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 /* 1675 * Find the entry prior to the proposed starting address; if it's part 1676 * of an existing entry, this range is bogus. 1677 */ 1678 if (vm_map_lookup_entry(map, start, &prev_entry)) 1679 return (KERN_NO_SPACE); 1680 1681 /* 1682 * Assert that the next entry doesn't overlap the end point. 1683 */ 1684 next_entry = vm_map_entry_succ(prev_entry); 1685 if (next_entry->start < end) 1686 return (KERN_NO_SPACE); 1687 1688 if ((cow & MAP_CREATE_GUARD) != 0 && (object != NULL || 1689 max != VM_PROT_NONE)) 1690 return (KERN_INVALID_ARGUMENT); 1691 1692 protoeflags = 0; 1693 if (cow & MAP_COPY_ON_WRITE) 1694 protoeflags |= MAP_ENTRY_COW | MAP_ENTRY_NEEDS_COPY; 1695 if (cow & MAP_NOFAULT) 1696 protoeflags |= MAP_ENTRY_NOFAULT; 1697 if (cow & MAP_DISABLE_SYNCER) 1698 protoeflags |= MAP_ENTRY_NOSYNC; 1699 if (cow & MAP_DISABLE_COREDUMP) 1700 protoeflags |= MAP_ENTRY_NOCOREDUMP; 1701 if (cow & MAP_STACK_GROWS_DOWN) 1702 protoeflags |= MAP_ENTRY_GROWS_DOWN; 1703 if (cow & MAP_STACK_GROWS_UP) 1704 protoeflags |= MAP_ENTRY_GROWS_UP; 1705 if (cow & MAP_WRITECOUNT) 1706 protoeflags |= MAP_ENTRY_WRITECNT; 1707 if (cow & MAP_VN_EXEC) 1708 protoeflags |= MAP_ENTRY_VN_EXEC; 1709 if ((cow & MAP_CREATE_GUARD) != 0) 1710 protoeflags |= MAP_ENTRY_GUARD; 1711 if ((cow & MAP_CREATE_STACK_GAP_DN) != 0) 1712 protoeflags |= MAP_ENTRY_STACK_GAP_DN; 1713 if ((cow & MAP_CREATE_STACK_GAP_UP) != 0) 1714 protoeflags |= MAP_ENTRY_STACK_GAP_UP; 1715 if (cow & MAP_INHERIT_SHARE) 1716 inheritance = VM_INHERIT_SHARE; 1717 else 1718 inheritance = VM_INHERIT_DEFAULT; 1719 if ((cow & MAP_SPLIT_BOUNDARY_MASK) != 0) { 1720 /* This magically ignores index 0, for usual page size. */ 1721 bidx = (cow & MAP_SPLIT_BOUNDARY_MASK) >> 1722 MAP_SPLIT_BOUNDARY_SHIFT; 1723 if (bidx >= MAXPAGESIZES) 1724 return (KERN_INVALID_ARGUMENT); 1725 bdry = pagesizes[bidx] - 1; 1726 if ((start & bdry) != 0 || (end & bdry) != 0) 1727 return (KERN_INVALID_ARGUMENT); 1728 protoeflags |= bidx << MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 1729 } 1730 1731 cred = NULL; 1732 if ((cow & (MAP_ACC_NO_CHARGE | MAP_NOFAULT | MAP_CREATE_GUARD)) != 0) 1733 goto charged; 1734 if ((cow & MAP_ACC_CHARGED) || ((prot & VM_PROT_WRITE) && 1735 ((protoeflags & MAP_ENTRY_NEEDS_COPY) || object == NULL))) { 1736 if (!(cow & MAP_ACC_CHARGED) && !swap_reserve(end - start)) 1737 return (KERN_RESOURCE_SHORTAGE); 1738 KASSERT(object == NULL || 1739 (protoeflags & MAP_ENTRY_NEEDS_COPY) != 0 || 1740 object->cred == NULL, 1741 ("overcommit: vm_map_insert o %p", object)); 1742 cred = curthread->td_ucred; 1743 } 1744 1745 charged: 1746 /* Expand the kernel pmap, if necessary. */ 1747 if (map == kernel_map && end > kernel_vm_end) 1748 pmap_growkernel(end); 1749 if (object != NULL) { 1750 /* 1751 * OBJ_ONEMAPPING must be cleared unless this mapping 1752 * is trivially proven to be the only mapping for any 1753 * of the object's pages. (Object granularity 1754 * reference counting is insufficient to recognize 1755 * aliases with precision.) 1756 */ 1757 if ((object->flags & OBJ_ANON) != 0) { 1758 VM_OBJECT_WLOCK(object); 1759 if (object->ref_count > 1 || object->shadow_count != 0) 1760 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1761 VM_OBJECT_WUNLOCK(object); 1762 } 1763 } else if ((prev_entry->eflags & ~MAP_ENTRY_USER_WIRED) == 1764 protoeflags && 1765 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP | 1766 MAP_VN_EXEC)) == 0 && 1767 prev_entry->end == start && (prev_entry->cred == cred || 1768 (prev_entry->object.vm_object != NULL && 1769 prev_entry->object.vm_object->cred == cred)) && 1770 vm_object_coalesce(prev_entry->object.vm_object, 1771 prev_entry->offset, 1772 (vm_size_t)(prev_entry->end - prev_entry->start), 1773 (vm_size_t)(end - prev_entry->end), cred != NULL && 1774 (protoeflags & MAP_ENTRY_NEEDS_COPY) == 0)) { 1775 /* 1776 * We were able to extend the object. Determine if we 1777 * can extend the previous map entry to include the 1778 * new range as well. 1779 */ 1780 if (prev_entry->inheritance == inheritance && 1781 prev_entry->protection == prot && 1782 prev_entry->max_protection == max && 1783 prev_entry->wired_count == 0) { 1784 KASSERT((prev_entry->eflags & MAP_ENTRY_USER_WIRED) == 1785 0, ("prev_entry %p has incoherent wiring", 1786 prev_entry)); 1787 if ((prev_entry->eflags & MAP_ENTRY_GUARD) == 0) 1788 map->size += end - prev_entry->end; 1789 vm_map_entry_resize(map, prev_entry, 1790 end - prev_entry->end); 1791 vm_map_try_merge_entries(map, prev_entry, next_entry); 1792 return (KERN_SUCCESS); 1793 } 1794 1795 /* 1796 * If we can extend the object but cannot extend the 1797 * map entry, we have to create a new map entry. We 1798 * must bump the ref count on the extended object to 1799 * account for it. object may be NULL. 1800 */ 1801 object = prev_entry->object.vm_object; 1802 offset = prev_entry->offset + 1803 (prev_entry->end - prev_entry->start); 1804 vm_object_reference(object); 1805 if (cred != NULL && object != NULL && object->cred != NULL && 1806 !(prev_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 1807 /* Object already accounts for this uid. */ 1808 cred = NULL; 1809 } 1810 } 1811 if (cred != NULL) 1812 crhold(cred); 1813 1814 /* 1815 * Create a new entry 1816 */ 1817 new_entry = vm_map_entry_create(map); 1818 new_entry->start = start; 1819 new_entry->end = end; 1820 new_entry->cred = NULL; 1821 1822 new_entry->eflags = protoeflags; 1823 new_entry->object.vm_object = object; 1824 new_entry->offset = offset; 1825 1826 new_entry->inheritance = inheritance; 1827 new_entry->protection = prot; 1828 new_entry->max_protection = max; 1829 new_entry->wired_count = 0; 1830 new_entry->wiring_thread = NULL; 1831 new_entry->read_ahead = VM_FAULT_READ_AHEAD_INIT; 1832 new_entry->next_read = start; 1833 1834 KASSERT(cred == NULL || !ENTRY_CHARGED(new_entry), 1835 ("overcommit: vm_map_insert leaks vm_map %p", new_entry)); 1836 new_entry->cred = cred; 1837 1838 /* 1839 * Insert the new entry into the list 1840 */ 1841 vm_map_entry_link(map, new_entry); 1842 if ((new_entry->eflags & MAP_ENTRY_GUARD) == 0) 1843 map->size += new_entry->end - new_entry->start; 1844 1845 /* 1846 * Try to coalesce the new entry with both the previous and next 1847 * entries in the list. Previously, we only attempted to coalesce 1848 * with the previous entry when object is NULL. Here, we handle the 1849 * other cases, which are less common. 1850 */ 1851 vm_map_try_merge_entries(map, prev_entry, new_entry); 1852 vm_map_try_merge_entries(map, new_entry, next_entry); 1853 1854 if ((cow & (MAP_PREFAULT | MAP_PREFAULT_PARTIAL)) != 0) { 1855 vm_map_pmap_enter(map, start, prot, object, OFF_TO_IDX(offset), 1856 end - start, cow & MAP_PREFAULT_PARTIAL); 1857 } 1858 1859 return (KERN_SUCCESS); 1860 } 1861 1862 /* 1863 * vm_map_findspace: 1864 * 1865 * Find the first fit (lowest VM address) for "length" free bytes 1866 * beginning at address >= start in the given map. 1867 * 1868 * In a vm_map_entry, "max_free" is the maximum amount of 1869 * contiguous free space between an entry in its subtree and a 1870 * neighbor of that entry. This allows finding a free region in 1871 * one path down the tree, so O(log n) amortized with splay 1872 * trees. 1873 * 1874 * The map must be locked, and leaves it so. 1875 * 1876 * Returns: starting address if sufficient space, 1877 * vm_map_max(map)-length+1 if insufficient space. 1878 */ 1879 vm_offset_t 1880 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length) 1881 { 1882 vm_map_entry_t header, llist, rlist, root, y; 1883 vm_size_t left_length, max_free_left, max_free_right; 1884 vm_offset_t gap_end; 1885 1886 VM_MAP_ASSERT_LOCKED(map); 1887 1888 /* 1889 * Request must fit within min/max VM address and must avoid 1890 * address wrap. 1891 */ 1892 start = MAX(start, vm_map_min(map)); 1893 if (start >= vm_map_max(map) || length > vm_map_max(map) - start) 1894 return (vm_map_max(map) - length + 1); 1895 1896 /* Empty tree means wide open address space. */ 1897 if (map->root == NULL) 1898 return (start); 1899 1900 /* 1901 * After splay_split, if start is within an entry, push it to the start 1902 * of the following gap. If rlist is at the end of the gap containing 1903 * start, save the end of that gap in gap_end to see if the gap is big 1904 * enough; otherwise set gap_end to start skip gap-checking and move 1905 * directly to a search of the right subtree. 1906 */ 1907 header = &map->header; 1908 root = vm_map_splay_split(map, start, length, &llist, &rlist); 1909 gap_end = rlist->start; 1910 if (root != NULL) { 1911 start = root->end; 1912 if (root->right != rlist) 1913 gap_end = start; 1914 max_free_left = vm_map_splay_merge_left(header, root, llist); 1915 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1916 } else if (rlist != header) { 1917 root = rlist; 1918 rlist = root->left; 1919 max_free_left = vm_map_splay_merge_pred(header, root, llist); 1920 max_free_right = vm_map_splay_merge_right(header, root, rlist); 1921 } else { 1922 root = llist; 1923 llist = root->right; 1924 max_free_left = vm_map_splay_merge_left(header, root, llist); 1925 max_free_right = vm_map_splay_merge_succ(header, root, rlist); 1926 } 1927 root->max_free = vm_size_max(max_free_left, max_free_right); 1928 map->root = root; 1929 VM_MAP_ASSERT_CONSISTENT(map); 1930 if (length <= gap_end - start) 1931 return (start); 1932 1933 /* With max_free, can immediately tell if no solution. */ 1934 if (root->right == header || length > root->right->max_free) 1935 return (vm_map_max(map) - length + 1); 1936 1937 /* 1938 * Splay for the least large-enough gap in the right subtree. 1939 */ 1940 llist = rlist = header; 1941 for (left_length = 0;; 1942 left_length = vm_map_entry_max_free_left(root, llist)) { 1943 if (length <= left_length) 1944 SPLAY_LEFT_STEP(root, y, llist, rlist, 1945 length <= vm_map_entry_max_free_left(y, llist)); 1946 else 1947 SPLAY_RIGHT_STEP(root, y, llist, rlist, 1948 length > vm_map_entry_max_free_left(y, root)); 1949 if (root == NULL) 1950 break; 1951 } 1952 root = llist; 1953 llist = root->right; 1954 max_free_left = vm_map_splay_merge_left(header, root, llist); 1955 if (rlist == header) { 1956 root->max_free = vm_size_max(max_free_left, 1957 vm_map_splay_merge_succ(header, root, rlist)); 1958 } else { 1959 y = rlist; 1960 rlist = y->left; 1961 y->max_free = vm_size_max( 1962 vm_map_splay_merge_pred(root, y, root), 1963 vm_map_splay_merge_right(header, y, rlist)); 1964 root->max_free = vm_size_max(max_free_left, y->max_free); 1965 } 1966 map->root = root; 1967 VM_MAP_ASSERT_CONSISTENT(map); 1968 return (root->end); 1969 } 1970 1971 int 1972 vm_map_fixed(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1973 vm_offset_t start, vm_size_t length, vm_prot_t prot, 1974 vm_prot_t max, int cow) 1975 { 1976 vm_offset_t end; 1977 int result; 1978 1979 end = start + length; 1980 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 1981 object == NULL, 1982 ("vm_map_fixed: non-NULL backing object for stack")); 1983 vm_map_lock(map); 1984 VM_MAP_RANGE_CHECK(map, start, end); 1985 if ((cow & MAP_CHECK_EXCL) == 0) { 1986 result = vm_map_delete(map, start, end); 1987 if (result != KERN_SUCCESS) 1988 goto out; 1989 } 1990 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 1991 result = vm_map_stack_locked(map, start, length, sgrowsiz, 1992 prot, max, cow); 1993 } else { 1994 result = vm_map_insert(map, object, offset, start, end, 1995 prot, max, cow); 1996 } 1997 out: 1998 vm_map_unlock(map); 1999 return (result); 2000 } 2001 2002 static const int aslr_pages_rnd_64[2] = {0x1000, 0x10}; 2003 static const int aslr_pages_rnd_32[2] = {0x100, 0x4}; 2004 2005 static int cluster_anon = 1; 2006 SYSCTL_INT(_vm, OID_AUTO, cluster_anon, CTLFLAG_RW, 2007 &cluster_anon, 0, 2008 "Cluster anonymous mappings: 0 = no, 1 = yes if no hint, 2 = always"); 2009 2010 static bool 2011 clustering_anon_allowed(vm_offset_t addr) 2012 { 2013 2014 switch (cluster_anon) { 2015 case 0: 2016 return (false); 2017 case 1: 2018 return (addr == 0); 2019 case 2: 2020 default: 2021 return (true); 2022 } 2023 } 2024 2025 static long aslr_restarts; 2026 SYSCTL_LONG(_vm, OID_AUTO, aslr_restarts, CTLFLAG_RD, 2027 &aslr_restarts, 0, 2028 "Number of aslr failures"); 2029 2030 /* 2031 * Searches for the specified amount of free space in the given map with the 2032 * specified alignment. Performs an address-ordered, first-fit search from 2033 * the given address "*addr", with an optional upper bound "max_addr". If the 2034 * parameter "alignment" is zero, then the alignment is computed from the 2035 * given (object, offset) pair so as to enable the greatest possible use of 2036 * superpage mappings. Returns KERN_SUCCESS and the address of the free space 2037 * in "*addr" if successful. Otherwise, returns KERN_NO_SPACE. 2038 * 2039 * The map must be locked. Initially, there must be at least "length" bytes 2040 * of free space at the given address. 2041 */ 2042 static int 2043 vm_map_alignspace(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2044 vm_offset_t *addr, vm_size_t length, vm_offset_t max_addr, 2045 vm_offset_t alignment) 2046 { 2047 vm_offset_t aligned_addr, free_addr; 2048 2049 VM_MAP_ASSERT_LOCKED(map); 2050 free_addr = *addr; 2051 KASSERT(free_addr == vm_map_findspace(map, free_addr, length), 2052 ("caller failed to provide space %#jx at address %p", 2053 (uintmax_t)length, (void *)free_addr)); 2054 for (;;) { 2055 /* 2056 * At the start of every iteration, the free space at address 2057 * "*addr" is at least "length" bytes. 2058 */ 2059 if (alignment == 0) 2060 pmap_align_superpage(object, offset, addr, length); 2061 else if ((*addr & (alignment - 1)) != 0) { 2062 *addr &= ~(alignment - 1); 2063 *addr += alignment; 2064 } 2065 aligned_addr = *addr; 2066 if (aligned_addr == free_addr) { 2067 /* 2068 * Alignment did not change "*addr", so "*addr" must 2069 * still provide sufficient free space. 2070 */ 2071 return (KERN_SUCCESS); 2072 } 2073 2074 /* 2075 * Test for address wrap on "*addr". A wrapped "*addr" could 2076 * be a valid address, in which case vm_map_findspace() cannot 2077 * be relied upon to fail. 2078 */ 2079 if (aligned_addr < free_addr) 2080 return (KERN_NO_SPACE); 2081 *addr = vm_map_findspace(map, aligned_addr, length); 2082 if (*addr + length > vm_map_max(map) || 2083 (max_addr != 0 && *addr + length > max_addr)) 2084 return (KERN_NO_SPACE); 2085 free_addr = *addr; 2086 if (free_addr == aligned_addr) { 2087 /* 2088 * If a successful call to vm_map_findspace() did not 2089 * change "*addr", then "*addr" must still be aligned 2090 * and provide sufficient free space. 2091 */ 2092 return (KERN_SUCCESS); 2093 } 2094 } 2095 } 2096 2097 int 2098 vm_map_find_aligned(vm_map_t map, vm_offset_t *addr, vm_size_t length, 2099 vm_offset_t max_addr, vm_offset_t alignment) 2100 { 2101 /* XXXKIB ASLR eh ? */ 2102 *addr = vm_map_findspace(map, *addr, length); 2103 if (*addr + length > vm_map_max(map) || 2104 (max_addr != 0 && *addr + length > max_addr)) 2105 return (KERN_NO_SPACE); 2106 return (vm_map_alignspace(map, NULL, 0, addr, length, max_addr, 2107 alignment)); 2108 } 2109 2110 /* 2111 * vm_map_find finds an unallocated region in the target address 2112 * map with the given length. The search is defined to be 2113 * first-fit from the specified address; the region found is 2114 * returned in the same parameter. 2115 * 2116 * If object is non-NULL, ref count must be bumped by caller 2117 * prior to making call to account for the new entry. 2118 */ 2119 int 2120 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2121 vm_offset_t *addr, /* IN/OUT */ 2122 vm_size_t length, vm_offset_t max_addr, int find_space, 2123 vm_prot_t prot, vm_prot_t max, int cow) 2124 { 2125 vm_offset_t alignment, curr_min_addr, min_addr; 2126 int gap, pidx, rv, try; 2127 bool cluster, en_aslr, update_anon; 2128 2129 KASSERT((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0 || 2130 object == NULL, 2131 ("vm_map_find: non-NULL backing object for stack")); 2132 MPASS((cow & MAP_REMAP) == 0 || (find_space == VMFS_NO_SPACE && 2133 (cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) == 0)); 2134 if (find_space == VMFS_OPTIMAL_SPACE && (object == NULL || 2135 (object->flags & OBJ_COLORED) == 0)) 2136 find_space = VMFS_ANY_SPACE; 2137 if (find_space >> 8 != 0) { 2138 KASSERT((find_space & 0xff) == 0, ("bad VMFS flags")); 2139 alignment = (vm_offset_t)1 << (find_space >> 8); 2140 } else 2141 alignment = 0; 2142 en_aslr = (map->flags & MAP_ASLR) != 0; 2143 update_anon = cluster = clustering_anon_allowed(*addr) && 2144 (map->flags & MAP_IS_SUB_MAP) == 0 && max_addr == 0 && 2145 find_space != VMFS_NO_SPACE && object == NULL && 2146 (cow & (MAP_INHERIT_SHARE | MAP_STACK_GROWS_UP | 2147 MAP_STACK_GROWS_DOWN)) == 0 && prot != PROT_NONE; 2148 curr_min_addr = min_addr = *addr; 2149 if (en_aslr && min_addr == 0 && !cluster && 2150 find_space != VMFS_NO_SPACE && 2151 (map->flags & MAP_ASLR_IGNSTART) != 0) 2152 curr_min_addr = min_addr = vm_map_min(map); 2153 try = 0; 2154 vm_map_lock(map); 2155 if (cluster) { 2156 curr_min_addr = map->anon_loc; 2157 if (curr_min_addr == 0) 2158 cluster = false; 2159 } 2160 if (find_space != VMFS_NO_SPACE) { 2161 KASSERT(find_space == VMFS_ANY_SPACE || 2162 find_space == VMFS_OPTIMAL_SPACE || 2163 find_space == VMFS_SUPER_SPACE || 2164 alignment != 0, ("unexpected VMFS flag")); 2165 again: 2166 /* 2167 * When creating an anonymous mapping, try clustering 2168 * with an existing anonymous mapping first. 2169 * 2170 * We make up to two attempts to find address space 2171 * for a given find_space value. The first attempt may 2172 * apply randomization or may cluster with an existing 2173 * anonymous mapping. If this first attempt fails, 2174 * perform a first-fit search of the available address 2175 * space. 2176 * 2177 * If all tries failed, and find_space is 2178 * VMFS_OPTIMAL_SPACE, fallback to VMFS_ANY_SPACE. 2179 * Again enable clustering and randomization. 2180 */ 2181 try++; 2182 MPASS(try <= 2); 2183 2184 if (try == 2) { 2185 /* 2186 * Second try: we failed either to find a 2187 * suitable region for randomizing the 2188 * allocation, or to cluster with an existing 2189 * mapping. Retry with free run. 2190 */ 2191 curr_min_addr = (map->flags & MAP_ASLR_IGNSTART) != 0 ? 2192 vm_map_min(map) : min_addr; 2193 atomic_add_long(&aslr_restarts, 1); 2194 } 2195 2196 if (try == 1 && en_aslr && !cluster) { 2197 /* 2198 * Find space for allocation, including 2199 * gap needed for later randomization. 2200 */ 2201 pidx = MAXPAGESIZES > 1 && pagesizes[1] != 0 && 2202 (find_space == VMFS_SUPER_SPACE || find_space == 2203 VMFS_OPTIMAL_SPACE) ? 1 : 0; 2204 gap = vm_map_max(map) > MAP_32BIT_MAX_ADDR && 2205 (max_addr == 0 || max_addr > MAP_32BIT_MAX_ADDR) ? 2206 aslr_pages_rnd_64[pidx] : aslr_pages_rnd_32[pidx]; 2207 *addr = vm_map_findspace(map, curr_min_addr, 2208 length + gap * pagesizes[pidx]); 2209 if (*addr + length + gap * pagesizes[pidx] > 2210 vm_map_max(map)) 2211 goto again; 2212 /* And randomize the start address. */ 2213 *addr += (arc4random() % gap) * pagesizes[pidx]; 2214 if (max_addr != 0 && *addr + length > max_addr) 2215 goto again; 2216 } else { 2217 *addr = vm_map_findspace(map, curr_min_addr, length); 2218 if (*addr + length > vm_map_max(map) || 2219 (max_addr != 0 && *addr + length > max_addr)) { 2220 if (cluster) { 2221 cluster = false; 2222 MPASS(try == 1); 2223 goto again; 2224 } 2225 rv = KERN_NO_SPACE; 2226 goto done; 2227 } 2228 } 2229 2230 if (find_space != VMFS_ANY_SPACE && 2231 (rv = vm_map_alignspace(map, object, offset, addr, length, 2232 max_addr, alignment)) != KERN_SUCCESS) { 2233 if (find_space == VMFS_OPTIMAL_SPACE) { 2234 find_space = VMFS_ANY_SPACE; 2235 curr_min_addr = min_addr; 2236 cluster = update_anon; 2237 try = 0; 2238 goto again; 2239 } 2240 goto done; 2241 } 2242 } else if ((cow & MAP_REMAP) != 0) { 2243 if (!vm_map_range_valid(map, *addr, *addr + length)) { 2244 rv = KERN_INVALID_ADDRESS; 2245 goto done; 2246 } 2247 rv = vm_map_delete(map, *addr, *addr + length); 2248 if (rv != KERN_SUCCESS) 2249 goto done; 2250 } 2251 if ((cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP)) != 0) { 2252 rv = vm_map_stack_locked(map, *addr, length, sgrowsiz, prot, 2253 max, cow); 2254 } else { 2255 rv = vm_map_insert(map, object, offset, *addr, *addr + length, 2256 prot, max, cow); 2257 } 2258 if (rv == KERN_SUCCESS && update_anon) 2259 map->anon_loc = *addr + length; 2260 done: 2261 vm_map_unlock(map); 2262 return (rv); 2263 } 2264 2265 /* 2266 * vm_map_find_min() is a variant of vm_map_find() that takes an 2267 * additional parameter (min_addr) and treats the given address 2268 * (*addr) differently. Specifically, it treats *addr as a hint 2269 * and not as the minimum address where the mapping is created. 2270 * 2271 * This function works in two phases. First, it tries to 2272 * allocate above the hint. If that fails and the hint is 2273 * greater than min_addr, it performs a second pass, replacing 2274 * the hint with min_addr as the minimum address for the 2275 * allocation. 2276 */ 2277 int 2278 vm_map_find_min(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 2279 vm_offset_t *addr, vm_size_t length, vm_offset_t min_addr, 2280 vm_offset_t max_addr, int find_space, vm_prot_t prot, vm_prot_t max, 2281 int cow) 2282 { 2283 vm_offset_t hint; 2284 int rv; 2285 2286 hint = *addr; 2287 for (;;) { 2288 rv = vm_map_find(map, object, offset, addr, length, max_addr, 2289 find_space, prot, max, cow); 2290 if (rv == KERN_SUCCESS || min_addr >= hint) 2291 return (rv); 2292 *addr = hint = min_addr; 2293 } 2294 } 2295 2296 /* 2297 * A map entry with any of the following flags set must not be merged with 2298 * another entry. 2299 */ 2300 #define MAP_ENTRY_NOMERGE_MASK (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP | \ 2301 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_IS_SUB_MAP | MAP_ENTRY_VN_EXEC) 2302 2303 static bool 2304 vm_map_mergeable_neighbors(vm_map_entry_t prev, vm_map_entry_t entry) 2305 { 2306 2307 KASSERT((prev->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 || 2308 (entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0, 2309 ("vm_map_mergeable_neighbors: neither %p nor %p are mergeable", 2310 prev, entry)); 2311 return (prev->end == entry->start && 2312 prev->object.vm_object == entry->object.vm_object && 2313 (prev->object.vm_object == NULL || 2314 prev->offset + (prev->end - prev->start) == entry->offset) && 2315 prev->eflags == entry->eflags && 2316 prev->protection == entry->protection && 2317 prev->max_protection == entry->max_protection && 2318 prev->inheritance == entry->inheritance && 2319 prev->wired_count == entry->wired_count && 2320 prev->cred == entry->cred); 2321 } 2322 2323 static void 2324 vm_map_merged_neighbor_dispose(vm_map_t map, vm_map_entry_t entry) 2325 { 2326 2327 /* 2328 * If the backing object is a vnode object, vm_object_deallocate() 2329 * calls vrele(). However, vrele() does not lock the vnode because 2330 * the vnode has additional references. Thus, the map lock can be 2331 * kept without causing a lock-order reversal with the vnode lock. 2332 * 2333 * Since we count the number of virtual page mappings in 2334 * object->un_pager.vnp.writemappings, the writemappings value 2335 * should not be adjusted when the entry is disposed of. 2336 */ 2337 if (entry->object.vm_object != NULL) 2338 vm_object_deallocate(entry->object.vm_object); 2339 if (entry->cred != NULL) 2340 crfree(entry->cred); 2341 vm_map_entry_dispose(map, entry); 2342 } 2343 2344 /* 2345 * vm_map_try_merge_entries: 2346 * 2347 * Compare the given map entry to its predecessor, and merge its precessor 2348 * into it if possible. The entry remains valid, and may be extended. 2349 * The predecessor may be deleted. 2350 * 2351 * The map must be locked. 2352 */ 2353 void 2354 vm_map_try_merge_entries(vm_map_t map, vm_map_entry_t prev_entry, 2355 vm_map_entry_t entry) 2356 { 2357 2358 VM_MAP_ASSERT_LOCKED(map); 2359 if ((entry->eflags & MAP_ENTRY_NOMERGE_MASK) == 0 && 2360 vm_map_mergeable_neighbors(prev_entry, entry)) { 2361 vm_map_entry_unlink(map, prev_entry, UNLINK_MERGE_NEXT); 2362 vm_map_merged_neighbor_dispose(map, prev_entry); 2363 } 2364 } 2365 2366 /* 2367 * vm_map_entry_back: 2368 * 2369 * Allocate an object to back a map entry. 2370 */ 2371 static inline void 2372 vm_map_entry_back(vm_map_entry_t entry) 2373 { 2374 vm_object_t object; 2375 2376 KASSERT(entry->object.vm_object == NULL, 2377 ("map entry %p has backing object", entry)); 2378 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2379 ("map entry %p is a submap", entry)); 2380 object = vm_object_allocate_anon(atop(entry->end - entry->start), NULL, 2381 entry->cred, entry->end - entry->start); 2382 entry->object.vm_object = object; 2383 entry->offset = 0; 2384 entry->cred = NULL; 2385 } 2386 2387 /* 2388 * vm_map_entry_charge_object 2389 * 2390 * If there is no object backing this entry, create one. Otherwise, if 2391 * the entry has cred, give it to the backing object. 2392 */ 2393 static inline void 2394 vm_map_entry_charge_object(vm_map_t map, vm_map_entry_t entry) 2395 { 2396 2397 VM_MAP_ASSERT_LOCKED(map); 2398 KASSERT((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0, 2399 ("map entry %p is a submap", entry)); 2400 if (entry->object.vm_object == NULL && !map->system_map && 2401 (entry->eflags & MAP_ENTRY_GUARD) == 0) 2402 vm_map_entry_back(entry); 2403 else if (entry->object.vm_object != NULL && 2404 ((entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) && 2405 entry->cred != NULL) { 2406 VM_OBJECT_WLOCK(entry->object.vm_object); 2407 KASSERT(entry->object.vm_object->cred == NULL, 2408 ("OVERCOMMIT: %s: both cred e %p", __func__, entry)); 2409 entry->object.vm_object->cred = entry->cred; 2410 entry->object.vm_object->charge = entry->end - entry->start; 2411 VM_OBJECT_WUNLOCK(entry->object.vm_object); 2412 entry->cred = NULL; 2413 } 2414 } 2415 2416 /* 2417 * vm_map_entry_clone 2418 * 2419 * Create a duplicate map entry for clipping. 2420 */ 2421 static vm_map_entry_t 2422 vm_map_entry_clone(vm_map_t map, vm_map_entry_t entry) 2423 { 2424 vm_map_entry_t new_entry; 2425 2426 VM_MAP_ASSERT_LOCKED(map); 2427 2428 /* 2429 * Create a backing object now, if none exists, so that more individual 2430 * objects won't be created after the map entry is split. 2431 */ 2432 vm_map_entry_charge_object(map, entry); 2433 2434 /* Clone the entry. */ 2435 new_entry = vm_map_entry_create(map); 2436 *new_entry = *entry; 2437 if (new_entry->cred != NULL) 2438 crhold(entry->cred); 2439 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 2440 vm_object_reference(new_entry->object.vm_object); 2441 vm_map_entry_set_vnode_text(new_entry, true); 2442 /* 2443 * The object->un_pager.vnp.writemappings for the object of 2444 * MAP_ENTRY_WRITECNT type entry shall be kept as is here. The 2445 * virtual pages are re-distributed among the clipped entries, 2446 * so the sum is left the same. 2447 */ 2448 } 2449 return (new_entry); 2450 } 2451 2452 /* 2453 * vm_map_clip_start: [ internal use only ] 2454 * 2455 * Asserts that the given entry begins at or after 2456 * the specified address; if necessary, 2457 * it splits the entry into two. 2458 */ 2459 static int 2460 vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t startaddr) 2461 { 2462 vm_map_entry_t new_entry; 2463 int bdry_idx; 2464 2465 if (!map->system_map) 2466 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2467 "%s: map %p entry %p start 0x%jx", __func__, map, entry, 2468 (uintmax_t)startaddr); 2469 2470 if (startaddr <= entry->start) 2471 return (KERN_SUCCESS); 2472 2473 VM_MAP_ASSERT_LOCKED(map); 2474 KASSERT(entry->end > startaddr && entry->start < startaddr, 2475 ("%s: invalid clip of entry %p", __func__, entry)); 2476 2477 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 2478 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 2479 if (bdry_idx != 0) { 2480 if ((startaddr & (pagesizes[bdry_idx] - 1)) != 0) 2481 return (KERN_INVALID_ARGUMENT); 2482 } 2483 2484 new_entry = vm_map_entry_clone(map, entry); 2485 2486 /* 2487 * Split off the front portion. Insert the new entry BEFORE this one, 2488 * so that this entry has the specified starting address. 2489 */ 2490 new_entry->end = startaddr; 2491 vm_map_entry_link(map, new_entry); 2492 return (KERN_SUCCESS); 2493 } 2494 2495 /* 2496 * vm_map_lookup_clip_start: 2497 * 2498 * Find the entry at or just after 'start', and clip it if 'start' is in 2499 * the interior of the entry. Return entry after 'start', and in 2500 * prev_entry set the entry before 'start'. 2501 */ 2502 static int 2503 vm_map_lookup_clip_start(vm_map_t map, vm_offset_t start, 2504 vm_map_entry_t *res_entry, vm_map_entry_t *prev_entry) 2505 { 2506 vm_map_entry_t entry; 2507 int rv; 2508 2509 if (!map->system_map) 2510 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2511 "%s: map %p start 0x%jx prev %p", __func__, map, 2512 (uintmax_t)start, prev_entry); 2513 2514 if (vm_map_lookup_entry(map, start, prev_entry)) { 2515 entry = *prev_entry; 2516 rv = vm_map_clip_start(map, entry, start); 2517 if (rv != KERN_SUCCESS) 2518 return (rv); 2519 *prev_entry = vm_map_entry_pred(entry); 2520 } else 2521 entry = vm_map_entry_succ(*prev_entry); 2522 *res_entry = entry; 2523 return (KERN_SUCCESS); 2524 } 2525 2526 /* 2527 * vm_map_clip_end: [ internal use only ] 2528 * 2529 * Asserts that the given entry ends at or before 2530 * the specified address; if necessary, 2531 * it splits the entry into two. 2532 */ 2533 static int 2534 vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t endaddr) 2535 { 2536 vm_map_entry_t new_entry; 2537 int bdry_idx; 2538 2539 if (!map->system_map) 2540 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, 2541 "%s: map %p entry %p end 0x%jx", __func__, map, entry, 2542 (uintmax_t)endaddr); 2543 2544 if (endaddr >= entry->end) 2545 return (KERN_SUCCESS); 2546 2547 VM_MAP_ASSERT_LOCKED(map); 2548 KASSERT(entry->start < endaddr && entry->end > endaddr, 2549 ("%s: invalid clip of entry %p", __func__, entry)); 2550 2551 bdry_idx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 2552 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 2553 if (bdry_idx != 0) { 2554 if ((endaddr & (pagesizes[bdry_idx] - 1)) != 0) 2555 return (KERN_INVALID_ARGUMENT); 2556 } 2557 2558 new_entry = vm_map_entry_clone(map, entry); 2559 2560 /* 2561 * Split off the back portion. Insert the new entry AFTER this one, 2562 * so that this entry has the specified ending address. 2563 */ 2564 new_entry->start = endaddr; 2565 vm_map_entry_link(map, new_entry); 2566 2567 return (KERN_SUCCESS); 2568 } 2569 2570 /* 2571 * vm_map_submap: [ kernel use only ] 2572 * 2573 * Mark the given range as handled by a subordinate map. 2574 * 2575 * This range must have been created with vm_map_find, 2576 * and no other operations may have been performed on this 2577 * range prior to calling vm_map_submap. 2578 * 2579 * Only a limited number of operations can be performed 2580 * within this rage after calling vm_map_submap: 2581 * vm_fault 2582 * [Don't try vm_map_copy!] 2583 * 2584 * To remove a submapping, one must first remove the 2585 * range from the superior map, and then destroy the 2586 * submap (if desired). [Better yet, don't try it.] 2587 */ 2588 int 2589 vm_map_submap( 2590 vm_map_t map, 2591 vm_offset_t start, 2592 vm_offset_t end, 2593 vm_map_t submap) 2594 { 2595 vm_map_entry_t entry; 2596 int result; 2597 2598 result = KERN_INVALID_ARGUMENT; 2599 2600 vm_map_lock(submap); 2601 submap->flags |= MAP_IS_SUB_MAP; 2602 vm_map_unlock(submap); 2603 2604 vm_map_lock(map); 2605 VM_MAP_RANGE_CHECK(map, start, end); 2606 if (vm_map_lookup_entry(map, start, &entry) && entry->end >= end && 2607 (entry->eflags & MAP_ENTRY_COW) == 0 && 2608 entry->object.vm_object == NULL) { 2609 result = vm_map_clip_start(map, entry, start); 2610 if (result != KERN_SUCCESS) 2611 goto unlock; 2612 result = vm_map_clip_end(map, entry, end); 2613 if (result != KERN_SUCCESS) 2614 goto unlock; 2615 entry->object.sub_map = submap; 2616 entry->eflags |= MAP_ENTRY_IS_SUB_MAP; 2617 result = KERN_SUCCESS; 2618 } 2619 unlock: 2620 vm_map_unlock(map); 2621 2622 if (result != KERN_SUCCESS) { 2623 vm_map_lock(submap); 2624 submap->flags &= ~MAP_IS_SUB_MAP; 2625 vm_map_unlock(submap); 2626 } 2627 return (result); 2628 } 2629 2630 /* 2631 * The maximum number of pages to map if MAP_PREFAULT_PARTIAL is specified 2632 */ 2633 #define MAX_INIT_PT 96 2634 2635 /* 2636 * vm_map_pmap_enter: 2637 * 2638 * Preload the specified map's pmap with mappings to the specified 2639 * object's memory-resident pages. No further physical pages are 2640 * allocated, and no further virtual pages are retrieved from secondary 2641 * storage. If the specified flags include MAP_PREFAULT_PARTIAL, then a 2642 * limited number of page mappings are created at the low-end of the 2643 * specified address range. (For this purpose, a superpage mapping 2644 * counts as one page mapping.) Otherwise, all resident pages within 2645 * the specified address range are mapped. 2646 */ 2647 static void 2648 vm_map_pmap_enter(vm_map_t map, vm_offset_t addr, vm_prot_t prot, 2649 vm_object_t object, vm_pindex_t pindex, vm_size_t size, int flags) 2650 { 2651 vm_offset_t start; 2652 vm_page_t p, p_start; 2653 vm_pindex_t mask, psize, threshold, tmpidx; 2654 2655 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 || object == NULL) 2656 return; 2657 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2658 VM_OBJECT_WLOCK(object); 2659 if (object->type == OBJT_DEVICE || object->type == OBJT_SG) { 2660 pmap_object_init_pt(map->pmap, addr, object, pindex, 2661 size); 2662 VM_OBJECT_WUNLOCK(object); 2663 return; 2664 } 2665 VM_OBJECT_LOCK_DOWNGRADE(object); 2666 } else 2667 VM_OBJECT_RLOCK(object); 2668 2669 psize = atop(size); 2670 if (psize + pindex > object->size) { 2671 if (pindex >= object->size) { 2672 VM_OBJECT_RUNLOCK(object); 2673 return; 2674 } 2675 psize = object->size - pindex; 2676 } 2677 2678 start = 0; 2679 p_start = NULL; 2680 threshold = MAX_INIT_PT; 2681 2682 p = vm_page_find_least(object, pindex); 2683 /* 2684 * Assert: the variable p is either (1) the page with the 2685 * least pindex greater than or equal to the parameter pindex 2686 * or (2) NULL. 2687 */ 2688 for (; 2689 p != NULL && (tmpidx = p->pindex - pindex) < psize; 2690 p = TAILQ_NEXT(p, listq)) { 2691 /* 2692 * don't allow an madvise to blow away our really 2693 * free pages allocating pv entries. 2694 */ 2695 if (((flags & MAP_PREFAULT_MADVISE) != 0 && 2696 vm_page_count_severe()) || 2697 ((flags & MAP_PREFAULT_PARTIAL) != 0 && 2698 tmpidx >= threshold)) { 2699 psize = tmpidx; 2700 break; 2701 } 2702 if (vm_page_all_valid(p)) { 2703 if (p_start == NULL) { 2704 start = addr + ptoa(tmpidx); 2705 p_start = p; 2706 } 2707 /* Jump ahead if a superpage mapping is possible. */ 2708 if (p->psind > 0 && ((addr + ptoa(tmpidx)) & 2709 (pagesizes[p->psind] - 1)) == 0) { 2710 mask = atop(pagesizes[p->psind]) - 1; 2711 if (tmpidx + mask < psize && 2712 vm_page_ps_test(p, PS_ALL_VALID, NULL)) { 2713 p += mask; 2714 threshold += mask; 2715 } 2716 } 2717 } else if (p_start != NULL) { 2718 pmap_enter_object(map->pmap, start, addr + 2719 ptoa(tmpidx), p_start, prot); 2720 p_start = NULL; 2721 } 2722 } 2723 if (p_start != NULL) 2724 pmap_enter_object(map->pmap, start, addr + ptoa(psize), 2725 p_start, prot); 2726 VM_OBJECT_RUNLOCK(object); 2727 } 2728 2729 /* 2730 * vm_map_protect: 2731 * 2732 * Sets the protection of the specified address 2733 * region in the target map. If "set_max" is 2734 * specified, the maximum protection is to be set; 2735 * otherwise, only the current protection is affected. 2736 */ 2737 int 2738 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 2739 vm_prot_t new_prot, boolean_t set_max) 2740 { 2741 vm_map_entry_t entry, first_entry, in_tran, prev_entry; 2742 vm_object_t obj; 2743 struct ucred *cred; 2744 vm_prot_t old_prot; 2745 int rv; 2746 2747 if (start == end) 2748 return (KERN_SUCCESS); 2749 2750 again: 2751 in_tran = NULL; 2752 vm_map_lock(map); 2753 2754 /* 2755 * Ensure that we are not concurrently wiring pages. vm_map_wire() may 2756 * need to fault pages into the map and will drop the map lock while 2757 * doing so, and the VM object may end up in an inconsistent state if we 2758 * update the protection on the map entry in between faults. 2759 */ 2760 vm_map_wait_busy(map); 2761 2762 VM_MAP_RANGE_CHECK(map, start, end); 2763 2764 if (!vm_map_lookup_entry(map, start, &first_entry)) 2765 first_entry = vm_map_entry_succ(first_entry); 2766 2767 /* 2768 * Make a first pass to check for protection violations. 2769 */ 2770 for (entry = first_entry; entry->start < end; 2771 entry = vm_map_entry_succ(entry)) { 2772 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 2773 continue; 2774 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 2775 vm_map_unlock(map); 2776 return (KERN_INVALID_ARGUMENT); 2777 } 2778 if ((new_prot & entry->max_protection) != new_prot) { 2779 vm_map_unlock(map); 2780 return (KERN_PROTECTION_FAILURE); 2781 } 2782 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0) 2783 in_tran = entry; 2784 } 2785 2786 /* 2787 * Postpone the operation until all in-transition map entries have 2788 * stabilized. An in-transition entry might already have its pages 2789 * wired and wired_count incremented, but not yet have its 2790 * MAP_ENTRY_USER_WIRED flag set. In which case, we would fail to call 2791 * vm_fault_copy_entry() in the final loop below. 2792 */ 2793 if (in_tran != NULL) { 2794 in_tran->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2795 vm_map_unlock_and_wait(map, 0); 2796 goto again; 2797 } 2798 2799 /* 2800 * Before changing the protections, try to reserve swap space for any 2801 * private (i.e., copy-on-write) mappings that are transitioning from 2802 * read-only to read/write access. If a reservation fails, break out 2803 * of this loop early and let the next loop simplify the entries, since 2804 * some may now be mergeable. 2805 */ 2806 rv = vm_map_clip_start(map, first_entry, start); 2807 if (rv != KERN_SUCCESS) { 2808 vm_map_unlock(map); 2809 return (rv); 2810 } 2811 for (entry = first_entry; entry->start < end; 2812 entry = vm_map_entry_succ(entry)) { 2813 rv = vm_map_clip_end(map, entry, end); 2814 if (rv != KERN_SUCCESS) { 2815 vm_map_unlock(map); 2816 return (rv); 2817 } 2818 2819 if (set_max || 2820 ((new_prot & ~entry->protection) & VM_PROT_WRITE) == 0 || 2821 ENTRY_CHARGED(entry) || 2822 (entry->eflags & MAP_ENTRY_GUARD) != 0) { 2823 continue; 2824 } 2825 2826 cred = curthread->td_ucred; 2827 obj = entry->object.vm_object; 2828 2829 if (obj == NULL || 2830 (entry->eflags & MAP_ENTRY_NEEDS_COPY) != 0) { 2831 if (!swap_reserve(entry->end - entry->start)) { 2832 rv = KERN_RESOURCE_SHORTAGE; 2833 end = entry->end; 2834 break; 2835 } 2836 crhold(cred); 2837 entry->cred = cred; 2838 continue; 2839 } 2840 2841 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) 2842 continue; 2843 VM_OBJECT_WLOCK(obj); 2844 if (obj->type != OBJT_DEFAULT && obj->type != OBJT_SWAP) { 2845 VM_OBJECT_WUNLOCK(obj); 2846 continue; 2847 } 2848 2849 /* 2850 * Charge for the whole object allocation now, since 2851 * we cannot distinguish between non-charged and 2852 * charged clipped mapping of the same object later. 2853 */ 2854 KASSERT(obj->charge == 0, 2855 ("vm_map_protect: object %p overcharged (entry %p)", 2856 obj, entry)); 2857 if (!swap_reserve(ptoa(obj->size))) { 2858 VM_OBJECT_WUNLOCK(obj); 2859 rv = KERN_RESOURCE_SHORTAGE; 2860 end = entry->end; 2861 break; 2862 } 2863 2864 crhold(cred); 2865 obj->cred = cred; 2866 obj->charge = ptoa(obj->size); 2867 VM_OBJECT_WUNLOCK(obj); 2868 } 2869 2870 /* 2871 * If enough swap space was available, go back and fix up protections. 2872 * Otherwise, just simplify entries, since some may have been modified. 2873 * [Note that clipping is not necessary the second time.] 2874 */ 2875 for (prev_entry = vm_map_entry_pred(first_entry), entry = first_entry; 2876 entry->start < end; 2877 vm_map_try_merge_entries(map, prev_entry, entry), 2878 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 2879 if (rv != KERN_SUCCESS || 2880 (entry->eflags & MAP_ENTRY_GUARD) != 0) 2881 continue; 2882 2883 old_prot = entry->protection; 2884 2885 if (set_max) 2886 entry->protection = 2887 (entry->max_protection = new_prot) & 2888 old_prot; 2889 else 2890 entry->protection = new_prot; 2891 2892 /* 2893 * For user wired map entries, the normal lazy evaluation of 2894 * write access upgrades through soft page faults is 2895 * undesirable. Instead, immediately copy any pages that are 2896 * copy-on-write and enable write access in the physical map. 2897 */ 2898 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0 && 2899 (entry->protection & VM_PROT_WRITE) != 0 && 2900 (old_prot & VM_PROT_WRITE) == 0) 2901 vm_fault_copy_entry(map, map, entry, entry, NULL); 2902 2903 /* 2904 * When restricting access, update the physical map. Worry 2905 * about copy-on-write here. 2906 */ 2907 if ((old_prot & ~entry->protection) != 0) { 2908 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 2909 VM_PROT_ALL) 2910 pmap_protect(map->pmap, entry->start, 2911 entry->end, 2912 entry->protection & MASK(entry)); 2913 #undef MASK 2914 } 2915 } 2916 vm_map_try_merge_entries(map, prev_entry, entry); 2917 vm_map_unlock(map); 2918 return (rv); 2919 } 2920 2921 /* 2922 * vm_map_madvise: 2923 * 2924 * This routine traverses a processes map handling the madvise 2925 * system call. Advisories are classified as either those effecting 2926 * the vm_map_entry structure, or those effecting the underlying 2927 * objects. 2928 */ 2929 int 2930 vm_map_madvise( 2931 vm_map_t map, 2932 vm_offset_t start, 2933 vm_offset_t end, 2934 int behav) 2935 { 2936 vm_map_entry_t entry, prev_entry; 2937 int rv; 2938 bool modify_map; 2939 2940 /* 2941 * Some madvise calls directly modify the vm_map_entry, in which case 2942 * we need to use an exclusive lock on the map and we need to perform 2943 * various clipping operations. Otherwise we only need a read-lock 2944 * on the map. 2945 */ 2946 switch(behav) { 2947 case MADV_NORMAL: 2948 case MADV_SEQUENTIAL: 2949 case MADV_RANDOM: 2950 case MADV_NOSYNC: 2951 case MADV_AUTOSYNC: 2952 case MADV_NOCORE: 2953 case MADV_CORE: 2954 if (start == end) 2955 return (0); 2956 modify_map = true; 2957 vm_map_lock(map); 2958 break; 2959 case MADV_WILLNEED: 2960 case MADV_DONTNEED: 2961 case MADV_FREE: 2962 if (start == end) 2963 return (0); 2964 modify_map = false; 2965 vm_map_lock_read(map); 2966 break; 2967 default: 2968 return (EINVAL); 2969 } 2970 2971 /* 2972 * Locate starting entry and clip if necessary. 2973 */ 2974 VM_MAP_RANGE_CHECK(map, start, end); 2975 2976 if (modify_map) { 2977 /* 2978 * madvise behaviors that are implemented in the vm_map_entry. 2979 * 2980 * We clip the vm_map_entry so that behavioral changes are 2981 * limited to the specified address range. 2982 */ 2983 rv = vm_map_lookup_clip_start(map, start, &entry, &prev_entry); 2984 if (rv != KERN_SUCCESS) { 2985 vm_map_unlock(map); 2986 return (vm_mmap_to_errno(rv)); 2987 } 2988 2989 for (; entry->start < end; prev_entry = entry, 2990 entry = vm_map_entry_succ(entry)) { 2991 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 2992 continue; 2993 2994 rv = vm_map_clip_end(map, entry, end); 2995 if (rv != KERN_SUCCESS) { 2996 vm_map_unlock(map); 2997 return (vm_mmap_to_errno(rv)); 2998 } 2999 3000 switch (behav) { 3001 case MADV_NORMAL: 3002 vm_map_entry_set_behavior(entry, 3003 MAP_ENTRY_BEHAV_NORMAL); 3004 break; 3005 case MADV_SEQUENTIAL: 3006 vm_map_entry_set_behavior(entry, 3007 MAP_ENTRY_BEHAV_SEQUENTIAL); 3008 break; 3009 case MADV_RANDOM: 3010 vm_map_entry_set_behavior(entry, 3011 MAP_ENTRY_BEHAV_RANDOM); 3012 break; 3013 case MADV_NOSYNC: 3014 entry->eflags |= MAP_ENTRY_NOSYNC; 3015 break; 3016 case MADV_AUTOSYNC: 3017 entry->eflags &= ~MAP_ENTRY_NOSYNC; 3018 break; 3019 case MADV_NOCORE: 3020 entry->eflags |= MAP_ENTRY_NOCOREDUMP; 3021 break; 3022 case MADV_CORE: 3023 entry->eflags &= ~MAP_ENTRY_NOCOREDUMP; 3024 break; 3025 default: 3026 break; 3027 } 3028 vm_map_try_merge_entries(map, prev_entry, entry); 3029 } 3030 vm_map_try_merge_entries(map, prev_entry, entry); 3031 vm_map_unlock(map); 3032 } else { 3033 vm_pindex_t pstart, pend; 3034 3035 /* 3036 * madvise behaviors that are implemented in the underlying 3037 * vm_object. 3038 * 3039 * Since we don't clip the vm_map_entry, we have to clip 3040 * the vm_object pindex and count. 3041 */ 3042 if (!vm_map_lookup_entry(map, start, &entry)) 3043 entry = vm_map_entry_succ(entry); 3044 for (; entry->start < end; 3045 entry = vm_map_entry_succ(entry)) { 3046 vm_offset_t useEnd, useStart; 3047 3048 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 3049 continue; 3050 3051 /* 3052 * MADV_FREE would otherwise rewind time to 3053 * the creation of the shadow object. Because 3054 * we hold the VM map read-locked, neither the 3055 * entry's object nor the presence of a 3056 * backing object can change. 3057 */ 3058 if (behav == MADV_FREE && 3059 entry->object.vm_object != NULL && 3060 entry->object.vm_object->backing_object != NULL) 3061 continue; 3062 3063 pstart = OFF_TO_IDX(entry->offset); 3064 pend = pstart + atop(entry->end - entry->start); 3065 useStart = entry->start; 3066 useEnd = entry->end; 3067 3068 if (entry->start < start) { 3069 pstart += atop(start - entry->start); 3070 useStart = start; 3071 } 3072 if (entry->end > end) { 3073 pend -= atop(entry->end - end); 3074 useEnd = end; 3075 } 3076 3077 if (pstart >= pend) 3078 continue; 3079 3080 /* 3081 * Perform the pmap_advise() before clearing 3082 * PGA_REFERENCED in vm_page_advise(). Otherwise, a 3083 * concurrent pmap operation, such as pmap_remove(), 3084 * could clear a reference in the pmap and set 3085 * PGA_REFERENCED on the page before the pmap_advise() 3086 * had completed. Consequently, the page would appear 3087 * referenced based upon an old reference that 3088 * occurred before this pmap_advise() ran. 3089 */ 3090 if (behav == MADV_DONTNEED || behav == MADV_FREE) 3091 pmap_advise(map->pmap, useStart, useEnd, 3092 behav); 3093 3094 vm_object_madvise(entry->object.vm_object, pstart, 3095 pend, behav); 3096 3097 /* 3098 * Pre-populate paging structures in the 3099 * WILLNEED case. For wired entries, the 3100 * paging structures are already populated. 3101 */ 3102 if (behav == MADV_WILLNEED && 3103 entry->wired_count == 0) { 3104 vm_map_pmap_enter(map, 3105 useStart, 3106 entry->protection, 3107 entry->object.vm_object, 3108 pstart, 3109 ptoa(pend - pstart), 3110 MAP_PREFAULT_MADVISE 3111 ); 3112 } 3113 } 3114 vm_map_unlock_read(map); 3115 } 3116 return (0); 3117 } 3118 3119 /* 3120 * vm_map_inherit: 3121 * 3122 * Sets the inheritance of the specified address 3123 * range in the target map. Inheritance 3124 * affects how the map will be shared with 3125 * child maps at the time of vmspace_fork. 3126 */ 3127 int 3128 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 3129 vm_inherit_t new_inheritance) 3130 { 3131 vm_map_entry_t entry, lentry, prev_entry, start_entry; 3132 int rv; 3133 3134 switch (new_inheritance) { 3135 case VM_INHERIT_NONE: 3136 case VM_INHERIT_COPY: 3137 case VM_INHERIT_SHARE: 3138 case VM_INHERIT_ZERO: 3139 break; 3140 default: 3141 return (KERN_INVALID_ARGUMENT); 3142 } 3143 if (start == end) 3144 return (KERN_SUCCESS); 3145 vm_map_lock(map); 3146 VM_MAP_RANGE_CHECK(map, start, end); 3147 rv = vm_map_lookup_clip_start(map, start, &start_entry, &prev_entry); 3148 if (rv != KERN_SUCCESS) 3149 goto unlock; 3150 if (vm_map_lookup_entry(map, end - 1, &lentry)) { 3151 rv = vm_map_clip_end(map, lentry, end); 3152 if (rv != KERN_SUCCESS) 3153 goto unlock; 3154 } 3155 if (new_inheritance == VM_INHERIT_COPY) { 3156 for (entry = start_entry; entry->start < end; 3157 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3158 if ((entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) 3159 != 0) { 3160 rv = KERN_INVALID_ARGUMENT; 3161 goto unlock; 3162 } 3163 } 3164 } 3165 for (entry = start_entry; entry->start < end; prev_entry = entry, 3166 entry = vm_map_entry_succ(entry)) { 3167 KASSERT(entry->end <= end, ("non-clipped entry %p end %jx %jx", 3168 entry, (uintmax_t)entry->end, (uintmax_t)end)); 3169 if ((entry->eflags & MAP_ENTRY_GUARD) == 0 || 3170 new_inheritance != VM_INHERIT_ZERO) 3171 entry->inheritance = new_inheritance; 3172 vm_map_try_merge_entries(map, prev_entry, entry); 3173 } 3174 vm_map_try_merge_entries(map, prev_entry, entry); 3175 unlock: 3176 vm_map_unlock(map); 3177 return (rv); 3178 } 3179 3180 /* 3181 * vm_map_entry_in_transition: 3182 * 3183 * Release the map lock, and sleep until the entry is no longer in 3184 * transition. Awake and acquire the map lock. If the map changed while 3185 * another held the lock, lookup a possibly-changed entry at or after the 3186 * 'start' position of the old entry. 3187 */ 3188 static vm_map_entry_t 3189 vm_map_entry_in_transition(vm_map_t map, vm_offset_t in_start, 3190 vm_offset_t *io_end, bool holes_ok, vm_map_entry_t in_entry) 3191 { 3192 vm_map_entry_t entry; 3193 vm_offset_t start; 3194 u_int last_timestamp; 3195 3196 VM_MAP_ASSERT_LOCKED(map); 3197 KASSERT((in_entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3198 ("not in-tranition map entry %p", in_entry)); 3199 /* 3200 * We have not yet clipped the entry. 3201 */ 3202 start = MAX(in_start, in_entry->start); 3203 in_entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3204 last_timestamp = map->timestamp; 3205 if (vm_map_unlock_and_wait(map, 0)) { 3206 /* 3207 * Allow interruption of user wiring/unwiring? 3208 */ 3209 } 3210 vm_map_lock(map); 3211 if (last_timestamp + 1 == map->timestamp) 3212 return (in_entry); 3213 3214 /* 3215 * Look again for the entry because the map was modified while it was 3216 * unlocked. Specifically, the entry may have been clipped, merged, or 3217 * deleted. 3218 */ 3219 if (!vm_map_lookup_entry(map, start, &entry)) { 3220 if (!holes_ok) { 3221 *io_end = start; 3222 return (NULL); 3223 } 3224 entry = vm_map_entry_succ(entry); 3225 } 3226 return (entry); 3227 } 3228 3229 /* 3230 * vm_map_unwire: 3231 * 3232 * Implements both kernel and user unwiring. 3233 */ 3234 int 3235 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t end, 3236 int flags) 3237 { 3238 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3239 int rv; 3240 bool holes_ok, need_wakeup, user_unwire; 3241 3242 if (start == end) 3243 return (KERN_SUCCESS); 3244 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3245 user_unwire = (flags & VM_MAP_WIRE_USER) != 0; 3246 vm_map_lock(map); 3247 VM_MAP_RANGE_CHECK(map, start, end); 3248 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3249 if (holes_ok) 3250 first_entry = vm_map_entry_succ(first_entry); 3251 else { 3252 vm_map_unlock(map); 3253 return (KERN_INVALID_ADDRESS); 3254 } 3255 } 3256 rv = KERN_SUCCESS; 3257 for (entry = first_entry; entry->start < end; entry = next_entry) { 3258 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3259 /* 3260 * We have not yet clipped the entry. 3261 */ 3262 next_entry = vm_map_entry_in_transition(map, start, 3263 &end, holes_ok, entry); 3264 if (next_entry == NULL) { 3265 if (entry == first_entry) { 3266 vm_map_unlock(map); 3267 return (KERN_INVALID_ADDRESS); 3268 } 3269 rv = KERN_INVALID_ADDRESS; 3270 break; 3271 } 3272 first_entry = (entry == first_entry) ? 3273 next_entry : NULL; 3274 continue; 3275 } 3276 rv = vm_map_clip_start(map, entry, start); 3277 if (rv != KERN_SUCCESS) 3278 break; 3279 rv = vm_map_clip_end(map, entry, end); 3280 if (rv != KERN_SUCCESS) 3281 break; 3282 3283 /* 3284 * Mark the entry in case the map lock is released. (See 3285 * above.) 3286 */ 3287 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3288 entry->wiring_thread == NULL, 3289 ("owned map entry %p", entry)); 3290 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3291 entry->wiring_thread = curthread; 3292 next_entry = vm_map_entry_succ(entry); 3293 /* 3294 * Check the map for holes in the specified region. 3295 * If holes_ok, skip this check. 3296 */ 3297 if (!holes_ok && 3298 entry->end < end && next_entry->start > entry->end) { 3299 end = entry->end; 3300 rv = KERN_INVALID_ADDRESS; 3301 break; 3302 } 3303 /* 3304 * If system unwiring, require that the entry is system wired. 3305 */ 3306 if (!user_unwire && 3307 vm_map_entry_system_wired_count(entry) == 0) { 3308 end = entry->end; 3309 rv = KERN_INVALID_ARGUMENT; 3310 break; 3311 } 3312 } 3313 need_wakeup = false; 3314 if (first_entry == NULL && 3315 !vm_map_lookup_entry(map, start, &first_entry)) { 3316 KASSERT(holes_ok, ("vm_map_unwire: lookup failed")); 3317 prev_entry = first_entry; 3318 entry = vm_map_entry_succ(first_entry); 3319 } else { 3320 prev_entry = vm_map_entry_pred(first_entry); 3321 entry = first_entry; 3322 } 3323 for (; entry->start < end; 3324 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3325 /* 3326 * If holes_ok was specified, an empty 3327 * space in the unwired region could have been mapped 3328 * while the map lock was dropped for draining 3329 * MAP_ENTRY_IN_TRANSITION. Moreover, another thread 3330 * could be simultaneously wiring this new mapping 3331 * entry. Detect these cases and skip any entries 3332 * marked as in transition by us. 3333 */ 3334 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3335 entry->wiring_thread != curthread) { 3336 KASSERT(holes_ok, 3337 ("vm_map_unwire: !HOLESOK and new/changed entry")); 3338 continue; 3339 } 3340 3341 if (rv == KERN_SUCCESS && (!user_unwire || 3342 (entry->eflags & MAP_ENTRY_USER_WIRED))) { 3343 if (entry->wired_count == 1) 3344 vm_map_entry_unwire(map, entry); 3345 else 3346 entry->wired_count--; 3347 if (user_unwire) 3348 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3349 } 3350 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3351 ("vm_map_unwire: in-transition flag missing %p", entry)); 3352 KASSERT(entry->wiring_thread == curthread, 3353 ("vm_map_unwire: alien wire %p", entry)); 3354 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 3355 entry->wiring_thread = NULL; 3356 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3357 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3358 need_wakeup = true; 3359 } 3360 vm_map_try_merge_entries(map, prev_entry, entry); 3361 } 3362 vm_map_try_merge_entries(map, prev_entry, entry); 3363 vm_map_unlock(map); 3364 if (need_wakeup) 3365 vm_map_wakeup(map); 3366 return (rv); 3367 } 3368 3369 static void 3370 vm_map_wire_user_count_sub(u_long npages) 3371 { 3372 3373 atomic_subtract_long(&vm_user_wire_count, npages); 3374 } 3375 3376 static bool 3377 vm_map_wire_user_count_add(u_long npages) 3378 { 3379 u_long wired; 3380 3381 wired = vm_user_wire_count; 3382 do { 3383 if (npages + wired > vm_page_max_user_wired) 3384 return (false); 3385 } while (!atomic_fcmpset_long(&vm_user_wire_count, &wired, 3386 npages + wired)); 3387 3388 return (true); 3389 } 3390 3391 /* 3392 * vm_map_wire_entry_failure: 3393 * 3394 * Handle a wiring failure on the given entry. 3395 * 3396 * The map should be locked. 3397 */ 3398 static void 3399 vm_map_wire_entry_failure(vm_map_t map, vm_map_entry_t entry, 3400 vm_offset_t failed_addr) 3401 { 3402 3403 VM_MAP_ASSERT_LOCKED(map); 3404 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 && 3405 entry->wired_count == 1, 3406 ("vm_map_wire_entry_failure: entry %p isn't being wired", entry)); 3407 KASSERT(failed_addr < entry->end, 3408 ("vm_map_wire_entry_failure: entry %p was fully wired", entry)); 3409 3410 /* 3411 * If any pages at the start of this entry were successfully wired, 3412 * then unwire them. 3413 */ 3414 if (failed_addr > entry->start) { 3415 pmap_unwire(map->pmap, entry->start, failed_addr); 3416 vm_object_unwire(entry->object.vm_object, entry->offset, 3417 failed_addr - entry->start, PQ_ACTIVE); 3418 } 3419 3420 /* 3421 * Assign an out-of-range value to represent the failure to wire this 3422 * entry. 3423 */ 3424 entry->wired_count = -1; 3425 } 3426 3427 int 3428 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3429 { 3430 int rv; 3431 3432 vm_map_lock(map); 3433 rv = vm_map_wire_locked(map, start, end, flags); 3434 vm_map_unlock(map); 3435 return (rv); 3436 } 3437 3438 /* 3439 * vm_map_wire_locked: 3440 * 3441 * Implements both kernel and user wiring. Returns with the map locked, 3442 * the map lock may be dropped. 3443 */ 3444 int 3445 vm_map_wire_locked(vm_map_t map, vm_offset_t start, vm_offset_t end, int flags) 3446 { 3447 vm_map_entry_t entry, first_entry, next_entry, prev_entry; 3448 vm_offset_t faddr, saved_end, saved_start; 3449 u_long incr, npages; 3450 u_int bidx, last_timestamp; 3451 int rv; 3452 bool holes_ok, need_wakeup, user_wire; 3453 vm_prot_t prot; 3454 3455 VM_MAP_ASSERT_LOCKED(map); 3456 3457 if (start == end) 3458 return (KERN_SUCCESS); 3459 prot = 0; 3460 if (flags & VM_MAP_WIRE_WRITE) 3461 prot |= VM_PROT_WRITE; 3462 holes_ok = (flags & VM_MAP_WIRE_HOLESOK) != 0; 3463 user_wire = (flags & VM_MAP_WIRE_USER) != 0; 3464 VM_MAP_RANGE_CHECK(map, start, end); 3465 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3466 if (holes_ok) 3467 first_entry = vm_map_entry_succ(first_entry); 3468 else 3469 return (KERN_INVALID_ADDRESS); 3470 } 3471 for (entry = first_entry; entry->start < end; entry = next_entry) { 3472 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 3473 /* 3474 * We have not yet clipped the entry. 3475 */ 3476 next_entry = vm_map_entry_in_transition(map, start, 3477 &end, holes_ok, entry); 3478 if (next_entry == NULL) { 3479 if (entry == first_entry) 3480 return (KERN_INVALID_ADDRESS); 3481 rv = KERN_INVALID_ADDRESS; 3482 goto done; 3483 } 3484 first_entry = (entry == first_entry) ? 3485 next_entry : NULL; 3486 continue; 3487 } 3488 rv = vm_map_clip_start(map, entry, start); 3489 if (rv != KERN_SUCCESS) 3490 goto done; 3491 rv = vm_map_clip_end(map, entry, end); 3492 if (rv != KERN_SUCCESS) 3493 goto done; 3494 3495 /* 3496 * Mark the entry in case the map lock is released. (See 3497 * above.) 3498 */ 3499 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 && 3500 entry->wiring_thread == NULL, 3501 ("owned map entry %p", entry)); 3502 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 3503 entry->wiring_thread = curthread; 3504 if ((entry->protection & (VM_PROT_READ | VM_PROT_EXECUTE)) == 0 3505 || (entry->protection & prot) != prot) { 3506 entry->eflags |= MAP_ENTRY_WIRE_SKIPPED; 3507 if (!holes_ok) { 3508 end = entry->end; 3509 rv = KERN_INVALID_ADDRESS; 3510 goto done; 3511 } 3512 } else if (entry->wired_count == 0) { 3513 entry->wired_count++; 3514 3515 npages = atop(entry->end - entry->start); 3516 if (user_wire && !vm_map_wire_user_count_add(npages)) { 3517 vm_map_wire_entry_failure(map, entry, 3518 entry->start); 3519 end = entry->end; 3520 rv = KERN_RESOURCE_SHORTAGE; 3521 goto done; 3522 } 3523 3524 /* 3525 * Release the map lock, relying on the in-transition 3526 * mark. Mark the map busy for fork. 3527 */ 3528 saved_start = entry->start; 3529 saved_end = entry->end; 3530 last_timestamp = map->timestamp; 3531 bidx = (entry->eflags & MAP_ENTRY_SPLIT_BOUNDARY_MASK) 3532 >> MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 3533 incr = pagesizes[bidx]; 3534 vm_map_busy(map); 3535 vm_map_unlock(map); 3536 3537 for (faddr = saved_start; faddr < saved_end; 3538 faddr += incr) { 3539 /* 3540 * Simulate a fault to get the page and enter 3541 * it into the physical map. 3542 */ 3543 rv = vm_fault(map, faddr, VM_PROT_NONE, 3544 VM_FAULT_WIRE, NULL); 3545 if (rv != KERN_SUCCESS) 3546 break; 3547 } 3548 vm_map_lock(map); 3549 vm_map_unbusy(map); 3550 if (last_timestamp + 1 != map->timestamp) { 3551 /* 3552 * Look again for the entry because the map was 3553 * modified while it was unlocked. The entry 3554 * may have been clipped, but NOT merged or 3555 * deleted. 3556 */ 3557 if (!vm_map_lookup_entry(map, saved_start, 3558 &next_entry)) 3559 KASSERT(false, 3560 ("vm_map_wire: lookup failed")); 3561 first_entry = (entry == first_entry) ? 3562 next_entry : NULL; 3563 for (entry = next_entry; entry->end < saved_end; 3564 entry = vm_map_entry_succ(entry)) { 3565 /* 3566 * In case of failure, handle entries 3567 * that were not fully wired here; 3568 * fully wired entries are handled 3569 * later. 3570 */ 3571 if (rv != KERN_SUCCESS && 3572 faddr < entry->end) 3573 vm_map_wire_entry_failure(map, 3574 entry, faddr); 3575 } 3576 } 3577 if (rv != KERN_SUCCESS) { 3578 vm_map_wire_entry_failure(map, entry, faddr); 3579 if (user_wire) 3580 vm_map_wire_user_count_sub(npages); 3581 end = entry->end; 3582 goto done; 3583 } 3584 } else if (!user_wire || 3585 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3586 entry->wired_count++; 3587 } 3588 /* 3589 * Check the map for holes in the specified region. 3590 * If holes_ok was specified, skip this check. 3591 */ 3592 next_entry = vm_map_entry_succ(entry); 3593 if (!holes_ok && 3594 entry->end < end && next_entry->start > entry->end) { 3595 end = entry->end; 3596 rv = KERN_INVALID_ADDRESS; 3597 goto done; 3598 } 3599 } 3600 rv = KERN_SUCCESS; 3601 done: 3602 need_wakeup = false; 3603 if (first_entry == NULL && 3604 !vm_map_lookup_entry(map, start, &first_entry)) { 3605 KASSERT(holes_ok, ("vm_map_wire: lookup failed")); 3606 prev_entry = first_entry; 3607 entry = vm_map_entry_succ(first_entry); 3608 } else { 3609 prev_entry = vm_map_entry_pred(first_entry); 3610 entry = first_entry; 3611 } 3612 for (; entry->start < end; 3613 prev_entry = entry, entry = vm_map_entry_succ(entry)) { 3614 /* 3615 * If holes_ok was specified, an empty 3616 * space in the unwired region could have been mapped 3617 * while the map lock was dropped for faulting in the 3618 * pages or draining MAP_ENTRY_IN_TRANSITION. 3619 * Moreover, another thread could be simultaneously 3620 * wiring this new mapping entry. Detect these cases 3621 * and skip any entries marked as in transition not by us. 3622 * 3623 * Another way to get an entry not marked with 3624 * MAP_ENTRY_IN_TRANSITION is after failed clipping, 3625 * which set rv to KERN_INVALID_ARGUMENT. 3626 */ 3627 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) == 0 || 3628 entry->wiring_thread != curthread) { 3629 KASSERT(holes_ok || rv == KERN_INVALID_ARGUMENT, 3630 ("vm_map_wire: !HOLESOK and new/changed entry")); 3631 continue; 3632 } 3633 3634 if ((entry->eflags & MAP_ENTRY_WIRE_SKIPPED) != 0) { 3635 /* do nothing */ 3636 } else if (rv == KERN_SUCCESS) { 3637 if (user_wire) 3638 entry->eflags |= MAP_ENTRY_USER_WIRED; 3639 } else if (entry->wired_count == -1) { 3640 /* 3641 * Wiring failed on this entry. Thus, unwiring is 3642 * unnecessary. 3643 */ 3644 entry->wired_count = 0; 3645 } else if (!user_wire || 3646 (entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 3647 /* 3648 * Undo the wiring. Wiring succeeded on this entry 3649 * but failed on a later entry. 3650 */ 3651 if (entry->wired_count == 1) { 3652 vm_map_entry_unwire(map, entry); 3653 if (user_wire) 3654 vm_map_wire_user_count_sub( 3655 atop(entry->end - entry->start)); 3656 } else 3657 entry->wired_count--; 3658 } 3659 KASSERT((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0, 3660 ("vm_map_wire: in-transition flag missing %p", entry)); 3661 KASSERT(entry->wiring_thread == curthread, 3662 ("vm_map_wire: alien wire %p", entry)); 3663 entry->eflags &= ~(MAP_ENTRY_IN_TRANSITION | 3664 MAP_ENTRY_WIRE_SKIPPED); 3665 entry->wiring_thread = NULL; 3666 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 3667 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 3668 need_wakeup = true; 3669 } 3670 vm_map_try_merge_entries(map, prev_entry, entry); 3671 } 3672 vm_map_try_merge_entries(map, prev_entry, entry); 3673 if (need_wakeup) 3674 vm_map_wakeup(map); 3675 return (rv); 3676 } 3677 3678 /* 3679 * vm_map_sync 3680 * 3681 * Push any dirty cached pages in the address range to their pager. 3682 * If syncio is TRUE, dirty pages are written synchronously. 3683 * If invalidate is TRUE, any cached pages are freed as well. 3684 * 3685 * If the size of the region from start to end is zero, we are 3686 * supposed to flush all modified pages within the region containing 3687 * start. Unfortunately, a region can be split or coalesced with 3688 * neighboring regions, making it difficult to determine what the 3689 * original region was. Therefore, we approximate this requirement by 3690 * flushing the current region containing start. 3691 * 3692 * Returns an error if any part of the specified range is not mapped. 3693 */ 3694 int 3695 vm_map_sync( 3696 vm_map_t map, 3697 vm_offset_t start, 3698 vm_offset_t end, 3699 boolean_t syncio, 3700 boolean_t invalidate) 3701 { 3702 vm_map_entry_t entry, first_entry, next_entry; 3703 vm_size_t size; 3704 vm_object_t object; 3705 vm_ooffset_t offset; 3706 unsigned int last_timestamp; 3707 int bdry_idx; 3708 boolean_t failed; 3709 3710 vm_map_lock_read(map); 3711 VM_MAP_RANGE_CHECK(map, start, end); 3712 if (!vm_map_lookup_entry(map, start, &first_entry)) { 3713 vm_map_unlock_read(map); 3714 return (KERN_INVALID_ADDRESS); 3715 } else if (start == end) { 3716 start = first_entry->start; 3717 end = first_entry->end; 3718 } 3719 3720 /* 3721 * Make a first pass to check for user-wired memory, holes, 3722 * and partial invalidation of largepage mappings. 3723 */ 3724 for (entry = first_entry; entry->start < end; entry = next_entry) { 3725 if (invalidate) { 3726 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) { 3727 vm_map_unlock_read(map); 3728 return (KERN_INVALID_ARGUMENT); 3729 } 3730 bdry_idx = (entry->eflags & 3731 MAP_ENTRY_SPLIT_BOUNDARY_MASK) >> 3732 MAP_ENTRY_SPLIT_BOUNDARY_SHIFT; 3733 if (bdry_idx != 0 && 3734 ((start & (pagesizes[bdry_idx] - 1)) != 0 || 3735 (end & (pagesizes[bdry_idx] - 1)) != 0)) { 3736 vm_map_unlock_read(map); 3737 return (KERN_INVALID_ARGUMENT); 3738 } 3739 } 3740 next_entry = vm_map_entry_succ(entry); 3741 if (end > entry->end && 3742 entry->end != next_entry->start) { 3743 vm_map_unlock_read(map); 3744 return (KERN_INVALID_ADDRESS); 3745 } 3746 } 3747 3748 if (invalidate) 3749 pmap_remove(map->pmap, start, end); 3750 failed = FALSE; 3751 3752 /* 3753 * Make a second pass, cleaning/uncaching pages from the indicated 3754 * objects as we go. 3755 */ 3756 for (entry = first_entry; entry->start < end;) { 3757 offset = entry->offset + (start - entry->start); 3758 size = (end <= entry->end ? end : entry->end) - start; 3759 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) { 3760 vm_map_t smap; 3761 vm_map_entry_t tentry; 3762 vm_size_t tsize; 3763 3764 smap = entry->object.sub_map; 3765 vm_map_lock_read(smap); 3766 (void) vm_map_lookup_entry(smap, offset, &tentry); 3767 tsize = tentry->end - offset; 3768 if (tsize < size) 3769 size = tsize; 3770 object = tentry->object.vm_object; 3771 offset = tentry->offset + (offset - tentry->start); 3772 vm_map_unlock_read(smap); 3773 } else { 3774 object = entry->object.vm_object; 3775 } 3776 vm_object_reference(object); 3777 last_timestamp = map->timestamp; 3778 vm_map_unlock_read(map); 3779 if (!vm_object_sync(object, offset, size, syncio, invalidate)) 3780 failed = TRUE; 3781 start += size; 3782 vm_object_deallocate(object); 3783 vm_map_lock_read(map); 3784 if (last_timestamp == map->timestamp || 3785 !vm_map_lookup_entry(map, start, &entry)) 3786 entry = vm_map_entry_succ(entry); 3787 } 3788 3789 vm_map_unlock_read(map); 3790 return (failed ? KERN_FAILURE : KERN_SUCCESS); 3791 } 3792 3793 /* 3794 * vm_map_entry_unwire: [ internal use only ] 3795 * 3796 * Make the region specified by this entry pageable. 3797 * 3798 * The map in question should be locked. 3799 * [This is the reason for this routine's existence.] 3800 */ 3801 static void 3802 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 3803 { 3804 vm_size_t size; 3805 3806 VM_MAP_ASSERT_LOCKED(map); 3807 KASSERT(entry->wired_count > 0, 3808 ("vm_map_entry_unwire: entry %p isn't wired", entry)); 3809 3810 size = entry->end - entry->start; 3811 if ((entry->eflags & MAP_ENTRY_USER_WIRED) != 0) 3812 vm_map_wire_user_count_sub(atop(size)); 3813 pmap_unwire(map->pmap, entry->start, entry->end); 3814 vm_object_unwire(entry->object.vm_object, entry->offset, size, 3815 PQ_ACTIVE); 3816 entry->wired_count = 0; 3817 } 3818 3819 static void 3820 vm_map_entry_deallocate(vm_map_entry_t entry, boolean_t system_map) 3821 { 3822 3823 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) 3824 vm_object_deallocate(entry->object.vm_object); 3825 uma_zfree(system_map ? kmapentzone : mapentzone, entry); 3826 } 3827 3828 /* 3829 * vm_map_entry_delete: [ internal use only ] 3830 * 3831 * Deallocate the given entry from the target map. 3832 */ 3833 static void 3834 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry) 3835 { 3836 vm_object_t object; 3837 vm_pindex_t offidxstart, offidxend, size1; 3838 vm_size_t size; 3839 3840 vm_map_entry_unlink(map, entry, UNLINK_MERGE_NONE); 3841 object = entry->object.vm_object; 3842 3843 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) { 3844 MPASS(entry->cred == NULL); 3845 MPASS((entry->eflags & MAP_ENTRY_IS_SUB_MAP) == 0); 3846 MPASS(object == NULL); 3847 vm_map_entry_deallocate(entry, map->system_map); 3848 return; 3849 } 3850 3851 size = entry->end - entry->start; 3852 map->size -= size; 3853 3854 if (entry->cred != NULL) { 3855 swap_release_by_cred(size, entry->cred); 3856 crfree(entry->cred); 3857 } 3858 3859 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || object == NULL) { 3860 entry->object.vm_object = NULL; 3861 } else if ((object->flags & OBJ_ANON) != 0 || 3862 object == kernel_object) { 3863 KASSERT(entry->cred == NULL || object->cred == NULL || 3864 (entry->eflags & MAP_ENTRY_NEEDS_COPY), 3865 ("OVERCOMMIT vm_map_entry_delete: both cred %p", entry)); 3866 offidxstart = OFF_TO_IDX(entry->offset); 3867 offidxend = offidxstart + atop(size); 3868 VM_OBJECT_WLOCK(object); 3869 if (object->ref_count != 1 && 3870 ((object->flags & OBJ_ONEMAPPING) != 0 || 3871 object == kernel_object)) { 3872 vm_object_collapse(object); 3873 3874 /* 3875 * The option OBJPR_NOTMAPPED can be passed here 3876 * because vm_map_delete() already performed 3877 * pmap_remove() on the only mapping to this range 3878 * of pages. 3879 */ 3880 vm_object_page_remove(object, offidxstart, offidxend, 3881 OBJPR_NOTMAPPED); 3882 if (offidxend >= object->size && 3883 offidxstart < object->size) { 3884 size1 = object->size; 3885 object->size = offidxstart; 3886 if (object->cred != NULL) { 3887 size1 -= object->size; 3888 KASSERT(object->charge >= ptoa(size1), 3889 ("object %p charge < 0", object)); 3890 swap_release_by_cred(ptoa(size1), 3891 object->cred); 3892 object->charge -= ptoa(size1); 3893 } 3894 } 3895 } 3896 VM_OBJECT_WUNLOCK(object); 3897 } 3898 if (map->system_map) 3899 vm_map_entry_deallocate(entry, TRUE); 3900 else { 3901 entry->defer_next = curthread->td_map_def_user; 3902 curthread->td_map_def_user = entry; 3903 } 3904 } 3905 3906 /* 3907 * vm_map_delete: [ internal use only ] 3908 * 3909 * Deallocates the given address range from the target 3910 * map. 3911 */ 3912 int 3913 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end) 3914 { 3915 vm_map_entry_t entry, next_entry, scratch_entry; 3916 int rv; 3917 3918 VM_MAP_ASSERT_LOCKED(map); 3919 3920 if (start == end) 3921 return (KERN_SUCCESS); 3922 3923 /* 3924 * Find the start of the region, and clip it. 3925 * Step through all entries in this region. 3926 */ 3927 rv = vm_map_lookup_clip_start(map, start, &entry, &scratch_entry); 3928 if (rv != KERN_SUCCESS) 3929 return (rv); 3930 for (; entry->start < end; entry = next_entry) { 3931 /* 3932 * Wait for wiring or unwiring of an entry to complete. 3933 * Also wait for any system wirings to disappear on 3934 * user maps. 3935 */ 3936 if ((entry->eflags & MAP_ENTRY_IN_TRANSITION) != 0 || 3937 (vm_map_pmap(map) != kernel_pmap && 3938 vm_map_entry_system_wired_count(entry) != 0)) { 3939 unsigned int last_timestamp; 3940 vm_offset_t saved_start; 3941 3942 saved_start = entry->start; 3943 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 3944 last_timestamp = map->timestamp; 3945 (void) vm_map_unlock_and_wait(map, 0); 3946 vm_map_lock(map); 3947 if (last_timestamp + 1 != map->timestamp) { 3948 /* 3949 * Look again for the entry because the map was 3950 * modified while it was unlocked. 3951 * Specifically, the entry may have been 3952 * clipped, merged, or deleted. 3953 */ 3954 rv = vm_map_lookup_clip_start(map, saved_start, 3955 &next_entry, &scratch_entry); 3956 if (rv != KERN_SUCCESS) 3957 break; 3958 } else 3959 next_entry = entry; 3960 continue; 3961 } 3962 3963 /* XXXKIB or delete to the upper superpage boundary ? */ 3964 rv = vm_map_clip_end(map, entry, end); 3965 if (rv != KERN_SUCCESS) 3966 break; 3967 next_entry = vm_map_entry_succ(entry); 3968 3969 /* 3970 * Unwire before removing addresses from the pmap; otherwise, 3971 * unwiring will put the entries back in the pmap. 3972 */ 3973 if (entry->wired_count != 0) 3974 vm_map_entry_unwire(map, entry); 3975 3976 /* 3977 * Remove mappings for the pages, but only if the 3978 * mappings could exist. For instance, it does not 3979 * make sense to call pmap_remove() for guard entries. 3980 */ 3981 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 3982 entry->object.vm_object != NULL) 3983 pmap_remove(map->pmap, entry->start, entry->end); 3984 3985 if (entry->end == map->anon_loc) 3986 map->anon_loc = entry->start; 3987 3988 /* 3989 * Delete the entry only after removing all pmap 3990 * entries pointing to its pages. (Otherwise, its 3991 * page frames may be reallocated, and any modify bits 3992 * will be set in the wrong object!) 3993 */ 3994 vm_map_entry_delete(map, entry); 3995 } 3996 return (rv); 3997 } 3998 3999 /* 4000 * vm_map_remove: 4001 * 4002 * Remove the given address range from the target map. 4003 * This is the exported form of vm_map_delete. 4004 */ 4005 int 4006 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 4007 { 4008 int result; 4009 4010 vm_map_lock(map); 4011 VM_MAP_RANGE_CHECK(map, start, end); 4012 result = vm_map_delete(map, start, end); 4013 vm_map_unlock(map); 4014 return (result); 4015 } 4016 4017 /* 4018 * vm_map_check_protection: 4019 * 4020 * Assert that the target map allows the specified privilege on the 4021 * entire address region given. The entire region must be allocated. 4022 * 4023 * WARNING! This code does not and should not check whether the 4024 * contents of the region is accessible. For example a smaller file 4025 * might be mapped into a larger address space. 4026 * 4027 * NOTE! This code is also called by munmap(). 4028 * 4029 * The map must be locked. A read lock is sufficient. 4030 */ 4031 boolean_t 4032 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 4033 vm_prot_t protection) 4034 { 4035 vm_map_entry_t entry; 4036 vm_map_entry_t tmp_entry; 4037 4038 if (!vm_map_lookup_entry(map, start, &tmp_entry)) 4039 return (FALSE); 4040 entry = tmp_entry; 4041 4042 while (start < end) { 4043 /* 4044 * No holes allowed! 4045 */ 4046 if (start < entry->start) 4047 return (FALSE); 4048 /* 4049 * Check protection associated with entry. 4050 */ 4051 if ((entry->protection & protection) != protection) 4052 return (FALSE); 4053 /* go to next entry */ 4054 start = entry->end; 4055 entry = vm_map_entry_succ(entry); 4056 } 4057 return (TRUE); 4058 } 4059 4060 /* 4061 * 4062 * vm_map_copy_swap_object: 4063 * 4064 * Copies a swap-backed object from an existing map entry to a 4065 * new one. Carries forward the swap charge. May change the 4066 * src object on return. 4067 */ 4068 static void 4069 vm_map_copy_swap_object(vm_map_entry_t src_entry, vm_map_entry_t dst_entry, 4070 vm_offset_t size, vm_ooffset_t *fork_charge) 4071 { 4072 vm_object_t src_object; 4073 struct ucred *cred; 4074 int charged; 4075 4076 src_object = src_entry->object.vm_object; 4077 charged = ENTRY_CHARGED(src_entry); 4078 if ((src_object->flags & OBJ_ANON) != 0) { 4079 VM_OBJECT_WLOCK(src_object); 4080 vm_object_collapse(src_object); 4081 if ((src_object->flags & OBJ_ONEMAPPING) != 0) { 4082 vm_object_split(src_entry); 4083 src_object = src_entry->object.vm_object; 4084 } 4085 vm_object_reference_locked(src_object); 4086 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 4087 VM_OBJECT_WUNLOCK(src_object); 4088 } else 4089 vm_object_reference(src_object); 4090 if (src_entry->cred != NULL && 4091 !(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 4092 KASSERT(src_object->cred == NULL, 4093 ("OVERCOMMIT: vm_map_copy_anon_entry: cred %p", 4094 src_object)); 4095 src_object->cred = src_entry->cred; 4096 src_object->charge = size; 4097 } 4098 dst_entry->object.vm_object = src_object; 4099 if (charged) { 4100 cred = curthread->td_ucred; 4101 crhold(cred); 4102 dst_entry->cred = cred; 4103 *fork_charge += size; 4104 if (!(src_entry->eflags & MAP_ENTRY_NEEDS_COPY)) { 4105 crhold(cred); 4106 src_entry->cred = cred; 4107 *fork_charge += size; 4108 } 4109 } 4110 } 4111 4112 /* 4113 * vm_map_copy_entry: 4114 * 4115 * Copies the contents of the source entry to the destination 4116 * entry. The entries *must* be aligned properly. 4117 */ 4118 static void 4119 vm_map_copy_entry( 4120 vm_map_t src_map, 4121 vm_map_t dst_map, 4122 vm_map_entry_t src_entry, 4123 vm_map_entry_t dst_entry, 4124 vm_ooffset_t *fork_charge) 4125 { 4126 vm_object_t src_object; 4127 vm_map_entry_t fake_entry; 4128 vm_offset_t size; 4129 4130 VM_MAP_ASSERT_LOCKED(dst_map); 4131 4132 if ((dst_entry->eflags|src_entry->eflags) & MAP_ENTRY_IS_SUB_MAP) 4133 return; 4134 4135 if (src_entry->wired_count == 0 || 4136 (src_entry->protection & VM_PROT_WRITE) == 0) { 4137 /* 4138 * If the source entry is marked needs_copy, it is already 4139 * write-protected. 4140 */ 4141 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0 && 4142 (src_entry->protection & VM_PROT_WRITE) != 0) { 4143 pmap_protect(src_map->pmap, 4144 src_entry->start, 4145 src_entry->end, 4146 src_entry->protection & ~VM_PROT_WRITE); 4147 } 4148 4149 /* 4150 * Make a copy of the object. 4151 */ 4152 size = src_entry->end - src_entry->start; 4153 if ((src_object = src_entry->object.vm_object) != NULL) { 4154 if (src_object->type == OBJT_DEFAULT || 4155 src_object->type == OBJT_SWAP) { 4156 vm_map_copy_swap_object(src_entry, dst_entry, 4157 size, fork_charge); 4158 /* May have split/collapsed, reload obj. */ 4159 src_object = src_entry->object.vm_object; 4160 } else { 4161 vm_object_reference(src_object); 4162 dst_entry->object.vm_object = src_object; 4163 } 4164 src_entry->eflags |= MAP_ENTRY_COW | 4165 MAP_ENTRY_NEEDS_COPY; 4166 dst_entry->eflags |= MAP_ENTRY_COW | 4167 MAP_ENTRY_NEEDS_COPY; 4168 dst_entry->offset = src_entry->offset; 4169 if (src_entry->eflags & MAP_ENTRY_WRITECNT) { 4170 /* 4171 * MAP_ENTRY_WRITECNT cannot 4172 * indicate write reference from 4173 * src_entry, since the entry is 4174 * marked as needs copy. Allocate a 4175 * fake entry that is used to 4176 * decrement object->un_pager writecount 4177 * at the appropriate time. Attach 4178 * fake_entry to the deferred list. 4179 */ 4180 fake_entry = vm_map_entry_create(dst_map); 4181 fake_entry->eflags = MAP_ENTRY_WRITECNT; 4182 src_entry->eflags &= ~MAP_ENTRY_WRITECNT; 4183 vm_object_reference(src_object); 4184 fake_entry->object.vm_object = src_object; 4185 fake_entry->start = src_entry->start; 4186 fake_entry->end = src_entry->end; 4187 fake_entry->defer_next = 4188 curthread->td_map_def_user; 4189 curthread->td_map_def_user = fake_entry; 4190 } 4191 4192 pmap_copy(dst_map->pmap, src_map->pmap, 4193 dst_entry->start, dst_entry->end - dst_entry->start, 4194 src_entry->start); 4195 } else { 4196 dst_entry->object.vm_object = NULL; 4197 dst_entry->offset = 0; 4198 if (src_entry->cred != NULL) { 4199 dst_entry->cred = curthread->td_ucred; 4200 crhold(dst_entry->cred); 4201 *fork_charge += size; 4202 } 4203 } 4204 } else { 4205 /* 4206 * We don't want to make writeable wired pages copy-on-write. 4207 * Immediately copy these pages into the new map by simulating 4208 * page faults. The new pages are pageable. 4209 */ 4210 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry, 4211 fork_charge); 4212 } 4213 } 4214 4215 /* 4216 * vmspace_map_entry_forked: 4217 * Update the newly-forked vmspace each time a map entry is inherited 4218 * or copied. The values for vm_dsize and vm_tsize are approximate 4219 * (and mostly-obsolete ideas in the face of mmap(2) et al.) 4220 */ 4221 static void 4222 vmspace_map_entry_forked(const struct vmspace *vm1, struct vmspace *vm2, 4223 vm_map_entry_t entry) 4224 { 4225 vm_size_t entrysize; 4226 vm_offset_t newend; 4227 4228 if ((entry->eflags & MAP_ENTRY_GUARD) != 0) 4229 return; 4230 entrysize = entry->end - entry->start; 4231 vm2->vm_map.size += entrysize; 4232 if (entry->eflags & (MAP_ENTRY_GROWS_DOWN | MAP_ENTRY_GROWS_UP)) { 4233 vm2->vm_ssize += btoc(entrysize); 4234 } else if (entry->start >= (vm_offset_t)vm1->vm_daddr && 4235 entry->start < (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)) { 4236 newend = MIN(entry->end, 4237 (vm_offset_t)vm1->vm_daddr + ctob(vm1->vm_dsize)); 4238 vm2->vm_dsize += btoc(newend - entry->start); 4239 } else if (entry->start >= (vm_offset_t)vm1->vm_taddr && 4240 entry->start < (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)) { 4241 newend = MIN(entry->end, 4242 (vm_offset_t)vm1->vm_taddr + ctob(vm1->vm_tsize)); 4243 vm2->vm_tsize += btoc(newend - entry->start); 4244 } 4245 } 4246 4247 /* 4248 * vmspace_fork: 4249 * Create a new process vmspace structure and vm_map 4250 * based on those of an existing process. The new map 4251 * is based on the old map, according to the inheritance 4252 * values on the regions in that map. 4253 * 4254 * XXX It might be worth coalescing the entries added to the new vmspace. 4255 * 4256 * The source map must not be locked. 4257 */ 4258 struct vmspace * 4259 vmspace_fork(struct vmspace *vm1, vm_ooffset_t *fork_charge) 4260 { 4261 struct vmspace *vm2; 4262 vm_map_t new_map, old_map; 4263 vm_map_entry_t new_entry, old_entry; 4264 vm_object_t object; 4265 int error, locked; 4266 vm_inherit_t inh; 4267 4268 old_map = &vm1->vm_map; 4269 /* Copy immutable fields of vm1 to vm2. */ 4270 vm2 = vmspace_alloc(vm_map_min(old_map), vm_map_max(old_map), 4271 pmap_pinit); 4272 if (vm2 == NULL) 4273 return (NULL); 4274 4275 vm2->vm_taddr = vm1->vm_taddr; 4276 vm2->vm_daddr = vm1->vm_daddr; 4277 vm2->vm_maxsaddr = vm1->vm_maxsaddr; 4278 vm_map_lock(old_map); 4279 if (old_map->busy) 4280 vm_map_wait_busy(old_map); 4281 new_map = &vm2->vm_map; 4282 locked = vm_map_trylock(new_map); /* trylock to silence WITNESS */ 4283 KASSERT(locked, ("vmspace_fork: lock failed")); 4284 4285 error = pmap_vmspace_copy(new_map->pmap, old_map->pmap); 4286 if (error != 0) { 4287 sx_xunlock(&old_map->lock); 4288 sx_xunlock(&new_map->lock); 4289 vm_map_process_deferred(); 4290 vmspace_free(vm2); 4291 return (NULL); 4292 } 4293 4294 new_map->anon_loc = old_map->anon_loc; 4295 new_map->flags |= old_map->flags & (MAP_ASLR | MAP_ASLR_IGNSTART); 4296 4297 VM_MAP_ENTRY_FOREACH(old_entry, old_map) { 4298 if ((old_entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0) 4299 panic("vm_map_fork: encountered a submap"); 4300 4301 inh = old_entry->inheritance; 4302 if ((old_entry->eflags & MAP_ENTRY_GUARD) != 0 && 4303 inh != VM_INHERIT_NONE) 4304 inh = VM_INHERIT_COPY; 4305 4306 switch (inh) { 4307 case VM_INHERIT_NONE: 4308 break; 4309 4310 case VM_INHERIT_SHARE: 4311 /* 4312 * Clone the entry, creating the shared object if 4313 * necessary. 4314 */ 4315 object = old_entry->object.vm_object; 4316 if (object == NULL) { 4317 vm_map_entry_back(old_entry); 4318 object = old_entry->object.vm_object; 4319 } 4320 4321 /* 4322 * Add the reference before calling vm_object_shadow 4323 * to insure that a shadow object is created. 4324 */ 4325 vm_object_reference(object); 4326 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 4327 vm_object_shadow(&old_entry->object.vm_object, 4328 &old_entry->offset, 4329 old_entry->end - old_entry->start, 4330 old_entry->cred, 4331 /* Transfer the second reference too. */ 4332 true); 4333 old_entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 4334 old_entry->cred = NULL; 4335 4336 /* 4337 * As in vm_map_merged_neighbor_dispose(), 4338 * the vnode lock will not be acquired in 4339 * this call to vm_object_deallocate(). 4340 */ 4341 vm_object_deallocate(object); 4342 object = old_entry->object.vm_object; 4343 } else { 4344 VM_OBJECT_WLOCK(object); 4345 vm_object_clear_flag(object, OBJ_ONEMAPPING); 4346 if (old_entry->cred != NULL) { 4347 KASSERT(object->cred == NULL, 4348 ("vmspace_fork both cred")); 4349 object->cred = old_entry->cred; 4350 object->charge = old_entry->end - 4351 old_entry->start; 4352 old_entry->cred = NULL; 4353 } 4354 4355 /* 4356 * Assert the correct state of the vnode 4357 * v_writecount while the object is locked, to 4358 * not relock it later for the assertion 4359 * correctness. 4360 */ 4361 if (old_entry->eflags & MAP_ENTRY_WRITECNT && 4362 object->type == OBJT_VNODE) { 4363 KASSERT(((struct vnode *)object-> 4364 handle)->v_writecount > 0, 4365 ("vmspace_fork: v_writecount %p", 4366 object)); 4367 KASSERT(object->un_pager.vnp. 4368 writemappings > 0, 4369 ("vmspace_fork: vnp.writecount %p", 4370 object)); 4371 } 4372 VM_OBJECT_WUNLOCK(object); 4373 } 4374 4375 /* 4376 * Clone the entry, referencing the shared object. 4377 */ 4378 new_entry = vm_map_entry_create(new_map); 4379 *new_entry = *old_entry; 4380 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4381 MAP_ENTRY_IN_TRANSITION); 4382 new_entry->wiring_thread = NULL; 4383 new_entry->wired_count = 0; 4384 if (new_entry->eflags & MAP_ENTRY_WRITECNT) { 4385 vm_pager_update_writecount(object, 4386 new_entry->start, new_entry->end); 4387 } 4388 vm_map_entry_set_vnode_text(new_entry, true); 4389 4390 /* 4391 * Insert the entry into the new map -- we know we're 4392 * inserting at the end of the new map. 4393 */ 4394 vm_map_entry_link(new_map, new_entry); 4395 vmspace_map_entry_forked(vm1, vm2, new_entry); 4396 4397 /* 4398 * Update the physical map 4399 */ 4400 pmap_copy(new_map->pmap, old_map->pmap, 4401 new_entry->start, 4402 (old_entry->end - old_entry->start), 4403 old_entry->start); 4404 break; 4405 4406 case VM_INHERIT_COPY: 4407 /* 4408 * Clone the entry and link into the map. 4409 */ 4410 new_entry = vm_map_entry_create(new_map); 4411 *new_entry = *old_entry; 4412 /* 4413 * Copied entry is COW over the old object. 4414 */ 4415 new_entry->eflags &= ~(MAP_ENTRY_USER_WIRED | 4416 MAP_ENTRY_IN_TRANSITION | MAP_ENTRY_WRITECNT); 4417 new_entry->wiring_thread = NULL; 4418 new_entry->wired_count = 0; 4419 new_entry->object.vm_object = NULL; 4420 new_entry->cred = NULL; 4421 vm_map_entry_link(new_map, new_entry); 4422 vmspace_map_entry_forked(vm1, vm2, new_entry); 4423 vm_map_copy_entry(old_map, new_map, old_entry, 4424 new_entry, fork_charge); 4425 vm_map_entry_set_vnode_text(new_entry, true); 4426 break; 4427 4428 case VM_INHERIT_ZERO: 4429 /* 4430 * Create a new anonymous mapping entry modelled from 4431 * the old one. 4432 */ 4433 new_entry = vm_map_entry_create(new_map); 4434 memset(new_entry, 0, sizeof(*new_entry)); 4435 4436 new_entry->start = old_entry->start; 4437 new_entry->end = old_entry->end; 4438 new_entry->eflags = old_entry->eflags & 4439 ~(MAP_ENTRY_USER_WIRED | MAP_ENTRY_IN_TRANSITION | 4440 MAP_ENTRY_WRITECNT | MAP_ENTRY_VN_EXEC | 4441 MAP_ENTRY_SPLIT_BOUNDARY_MASK); 4442 new_entry->protection = old_entry->protection; 4443 new_entry->max_protection = old_entry->max_protection; 4444 new_entry->inheritance = VM_INHERIT_ZERO; 4445 4446 vm_map_entry_link(new_map, new_entry); 4447 vmspace_map_entry_forked(vm1, vm2, new_entry); 4448 4449 new_entry->cred = curthread->td_ucred; 4450 crhold(new_entry->cred); 4451 *fork_charge += (new_entry->end - new_entry->start); 4452 4453 break; 4454 } 4455 } 4456 /* 4457 * Use inlined vm_map_unlock() to postpone handling the deferred 4458 * map entries, which cannot be done until both old_map and 4459 * new_map locks are released. 4460 */ 4461 sx_xunlock(&old_map->lock); 4462 sx_xunlock(&new_map->lock); 4463 vm_map_process_deferred(); 4464 4465 return (vm2); 4466 } 4467 4468 /* 4469 * Create a process's stack for exec_new_vmspace(). This function is never 4470 * asked to wire the newly created stack. 4471 */ 4472 int 4473 vm_map_stack(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4474 vm_prot_t prot, vm_prot_t max, int cow) 4475 { 4476 vm_size_t growsize, init_ssize; 4477 rlim_t vmemlim; 4478 int rv; 4479 4480 MPASS((map->flags & MAP_WIREFUTURE) == 0); 4481 growsize = sgrowsiz; 4482 init_ssize = (max_ssize < growsize) ? max_ssize : growsize; 4483 vm_map_lock(map); 4484 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4485 /* If we would blow our VMEM resource limit, no go */ 4486 if (map->size + init_ssize > vmemlim) { 4487 rv = KERN_NO_SPACE; 4488 goto out; 4489 } 4490 rv = vm_map_stack_locked(map, addrbos, max_ssize, growsize, prot, 4491 max, cow); 4492 out: 4493 vm_map_unlock(map); 4494 return (rv); 4495 } 4496 4497 static int stack_guard_page = 1; 4498 SYSCTL_INT(_security_bsd, OID_AUTO, stack_guard_page, CTLFLAG_RWTUN, 4499 &stack_guard_page, 0, 4500 "Specifies the number of guard pages for a stack that grows"); 4501 4502 static int 4503 vm_map_stack_locked(vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 4504 vm_size_t growsize, vm_prot_t prot, vm_prot_t max, int cow) 4505 { 4506 vm_map_entry_t new_entry, prev_entry; 4507 vm_offset_t bot, gap_bot, gap_top, top; 4508 vm_size_t init_ssize, sgp; 4509 int orient, rv; 4510 4511 /* 4512 * The stack orientation is piggybacked with the cow argument. 4513 * Extract it into orient and mask the cow argument so that we 4514 * don't pass it around further. 4515 */ 4516 orient = cow & (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP); 4517 KASSERT(orient != 0, ("No stack grow direction")); 4518 KASSERT(orient != (MAP_STACK_GROWS_DOWN | MAP_STACK_GROWS_UP), 4519 ("bi-dir stack")); 4520 4521 if (max_ssize == 0 || 4522 !vm_map_range_valid(map, addrbos, addrbos + max_ssize)) 4523 return (KERN_INVALID_ADDRESS); 4524 sgp = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4525 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4526 (vm_size_t)stack_guard_page * PAGE_SIZE; 4527 if (sgp >= max_ssize) 4528 return (KERN_INVALID_ARGUMENT); 4529 4530 init_ssize = growsize; 4531 if (max_ssize < init_ssize + sgp) 4532 init_ssize = max_ssize - sgp; 4533 4534 /* If addr is already mapped, no go */ 4535 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) 4536 return (KERN_NO_SPACE); 4537 4538 /* 4539 * If we can't accommodate max_ssize in the current mapping, no go. 4540 */ 4541 if (vm_map_entry_succ(prev_entry)->start < addrbos + max_ssize) 4542 return (KERN_NO_SPACE); 4543 4544 /* 4545 * We initially map a stack of only init_ssize. We will grow as 4546 * needed later. Depending on the orientation of the stack (i.e. 4547 * the grow direction) we either map at the top of the range, the 4548 * bottom of the range or in the middle. 4549 * 4550 * Note: we would normally expect prot and max to be VM_PROT_ALL, 4551 * and cow to be 0. Possibly we should eliminate these as input 4552 * parameters, and just pass these values here in the insert call. 4553 */ 4554 if (orient == MAP_STACK_GROWS_DOWN) { 4555 bot = addrbos + max_ssize - init_ssize; 4556 top = bot + init_ssize; 4557 gap_bot = addrbos; 4558 gap_top = bot; 4559 } else /* if (orient == MAP_STACK_GROWS_UP) */ { 4560 bot = addrbos; 4561 top = bot + init_ssize; 4562 gap_bot = top; 4563 gap_top = addrbos + max_ssize; 4564 } 4565 rv = vm_map_insert(map, NULL, 0, bot, top, prot, max, cow); 4566 if (rv != KERN_SUCCESS) 4567 return (rv); 4568 new_entry = vm_map_entry_succ(prev_entry); 4569 KASSERT(new_entry->end == top || new_entry->start == bot, 4570 ("Bad entry start/end for new stack entry")); 4571 KASSERT((orient & MAP_STACK_GROWS_DOWN) == 0 || 4572 (new_entry->eflags & MAP_ENTRY_GROWS_DOWN) != 0, 4573 ("new entry lacks MAP_ENTRY_GROWS_DOWN")); 4574 KASSERT((orient & MAP_STACK_GROWS_UP) == 0 || 4575 (new_entry->eflags & MAP_ENTRY_GROWS_UP) != 0, 4576 ("new entry lacks MAP_ENTRY_GROWS_UP")); 4577 if (gap_bot == gap_top) 4578 return (KERN_SUCCESS); 4579 rv = vm_map_insert(map, NULL, 0, gap_bot, gap_top, VM_PROT_NONE, 4580 VM_PROT_NONE, MAP_CREATE_GUARD | (orient == MAP_STACK_GROWS_DOWN ? 4581 MAP_CREATE_STACK_GAP_DN : MAP_CREATE_STACK_GAP_UP)); 4582 if (rv == KERN_SUCCESS) { 4583 /* 4584 * Gap can never successfully handle a fault, so 4585 * read-ahead logic is never used for it. Re-use 4586 * next_read of the gap entry to store 4587 * stack_guard_page for vm_map_growstack(). 4588 */ 4589 if (orient == MAP_STACK_GROWS_DOWN) 4590 vm_map_entry_pred(new_entry)->next_read = sgp; 4591 else 4592 vm_map_entry_succ(new_entry)->next_read = sgp; 4593 } else { 4594 (void)vm_map_delete(map, bot, top); 4595 } 4596 return (rv); 4597 } 4598 4599 /* 4600 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if we 4601 * successfully grow the stack. 4602 */ 4603 static int 4604 vm_map_growstack(vm_map_t map, vm_offset_t addr, vm_map_entry_t gap_entry) 4605 { 4606 vm_map_entry_t stack_entry; 4607 struct proc *p; 4608 struct vmspace *vm; 4609 struct ucred *cred; 4610 vm_offset_t gap_end, gap_start, grow_start; 4611 vm_size_t grow_amount, guard, max_grow; 4612 rlim_t lmemlim, stacklim, vmemlim; 4613 int rv, rv1; 4614 bool gap_deleted, grow_down, is_procstack; 4615 #ifdef notyet 4616 uint64_t limit; 4617 #endif 4618 #ifdef RACCT 4619 int error; 4620 #endif 4621 4622 p = curproc; 4623 vm = p->p_vmspace; 4624 4625 /* 4626 * Disallow stack growth when the access is performed by a 4627 * debugger or AIO daemon. The reason is that the wrong 4628 * resource limits are applied. 4629 */ 4630 if (p != initproc && (map != &p->p_vmspace->vm_map || 4631 p->p_textvp == NULL)) 4632 return (KERN_FAILURE); 4633 4634 MPASS(!map->system_map); 4635 4636 lmemlim = lim_cur(curthread, RLIMIT_MEMLOCK); 4637 stacklim = lim_cur(curthread, RLIMIT_STACK); 4638 vmemlim = lim_cur(curthread, RLIMIT_VMEM); 4639 retry: 4640 /* If addr is not in a hole for a stack grow area, no need to grow. */ 4641 if (gap_entry == NULL && !vm_map_lookup_entry(map, addr, &gap_entry)) 4642 return (KERN_FAILURE); 4643 if ((gap_entry->eflags & MAP_ENTRY_GUARD) == 0) 4644 return (KERN_SUCCESS); 4645 if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_DN) != 0) { 4646 stack_entry = vm_map_entry_succ(gap_entry); 4647 if ((stack_entry->eflags & MAP_ENTRY_GROWS_DOWN) == 0 || 4648 stack_entry->start != gap_entry->end) 4649 return (KERN_FAILURE); 4650 grow_amount = round_page(stack_entry->start - addr); 4651 grow_down = true; 4652 } else if ((gap_entry->eflags & MAP_ENTRY_STACK_GAP_UP) != 0) { 4653 stack_entry = vm_map_entry_pred(gap_entry); 4654 if ((stack_entry->eflags & MAP_ENTRY_GROWS_UP) == 0 || 4655 stack_entry->end != gap_entry->start) 4656 return (KERN_FAILURE); 4657 grow_amount = round_page(addr + 1 - stack_entry->end); 4658 grow_down = false; 4659 } else { 4660 return (KERN_FAILURE); 4661 } 4662 guard = ((curproc->p_flag2 & P2_STKGAP_DISABLE) != 0 || 4663 (curproc->p_fctl0 & NT_FREEBSD_FCTL_STKGAP_DISABLE) != 0) ? 0 : 4664 gap_entry->next_read; 4665 max_grow = gap_entry->end - gap_entry->start; 4666 if (guard > max_grow) 4667 return (KERN_NO_SPACE); 4668 max_grow -= guard; 4669 if (grow_amount > max_grow) 4670 return (KERN_NO_SPACE); 4671 4672 /* 4673 * If this is the main process stack, see if we're over the stack 4674 * limit. 4675 */ 4676 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr && 4677 addr < (vm_offset_t)p->p_sysent->sv_usrstack; 4678 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) 4679 return (KERN_NO_SPACE); 4680 4681 #ifdef RACCT 4682 if (racct_enable) { 4683 PROC_LOCK(p); 4684 if (is_procstack && racct_set(p, RACCT_STACK, 4685 ctob(vm->vm_ssize) + grow_amount)) { 4686 PROC_UNLOCK(p); 4687 return (KERN_NO_SPACE); 4688 } 4689 PROC_UNLOCK(p); 4690 } 4691 #endif 4692 4693 grow_amount = roundup(grow_amount, sgrowsiz); 4694 if (grow_amount > max_grow) 4695 grow_amount = max_grow; 4696 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > stacklim)) { 4697 grow_amount = trunc_page((vm_size_t)stacklim) - 4698 ctob(vm->vm_ssize); 4699 } 4700 4701 #ifdef notyet 4702 PROC_LOCK(p); 4703 limit = racct_get_available(p, RACCT_STACK); 4704 PROC_UNLOCK(p); 4705 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > limit)) 4706 grow_amount = limit - ctob(vm->vm_ssize); 4707 #endif 4708 4709 if (!old_mlock && (map->flags & MAP_WIREFUTURE) != 0) { 4710 if (ptoa(pmap_wired_count(map->pmap)) + grow_amount > lmemlim) { 4711 rv = KERN_NO_SPACE; 4712 goto out; 4713 } 4714 #ifdef RACCT 4715 if (racct_enable) { 4716 PROC_LOCK(p); 4717 if (racct_set(p, RACCT_MEMLOCK, 4718 ptoa(pmap_wired_count(map->pmap)) + grow_amount)) { 4719 PROC_UNLOCK(p); 4720 rv = KERN_NO_SPACE; 4721 goto out; 4722 } 4723 PROC_UNLOCK(p); 4724 } 4725 #endif 4726 } 4727 4728 /* If we would blow our VMEM resource limit, no go */ 4729 if (map->size + grow_amount > vmemlim) { 4730 rv = KERN_NO_SPACE; 4731 goto out; 4732 } 4733 #ifdef RACCT 4734 if (racct_enable) { 4735 PROC_LOCK(p); 4736 if (racct_set(p, RACCT_VMEM, map->size + grow_amount)) { 4737 PROC_UNLOCK(p); 4738 rv = KERN_NO_SPACE; 4739 goto out; 4740 } 4741 PROC_UNLOCK(p); 4742 } 4743 #endif 4744 4745 if (vm_map_lock_upgrade(map)) { 4746 gap_entry = NULL; 4747 vm_map_lock_read(map); 4748 goto retry; 4749 } 4750 4751 if (grow_down) { 4752 grow_start = gap_entry->end - grow_amount; 4753 if (gap_entry->start + grow_amount == gap_entry->end) { 4754 gap_start = gap_entry->start; 4755 gap_end = gap_entry->end; 4756 vm_map_entry_delete(map, gap_entry); 4757 gap_deleted = true; 4758 } else { 4759 MPASS(gap_entry->start < gap_entry->end - grow_amount); 4760 vm_map_entry_resize(map, gap_entry, -grow_amount); 4761 gap_deleted = false; 4762 } 4763 rv = vm_map_insert(map, NULL, 0, grow_start, 4764 grow_start + grow_amount, 4765 stack_entry->protection, stack_entry->max_protection, 4766 MAP_STACK_GROWS_DOWN); 4767 if (rv != KERN_SUCCESS) { 4768 if (gap_deleted) { 4769 rv1 = vm_map_insert(map, NULL, 0, gap_start, 4770 gap_end, VM_PROT_NONE, VM_PROT_NONE, 4771 MAP_CREATE_GUARD | MAP_CREATE_STACK_GAP_DN); 4772 MPASS(rv1 == KERN_SUCCESS); 4773 } else 4774 vm_map_entry_resize(map, gap_entry, 4775 grow_amount); 4776 } 4777 } else { 4778 grow_start = stack_entry->end; 4779 cred = stack_entry->cred; 4780 if (cred == NULL && stack_entry->object.vm_object != NULL) 4781 cred = stack_entry->object.vm_object->cred; 4782 if (cred != NULL && !swap_reserve_by_cred(grow_amount, cred)) 4783 rv = KERN_NO_SPACE; 4784 /* Grow the underlying object if applicable. */ 4785 else if (stack_entry->object.vm_object == NULL || 4786 vm_object_coalesce(stack_entry->object.vm_object, 4787 stack_entry->offset, 4788 (vm_size_t)(stack_entry->end - stack_entry->start), 4789 grow_amount, cred != NULL)) { 4790 if (gap_entry->start + grow_amount == gap_entry->end) { 4791 vm_map_entry_delete(map, gap_entry); 4792 vm_map_entry_resize(map, stack_entry, 4793 grow_amount); 4794 } else { 4795 gap_entry->start += grow_amount; 4796 stack_entry->end += grow_amount; 4797 } 4798 map->size += grow_amount; 4799 rv = KERN_SUCCESS; 4800 } else 4801 rv = KERN_FAILURE; 4802 } 4803 if (rv == KERN_SUCCESS && is_procstack) 4804 vm->vm_ssize += btoc(grow_amount); 4805 4806 /* 4807 * Heed the MAP_WIREFUTURE flag if it was set for this process. 4808 */ 4809 if (rv == KERN_SUCCESS && (map->flags & MAP_WIREFUTURE) != 0) { 4810 rv = vm_map_wire_locked(map, grow_start, 4811 grow_start + grow_amount, 4812 VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); 4813 } 4814 vm_map_lock_downgrade(map); 4815 4816 out: 4817 #ifdef RACCT 4818 if (racct_enable && rv != KERN_SUCCESS) { 4819 PROC_LOCK(p); 4820 error = racct_set(p, RACCT_VMEM, map->size); 4821 KASSERT(error == 0, ("decreasing RACCT_VMEM failed")); 4822 if (!old_mlock) { 4823 error = racct_set(p, RACCT_MEMLOCK, 4824 ptoa(pmap_wired_count(map->pmap))); 4825 KASSERT(error == 0, ("decreasing RACCT_MEMLOCK failed")); 4826 } 4827 error = racct_set(p, RACCT_STACK, ctob(vm->vm_ssize)); 4828 KASSERT(error == 0, ("decreasing RACCT_STACK failed")); 4829 PROC_UNLOCK(p); 4830 } 4831 #endif 4832 4833 return (rv); 4834 } 4835 4836 /* 4837 * Unshare the specified VM space for exec. If other processes are 4838 * mapped to it, then create a new one. The new vmspace is null. 4839 */ 4840 int 4841 vmspace_exec(struct proc *p, vm_offset_t minuser, vm_offset_t maxuser) 4842 { 4843 struct vmspace *oldvmspace = p->p_vmspace; 4844 struct vmspace *newvmspace; 4845 4846 KASSERT((curthread->td_pflags & TDP_EXECVMSPC) == 0, 4847 ("vmspace_exec recursed")); 4848 newvmspace = vmspace_alloc(minuser, maxuser, pmap_pinit); 4849 if (newvmspace == NULL) 4850 return (ENOMEM); 4851 newvmspace->vm_swrss = oldvmspace->vm_swrss; 4852 /* 4853 * This code is written like this for prototype purposes. The 4854 * goal is to avoid running down the vmspace here, but let the 4855 * other process's that are still using the vmspace to finally 4856 * run it down. Even though there is little or no chance of blocking 4857 * here, it is a good idea to keep this form for future mods. 4858 */ 4859 PROC_VMSPACE_LOCK(p); 4860 p->p_vmspace = newvmspace; 4861 PROC_VMSPACE_UNLOCK(p); 4862 if (p == curthread->td_proc) 4863 pmap_activate(curthread); 4864 curthread->td_pflags |= TDP_EXECVMSPC; 4865 return (0); 4866 } 4867 4868 /* 4869 * Unshare the specified VM space for forcing COW. This 4870 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 4871 */ 4872 int 4873 vmspace_unshare(struct proc *p) 4874 { 4875 struct vmspace *oldvmspace = p->p_vmspace; 4876 struct vmspace *newvmspace; 4877 vm_ooffset_t fork_charge; 4878 4879 if (refcount_load(&oldvmspace->vm_refcnt) == 1) 4880 return (0); 4881 fork_charge = 0; 4882 newvmspace = vmspace_fork(oldvmspace, &fork_charge); 4883 if (newvmspace == NULL) 4884 return (ENOMEM); 4885 if (!swap_reserve_by_cred(fork_charge, p->p_ucred)) { 4886 vmspace_free(newvmspace); 4887 return (ENOMEM); 4888 } 4889 PROC_VMSPACE_LOCK(p); 4890 p->p_vmspace = newvmspace; 4891 PROC_VMSPACE_UNLOCK(p); 4892 if (p == curthread->td_proc) 4893 pmap_activate(curthread); 4894 vmspace_free(oldvmspace); 4895 return (0); 4896 } 4897 4898 /* 4899 * vm_map_lookup: 4900 * 4901 * Finds the VM object, offset, and 4902 * protection for a given virtual address in the 4903 * specified map, assuming a page fault of the 4904 * type specified. 4905 * 4906 * Leaves the map in question locked for read; return 4907 * values are guaranteed until a vm_map_lookup_done 4908 * call is performed. Note that the map argument 4909 * is in/out; the returned map must be used in 4910 * the call to vm_map_lookup_done. 4911 * 4912 * A handle (out_entry) is returned for use in 4913 * vm_map_lookup_done, to make that fast. 4914 * 4915 * If a lookup is requested with "write protection" 4916 * specified, the map may be changed to perform virtual 4917 * copying operations, although the data referenced will 4918 * remain the same. 4919 */ 4920 int 4921 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 4922 vm_offset_t vaddr, 4923 vm_prot_t fault_typea, 4924 vm_map_entry_t *out_entry, /* OUT */ 4925 vm_object_t *object, /* OUT */ 4926 vm_pindex_t *pindex, /* OUT */ 4927 vm_prot_t *out_prot, /* OUT */ 4928 boolean_t *wired) /* OUT */ 4929 { 4930 vm_map_entry_t entry; 4931 vm_map_t map = *var_map; 4932 vm_prot_t prot; 4933 vm_prot_t fault_type; 4934 vm_object_t eobject; 4935 vm_size_t size; 4936 struct ucred *cred; 4937 4938 RetryLookup: 4939 4940 vm_map_lock_read(map); 4941 4942 RetryLookupLocked: 4943 /* 4944 * Lookup the faulting address. 4945 */ 4946 if (!vm_map_lookup_entry(map, vaddr, out_entry)) { 4947 vm_map_unlock_read(map); 4948 return (KERN_INVALID_ADDRESS); 4949 } 4950 4951 entry = *out_entry; 4952 4953 /* 4954 * Handle submaps. 4955 */ 4956 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 4957 vm_map_t old_map = map; 4958 4959 *var_map = map = entry->object.sub_map; 4960 vm_map_unlock_read(old_map); 4961 goto RetryLookup; 4962 } 4963 4964 /* 4965 * Check whether this task is allowed to have this page. 4966 */ 4967 prot = entry->protection; 4968 if ((fault_typea & VM_PROT_FAULT_LOOKUP) != 0) { 4969 fault_typea &= ~VM_PROT_FAULT_LOOKUP; 4970 if (prot == VM_PROT_NONE && map != kernel_map && 4971 (entry->eflags & MAP_ENTRY_GUARD) != 0 && 4972 (entry->eflags & (MAP_ENTRY_STACK_GAP_DN | 4973 MAP_ENTRY_STACK_GAP_UP)) != 0 && 4974 vm_map_growstack(map, vaddr, entry) == KERN_SUCCESS) 4975 goto RetryLookupLocked; 4976 } 4977 fault_type = fault_typea & VM_PROT_ALL; 4978 if ((fault_type & prot) != fault_type || prot == VM_PROT_NONE) { 4979 vm_map_unlock_read(map); 4980 return (KERN_PROTECTION_FAILURE); 4981 } 4982 KASSERT((prot & VM_PROT_WRITE) == 0 || (entry->eflags & 4983 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY)) != 4984 (MAP_ENTRY_USER_WIRED | MAP_ENTRY_NEEDS_COPY), 4985 ("entry %p flags %x", entry, entry->eflags)); 4986 if ((fault_typea & VM_PROT_COPY) != 0 && 4987 (entry->max_protection & VM_PROT_WRITE) == 0 && 4988 (entry->eflags & MAP_ENTRY_COW) == 0) { 4989 vm_map_unlock_read(map); 4990 return (KERN_PROTECTION_FAILURE); 4991 } 4992 4993 /* 4994 * If this page is not pageable, we have to get it for all possible 4995 * accesses. 4996 */ 4997 *wired = (entry->wired_count != 0); 4998 if (*wired) 4999 fault_type = entry->protection; 5000 size = entry->end - entry->start; 5001 5002 /* 5003 * If the entry was copy-on-write, we either ... 5004 */ 5005 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 5006 /* 5007 * If we want to write the page, we may as well handle that 5008 * now since we've got the map locked. 5009 * 5010 * If we don't need to write the page, we just demote the 5011 * permissions allowed. 5012 */ 5013 if ((fault_type & VM_PROT_WRITE) != 0 || 5014 (fault_typea & VM_PROT_COPY) != 0) { 5015 /* 5016 * Make a new object, and place it in the object 5017 * chain. Note that no new references have appeared 5018 * -- one just moved from the map to the new 5019 * object. 5020 */ 5021 if (vm_map_lock_upgrade(map)) 5022 goto RetryLookup; 5023 5024 if (entry->cred == NULL) { 5025 /* 5026 * The debugger owner is charged for 5027 * the memory. 5028 */ 5029 cred = curthread->td_ucred; 5030 crhold(cred); 5031 if (!swap_reserve_by_cred(size, cred)) { 5032 crfree(cred); 5033 vm_map_unlock(map); 5034 return (KERN_RESOURCE_SHORTAGE); 5035 } 5036 entry->cred = cred; 5037 } 5038 eobject = entry->object.vm_object; 5039 vm_object_shadow(&entry->object.vm_object, 5040 &entry->offset, size, entry->cred, false); 5041 if (eobject == entry->object.vm_object) { 5042 /* 5043 * The object was not shadowed. 5044 */ 5045 swap_release_by_cred(size, entry->cred); 5046 crfree(entry->cred); 5047 } 5048 entry->cred = NULL; 5049 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 5050 5051 vm_map_lock_downgrade(map); 5052 } else { 5053 /* 5054 * We're attempting to read a copy-on-write page -- 5055 * don't allow writes. 5056 */ 5057 prot &= ~VM_PROT_WRITE; 5058 } 5059 } 5060 5061 /* 5062 * Create an object if necessary. 5063 */ 5064 if (entry->object.vm_object == NULL && !map->system_map) { 5065 if (vm_map_lock_upgrade(map)) 5066 goto RetryLookup; 5067 entry->object.vm_object = vm_object_allocate_anon(atop(size), 5068 NULL, entry->cred, entry->cred != NULL ? size : 0); 5069 entry->offset = 0; 5070 entry->cred = NULL; 5071 vm_map_lock_downgrade(map); 5072 } 5073 5074 /* 5075 * Return the object/offset from this entry. If the entry was 5076 * copy-on-write or empty, it has been fixed up. 5077 */ 5078 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 5079 *object = entry->object.vm_object; 5080 5081 *out_prot = prot; 5082 return (KERN_SUCCESS); 5083 } 5084 5085 /* 5086 * vm_map_lookup_locked: 5087 * 5088 * Lookup the faulting address. A version of vm_map_lookup that returns 5089 * KERN_FAILURE instead of blocking on map lock or memory allocation. 5090 */ 5091 int 5092 vm_map_lookup_locked(vm_map_t *var_map, /* IN/OUT */ 5093 vm_offset_t vaddr, 5094 vm_prot_t fault_typea, 5095 vm_map_entry_t *out_entry, /* OUT */ 5096 vm_object_t *object, /* OUT */ 5097 vm_pindex_t *pindex, /* OUT */ 5098 vm_prot_t *out_prot, /* OUT */ 5099 boolean_t *wired) /* OUT */ 5100 { 5101 vm_map_entry_t entry; 5102 vm_map_t map = *var_map; 5103 vm_prot_t prot; 5104 vm_prot_t fault_type = fault_typea; 5105 5106 /* 5107 * Lookup the faulting address. 5108 */ 5109 if (!vm_map_lookup_entry(map, vaddr, out_entry)) 5110 return (KERN_INVALID_ADDRESS); 5111 5112 entry = *out_entry; 5113 5114 /* 5115 * Fail if the entry refers to a submap. 5116 */ 5117 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) 5118 return (KERN_FAILURE); 5119 5120 /* 5121 * Check whether this task is allowed to have this page. 5122 */ 5123 prot = entry->protection; 5124 fault_type &= VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE; 5125 if ((fault_type & prot) != fault_type) 5126 return (KERN_PROTECTION_FAILURE); 5127 5128 /* 5129 * If this page is not pageable, we have to get it for all possible 5130 * accesses. 5131 */ 5132 *wired = (entry->wired_count != 0); 5133 if (*wired) 5134 fault_type = entry->protection; 5135 5136 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 5137 /* 5138 * Fail if the entry was copy-on-write for a write fault. 5139 */ 5140 if (fault_type & VM_PROT_WRITE) 5141 return (KERN_FAILURE); 5142 /* 5143 * We're attempting to read a copy-on-write page -- 5144 * don't allow writes. 5145 */ 5146 prot &= ~VM_PROT_WRITE; 5147 } 5148 5149 /* 5150 * Fail if an object should be created. 5151 */ 5152 if (entry->object.vm_object == NULL && !map->system_map) 5153 return (KERN_FAILURE); 5154 5155 /* 5156 * Return the object/offset from this entry. If the entry was 5157 * copy-on-write or empty, it has been fixed up. 5158 */ 5159 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 5160 *object = entry->object.vm_object; 5161 5162 *out_prot = prot; 5163 return (KERN_SUCCESS); 5164 } 5165 5166 /* 5167 * vm_map_lookup_done: 5168 * 5169 * Releases locks acquired by a vm_map_lookup 5170 * (according to the handle returned by that lookup). 5171 */ 5172 void 5173 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry) 5174 { 5175 /* 5176 * Unlock the main-level map 5177 */ 5178 vm_map_unlock_read(map); 5179 } 5180 5181 vm_offset_t 5182 vm_map_max_KBI(const struct vm_map *map) 5183 { 5184 5185 return (vm_map_max(map)); 5186 } 5187 5188 vm_offset_t 5189 vm_map_min_KBI(const struct vm_map *map) 5190 { 5191 5192 return (vm_map_min(map)); 5193 } 5194 5195 pmap_t 5196 vm_map_pmap_KBI(vm_map_t map) 5197 { 5198 5199 return (map->pmap); 5200 } 5201 5202 bool 5203 vm_map_range_valid_KBI(vm_map_t map, vm_offset_t start, vm_offset_t end) 5204 { 5205 5206 return (vm_map_range_valid(map, start, end)); 5207 } 5208 5209 #ifdef INVARIANTS 5210 static void 5211 _vm_map_assert_consistent(vm_map_t map, int check) 5212 { 5213 vm_map_entry_t entry, prev; 5214 vm_map_entry_t cur, header, lbound, ubound; 5215 vm_size_t max_left, max_right; 5216 5217 #ifdef DIAGNOSTIC 5218 ++map->nupdates; 5219 #endif 5220 if (enable_vmmap_check != check) 5221 return; 5222 5223 header = prev = &map->header; 5224 VM_MAP_ENTRY_FOREACH(entry, map) { 5225 KASSERT(prev->end <= entry->start, 5226 ("map %p prev->end = %jx, start = %jx", map, 5227 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5228 KASSERT(entry->start < entry->end, 5229 ("map %p start = %jx, end = %jx", map, 5230 (uintmax_t)entry->start, (uintmax_t)entry->end)); 5231 KASSERT(entry->left == header || 5232 entry->left->start < entry->start, 5233 ("map %p left->start = %jx, start = %jx", map, 5234 (uintmax_t)entry->left->start, (uintmax_t)entry->start)); 5235 KASSERT(entry->right == header || 5236 entry->start < entry->right->start, 5237 ("map %p start = %jx, right->start = %jx", map, 5238 (uintmax_t)entry->start, (uintmax_t)entry->right->start)); 5239 cur = map->root; 5240 lbound = ubound = header; 5241 for (;;) { 5242 if (entry->start < cur->start) { 5243 ubound = cur; 5244 cur = cur->left; 5245 KASSERT(cur != lbound, 5246 ("map %p cannot find %jx", 5247 map, (uintmax_t)entry->start)); 5248 } else if (cur->end <= entry->start) { 5249 lbound = cur; 5250 cur = cur->right; 5251 KASSERT(cur != ubound, 5252 ("map %p cannot find %jx", 5253 map, (uintmax_t)entry->start)); 5254 } else { 5255 KASSERT(cur == entry, 5256 ("map %p cannot find %jx", 5257 map, (uintmax_t)entry->start)); 5258 break; 5259 } 5260 } 5261 max_left = vm_map_entry_max_free_left(entry, lbound); 5262 max_right = vm_map_entry_max_free_right(entry, ubound); 5263 KASSERT(entry->max_free == vm_size_max(max_left, max_right), 5264 ("map %p max = %jx, max_left = %jx, max_right = %jx", map, 5265 (uintmax_t)entry->max_free, 5266 (uintmax_t)max_left, (uintmax_t)max_right)); 5267 prev = entry; 5268 } 5269 KASSERT(prev->end <= entry->start, 5270 ("map %p prev->end = %jx, start = %jx", map, 5271 (uintmax_t)prev->end, (uintmax_t)entry->start)); 5272 } 5273 #endif 5274 5275 #include "opt_ddb.h" 5276 #ifdef DDB 5277 #include <sys/kernel.h> 5278 5279 #include <ddb/ddb.h> 5280 5281 static void 5282 vm_map_print(vm_map_t map) 5283 { 5284 vm_map_entry_t entry, prev; 5285 5286 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 5287 (void *)map, 5288 (void *)map->pmap, map->nentries, map->timestamp); 5289 5290 db_indent += 2; 5291 prev = &map->header; 5292 VM_MAP_ENTRY_FOREACH(entry, map) { 5293 db_iprintf("map entry %p: start=%p, end=%p, eflags=%#x, \n", 5294 (void *)entry, (void *)entry->start, (void *)entry->end, 5295 entry->eflags); 5296 { 5297 static const char * const inheritance_name[4] = 5298 {"share", "copy", "none", "donate_copy"}; 5299 5300 db_iprintf(" prot=%x/%x/%s", 5301 entry->protection, 5302 entry->max_protection, 5303 inheritance_name[(int)(unsigned char) 5304 entry->inheritance]); 5305 if (entry->wired_count != 0) 5306 db_printf(", wired"); 5307 } 5308 if (entry->eflags & MAP_ENTRY_IS_SUB_MAP) { 5309 db_printf(", share=%p, offset=0x%jx\n", 5310 (void *)entry->object.sub_map, 5311 (uintmax_t)entry->offset); 5312 if (prev == &map->header || 5313 prev->object.sub_map != 5314 entry->object.sub_map) { 5315 db_indent += 2; 5316 vm_map_print((vm_map_t)entry->object.sub_map); 5317 db_indent -= 2; 5318 } 5319 } else { 5320 if (entry->cred != NULL) 5321 db_printf(", ruid %d", entry->cred->cr_ruid); 5322 db_printf(", object=%p, offset=0x%jx", 5323 (void *)entry->object.vm_object, 5324 (uintmax_t)entry->offset); 5325 if (entry->object.vm_object && entry->object.vm_object->cred) 5326 db_printf(", obj ruid %d charge %jx", 5327 entry->object.vm_object->cred->cr_ruid, 5328 (uintmax_t)entry->object.vm_object->charge); 5329 if (entry->eflags & MAP_ENTRY_COW) 5330 db_printf(", copy (%s)", 5331 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 5332 db_printf("\n"); 5333 5334 if (prev == &map->header || 5335 prev->object.vm_object != 5336 entry->object.vm_object) { 5337 db_indent += 2; 5338 vm_object_print((db_expr_t)(intptr_t) 5339 entry->object.vm_object, 5340 0, 0, (char *)0); 5341 db_indent -= 2; 5342 } 5343 } 5344 prev = entry; 5345 } 5346 db_indent -= 2; 5347 } 5348 5349 DB_SHOW_COMMAND(map, map) 5350 { 5351 5352 if (!have_addr) { 5353 db_printf("usage: show map <addr>\n"); 5354 return; 5355 } 5356 vm_map_print((vm_map_t)addr); 5357 } 5358 5359 DB_SHOW_COMMAND(procvm, procvm) 5360 { 5361 struct proc *p; 5362 5363 if (have_addr) { 5364 p = db_lookup_proc(addr); 5365 } else { 5366 p = curproc; 5367 } 5368 5369 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 5370 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 5371 (void *)vmspace_pmap(p->p_vmspace)); 5372 5373 vm_map_print((vm_map_t)&p->p_vmspace->vm_map); 5374 } 5375 5376 #endif /* DDB */ 5377