offset from this entry. If the entry was 5182 * copy-on-write or empty, it has been fixed up. 5183 */ 5184 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 5185 *object = entry->object.vm_object; 5186 5187 *out_prot = prot; 5188 return (KERN_SUCCESS); 5189 } 5190 5191 /* 5192 * vm_map_lookup_done: 5193 * 5194 * Releases locks acquired by a vm_map_lookup 5195 * (according to the handle returned by that lookup). 5196 */ 5197 void 5198 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry) 5199 { 5200 /* 5201 * Unlock the main-level map 5202 */ 5203 vm_map_unlock_read(map); 5204 } 5205 5206 vm_offset_t 5207 vm_map_max_KBI(const struct vm_map *map) 5208 { 5209 5210 return (vm_map_max(map)); 5211 } 5212 5213 vm_offset_t 5214 vm_map_min_KBI(const struct vm_map *map) 5215 { 5216 5217 return (vm_map_min(map)); 5218 } 5219 5220 pmap_t 5221 vm_map_pmap_KBI(vm_map_t map) 5222 { 5223 5224 return (map->pmap); 5225 } 5226 5227 bool 5228 vm_map_range_valid_KBI(vm_map_t map, vm_offset_t start, vm_offset_t end) 5229 { 5230 5231 return (vm_map_range_valid(map, start, end)); 5232 } 5233 5234 #ifdef INVARIANTS 5235 static void 5236 _vm_map_assert_consistent(vm_map_t map, int check) 5237 { 5238 vm_map_entry_t entry, prev; 5239 vm_map_entry_t cur, header, lbound, ubound; 5240 vm_size_t max_left, max_right; 5241 5242 #ifdef DIAGNOSTIC 5243 ++map->nupdates; 5244 #endif 5245 if (enable_vmmap_check != check) 5246 return; 5247 5248 header = prev = &map->header; 5249 VM_MAP_ENTRY_FOREACH(entry, map) { 5250 KASSERT(prev->end <= entry->start, 5251 ("map %p prev->end = %jx, start = %jx", map, 5252 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5253 KASSERT(entry->start < entry->end, 5254 ("map %p start = %jx, end = %jx", map, 5255 (uintmax_t)entry->start, (uintmax_t)entry->end)); 5256 KASSERT(entry->left == header || 5257 entry->left->start < entry->start, 5258 ("map %p left->start = %jx, start = %jx", map, 5259 (uintmax_t)entry->left->start, (uintmax_t)entry->start)); 5260 KASSERT(entry->right == header || 5261 entry->start < entry->right->start, 5262 ("map %p start = %jx, right->start = %jx", map, 5263 (uintmax_t)entry->start, (uintmax_t)entry->right->start)); 5264 cur = map->root; 5265 lbound = ubound = header; 5266 for (;;) { 5267 if (entry->start < cur->start) { 5268 ubound = cur; 5269 cur = cur->left; 5270 KASSERT(cur != lbound, 5271 ("map %p cannot find %jx", 5272 map, (uintmax_t)entry->start)); 5273 } else if (cur->end <= entry->start) { 5274 lbound = cur; 5275 cur = cur->right; 5276 KASSERT(cur != ubound, 5277 ("map %p cannot find %jx", 5278 map, (uintmax_t)entry->start)); 5279 } else { 5280 KASSERT(cur == entry, 5281 ("map %p cannot find %jx", 5282 map, (uintmax_t)entry->start)); 5283 break; 5284 } 5285 } 5286 max_left = vm_map_entry_max_free_left(entry, lbound); 5287 max_right = vm_map_entry_max_free_right(entry, ubound); 5288 KASSERT(entry->max_free == vm_size_max(max_left, max_right), 5289 ("map %p max = %jx, max_left = %jx, max_right = %jx", map, 5290 (uintmax_t)entry->max_free, 5291 (uintmax_t)max_left, (uintmax_t)max_right)); 5292 prev = entry; 5293 } 5294 KASSERT(prev->end <= entry->start, 5295 ("map %p prev->end = %jx, start = %jx", map, 5296 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5297 } 5298 #endif 5299 5300 #include "opt_ddb.h" 5301 #ifdef DDB 5302 #include <sys/kernel.h> 5303 5304 #include <ddb/ddb.h> 5305 5306 static void 5307 vm_map_print(vm_map_t map) 5308 { 5309 vm_map_entry_t entry, prev; 5310 5311 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 5312 (void *)map, 5313 (void *)map->pmap, map->nentries, map->timestamp); 5314 5315 db_indent += 2; 5316 prev = &map->header; 5317 VM_MAP_ENTRY_FOREACH(entry, map) { 5318 db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n", 5319 (void *)entry, (void *)entry->start, (void *)entry->end, 5320 entry->eflags); 5321 { 5322 static const char * const inheritance_name[4] = 5323 {"share", "copy", "none", "donate_copy"}; 5324 5325 db_iprintf(" prot=%x/%x/%s", 5326 entry->protection, 5327 entry->max_protection, 5328 inheritance_name[(int)(unsigned char) 5329 entry->inheritance]); 5330 if (entry->wired_count != 0) 5331 db_printf(", wired"); 5332 } 5333 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 5334 db_printf(", share=%p, offset=0x%jx\n", 5335 (void *)entry->object.sub_map, 5336 (uintmax_t)entry->offset); 5337 if (prev == &map->header || 5338 prev->object.sub_map != 5339 entry->object.sub_map) { 5340 db_indent += 2; 5341 vm_map_print((vm_map_t)entry->object.sub_map); 5342 db_indent -= 2; 5343 } 5344 } else { 5345 if (entry->cred != NULL) 5346 db_printf(", ruid %d", entry->cred->cr_ruid); 5347 db_printf(", object=%p, offset=0x%jx", 5348 (void *)entry->object.vm_object, 5349 (uintmax_t)entry->offset); 5350 if (entry->object.vm_object && entry->object.vm_object->cred) 5351 db_printf(", obj ruid %d charge %jx", 5352 entry->object.vm_object->cred->cr_ruid, 5353 (uintmax_t)entry->object.vm_object->charge); 5354 if (entry->eflags & MAP_ENTRY_COW) 5355 db_printf(", copy (%s)", 5356 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 5357 db_printf("\n"); 5358 5359 if (prev == &map->header || 5360 prev->object.vm_object != 5361 entry->object.vm_object) { 5362 db_indent += 2; 5363 vm_object_print((db_expr_t)(intptr_t) 5364 entry->object.vm_object, 5365 0, 0, (char *)0); 5366 db_indent -= 2; 5367 } 5368 } 5369 prev = entry; 5370 } 5371 db_indent -= 2; 5372 } 5373 5374 DB_SHOW_COMMAND(map, map) 5375 { 5376 5377 if (!have_addr) { 5378 db_printf("usage: show map <addr>\n"); 5379 return; 5380 } 5381 vm_map_print((vm_map_t)addr); 5382 } 5383 5384 DB_SHOW_COMMAND(procvm, procvm) 5385 { 5386 struct proc *p; 5387 5388 if (have_addr) { 5389 p = db_lookup_proc(addr); 5390 } else { 5391 p = curproc; 5392 } 5393 5394 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 5395 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 5396 (void *)vmspace_pmap(p->p_vmspace)); 5397 5398 vm_map_print((vm_map_t)&p->p_vmspace->vm_map); 5399 } 5400 5401 #endif /* DDB */ 5402