1 /* 2 * (MPSAFE) 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. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * from: @(#)vm_map.c 8.3 (Berkeley) 1/12/94 39 * 40 * 41 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 42 * All rights reserved. 43 * 44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 45 * 46 * Permission to use, copy, modify and distribute this software and 47 * its documentation is hereby granted, provided that both the copyright 48 * notice and this permission notice appear in all copies of the 49 * software, derivative works or modified versions, and any portions 50 * thereof, and that both notices appear in supporting documentation. 51 * 52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 55 * 56 * Carnegie Mellon requests users of this software to return to 57 * 58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 59 * School of Computer Science 60 * Carnegie Mellon University 61 * Pittsburgh PA 15213-3890 62 * 63 * any improvements or extensions that they make and grant Carnegie the 64 * rights to redistribute these changes. 65 * 66 * $FreeBSD: src/sys/vm/vm_map.c,v 1.187.2.19 2003/05/27 00:47:02 alc Exp $ 67 */ 68 69 /* 70 * Virtual memory mapping module. 71 */ 72 73 #include <sys/param.h> 74 #include <sys/systm.h> 75 #include <sys/kernel.h> 76 #include <sys/proc.h> 77 #include <sys/serialize.h> 78 #include <sys/lock.h> 79 #include <sys/vmmeter.h> 80 #include <sys/mman.h> 81 #include <sys/vnode.h> 82 #include <sys/resourcevar.h> 83 #include <sys/shm.h> 84 #include <sys/tree.h> 85 #include <sys/malloc.h> 86 87 #include <vm/vm.h> 88 #include <vm/vm_param.h> 89 #include <vm/pmap.h> 90 #include <vm/vm_map.h> 91 #include <vm/vm_page.h> 92 #include <vm/vm_object.h> 93 #include <vm/vm_pager.h> 94 #include <vm/vm_kern.h> 95 #include <vm/vm_extern.h> 96 #include <vm/swap_pager.h> 97 #include <vm/vm_zone.h> 98 99 #include <sys/thread2.h> 100 #include <sys/sysref2.h> 101 #include <sys/random.h> 102 #include <sys/sysctl.h> 103 104 /* 105 * Virtual memory maps provide for the mapping, protection, and sharing 106 * of virtual memory objects. In addition, this module provides for an 107 * efficient virtual copy of 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 entries. 112 * A hint and a RB tree is used to speed-up lookups. 113 * 114 * Callers looking to modify maps specify start/end addresses which cause 115 * the related map entry to be clipped if necessary, and then later 116 * recombined if the pieces remained compatible. 117 * 118 * Virtual copy operations are performed by copying VM object references 119 * from one map to another, and then marking both regions as copy-on-write. 120 */ 121 static void vmspace_terminate(struct vmspace *vm); 122 static void vmspace_lock(struct vmspace *vm); 123 static void vmspace_unlock(struct vmspace *vm); 124 static void vmspace_dtor(void *obj, void *private); 125 126 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore"); 127 128 struct sysref_class vmspace_sysref_class = { 129 .name = "vmspace", 130 .mtype = M_VMSPACE, 131 .proto = SYSREF_PROTO_VMSPACE, 132 .offset = offsetof(struct vmspace, vm_sysref), 133 .objsize = sizeof(struct vmspace), 134 .nom_cache = 32, 135 .flags = SRC_MANAGEDINIT, 136 .dtor = vmspace_dtor, 137 .ops = { 138 .terminate = (sysref_terminate_func_t)vmspace_terminate, 139 .lock = (sysref_lock_func_t)vmspace_lock, 140 .unlock = (sysref_lock_func_t)vmspace_unlock 141 } 142 }; 143 144 /* 145 * per-cpu page table cross mappings are initialized in early boot 146 * and might require a considerable number of vm_map_entry structures. 147 */ 148 #define VMEPERCPU (MAXCPU+1) 149 150 static struct vm_zone mapentzone_store, mapzone_store; 151 static vm_zone_t mapentzone, mapzone; 152 static struct vm_object mapentobj, mapobj; 153 154 static struct vm_map_entry map_entry_init[MAX_MAPENT]; 155 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU]; 156 static struct vm_map map_init[MAX_KMAP]; 157 158 static int randomize_mmap; 159 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0, 160 "Randomize mmap offsets"); 161 static int vm_map_relock_enable = 1; 162 SYSCTL_INT(_vm, OID_AUTO, map_relock_enable, CTLFLAG_RW, 163 &vm_map_relock_enable, 0, "Randomize mmap offsets"); 164 165 static void vm_map_entry_shadow(vm_map_entry_t entry, int addref); 166 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *); 167 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *); 168 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 169 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 170 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *); 171 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t); 172 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t, 173 vm_map_entry_t); 174 static void vm_map_unclip_range (vm_map_t map, vm_map_entry_t start_entry, vm_offset_t start, vm_offset_t end, int *count, int flags); 175 176 /* 177 * Initialize the vm_map module. Must be called before any other vm_map 178 * routines. 179 * 180 * Map and entry structures are allocated from the general purpose 181 * memory pool with some exceptions: 182 * 183 * - The kernel map is allocated statically. 184 * - Initial kernel map entries are allocated out of a static pool. 185 * 186 * These restrictions are necessary since malloc() uses the 187 * maps and requires map entries. 188 * 189 * Called from the low level boot code only. 190 */ 191 void 192 vm_map_startup(void) 193 { 194 mapzone = &mapzone_store; 195 zbootinit(mapzone, "MAP", sizeof (struct vm_map), 196 map_init, MAX_KMAP); 197 mapentzone = &mapentzone_store; 198 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry), 199 map_entry_init, MAX_MAPENT); 200 } 201 202 /* 203 * Called prior to any vmspace allocations. 204 * 205 * Called from the low level boot code only. 206 */ 207 void 208 vm_init2(void) 209 { 210 zinitna(mapentzone, &mapentobj, NULL, 0, 0, 211 ZONE_USE_RESERVE | ZONE_SPECIAL, 1); 212 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1); 213 pmap_init2(); 214 vm_object_init2(); 215 } 216 217 218 /* 219 * Red black tree functions 220 * 221 * The caller must hold the related map lock. 222 */ 223 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b); 224 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare); 225 226 /* a->start is address, and the only field has to be initialized */ 227 static int 228 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b) 229 { 230 if (a->start < b->start) 231 return(-1); 232 else if (a->start > b->start) 233 return(1); 234 return(0); 235 } 236 237 /* 238 * Allocate a vmspace structure, including a vm_map and pmap. 239 * Initialize numerous fields. While the initial allocation is zerod, 240 * subsequence reuse from the objcache leaves elements of the structure 241 * intact (particularly the pmap), so portions must be zerod. 242 * 243 * The structure is not considered activated until we call sysref_activate(). 244 * 245 * No requirements. 246 */ 247 struct vmspace * 248 vmspace_alloc(vm_offset_t min, vm_offset_t max) 249 { 250 struct vmspace *vm; 251 252 vm = sysref_alloc(&vmspace_sysref_class); 253 bzero(&vm->vm_startcopy, 254 (char *)&vm->vm_endcopy - (char *)&vm->vm_startcopy); 255 vm_map_init(&vm->vm_map, min, max, NULL); /* initializes token */ 256 257 /* 258 * Use a hold to prevent any additional racing hold from terminating 259 * the vmspace before we manage to activate it. This also acquires 260 * the token for safety. 261 */ 262 KKASSERT(vm->vm_holdcount == 0); 263 KKASSERT(vm->vm_exitingcnt == 0); 264 vmspace_hold(vm); 265 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */ 266 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */ 267 vm->vm_shm = NULL; 268 vm->vm_flags = 0; 269 cpu_vmspace_alloc(vm); 270 sysref_activate(&vm->vm_sysref); 271 vmspace_drop(vm); 272 273 return (vm); 274 } 275 276 /* 277 * Free a primary reference to a vmspace. This can trigger a 278 * stage-1 termination. 279 */ 280 void 281 vmspace_free(struct vmspace *vm) 282 { 283 /* 284 * We want all finalization to occur via vmspace_drop() so we 285 * need to hold the vm around the put. 286 */ 287 vmspace_hold(vm); 288 sysref_put(&vm->vm_sysref); 289 vmspace_drop(vm); 290 } 291 292 void 293 vmspace_ref(struct vmspace *vm) 294 { 295 sysref_get(&vm->vm_sysref); 296 } 297 298 void 299 vmspace_hold(struct vmspace *vm) 300 { 301 refcount_acquire(&vm->vm_holdcount); 302 lwkt_gettoken(&vm->vm_map.token); 303 } 304 305 void 306 vmspace_drop(struct vmspace *vm) 307 { 308 lwkt_reltoken(&vm->vm_map.token); 309 if (refcount_release(&vm->vm_holdcount)) { 310 if (vm->vm_exitingcnt == 0 && 311 sysref_isinactive(&vm->vm_sysref)) { 312 vmspace_terminate(vm); 313 } 314 } 315 } 316 317 /* 318 * dtor function - Some elements of the pmap are retained in the 319 * free-cached vmspaces to improve performance. We have to clean them up 320 * here before returning the vmspace to the memory pool. 321 * 322 * No requirements. 323 */ 324 static void 325 vmspace_dtor(void *obj, void *private) 326 { 327 struct vmspace *vm = obj; 328 329 pmap_puninit(vmspace_pmap(vm)); 330 } 331 332 /* 333 * Called in three cases: 334 * 335 * (1) When the last sysref is dropped and the vmspace becomes inactive. 336 * (holdcount will not be 0 because the vmspace is held through the op) 337 * 338 * (2) When exitingcount becomes 0 on the last reap 339 * (holdcount will not be 0 because the vmspace is held through the op) 340 * 341 * (3) When the holdcount becomes 0 in addition to the above two 342 * 343 * sysref will not scrap the object until we call sysref_put() once more 344 * after the last ref has been dropped. 345 * 346 * VMSPACE_EXIT1 flags the primary deactivation 347 * VMSPACE_EXIT2 flags the last reap 348 */ 349 static void 350 vmspace_terminate(struct vmspace *vm) 351 { 352 int count; 353 354 /* 355 * 356 */ 357 lwkt_gettoken(&vm->vm_map.token); 358 if ((vm->vm_flags & VMSPACE_EXIT1) == 0) { 359 vm->vm_flags |= VMSPACE_EXIT1; 360 shmexit(vm); 361 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 362 VM_MAX_USER_ADDRESS); 363 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 364 VM_MAX_USER_ADDRESS); 365 } 366 if ((vm->vm_flags & VMSPACE_EXIT2) == 0 && vm->vm_exitingcnt == 0) { 367 vm->vm_flags |= VMSPACE_EXIT2; 368 cpu_vmspace_free(vm); 369 shmexit(vm); 370 371 /* 372 * Lock the map, to wait out all other references to it. 373 * Delete all of the mappings and pages they hold, then call 374 * the pmap module to reclaim anything left. 375 */ 376 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 377 vm_map_lock(&vm->vm_map); 378 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset, 379 vm->vm_map.max_offset, &count); 380 vm_map_unlock(&vm->vm_map); 381 vm_map_entry_release(count); 382 383 lwkt_gettoken(&vmspace_pmap(vm)->pm_token); 384 pmap_release(vmspace_pmap(vm)); 385 lwkt_reltoken(&vmspace_pmap(vm)->pm_token); 386 } 387 388 lwkt_reltoken(&vm->vm_map.token); 389 if (vm->vm_exitingcnt == 0 && vm->vm_holdcount == 0) { 390 KKASSERT(vm->vm_flags & VMSPACE_EXIT1); 391 KKASSERT(vm->vm_flags & VMSPACE_EXIT2); 392 sysref_put(&vm->vm_sysref); 393 } 394 } 395 396 /* 397 * vmspaces are not currently locked. 398 */ 399 static void 400 vmspace_lock(struct vmspace *vm __unused) 401 { 402 } 403 404 static void 405 vmspace_unlock(struct vmspace *vm __unused) 406 { 407 } 408 409 /* 410 * This is called during exit indicating that the vmspace is no 411 * longer in used by an exiting process, but the process has not yet 412 * been reaped. 413 * 414 * No requirements. 415 */ 416 void 417 vmspace_exitbump(struct vmspace *vm) 418 { 419 vmspace_hold(vm); 420 ++vm->vm_exitingcnt; 421 vmspace_drop(vm); /* handles termination sequencing */ 422 } 423 424 /* 425 * Decrement the exitingcnt and issue the stage-2 termination if it becomes 426 * zero and the stage1 termination has already occured. 427 * 428 * No requirements. 429 */ 430 void 431 vmspace_exitfree(struct proc *p) 432 { 433 struct vmspace *vm; 434 435 vm = p->p_vmspace; 436 p->p_vmspace = NULL; 437 vmspace_hold(vm); 438 KKASSERT(vm->vm_exitingcnt > 0); 439 if (--vm->vm_exitingcnt == 0 && sysref_isinactive(&vm->vm_sysref)) 440 vmspace_terminate(vm); 441 vmspace_drop(vm); /* handles termination sequencing */ 442 } 443 444 /* 445 * Swap useage is determined by taking the proportional swap used by 446 * VM objects backing the VM map. To make up for fractional losses, 447 * if the VM object has any swap use at all the associated map entries 448 * count for at least 1 swap page. 449 * 450 * No requirements. 451 */ 452 int 453 vmspace_swap_count(struct vmspace *vm) 454 { 455 vm_map_t map = &vm->vm_map; 456 vm_map_entry_t cur; 457 vm_object_t object; 458 int count = 0; 459 int n; 460 461 vmspace_hold(vm); 462 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 463 switch(cur->maptype) { 464 case VM_MAPTYPE_NORMAL: 465 case VM_MAPTYPE_VPAGETABLE: 466 if ((object = cur->object.vm_object) == NULL) 467 break; 468 if (object->swblock_count) { 469 n = (cur->end - cur->start) / PAGE_SIZE; 470 count += object->swblock_count * 471 SWAP_META_PAGES * n / object->size + 1; 472 } 473 break; 474 default: 475 break; 476 } 477 } 478 vmspace_drop(vm); 479 480 return(count); 481 } 482 483 /* 484 * Calculate the approximate number of anonymous pages in use by 485 * this vmspace. To make up for fractional losses, we count each 486 * VM object as having at least 1 anonymous page. 487 * 488 * No requirements. 489 */ 490 int 491 vmspace_anonymous_count(struct vmspace *vm) 492 { 493 vm_map_t map = &vm->vm_map; 494 vm_map_entry_t cur; 495 vm_object_t object; 496 int count = 0; 497 498 vmspace_hold(vm); 499 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 500 switch(cur->maptype) { 501 case VM_MAPTYPE_NORMAL: 502 case VM_MAPTYPE_VPAGETABLE: 503 if ((object = cur->object.vm_object) == NULL) 504 break; 505 if (object->type != OBJT_DEFAULT && 506 object->type != OBJT_SWAP) { 507 break; 508 } 509 count += object->resident_page_count; 510 break; 511 default: 512 break; 513 } 514 } 515 vmspace_drop(vm); 516 517 return(count); 518 } 519 520 /* 521 * Creates and returns a new empty VM map with the given physical map 522 * structure, and having the given lower and upper address bounds. 523 * 524 * No requirements. 525 */ 526 vm_map_t 527 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max) 528 { 529 if (result == NULL) 530 result = zalloc(mapzone); 531 vm_map_init(result, min, max, pmap); 532 return (result); 533 } 534 535 /* 536 * Initialize an existing vm_map structure such as that in the vmspace 537 * structure. The pmap is initialized elsewhere. 538 * 539 * No requirements. 540 */ 541 void 542 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap) 543 { 544 map->header.next = map->header.prev = &map->header; 545 RB_INIT(&map->rb_root); 546 map->nentries = 0; 547 map->size = 0; 548 map->system_map = 0; 549 map->min_offset = min; 550 map->max_offset = max; 551 map->pmap = pmap; 552 map->first_free = &map->header; 553 map->hint = &map->header; 554 map->timestamp = 0; 555 map->flags = 0; 556 lwkt_token_init(&map->token, "vm_map"); 557 lockinit(&map->lock, "thrd_sleep", (hz + 9) / 10, 0); 558 TUNABLE_INT("vm.cache_vmspaces", &vmspace_sysref_class.nom_cache); 559 } 560 561 /* 562 * Shadow the vm_map_entry's object. This typically needs to be done when 563 * a write fault is taken on an entry which had previously been cloned by 564 * fork(). The shared object (which might be NULL) must become private so 565 * we add a shadow layer above it. 566 * 567 * Object allocation for anonymous mappings is defered as long as possible. 568 * When creating a shadow, however, the underlying object must be instantiated 569 * so it can be shared. 570 * 571 * If the map segment is governed by a virtual page table then it is 572 * possible to address offsets beyond the mapped area. Just allocate 573 * a maximally sized object for this case. 574 * 575 * The vm_map must be exclusively locked. 576 * No other requirements. 577 */ 578 static 579 void 580 vm_map_entry_shadow(vm_map_entry_t entry, int addref) 581 { 582 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 583 vm_object_shadow(&entry->object.vm_object, &entry->offset, 584 0x7FFFFFFF, addref); /* XXX */ 585 } else { 586 vm_object_shadow(&entry->object.vm_object, &entry->offset, 587 atop(entry->end - entry->start), addref); 588 } 589 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 590 } 591 592 /* 593 * Allocate an object for a vm_map_entry. 594 * 595 * Object allocation for anonymous mappings is defered as long as possible. 596 * This function is called when we can defer no longer, generally when a map 597 * entry might be split or forked or takes a page fault. 598 * 599 * If the map segment is governed by a virtual page table then it is 600 * possible to address offsets beyond the mapped area. Just allocate 601 * a maximally sized object for this case. 602 * 603 * The vm_map must be exclusively locked. 604 * No other requirements. 605 */ 606 void 607 vm_map_entry_allocate_object(vm_map_entry_t entry) 608 { 609 vm_object_t obj; 610 611 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 612 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */ 613 } else { 614 obj = vm_object_allocate(OBJT_DEFAULT, 615 atop(entry->end - entry->start)); 616 } 617 entry->object.vm_object = obj; 618 entry->offset = 0; 619 } 620 621 /* 622 * Set an initial negative count so the first attempt to reserve 623 * space preloads a bunch of vm_map_entry's for this cpu. Also 624 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 625 * map a new page for vm_map_entry structures. SMP systems are 626 * particularly sensitive. 627 * 628 * This routine is called in early boot so we cannot just call 629 * vm_map_entry_reserve(). 630 * 631 * Called from the low level boot code only (for each cpu) 632 */ 633 void 634 vm_map_entry_reserve_cpu_init(globaldata_t gd) 635 { 636 vm_map_entry_t entry; 637 int i; 638 639 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2; 640 entry = &cpu_map_entry_init[gd->gd_cpuid][0]; 641 for (i = 0; i < VMEPERCPU; ++i, ++entry) { 642 entry->next = gd->gd_vme_base; 643 gd->gd_vme_base = entry; 644 } 645 } 646 647 /* 648 * Reserves vm_map_entry structures so code later on can manipulate 649 * map_entry structures within a locked map without blocking trying 650 * to allocate a new vm_map_entry. 651 * 652 * No requirements. 653 */ 654 int 655 vm_map_entry_reserve(int count) 656 { 657 struct globaldata *gd = mycpu; 658 vm_map_entry_t entry; 659 660 /* 661 * Make sure we have enough structures in gd_vme_base to handle 662 * the reservation request. 663 * 664 * The critical section protects access to the per-cpu gd. 665 */ 666 crit_enter(); 667 while (gd->gd_vme_avail < count) { 668 entry = zalloc(mapentzone); 669 entry->next = gd->gd_vme_base; 670 gd->gd_vme_base = entry; 671 ++gd->gd_vme_avail; 672 } 673 gd->gd_vme_avail -= count; 674 crit_exit(); 675 676 return(count); 677 } 678 679 /* 680 * Releases previously reserved vm_map_entry structures that were not 681 * used. If we have too much junk in our per-cpu cache clean some of 682 * it out. 683 * 684 * No requirements. 685 */ 686 void 687 vm_map_entry_release(int count) 688 { 689 struct globaldata *gd = mycpu; 690 vm_map_entry_t entry; 691 692 crit_enter(); 693 gd->gd_vme_avail += count; 694 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 695 entry = gd->gd_vme_base; 696 KKASSERT(entry != NULL); 697 gd->gd_vme_base = entry->next; 698 --gd->gd_vme_avail; 699 crit_exit(); 700 zfree(mapentzone, entry); 701 crit_enter(); 702 } 703 crit_exit(); 704 } 705 706 /* 707 * Reserve map entry structures for use in kernel_map itself. These 708 * entries have *ALREADY* been reserved on a per-cpu basis when the map 709 * was inited. This function is used by zalloc() to avoid a recursion 710 * when zalloc() itself needs to allocate additional kernel memory. 711 * 712 * This function works like the normal reserve but does not load the 713 * vm_map_entry cache (because that would result in an infinite 714 * recursion). Note that gd_vme_avail may go negative. This is expected. 715 * 716 * Any caller of this function must be sure to renormalize after 717 * potentially eating entries to ensure that the reserve supply 718 * remains intact. 719 * 720 * No requirements. 721 */ 722 int 723 vm_map_entry_kreserve(int count) 724 { 725 struct globaldata *gd = mycpu; 726 727 crit_enter(); 728 gd->gd_vme_avail -= count; 729 crit_exit(); 730 KASSERT(gd->gd_vme_base != NULL, 731 ("no reserved entries left, gd_vme_avail = %d", 732 gd->gd_vme_avail)); 733 return(count); 734 } 735 736 /* 737 * Release previously reserved map entries for kernel_map. We do not 738 * attempt to clean up like the normal release function as this would 739 * cause an unnecessary (but probably not fatal) deep procedure call. 740 * 741 * No requirements. 742 */ 743 void 744 vm_map_entry_krelease(int count) 745 { 746 struct globaldata *gd = mycpu; 747 748 crit_enter(); 749 gd->gd_vme_avail += count; 750 crit_exit(); 751 } 752 753 /* 754 * Allocates a VM map entry for insertion. No entry fields are filled in. 755 * 756 * The entries should have previously been reserved. The reservation count 757 * is tracked in (*countp). 758 * 759 * No requirements. 760 */ 761 static vm_map_entry_t 762 vm_map_entry_create(vm_map_t map, int *countp) 763 { 764 struct globaldata *gd = mycpu; 765 vm_map_entry_t entry; 766 767 KKASSERT(*countp > 0); 768 --*countp; 769 crit_enter(); 770 entry = gd->gd_vme_base; 771 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 772 gd->gd_vme_base = entry->next; 773 crit_exit(); 774 775 return(entry); 776 } 777 778 /* 779 * Dispose of a vm_map_entry that is no longer being referenced. 780 * 781 * No requirements. 782 */ 783 static void 784 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 785 { 786 struct globaldata *gd = mycpu; 787 788 KKASSERT(map->hint != entry); 789 KKASSERT(map->first_free != entry); 790 791 ++*countp; 792 crit_enter(); 793 entry->next = gd->gd_vme_base; 794 gd->gd_vme_base = entry; 795 crit_exit(); 796 } 797 798 799 /* 800 * Insert/remove entries from maps. 801 * 802 * The related map must be exclusively locked. 803 * The caller must hold map->token 804 * No other requirements. 805 */ 806 static __inline void 807 vm_map_entry_link(vm_map_t map, 808 vm_map_entry_t after_where, 809 vm_map_entry_t entry) 810 { 811 ASSERT_VM_MAP_LOCKED(map); 812 813 map->nentries++; 814 entry->prev = after_where; 815 entry->next = after_where->next; 816 entry->next->prev = entry; 817 after_where->next = entry; 818 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry)) 819 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry); 820 } 821 822 static __inline void 823 vm_map_entry_unlink(vm_map_t map, 824 vm_map_entry_t entry) 825 { 826 vm_map_entry_t prev; 827 vm_map_entry_t next; 828 829 ASSERT_VM_MAP_LOCKED(map); 830 831 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 832 panic("vm_map_entry_unlink: attempt to mess with " 833 "locked entry! %p", entry); 834 } 835 prev = entry->prev; 836 next = entry->next; 837 next->prev = prev; 838 prev->next = next; 839 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry); 840 map->nentries--; 841 } 842 843 /* 844 * Finds the map entry containing (or immediately preceding) the specified 845 * address in the given map. The entry is returned in (*entry). 846 * 847 * The boolean result indicates whether the address is actually contained 848 * in the map. 849 * 850 * The related map must be locked. 851 * No other requirements. 852 */ 853 boolean_t 854 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry) 855 { 856 vm_map_entry_t tmp; 857 vm_map_entry_t last; 858 859 ASSERT_VM_MAP_LOCKED(map); 860 #if 0 861 /* 862 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive 863 * the hint code with the red-black lookup meets with system crashes 864 * and lockups. We do not yet know why. 865 * 866 * It is possible that the problem is related to the setting 867 * of the hint during map_entry deletion, in the code specified 868 * at the GGG comment later on in this file. 869 * 870 * YYY More likely it's because this function can be called with 871 * a shared lock on the map, resulting in map->hint updates possibly 872 * racing. Fixed now but untested. 873 */ 874 /* 875 * Quickly check the cached hint, there's a good chance of a match. 876 */ 877 tmp = map->hint; 878 cpu_ccfence(); 879 if (tmp != &map->header) { 880 if (address >= tmp->start && address < tmp->end) { 881 *entry = tmp; 882 return(TRUE); 883 } 884 } 885 #endif 886 887 /* 888 * Locate the record from the top of the tree. 'last' tracks the 889 * closest prior record and is returned if no match is found, which 890 * in binary tree terms means tracking the most recent right-branch 891 * taken. If there is no prior record, &map->header is returned. 892 */ 893 last = &map->header; 894 tmp = RB_ROOT(&map->rb_root); 895 896 while (tmp) { 897 if (address >= tmp->start) { 898 if (address < tmp->end) { 899 *entry = tmp; 900 map->hint = tmp; 901 return(TRUE); 902 } 903 last = tmp; 904 tmp = RB_RIGHT(tmp, rb_entry); 905 } else { 906 tmp = RB_LEFT(tmp, rb_entry); 907 } 908 } 909 *entry = last; 910 return (FALSE); 911 } 912 913 /* 914 * Inserts the given whole VM object into the target map at the specified 915 * address range. The object's size should match that of the address range. 916 * 917 * The map must be exclusively locked. 918 * The object must be held. 919 * The caller must have reserved sufficient vm_map_entry structures. 920 * 921 * If object is non-NULL, ref count must be bumped by caller prior to 922 * making call to account for the new entry. 923 */ 924 int 925 vm_map_insert(vm_map_t map, int *countp, 926 vm_object_t object, vm_ooffset_t offset, 927 vm_offset_t start, vm_offset_t end, 928 vm_maptype_t maptype, 929 vm_prot_t prot, vm_prot_t max, 930 int cow) 931 { 932 vm_map_entry_t new_entry; 933 vm_map_entry_t prev_entry; 934 vm_map_entry_t temp_entry; 935 vm_eflags_t protoeflags; 936 int must_drop = 0; 937 938 ASSERT_VM_MAP_LOCKED(map); 939 if (object) 940 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object)); 941 942 /* 943 * Check that the start and end points are not bogus. 944 */ 945 if ((start < map->min_offset) || (end > map->max_offset) || 946 (start >= end)) 947 return (KERN_INVALID_ADDRESS); 948 949 /* 950 * Find the entry prior to the proposed starting address; if it's part 951 * of an existing entry, this range is bogus. 952 */ 953 if (vm_map_lookup_entry(map, start, &temp_entry)) 954 return (KERN_NO_SPACE); 955 956 prev_entry = temp_entry; 957 958 /* 959 * Assert that the next entry doesn't overlap the end point. 960 */ 961 962 if ((prev_entry->next != &map->header) && 963 (prev_entry->next->start < end)) 964 return (KERN_NO_SPACE); 965 966 protoeflags = 0; 967 968 if (cow & MAP_COPY_ON_WRITE) 969 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 970 971 if (cow & MAP_NOFAULT) { 972 protoeflags |= MAP_ENTRY_NOFAULT; 973 974 KASSERT(object == NULL, 975 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 976 } 977 if (cow & MAP_DISABLE_SYNCER) 978 protoeflags |= MAP_ENTRY_NOSYNC; 979 if (cow & MAP_DISABLE_COREDUMP) 980 protoeflags |= MAP_ENTRY_NOCOREDUMP; 981 if (cow & MAP_IS_STACK) 982 protoeflags |= MAP_ENTRY_STACK; 983 if (cow & MAP_IS_KSTACK) 984 protoeflags |= MAP_ENTRY_KSTACK; 985 986 lwkt_gettoken(&map->token); 987 988 if (object) { 989 /* 990 * When object is non-NULL, it could be shared with another 991 * process. We have to set or clear OBJ_ONEMAPPING 992 * appropriately. 993 * 994 * NOTE: This flag is only applicable to DEFAULT and SWAP 995 * objects and will already be clear in other types 996 * of objects, so a shared object lock is ok for 997 * VNODE objects. 998 */ 999 if ((object->ref_count > 1) || (object->shadow_count != 0)) { 1000 vm_object_clear_flag(object, OBJ_ONEMAPPING); 1001 } 1002 } 1003 else if ((prev_entry != &map->header) && 1004 (prev_entry->eflags == protoeflags) && 1005 (prev_entry->end == start) && 1006 (prev_entry->wired_count == 0) && 1007 prev_entry->maptype == maptype && 1008 ((prev_entry->object.vm_object == NULL) || 1009 vm_object_coalesce(prev_entry->object.vm_object, 1010 OFF_TO_IDX(prev_entry->offset), 1011 (vm_size_t)(prev_entry->end - prev_entry->start), 1012 (vm_size_t)(end - prev_entry->end)))) { 1013 /* 1014 * We were able to extend the object. Determine if we 1015 * can extend the previous map entry to include the 1016 * new range as well. 1017 */ 1018 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 1019 (prev_entry->protection == prot) && 1020 (prev_entry->max_protection == max)) { 1021 map->size += (end - prev_entry->end); 1022 prev_entry->end = end; 1023 vm_map_simplify_entry(map, prev_entry, countp); 1024 lwkt_reltoken(&map->token); 1025 return (KERN_SUCCESS); 1026 } 1027 1028 /* 1029 * If we can extend the object but cannot extend the 1030 * map entry, we have to create a new map entry. We 1031 * must bump the ref count on the extended object to 1032 * account for it. object may be NULL. 1033 */ 1034 object = prev_entry->object.vm_object; 1035 offset = prev_entry->offset + 1036 (prev_entry->end - prev_entry->start); 1037 if (object) { 1038 vm_object_hold(object); 1039 vm_object_chain_wait(object); 1040 vm_object_reference_locked(object); 1041 must_drop = 1; 1042 } 1043 } 1044 1045 /* 1046 * NOTE: if conditionals fail, object can be NULL here. This occurs 1047 * in things like the buffer map where we manage kva but do not manage 1048 * backing objects. 1049 */ 1050 1051 /* 1052 * Create a new entry 1053 */ 1054 1055 new_entry = vm_map_entry_create(map, countp); 1056 new_entry->start = start; 1057 new_entry->end = end; 1058 1059 new_entry->maptype = maptype; 1060 new_entry->eflags = protoeflags; 1061 new_entry->object.vm_object = object; 1062 new_entry->offset = offset; 1063 new_entry->aux.master_pde = 0; 1064 1065 new_entry->inheritance = VM_INHERIT_DEFAULT; 1066 new_entry->protection = prot; 1067 new_entry->max_protection = max; 1068 new_entry->wired_count = 0; 1069 1070 /* 1071 * Insert the new entry into the list 1072 */ 1073 1074 vm_map_entry_link(map, prev_entry, new_entry); 1075 map->size += new_entry->end - new_entry->start; 1076 1077 /* 1078 * Update the free space hint. Entries cannot overlap. 1079 * An exact comparison is needed to avoid matching 1080 * against the map->header. 1081 */ 1082 if ((map->first_free == prev_entry) && 1083 (prev_entry->end == new_entry->start)) { 1084 map->first_free = new_entry; 1085 } 1086 1087 #if 0 1088 /* 1089 * Temporarily removed to avoid MAP_STACK panic, due to 1090 * MAP_STACK being a huge hack. Will be added back in 1091 * when MAP_STACK (and the user stack mapping) is fixed. 1092 */ 1093 /* 1094 * It may be possible to simplify the entry 1095 */ 1096 vm_map_simplify_entry(map, new_entry, countp); 1097 #endif 1098 1099 /* 1100 * Try to pre-populate the page table. Mappings governed by virtual 1101 * page tables cannot be prepopulated without a lot of work, so 1102 * don't try. 1103 */ 1104 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) && 1105 maptype != VM_MAPTYPE_VPAGETABLE) { 1106 int dorelock = 0; 1107 if (vm_map_relock_enable && (cow & MAP_PREFAULT_RELOCK)) { 1108 dorelock = 1; 1109 vm_object_lock_swap(); 1110 vm_object_drop(object); 1111 } 1112 pmap_object_init_pt(map->pmap, start, prot, 1113 object, OFF_TO_IDX(offset), end - start, 1114 cow & MAP_PREFAULT_PARTIAL); 1115 if (dorelock) { 1116 vm_object_hold(object); 1117 vm_object_lock_swap(); 1118 } 1119 } 1120 if (must_drop) 1121 vm_object_drop(object); 1122 1123 lwkt_reltoken(&map->token); 1124 return (KERN_SUCCESS); 1125 } 1126 1127 /* 1128 * Find sufficient space for `length' bytes in the given map, starting at 1129 * `start'. Returns 0 on success, 1 on no space. 1130 * 1131 * This function will returned an arbitrarily aligned pointer. If no 1132 * particular alignment is required you should pass align as 1. Note that 1133 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 1134 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 1135 * argument. 1136 * 1137 * 'align' should be a power of 2 but is not required to be. 1138 * 1139 * The map must be exclusively locked. 1140 * No other requirements. 1141 */ 1142 int 1143 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length, 1144 vm_size_t align, int flags, vm_offset_t *addr) 1145 { 1146 vm_map_entry_t entry, next; 1147 vm_offset_t end; 1148 vm_offset_t align_mask; 1149 1150 if (start < map->min_offset) 1151 start = map->min_offset; 1152 if (start > map->max_offset) 1153 return (1); 1154 1155 /* 1156 * If the alignment is not a power of 2 we will have to use 1157 * a mod/division, set align_mask to a special value. 1158 */ 1159 if ((align | (align - 1)) + 1 != (align << 1)) 1160 align_mask = (vm_offset_t)-1; 1161 else 1162 align_mask = align - 1; 1163 1164 /* 1165 * Look for the first possible address; if there's already something 1166 * at this address, we have to start after it. 1167 */ 1168 if (start == map->min_offset) { 1169 if ((entry = map->first_free) != &map->header) 1170 start = entry->end; 1171 } else { 1172 vm_map_entry_t tmp; 1173 1174 if (vm_map_lookup_entry(map, start, &tmp)) 1175 start = tmp->end; 1176 entry = tmp; 1177 } 1178 1179 /* 1180 * Look through the rest of the map, trying to fit a new region in the 1181 * gap between existing regions, or after the very last region. 1182 */ 1183 for (;; start = (entry = next)->end) { 1184 /* 1185 * Adjust the proposed start by the requested alignment, 1186 * be sure that we didn't wrap the address. 1187 */ 1188 if (align_mask == (vm_offset_t)-1) 1189 end = ((start + align - 1) / align) * align; 1190 else 1191 end = (start + align_mask) & ~align_mask; 1192 if (end < start) 1193 return (1); 1194 start = end; 1195 /* 1196 * Find the end of the proposed new region. Be sure we didn't 1197 * go beyond the end of the map, or wrap around the address. 1198 * Then check to see if this is the last entry or if the 1199 * proposed end fits in the gap between this and the next 1200 * entry. 1201 */ 1202 end = start + length; 1203 if (end > map->max_offset || end < start) 1204 return (1); 1205 next = entry->next; 1206 1207 /* 1208 * If the next entry's start address is beyond the desired 1209 * end address we may have found a good entry. 1210 * 1211 * If the next entry is a stack mapping we do not map into 1212 * the stack's reserved space. 1213 * 1214 * XXX continue to allow mapping into the stack's reserved 1215 * space if doing a MAP_STACK mapping inside a MAP_STACK 1216 * mapping, for backwards compatibility. But the caller 1217 * really should use MAP_STACK | MAP_TRYFIXED if they 1218 * want to do that. 1219 */ 1220 if (next == &map->header) 1221 break; 1222 if (next->start >= end) { 1223 if ((next->eflags & MAP_ENTRY_STACK) == 0) 1224 break; 1225 if (flags & MAP_STACK) 1226 break; 1227 if (next->start - next->aux.avail_ssize >= end) 1228 break; 1229 } 1230 } 1231 map->hint = entry; 1232 1233 /* 1234 * Grow the kernel_map if necessary. pmap_growkernel() will panic 1235 * if it fails. The kernel_map is locked and nothing can steal 1236 * our address space if pmap_growkernel() blocks. 1237 * 1238 * NOTE: This may be unconditionally called for kldload areas on 1239 * x86_64 because these do not bump kernel_vm_end (which would 1240 * fill 128G worth of page tables!). Therefore we must not 1241 * retry. 1242 */ 1243 if (map == &kernel_map) { 1244 vm_offset_t kstop; 1245 1246 kstop = round_page(start + length); 1247 if (kstop > kernel_vm_end) 1248 pmap_growkernel(start, kstop); 1249 } 1250 *addr = start; 1251 return (0); 1252 } 1253 1254 /* 1255 * vm_map_find finds an unallocated region in the target address map with 1256 * the given length and allocates it. The search is defined to be first-fit 1257 * from the specified address; the region found is returned in the same 1258 * parameter. 1259 * 1260 * If object is non-NULL, ref count must be bumped by caller 1261 * prior to making call to account for the new entry. 1262 * 1263 * No requirements. This function will lock the map temporarily. 1264 */ 1265 int 1266 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1267 vm_offset_t *addr, vm_size_t length, vm_size_t align, 1268 boolean_t fitit, 1269 vm_maptype_t maptype, 1270 vm_prot_t prot, vm_prot_t max, 1271 int cow) 1272 { 1273 vm_offset_t start; 1274 int result; 1275 int count; 1276 1277 start = *addr; 1278 1279 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1280 vm_map_lock(map); 1281 if (object) 1282 vm_object_hold(object); 1283 if (fitit) { 1284 if (vm_map_findspace(map, start, length, align, 0, addr)) { 1285 if (object) 1286 vm_object_drop(object); 1287 vm_map_unlock(map); 1288 vm_map_entry_release(count); 1289 return (KERN_NO_SPACE); 1290 } 1291 start = *addr; 1292 } 1293 result = vm_map_insert(map, &count, object, offset, 1294 start, start + length, 1295 maptype, 1296 prot, max, 1297 cow); 1298 if (object) 1299 vm_object_drop(object); 1300 vm_map_unlock(map); 1301 vm_map_entry_release(count); 1302 1303 return (result); 1304 } 1305 1306 /* 1307 * Simplify the given map entry by merging with either neighbor. This 1308 * routine also has the ability to merge with both neighbors. 1309 * 1310 * This routine guarentees that the passed entry remains valid (though 1311 * possibly extended). When merging, this routine may delete one or 1312 * both neighbors. No action is taken on entries which have their 1313 * in-transition flag set. 1314 * 1315 * The map must be exclusively locked. 1316 */ 1317 void 1318 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 1319 { 1320 vm_map_entry_t next, prev; 1321 vm_size_t prevsize, esize; 1322 1323 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1324 ++mycpu->gd_cnt.v_intrans_coll; 1325 return; 1326 } 1327 1328 if (entry->maptype == VM_MAPTYPE_SUBMAP) 1329 return; 1330 1331 prev = entry->prev; 1332 if (prev != &map->header) { 1333 prevsize = prev->end - prev->start; 1334 if ( (prev->end == entry->start) && 1335 (prev->maptype == entry->maptype) && 1336 (prev->object.vm_object == entry->object.vm_object) && 1337 (!prev->object.vm_object || 1338 (prev->offset + prevsize == entry->offset)) && 1339 (prev->eflags == entry->eflags) && 1340 (prev->protection == entry->protection) && 1341 (prev->max_protection == entry->max_protection) && 1342 (prev->inheritance == entry->inheritance) && 1343 (prev->wired_count == entry->wired_count)) { 1344 if (map->first_free == prev) 1345 map->first_free = entry; 1346 if (map->hint == prev) 1347 map->hint = entry; 1348 vm_map_entry_unlink(map, prev); 1349 entry->start = prev->start; 1350 entry->offset = prev->offset; 1351 if (prev->object.vm_object) 1352 vm_object_deallocate(prev->object.vm_object); 1353 vm_map_entry_dispose(map, prev, countp); 1354 } 1355 } 1356 1357 next = entry->next; 1358 if (next != &map->header) { 1359 esize = entry->end - entry->start; 1360 if ((entry->end == next->start) && 1361 (next->maptype == entry->maptype) && 1362 (next->object.vm_object == entry->object.vm_object) && 1363 (!entry->object.vm_object || 1364 (entry->offset + esize == next->offset)) && 1365 (next->eflags == entry->eflags) && 1366 (next->protection == entry->protection) && 1367 (next->max_protection == entry->max_protection) && 1368 (next->inheritance == entry->inheritance) && 1369 (next->wired_count == entry->wired_count)) { 1370 if (map->first_free == next) 1371 map->first_free = entry; 1372 if (map->hint == next) 1373 map->hint = entry; 1374 vm_map_entry_unlink(map, next); 1375 entry->end = next->end; 1376 if (next->object.vm_object) 1377 vm_object_deallocate(next->object.vm_object); 1378 vm_map_entry_dispose(map, next, countp); 1379 } 1380 } 1381 } 1382 1383 /* 1384 * Asserts that the given entry begins at or after the specified address. 1385 * If necessary, it splits the entry into two. 1386 */ 1387 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1388 { \ 1389 if (startaddr > entry->start) \ 1390 _vm_map_clip_start(map, entry, startaddr, countp); \ 1391 } 1392 1393 /* 1394 * This routine is called only when it is known that the entry must be split. 1395 * 1396 * The map must be exclusively locked. 1397 */ 1398 static void 1399 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, 1400 int *countp) 1401 { 1402 vm_map_entry_t new_entry; 1403 1404 /* 1405 * Split off the front portion -- note that we must insert the new 1406 * entry BEFORE this one, so that this entry has the specified 1407 * starting address. 1408 */ 1409 1410 vm_map_simplify_entry(map, entry, countp); 1411 1412 /* 1413 * If there is no object backing this entry, we might as well create 1414 * one now. If we defer it, an object can get created after the map 1415 * is clipped, and individual objects will be created for the split-up 1416 * map. This is a bit of a hack, but is also about the best place to 1417 * put this improvement. 1418 */ 1419 if (entry->object.vm_object == NULL && !map->system_map) { 1420 vm_map_entry_allocate_object(entry); 1421 } 1422 1423 new_entry = vm_map_entry_create(map, countp); 1424 *new_entry = *entry; 1425 1426 new_entry->end = start; 1427 entry->offset += (start - entry->start); 1428 entry->start = start; 1429 1430 vm_map_entry_link(map, entry->prev, new_entry); 1431 1432 switch(entry->maptype) { 1433 case VM_MAPTYPE_NORMAL: 1434 case VM_MAPTYPE_VPAGETABLE: 1435 if (new_entry->object.vm_object) { 1436 vm_object_hold(new_entry->object.vm_object); 1437 vm_object_chain_wait(new_entry->object.vm_object); 1438 vm_object_reference_locked(new_entry->object.vm_object); 1439 vm_object_drop(new_entry->object.vm_object); 1440 } 1441 break; 1442 default: 1443 break; 1444 } 1445 } 1446 1447 /* 1448 * Asserts that the given entry ends at or before the specified address. 1449 * If necessary, it splits the entry into two. 1450 * 1451 * The map must be exclusively locked. 1452 */ 1453 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1454 { \ 1455 if (endaddr < entry->end) \ 1456 _vm_map_clip_end(map, entry, endaddr, countp); \ 1457 } 1458 1459 /* 1460 * This routine is called only when it is known that the entry must be split. 1461 * 1462 * The map must be exclusively locked. 1463 */ 1464 static void 1465 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, 1466 int *countp) 1467 { 1468 vm_map_entry_t new_entry; 1469 1470 /* 1471 * If there is no object backing this entry, we might as well create 1472 * one now. If we defer it, an object can get created after the map 1473 * is clipped, and individual objects will be created for the split-up 1474 * map. This is a bit of a hack, but is also about the best place to 1475 * put this improvement. 1476 */ 1477 1478 if (entry->object.vm_object == NULL && !map->system_map) { 1479 vm_map_entry_allocate_object(entry); 1480 } 1481 1482 /* 1483 * Create a new entry and insert it AFTER the specified entry 1484 */ 1485 1486 new_entry = vm_map_entry_create(map, countp); 1487 *new_entry = *entry; 1488 1489 new_entry->start = entry->end = end; 1490 new_entry->offset += (end - entry->start); 1491 1492 vm_map_entry_link(map, entry, new_entry); 1493 1494 switch(entry->maptype) { 1495 case VM_MAPTYPE_NORMAL: 1496 case VM_MAPTYPE_VPAGETABLE: 1497 if (new_entry->object.vm_object) { 1498 vm_object_hold(new_entry->object.vm_object); 1499 vm_object_chain_wait(new_entry->object.vm_object); 1500 vm_object_reference_locked(new_entry->object.vm_object); 1501 vm_object_drop(new_entry->object.vm_object); 1502 } 1503 break; 1504 default: 1505 break; 1506 } 1507 } 1508 1509 /* 1510 * Asserts that the starting and ending region addresses fall within the 1511 * valid range for the map. 1512 */ 1513 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1514 { \ 1515 if (start < vm_map_min(map)) \ 1516 start = vm_map_min(map); \ 1517 if (end > vm_map_max(map)) \ 1518 end = vm_map_max(map); \ 1519 if (start > end) \ 1520 start = end; \ 1521 } 1522 1523 /* 1524 * Used to block when an in-transition collison occurs. The map 1525 * is unlocked for the sleep and relocked before the return. 1526 */ 1527 void 1528 vm_map_transition_wait(vm_map_t map) 1529 { 1530 tsleep_interlock(map, 0); 1531 vm_map_unlock(map); 1532 tsleep(map, PINTERLOCKED, "vment", 0); 1533 vm_map_lock(map); 1534 } 1535 1536 /* 1537 * When we do blocking operations with the map lock held it is 1538 * possible that a clip might have occured on our in-transit entry, 1539 * requiring an adjustment to the entry in our loop. These macros 1540 * help the pageable and clip_range code deal with the case. The 1541 * conditional costs virtually nothing if no clipping has occured. 1542 */ 1543 1544 #define CLIP_CHECK_BACK(entry, save_start) \ 1545 do { \ 1546 while (entry->start != save_start) { \ 1547 entry = entry->prev; \ 1548 KASSERT(entry != &map->header, ("bad entry clip")); \ 1549 } \ 1550 } while(0) 1551 1552 #define CLIP_CHECK_FWD(entry, save_end) \ 1553 do { \ 1554 while (entry->end != save_end) { \ 1555 entry = entry->next; \ 1556 KASSERT(entry != &map->header, ("bad entry clip")); \ 1557 } \ 1558 } while(0) 1559 1560 1561 /* 1562 * Clip the specified range and return the base entry. The 1563 * range may cover several entries starting at the returned base 1564 * and the first and last entry in the covering sequence will be 1565 * properly clipped to the requested start and end address. 1566 * 1567 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1568 * flag. 1569 * 1570 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1571 * covered by the requested range. 1572 * 1573 * The map must be exclusively locked on entry and will remain locked 1574 * on return. If no range exists or the range contains holes and you 1575 * specified that no holes were allowed, NULL will be returned. This 1576 * routine may temporarily unlock the map in order avoid a deadlock when 1577 * sleeping. 1578 */ 1579 static 1580 vm_map_entry_t 1581 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1582 int *countp, int flags) 1583 { 1584 vm_map_entry_t start_entry; 1585 vm_map_entry_t entry; 1586 1587 /* 1588 * Locate the entry and effect initial clipping. The in-transition 1589 * case does not occur very often so do not try to optimize it. 1590 */ 1591 again: 1592 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1593 return (NULL); 1594 entry = start_entry; 1595 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1596 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1597 ++mycpu->gd_cnt.v_intrans_coll; 1598 ++mycpu->gd_cnt.v_intrans_wait; 1599 vm_map_transition_wait(map); 1600 /* 1601 * entry and/or start_entry may have been clipped while 1602 * we slept, or may have gone away entirely. We have 1603 * to restart from the lookup. 1604 */ 1605 goto again; 1606 } 1607 1608 /* 1609 * Since we hold an exclusive map lock we do not have to restart 1610 * after clipping, even though clipping may block in zalloc. 1611 */ 1612 vm_map_clip_start(map, entry, start, countp); 1613 vm_map_clip_end(map, entry, end, countp); 1614 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1615 1616 /* 1617 * Scan entries covered by the range. When working on the next 1618 * entry a restart need only re-loop on the current entry which 1619 * we have already locked, since 'next' may have changed. Also, 1620 * even though entry is safe, it may have been clipped so we 1621 * have to iterate forwards through the clip after sleeping. 1622 */ 1623 while (entry->next != &map->header && entry->next->start < end) { 1624 vm_map_entry_t next = entry->next; 1625 1626 if (flags & MAP_CLIP_NO_HOLES) { 1627 if (next->start > entry->end) { 1628 vm_map_unclip_range(map, start_entry, 1629 start, entry->end, countp, flags); 1630 return(NULL); 1631 } 1632 } 1633 1634 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1635 vm_offset_t save_end = entry->end; 1636 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1637 ++mycpu->gd_cnt.v_intrans_coll; 1638 ++mycpu->gd_cnt.v_intrans_wait; 1639 vm_map_transition_wait(map); 1640 1641 /* 1642 * clips might have occured while we blocked. 1643 */ 1644 CLIP_CHECK_FWD(entry, save_end); 1645 CLIP_CHECK_BACK(start_entry, start); 1646 continue; 1647 } 1648 /* 1649 * No restart necessary even though clip_end may block, we 1650 * are holding the map lock. 1651 */ 1652 vm_map_clip_end(map, next, end, countp); 1653 next->eflags |= MAP_ENTRY_IN_TRANSITION; 1654 entry = next; 1655 } 1656 if (flags & MAP_CLIP_NO_HOLES) { 1657 if (entry->end != end) { 1658 vm_map_unclip_range(map, start_entry, 1659 start, entry->end, countp, flags); 1660 return(NULL); 1661 } 1662 } 1663 return(start_entry); 1664 } 1665 1666 /* 1667 * Undo the effect of vm_map_clip_range(). You should pass the same 1668 * flags and the same range that you passed to vm_map_clip_range(). 1669 * This code will clear the in-transition flag on the entries and 1670 * wake up anyone waiting. This code will also simplify the sequence 1671 * and attempt to merge it with entries before and after the sequence. 1672 * 1673 * The map must be locked on entry and will remain locked on return. 1674 * 1675 * Note that you should also pass the start_entry returned by 1676 * vm_map_clip_range(). However, if you block between the two calls 1677 * with the map unlocked please be aware that the start_entry may 1678 * have been clipped and you may need to scan it backwards to find 1679 * the entry corresponding with the original start address. You are 1680 * responsible for this, vm_map_unclip_range() expects the correct 1681 * start_entry to be passed to it and will KASSERT otherwise. 1682 */ 1683 static 1684 void 1685 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry, 1686 vm_offset_t start, vm_offset_t end, 1687 int *countp, int flags) 1688 { 1689 vm_map_entry_t entry; 1690 1691 entry = start_entry; 1692 1693 KASSERT(entry->start == start, ("unclip_range: illegal base entry")); 1694 while (entry != &map->header && entry->start < end) { 1695 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, 1696 ("in-transition flag not set during unclip on: %p", 1697 entry)); 1698 KASSERT(entry->end <= end, 1699 ("unclip_range: tail wasn't clipped")); 1700 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 1701 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 1702 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 1703 wakeup(map); 1704 } 1705 entry = entry->next; 1706 } 1707 1708 /* 1709 * Simplification does not block so there is no restart case. 1710 */ 1711 entry = start_entry; 1712 while (entry != &map->header && entry->start < end) { 1713 vm_map_simplify_entry(map, entry, countp); 1714 entry = entry->next; 1715 } 1716 } 1717 1718 /* 1719 * Mark the given range as handled by a subordinate map. 1720 * 1721 * This range must have been created with vm_map_find(), and no other 1722 * operations may have been performed on this range prior to calling 1723 * vm_map_submap(). 1724 * 1725 * Submappings cannot be removed. 1726 * 1727 * No requirements. 1728 */ 1729 int 1730 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 1731 { 1732 vm_map_entry_t entry; 1733 int result = KERN_INVALID_ARGUMENT; 1734 int count; 1735 1736 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1737 vm_map_lock(map); 1738 1739 VM_MAP_RANGE_CHECK(map, start, end); 1740 1741 if (vm_map_lookup_entry(map, start, &entry)) { 1742 vm_map_clip_start(map, entry, start, &count); 1743 } else { 1744 entry = entry->next; 1745 } 1746 1747 vm_map_clip_end(map, entry, end, &count); 1748 1749 if ((entry->start == start) && (entry->end == end) && 1750 ((entry->eflags & MAP_ENTRY_COW) == 0) && 1751 (entry->object.vm_object == NULL)) { 1752 entry->object.sub_map = submap; 1753 entry->maptype = VM_MAPTYPE_SUBMAP; 1754 result = KERN_SUCCESS; 1755 } 1756 vm_map_unlock(map); 1757 vm_map_entry_release(count); 1758 1759 return (result); 1760 } 1761 1762 /* 1763 * Sets the protection of the specified address region in the target map. 1764 * If "set_max" is specified, the maximum protection is to be set; 1765 * otherwise, only the current protection is affected. 1766 * 1767 * The protection is not applicable to submaps, but is applicable to normal 1768 * maps and maps governed by virtual page tables. For example, when operating 1769 * on a virtual page table our protection basically controls how COW occurs 1770 * on the backing object, whereas the virtual page table abstraction itself 1771 * is an abstraction for userland. 1772 * 1773 * No requirements. 1774 */ 1775 int 1776 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 1777 vm_prot_t new_prot, boolean_t set_max) 1778 { 1779 vm_map_entry_t current; 1780 vm_map_entry_t entry; 1781 int count; 1782 1783 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1784 vm_map_lock(map); 1785 1786 VM_MAP_RANGE_CHECK(map, start, end); 1787 1788 if (vm_map_lookup_entry(map, start, &entry)) { 1789 vm_map_clip_start(map, entry, start, &count); 1790 } else { 1791 entry = entry->next; 1792 } 1793 1794 /* 1795 * Make a first pass to check for protection violations. 1796 */ 1797 current = entry; 1798 while ((current != &map->header) && (current->start < end)) { 1799 if (current->maptype == VM_MAPTYPE_SUBMAP) { 1800 vm_map_unlock(map); 1801 vm_map_entry_release(count); 1802 return (KERN_INVALID_ARGUMENT); 1803 } 1804 if ((new_prot & current->max_protection) != new_prot) { 1805 vm_map_unlock(map); 1806 vm_map_entry_release(count); 1807 return (KERN_PROTECTION_FAILURE); 1808 } 1809 current = current->next; 1810 } 1811 1812 /* 1813 * Go back and fix up protections. [Note that clipping is not 1814 * necessary the second time.] 1815 */ 1816 current = entry; 1817 1818 while ((current != &map->header) && (current->start < end)) { 1819 vm_prot_t old_prot; 1820 1821 vm_map_clip_end(map, current, end, &count); 1822 1823 old_prot = current->protection; 1824 if (set_max) { 1825 current->protection = 1826 (current->max_protection = new_prot) & 1827 old_prot; 1828 } else { 1829 current->protection = new_prot; 1830 } 1831 1832 /* 1833 * Update physical map if necessary. Worry about copy-on-write 1834 * here -- CHECK THIS XXX 1835 */ 1836 1837 if (current->protection != old_prot) { 1838 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 1839 VM_PROT_ALL) 1840 1841 pmap_protect(map->pmap, current->start, 1842 current->end, 1843 current->protection & MASK(current)); 1844 #undef MASK 1845 } 1846 1847 vm_map_simplify_entry(map, current, &count); 1848 1849 current = current->next; 1850 } 1851 1852 vm_map_unlock(map); 1853 vm_map_entry_release(count); 1854 return (KERN_SUCCESS); 1855 } 1856 1857 /* 1858 * This routine traverses a processes map handling the madvise 1859 * system call. Advisories are classified as either those effecting 1860 * the vm_map_entry structure, or those effecting the underlying 1861 * objects. 1862 * 1863 * The <value> argument is used for extended madvise calls. 1864 * 1865 * No requirements. 1866 */ 1867 int 1868 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, 1869 int behav, off_t value) 1870 { 1871 vm_map_entry_t current, entry; 1872 int modify_map = 0; 1873 int error = 0; 1874 int count; 1875 1876 /* 1877 * Some madvise calls directly modify the vm_map_entry, in which case 1878 * we need to use an exclusive lock on the map and we need to perform 1879 * various clipping operations. Otherwise we only need a read-lock 1880 * on the map. 1881 */ 1882 1883 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1884 1885 switch(behav) { 1886 case MADV_NORMAL: 1887 case MADV_SEQUENTIAL: 1888 case MADV_RANDOM: 1889 case MADV_NOSYNC: 1890 case MADV_AUTOSYNC: 1891 case MADV_NOCORE: 1892 case MADV_CORE: 1893 case MADV_SETMAP: 1894 case MADV_INVAL: 1895 modify_map = 1; 1896 vm_map_lock(map); 1897 break; 1898 case MADV_WILLNEED: 1899 case MADV_DONTNEED: 1900 case MADV_FREE: 1901 vm_map_lock_read(map); 1902 break; 1903 default: 1904 vm_map_entry_release(count); 1905 return (EINVAL); 1906 } 1907 1908 /* 1909 * Locate starting entry and clip if necessary. 1910 */ 1911 1912 VM_MAP_RANGE_CHECK(map, start, end); 1913 1914 if (vm_map_lookup_entry(map, start, &entry)) { 1915 if (modify_map) 1916 vm_map_clip_start(map, entry, start, &count); 1917 } else { 1918 entry = entry->next; 1919 } 1920 1921 if (modify_map) { 1922 /* 1923 * madvise behaviors that are implemented in the vm_map_entry. 1924 * 1925 * We clip the vm_map_entry so that behavioral changes are 1926 * limited to the specified address range. 1927 */ 1928 for (current = entry; 1929 (current != &map->header) && (current->start < end); 1930 current = current->next 1931 ) { 1932 if (current->maptype == VM_MAPTYPE_SUBMAP) 1933 continue; 1934 1935 vm_map_clip_end(map, current, end, &count); 1936 1937 switch (behav) { 1938 case MADV_NORMAL: 1939 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 1940 break; 1941 case MADV_SEQUENTIAL: 1942 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 1943 break; 1944 case MADV_RANDOM: 1945 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 1946 break; 1947 case MADV_NOSYNC: 1948 current->eflags |= MAP_ENTRY_NOSYNC; 1949 break; 1950 case MADV_AUTOSYNC: 1951 current->eflags &= ~MAP_ENTRY_NOSYNC; 1952 break; 1953 case MADV_NOCORE: 1954 current->eflags |= MAP_ENTRY_NOCOREDUMP; 1955 break; 1956 case MADV_CORE: 1957 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 1958 break; 1959 case MADV_INVAL: 1960 /* 1961 * Invalidate the related pmap entries, used 1962 * to flush portions of the real kernel's 1963 * pmap when the caller has removed or 1964 * modified existing mappings in a virtual 1965 * page table. 1966 */ 1967 pmap_remove(map->pmap, 1968 current->start, current->end); 1969 break; 1970 case MADV_SETMAP: 1971 /* 1972 * Set the page directory page for a map 1973 * governed by a virtual page table. Mark 1974 * the entry as being governed by a virtual 1975 * page table if it is not. 1976 * 1977 * XXX the page directory page is stored 1978 * in the avail_ssize field if the map_entry. 1979 * 1980 * XXX the map simplification code does not 1981 * compare this field so weird things may 1982 * happen if you do not apply this function 1983 * to the entire mapping governed by the 1984 * virtual page table. 1985 */ 1986 if (current->maptype != VM_MAPTYPE_VPAGETABLE) { 1987 error = EINVAL; 1988 break; 1989 } 1990 current->aux.master_pde = value; 1991 pmap_remove(map->pmap, 1992 current->start, current->end); 1993 break; 1994 default: 1995 error = EINVAL; 1996 break; 1997 } 1998 vm_map_simplify_entry(map, current, &count); 1999 } 2000 vm_map_unlock(map); 2001 } else { 2002 vm_pindex_t pindex; 2003 int count; 2004 2005 /* 2006 * madvise behaviors that are implemented in the underlying 2007 * vm_object. 2008 * 2009 * Since we don't clip the vm_map_entry, we have to clip 2010 * the vm_object pindex and count. 2011 * 2012 * NOTE! We currently do not support these functions on 2013 * virtual page tables. 2014 */ 2015 for (current = entry; 2016 (current != &map->header) && (current->start < end); 2017 current = current->next 2018 ) { 2019 vm_offset_t useStart; 2020 2021 if (current->maptype != VM_MAPTYPE_NORMAL) 2022 continue; 2023 2024 pindex = OFF_TO_IDX(current->offset); 2025 count = atop(current->end - current->start); 2026 useStart = current->start; 2027 2028 if (current->start < start) { 2029 pindex += atop(start - current->start); 2030 count -= atop(start - current->start); 2031 useStart = start; 2032 } 2033 if (current->end > end) 2034 count -= atop(current->end - end); 2035 2036 if (count <= 0) 2037 continue; 2038 2039 vm_object_madvise(current->object.vm_object, 2040 pindex, count, behav); 2041 2042 /* 2043 * Try to populate the page table. Mappings governed 2044 * by virtual page tables cannot be pre-populated 2045 * without a lot of work so don't try. 2046 */ 2047 if (behav == MADV_WILLNEED && 2048 current->maptype != VM_MAPTYPE_VPAGETABLE) { 2049 pmap_object_init_pt( 2050 map->pmap, 2051 useStart, 2052 current->protection, 2053 current->object.vm_object, 2054 pindex, 2055 (count << PAGE_SHIFT), 2056 MAP_PREFAULT_MADVISE 2057 ); 2058 } 2059 } 2060 vm_map_unlock_read(map); 2061 } 2062 vm_map_entry_release(count); 2063 return(error); 2064 } 2065 2066 2067 /* 2068 * Sets the inheritance of the specified address range in the target map. 2069 * Inheritance affects how the map will be shared with child maps at the 2070 * time of vm_map_fork. 2071 */ 2072 int 2073 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 2074 vm_inherit_t new_inheritance) 2075 { 2076 vm_map_entry_t entry; 2077 vm_map_entry_t temp_entry; 2078 int count; 2079 2080 switch (new_inheritance) { 2081 case VM_INHERIT_NONE: 2082 case VM_INHERIT_COPY: 2083 case VM_INHERIT_SHARE: 2084 break; 2085 default: 2086 return (KERN_INVALID_ARGUMENT); 2087 } 2088 2089 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2090 vm_map_lock(map); 2091 2092 VM_MAP_RANGE_CHECK(map, start, end); 2093 2094 if (vm_map_lookup_entry(map, start, &temp_entry)) { 2095 entry = temp_entry; 2096 vm_map_clip_start(map, entry, start, &count); 2097 } else 2098 entry = temp_entry->next; 2099 2100 while ((entry != &map->header) && (entry->start < end)) { 2101 vm_map_clip_end(map, entry, end, &count); 2102 2103 entry->inheritance = new_inheritance; 2104 2105 vm_map_simplify_entry(map, entry, &count); 2106 2107 entry = entry->next; 2108 } 2109 vm_map_unlock(map); 2110 vm_map_entry_release(count); 2111 return (KERN_SUCCESS); 2112 } 2113 2114 /* 2115 * Implement the semantics of mlock 2116 */ 2117 int 2118 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 2119 boolean_t new_pageable) 2120 { 2121 vm_map_entry_t entry; 2122 vm_map_entry_t start_entry; 2123 vm_offset_t end; 2124 int rv = KERN_SUCCESS; 2125 int count; 2126 2127 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2128 vm_map_lock(map); 2129 VM_MAP_RANGE_CHECK(map, start, real_end); 2130 end = real_end; 2131 2132 start_entry = vm_map_clip_range(map, start, end, &count, 2133 MAP_CLIP_NO_HOLES); 2134 if (start_entry == NULL) { 2135 vm_map_unlock(map); 2136 vm_map_entry_release(count); 2137 return (KERN_INVALID_ADDRESS); 2138 } 2139 2140 if (new_pageable == 0) { 2141 entry = start_entry; 2142 while ((entry != &map->header) && (entry->start < end)) { 2143 vm_offset_t save_start; 2144 vm_offset_t save_end; 2145 2146 /* 2147 * Already user wired or hard wired (trivial cases) 2148 */ 2149 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 2150 entry = entry->next; 2151 continue; 2152 } 2153 if (entry->wired_count != 0) { 2154 entry->wired_count++; 2155 entry->eflags |= MAP_ENTRY_USER_WIRED; 2156 entry = entry->next; 2157 continue; 2158 } 2159 2160 /* 2161 * A new wiring requires instantiation of appropriate 2162 * management structures and the faulting in of the 2163 * page. 2164 */ 2165 if (entry->maptype != VM_MAPTYPE_SUBMAP) { 2166 int copyflag = entry->eflags & 2167 MAP_ENTRY_NEEDS_COPY; 2168 if (copyflag && ((entry->protection & 2169 VM_PROT_WRITE) != 0)) { 2170 vm_map_entry_shadow(entry, 0); 2171 } else if (entry->object.vm_object == NULL && 2172 !map->system_map) { 2173 vm_map_entry_allocate_object(entry); 2174 } 2175 } 2176 entry->wired_count++; 2177 entry->eflags |= MAP_ENTRY_USER_WIRED; 2178 2179 /* 2180 * Now fault in the area. Note that vm_fault_wire() 2181 * may release the map lock temporarily, it will be 2182 * relocked on return. The in-transition 2183 * flag protects the entries. 2184 */ 2185 save_start = entry->start; 2186 save_end = entry->end; 2187 rv = vm_fault_wire(map, entry, TRUE); 2188 if (rv) { 2189 CLIP_CHECK_BACK(entry, save_start); 2190 for (;;) { 2191 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 2192 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2193 entry->wired_count = 0; 2194 if (entry->end == save_end) 2195 break; 2196 entry = entry->next; 2197 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2198 } 2199 end = save_start; /* unwire the rest */ 2200 break; 2201 } 2202 /* 2203 * note that even though the entry might have been 2204 * clipped, the USER_WIRED flag we set prevents 2205 * duplication so we do not have to do a 2206 * clip check. 2207 */ 2208 entry = entry->next; 2209 } 2210 2211 /* 2212 * If we failed fall through to the unwiring section to 2213 * unwire what we had wired so far. 'end' has already 2214 * been adjusted. 2215 */ 2216 if (rv) 2217 new_pageable = 1; 2218 2219 /* 2220 * start_entry might have been clipped if we unlocked the 2221 * map and blocked. No matter how clipped it has gotten 2222 * there should be a fragment that is on our start boundary. 2223 */ 2224 CLIP_CHECK_BACK(start_entry, start); 2225 } 2226 2227 /* 2228 * Deal with the unwiring case. 2229 */ 2230 if (new_pageable) { 2231 /* 2232 * This is the unwiring case. We must first ensure that the 2233 * range to be unwired is really wired down. We know there 2234 * are no holes. 2235 */ 2236 entry = start_entry; 2237 while ((entry != &map->header) && (entry->start < end)) { 2238 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 2239 rv = KERN_INVALID_ARGUMENT; 2240 goto done; 2241 } 2242 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry)); 2243 entry = entry->next; 2244 } 2245 2246 /* 2247 * Now decrement the wiring count for each region. If a region 2248 * becomes completely unwired, unwire its physical pages and 2249 * mappings. 2250 */ 2251 /* 2252 * The map entries are processed in a loop, checking to 2253 * make sure the entry is wired and asserting it has a wired 2254 * count. However, another loop was inserted more-or-less in 2255 * the middle of the unwiring path. This loop picks up the 2256 * "entry" loop variable from the first loop without first 2257 * setting it to start_entry. Naturally, the secound loop 2258 * is never entered and the pages backing the entries are 2259 * never unwired. This can lead to a leak of wired pages. 2260 */ 2261 entry = start_entry; 2262 while ((entry != &map->header) && (entry->start < end)) { 2263 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 2264 ("expected USER_WIRED on entry %p", entry)); 2265 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2266 entry->wired_count--; 2267 if (entry->wired_count == 0) 2268 vm_fault_unwire(map, entry); 2269 entry = entry->next; 2270 } 2271 } 2272 done: 2273 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2274 MAP_CLIP_NO_HOLES); 2275 map->timestamp++; 2276 vm_map_unlock(map); 2277 vm_map_entry_release(count); 2278 return (rv); 2279 } 2280 2281 /* 2282 * Sets the pageability of the specified address range in the target map. 2283 * Regions specified as not pageable require locked-down physical 2284 * memory and physical page maps. 2285 * 2286 * The map must not be locked, but a reference must remain to the map 2287 * throughout the call. 2288 * 2289 * This function may be called via the zalloc path and must properly 2290 * reserve map entries for kernel_map. 2291 * 2292 * No requirements. 2293 */ 2294 int 2295 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags) 2296 { 2297 vm_map_entry_t entry; 2298 vm_map_entry_t start_entry; 2299 vm_offset_t end; 2300 int rv = KERN_SUCCESS; 2301 int count; 2302 2303 if (kmflags & KM_KRESERVE) 2304 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 2305 else 2306 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2307 vm_map_lock(map); 2308 VM_MAP_RANGE_CHECK(map, start, real_end); 2309 end = real_end; 2310 2311 start_entry = vm_map_clip_range(map, start, end, &count, 2312 MAP_CLIP_NO_HOLES); 2313 if (start_entry == NULL) { 2314 vm_map_unlock(map); 2315 rv = KERN_INVALID_ADDRESS; 2316 goto failure; 2317 } 2318 if ((kmflags & KM_PAGEABLE) == 0) { 2319 /* 2320 * Wiring. 2321 * 2322 * 1. Holding the write lock, we create any shadow or zero-fill 2323 * objects that need to be created. Then we clip each map 2324 * entry to the region to be wired and increment its wiring 2325 * count. We create objects before clipping the map entries 2326 * to avoid object proliferation. 2327 * 2328 * 2. We downgrade to a read lock, and call vm_fault_wire to 2329 * fault in the pages for any newly wired area (wired_count is 2330 * 1). 2331 * 2332 * Downgrading to a read lock for vm_fault_wire avoids a 2333 * possible deadlock with another process that may have faulted 2334 * on one of the pages to be wired (it would mark the page busy, 2335 * blocking us, then in turn block on the map lock that we 2336 * hold). Because of problems in the recursive lock package, 2337 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 2338 * any actions that require the write lock must be done 2339 * beforehand. Because we keep the read lock on the map, the 2340 * copy-on-write status of the entries we modify here cannot 2341 * change. 2342 */ 2343 entry = start_entry; 2344 while ((entry != &map->header) && (entry->start < end)) { 2345 /* 2346 * Trivial case if the entry is already wired 2347 */ 2348 if (entry->wired_count) { 2349 entry->wired_count++; 2350 entry = entry->next; 2351 continue; 2352 } 2353 2354 /* 2355 * The entry is being newly wired, we have to setup 2356 * appropriate management structures. A shadow 2357 * object is required for a copy-on-write region, 2358 * or a normal object for a zero-fill region. We 2359 * do not have to do this for entries that point to sub 2360 * maps because we won't hold the lock on the sub map. 2361 */ 2362 if (entry->maptype != VM_MAPTYPE_SUBMAP) { 2363 int copyflag = entry->eflags & 2364 MAP_ENTRY_NEEDS_COPY; 2365 if (copyflag && ((entry->protection & 2366 VM_PROT_WRITE) != 0)) { 2367 vm_map_entry_shadow(entry, 0); 2368 } else if (entry->object.vm_object == NULL && 2369 !map->system_map) { 2370 vm_map_entry_allocate_object(entry); 2371 } 2372 } 2373 2374 entry->wired_count++; 2375 entry = entry->next; 2376 } 2377 2378 /* 2379 * Pass 2. 2380 */ 2381 2382 /* 2383 * HACK HACK HACK HACK 2384 * 2385 * vm_fault_wire() temporarily unlocks the map to avoid 2386 * deadlocks. The in-transition flag from vm_map_clip_range 2387 * call should protect us from changes while the map is 2388 * unlocked. T 2389 * 2390 * NOTE: Previously this comment stated that clipping might 2391 * still occur while the entry is unlocked, but from 2392 * what I can tell it actually cannot. 2393 * 2394 * It is unclear whether the CLIP_CHECK_*() calls 2395 * are still needed but we keep them in anyway. 2396 * 2397 * HACK HACK HACK HACK 2398 */ 2399 2400 entry = start_entry; 2401 while (entry != &map->header && entry->start < end) { 2402 /* 2403 * If vm_fault_wire fails for any page we need to undo 2404 * what has been done. We decrement the wiring count 2405 * for those pages which have not yet been wired (now) 2406 * and unwire those that have (later). 2407 */ 2408 vm_offset_t save_start = entry->start; 2409 vm_offset_t save_end = entry->end; 2410 2411 if (entry->wired_count == 1) 2412 rv = vm_fault_wire(map, entry, FALSE); 2413 if (rv) { 2414 CLIP_CHECK_BACK(entry, save_start); 2415 for (;;) { 2416 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly")); 2417 entry->wired_count = 0; 2418 if (entry->end == save_end) 2419 break; 2420 entry = entry->next; 2421 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2422 } 2423 end = save_start; 2424 break; 2425 } 2426 CLIP_CHECK_FWD(entry, save_end); 2427 entry = entry->next; 2428 } 2429 2430 /* 2431 * If a failure occured undo everything by falling through 2432 * to the unwiring code. 'end' has already been adjusted 2433 * appropriately. 2434 */ 2435 if (rv) 2436 kmflags |= KM_PAGEABLE; 2437 2438 /* 2439 * start_entry is still IN_TRANSITION but may have been 2440 * clipped since vm_fault_wire() unlocks and relocks the 2441 * map. No matter how clipped it has gotten there should 2442 * be a fragment that is on our start boundary. 2443 */ 2444 CLIP_CHECK_BACK(start_entry, start); 2445 } 2446 2447 if (kmflags & KM_PAGEABLE) { 2448 /* 2449 * This is the unwiring case. We must first ensure that the 2450 * range to be unwired is really wired down. We know there 2451 * are no holes. 2452 */ 2453 entry = start_entry; 2454 while ((entry != &map->header) && (entry->start < end)) { 2455 if (entry->wired_count == 0) { 2456 rv = KERN_INVALID_ARGUMENT; 2457 goto done; 2458 } 2459 entry = entry->next; 2460 } 2461 2462 /* 2463 * Now decrement the wiring count for each region. If a region 2464 * becomes completely unwired, unwire its physical pages and 2465 * mappings. 2466 */ 2467 entry = start_entry; 2468 while ((entry != &map->header) && (entry->start < end)) { 2469 entry->wired_count--; 2470 if (entry->wired_count == 0) 2471 vm_fault_unwire(map, entry); 2472 entry = entry->next; 2473 } 2474 } 2475 done: 2476 vm_map_unclip_range(map, start_entry, start, real_end, 2477 &count, MAP_CLIP_NO_HOLES); 2478 map->timestamp++; 2479 vm_map_unlock(map); 2480 failure: 2481 if (kmflags & KM_KRESERVE) 2482 vm_map_entry_krelease(count); 2483 else 2484 vm_map_entry_release(count); 2485 return (rv); 2486 } 2487 2488 /* 2489 * Mark a newly allocated address range as wired but do not fault in 2490 * the pages. The caller is expected to load the pages into the object. 2491 * 2492 * The map must be locked on entry and will remain locked on return. 2493 * No other requirements. 2494 */ 2495 void 2496 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, 2497 int *countp) 2498 { 2499 vm_map_entry_t scan; 2500 vm_map_entry_t entry; 2501 2502 entry = vm_map_clip_range(map, addr, addr + size, 2503 countp, MAP_CLIP_NO_HOLES); 2504 for (scan = entry; 2505 scan != &map->header && scan->start < addr + size; 2506 scan = scan->next) { 2507 KKASSERT(scan->wired_count == 0); 2508 scan->wired_count = 1; 2509 } 2510 vm_map_unclip_range(map, entry, addr, addr + size, 2511 countp, MAP_CLIP_NO_HOLES); 2512 } 2513 2514 /* 2515 * Push any dirty cached pages in the address range to their pager. 2516 * If syncio is TRUE, dirty pages are written synchronously. 2517 * If invalidate is TRUE, any cached pages are freed as well. 2518 * 2519 * This routine is called by sys_msync() 2520 * 2521 * Returns an error if any part of the specified range is not mapped. 2522 * 2523 * No requirements. 2524 */ 2525 int 2526 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, 2527 boolean_t syncio, boolean_t invalidate) 2528 { 2529 vm_map_entry_t current; 2530 vm_map_entry_t entry; 2531 vm_size_t size; 2532 vm_object_t object; 2533 vm_object_t tobj; 2534 vm_ooffset_t offset; 2535 2536 vm_map_lock_read(map); 2537 VM_MAP_RANGE_CHECK(map, start, end); 2538 if (!vm_map_lookup_entry(map, start, &entry)) { 2539 vm_map_unlock_read(map); 2540 return (KERN_INVALID_ADDRESS); 2541 } 2542 lwkt_gettoken(&map->token); 2543 2544 /* 2545 * Make a first pass to check for holes. 2546 */ 2547 for (current = entry; current->start < end; current = current->next) { 2548 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2549 lwkt_reltoken(&map->token); 2550 vm_map_unlock_read(map); 2551 return (KERN_INVALID_ARGUMENT); 2552 } 2553 if (end > current->end && 2554 (current->next == &map->header || 2555 current->end != current->next->start)) { 2556 lwkt_reltoken(&map->token); 2557 vm_map_unlock_read(map); 2558 return (KERN_INVALID_ADDRESS); 2559 } 2560 } 2561 2562 if (invalidate) 2563 pmap_remove(vm_map_pmap(map), start, end); 2564 2565 /* 2566 * Make a second pass, cleaning/uncaching pages from the indicated 2567 * objects as we go. 2568 */ 2569 for (current = entry; current->start < end; current = current->next) { 2570 offset = current->offset + (start - current->start); 2571 size = (end <= current->end ? end : current->end) - start; 2572 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2573 vm_map_t smap; 2574 vm_map_entry_t tentry; 2575 vm_size_t tsize; 2576 2577 smap = current->object.sub_map; 2578 vm_map_lock_read(smap); 2579 vm_map_lookup_entry(smap, offset, &tentry); 2580 tsize = tentry->end - offset; 2581 if (tsize < size) 2582 size = tsize; 2583 object = tentry->object.vm_object; 2584 offset = tentry->offset + (offset - tentry->start); 2585 vm_map_unlock_read(smap); 2586 } else { 2587 object = current->object.vm_object; 2588 } 2589 2590 if (object) 2591 vm_object_hold(object); 2592 2593 /* 2594 * Note that there is absolutely no sense in writing out 2595 * anonymous objects, so we track down the vnode object 2596 * to write out. 2597 * We invalidate (remove) all pages from the address space 2598 * anyway, for semantic correctness. 2599 * 2600 * note: certain anonymous maps, such as MAP_NOSYNC maps, 2601 * may start out with a NULL object. 2602 */ 2603 while (object && (tobj = object->backing_object) != NULL) { 2604 vm_object_hold(tobj); 2605 if (tobj == object->backing_object) { 2606 vm_object_lock_swap(); 2607 offset += object->backing_object_offset; 2608 vm_object_drop(object); 2609 object = tobj; 2610 if (object->size < OFF_TO_IDX(offset + size)) 2611 size = IDX_TO_OFF(object->size) - 2612 offset; 2613 break; 2614 } 2615 vm_object_drop(tobj); 2616 } 2617 if (object && (object->type == OBJT_VNODE) && 2618 (current->protection & VM_PROT_WRITE) && 2619 (object->flags & OBJ_NOMSYNC) == 0) { 2620 /* 2621 * Flush pages if writing is allowed, invalidate them 2622 * if invalidation requested. Pages undergoing I/O 2623 * will be ignored by vm_object_page_remove(). 2624 * 2625 * We cannot lock the vnode and then wait for paging 2626 * to complete without deadlocking against vm_fault. 2627 * Instead we simply call vm_object_page_remove() and 2628 * allow it to block internally on a page-by-page 2629 * basis when it encounters pages undergoing async 2630 * I/O. 2631 */ 2632 int flags; 2633 2634 /* no chain wait needed for vnode objects */ 2635 vm_object_reference_locked(object); 2636 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY); 2637 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 2638 flags |= invalidate ? OBJPC_INVAL : 0; 2639 2640 /* 2641 * When operating on a virtual page table just 2642 * flush the whole object. XXX we probably ought 2643 * to 2644 */ 2645 switch(current->maptype) { 2646 case VM_MAPTYPE_NORMAL: 2647 vm_object_page_clean(object, 2648 OFF_TO_IDX(offset), 2649 OFF_TO_IDX(offset + size + PAGE_MASK), 2650 flags); 2651 break; 2652 case VM_MAPTYPE_VPAGETABLE: 2653 vm_object_page_clean(object, 0, 0, flags); 2654 break; 2655 } 2656 vn_unlock(((struct vnode *)object->handle)); 2657 vm_object_deallocate_locked(object); 2658 } 2659 if (object && invalidate && 2660 ((object->type == OBJT_VNODE) || 2661 (object->type == OBJT_DEVICE))) { 2662 int clean_only = 2663 (object->type == OBJT_DEVICE) ? FALSE : TRUE; 2664 /* no chain wait needed for vnode/device objects */ 2665 vm_object_reference_locked(object); 2666 switch(current->maptype) { 2667 case VM_MAPTYPE_NORMAL: 2668 vm_object_page_remove(object, 2669 OFF_TO_IDX(offset), 2670 OFF_TO_IDX(offset + size + PAGE_MASK), 2671 clean_only); 2672 break; 2673 case VM_MAPTYPE_VPAGETABLE: 2674 vm_object_page_remove(object, 0, 0, clean_only); 2675 break; 2676 } 2677 vm_object_deallocate_locked(object); 2678 } 2679 start += size; 2680 if (object) 2681 vm_object_drop(object); 2682 } 2683 2684 lwkt_reltoken(&map->token); 2685 vm_map_unlock_read(map); 2686 2687 return (KERN_SUCCESS); 2688 } 2689 2690 /* 2691 * Make the region specified by this entry pageable. 2692 * 2693 * The vm_map must be exclusively locked. 2694 */ 2695 static void 2696 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 2697 { 2698 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2699 entry->wired_count = 0; 2700 vm_fault_unwire(map, entry); 2701 } 2702 2703 /* 2704 * Deallocate the given entry from the target map. 2705 * 2706 * The vm_map must be exclusively locked. 2707 */ 2708 static void 2709 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 2710 { 2711 vm_map_entry_unlink(map, entry); 2712 map->size -= entry->end - entry->start; 2713 2714 switch(entry->maptype) { 2715 case VM_MAPTYPE_NORMAL: 2716 case VM_MAPTYPE_VPAGETABLE: 2717 vm_object_deallocate(entry->object.vm_object); 2718 break; 2719 default: 2720 break; 2721 } 2722 2723 vm_map_entry_dispose(map, entry, countp); 2724 } 2725 2726 /* 2727 * Deallocates the given address range from the target map. 2728 * 2729 * The vm_map must be exclusively locked. 2730 */ 2731 int 2732 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 2733 { 2734 vm_object_t object; 2735 vm_map_entry_t entry; 2736 vm_map_entry_t first_entry; 2737 2738 ASSERT_VM_MAP_LOCKED(map); 2739 lwkt_gettoken(&map->token); 2740 again: 2741 /* 2742 * Find the start of the region, and clip it. Set entry to point 2743 * at the first record containing the requested address or, if no 2744 * such record exists, the next record with a greater address. The 2745 * loop will run from this point until a record beyond the termination 2746 * address is encountered. 2747 * 2748 * map->hint must be adjusted to not point to anything we delete, 2749 * so set it to the entry prior to the one being deleted. 2750 * 2751 * GGG see other GGG comment. 2752 */ 2753 if (vm_map_lookup_entry(map, start, &first_entry)) { 2754 entry = first_entry; 2755 vm_map_clip_start(map, entry, start, countp); 2756 map->hint = entry->prev; /* possible problem XXX */ 2757 } else { 2758 map->hint = first_entry; /* possible problem XXX */ 2759 entry = first_entry->next; 2760 } 2761 2762 /* 2763 * If a hole opens up prior to the current first_free then 2764 * adjust first_free. As with map->hint, map->first_free 2765 * cannot be left set to anything we might delete. 2766 */ 2767 if (entry == &map->header) { 2768 map->first_free = &map->header; 2769 } else if (map->first_free->start >= start) { 2770 map->first_free = entry->prev; 2771 } 2772 2773 /* 2774 * Step through all entries in this region 2775 */ 2776 while ((entry != &map->header) && (entry->start < end)) { 2777 vm_map_entry_t next; 2778 vm_offset_t s, e; 2779 vm_pindex_t offidxstart, offidxend, count; 2780 2781 /* 2782 * If we hit an in-transition entry we have to sleep and 2783 * retry. It's easier (and not really slower) to just retry 2784 * since this case occurs so rarely and the hint is already 2785 * pointing at the right place. We have to reset the 2786 * start offset so as not to accidently delete an entry 2787 * another process just created in vacated space. 2788 */ 2789 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 2790 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2791 start = entry->start; 2792 ++mycpu->gd_cnt.v_intrans_coll; 2793 ++mycpu->gd_cnt.v_intrans_wait; 2794 vm_map_transition_wait(map); 2795 goto again; 2796 } 2797 vm_map_clip_end(map, entry, end, countp); 2798 2799 s = entry->start; 2800 e = entry->end; 2801 next = entry->next; 2802 2803 offidxstart = OFF_TO_IDX(entry->offset); 2804 count = OFF_TO_IDX(e - s); 2805 object = entry->object.vm_object; 2806 2807 /* 2808 * Unwire before removing addresses from the pmap; otherwise, 2809 * unwiring will put the entries back in the pmap. 2810 */ 2811 if (entry->wired_count != 0) 2812 vm_map_entry_unwire(map, entry); 2813 2814 offidxend = offidxstart + count; 2815 2816 if (object == &kernel_object) { 2817 vm_object_hold(object); 2818 vm_object_page_remove(object, offidxstart, 2819 offidxend, FALSE); 2820 vm_object_drop(object); 2821 } else if (object && object->type != OBJT_DEFAULT && 2822 object->type != OBJT_SWAP) { 2823 /* 2824 * vnode object routines cannot be chain-locked, 2825 * but since we aren't removing pages from the 2826 * object here we can use a shared hold. 2827 */ 2828 vm_object_hold_shared(object); 2829 pmap_remove(map->pmap, s, e); 2830 vm_object_drop(object); 2831 } else if (object) { 2832 vm_object_hold(object); 2833 vm_object_chain_acquire(object); 2834 pmap_remove(map->pmap, s, e); 2835 2836 if (object != NULL && 2837 object->ref_count != 1 && 2838 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == 2839 OBJ_ONEMAPPING && 2840 (object->type == OBJT_DEFAULT || 2841 object->type == OBJT_SWAP)) { 2842 vm_object_collapse(object, NULL); 2843 vm_object_page_remove(object, offidxstart, 2844 offidxend, FALSE); 2845 if (object->type == OBJT_SWAP) { 2846 swap_pager_freespace(object, 2847 offidxstart, 2848 count); 2849 } 2850 if (offidxend >= object->size && 2851 offidxstart < object->size) { 2852 object->size = offidxstart; 2853 } 2854 } 2855 vm_object_chain_release(object); 2856 vm_object_drop(object); 2857 } 2858 2859 /* 2860 * Delete the entry (which may delete the object) only after 2861 * removing all pmap entries pointing to its pages. 2862 * (Otherwise, its page frames may be reallocated, and any 2863 * modify bits will be set in the wrong object!) 2864 */ 2865 vm_map_entry_delete(map, entry, countp); 2866 entry = next; 2867 } 2868 lwkt_reltoken(&map->token); 2869 return (KERN_SUCCESS); 2870 } 2871 2872 /* 2873 * Remove the given address range from the target map. 2874 * This is the exported form of vm_map_delete. 2875 * 2876 * No requirements. 2877 */ 2878 int 2879 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 2880 { 2881 int result; 2882 int count; 2883 2884 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2885 vm_map_lock(map); 2886 VM_MAP_RANGE_CHECK(map, start, end); 2887 result = vm_map_delete(map, start, end, &count); 2888 vm_map_unlock(map); 2889 vm_map_entry_release(count); 2890 2891 return (result); 2892 } 2893 2894 /* 2895 * Assert that the target map allows the specified privilege on the 2896 * entire address region given. The entire region must be allocated. 2897 * 2898 * The caller must specify whether the vm_map is already locked or not. 2899 */ 2900 boolean_t 2901 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 2902 vm_prot_t protection, boolean_t have_lock) 2903 { 2904 vm_map_entry_t entry; 2905 vm_map_entry_t tmp_entry; 2906 boolean_t result; 2907 2908 if (have_lock == FALSE) 2909 vm_map_lock_read(map); 2910 2911 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 2912 if (have_lock == FALSE) 2913 vm_map_unlock_read(map); 2914 return (FALSE); 2915 } 2916 entry = tmp_entry; 2917 2918 result = TRUE; 2919 while (start < end) { 2920 if (entry == &map->header) { 2921 result = FALSE; 2922 break; 2923 } 2924 /* 2925 * No holes allowed! 2926 */ 2927 2928 if (start < entry->start) { 2929 result = FALSE; 2930 break; 2931 } 2932 /* 2933 * Check protection associated with entry. 2934 */ 2935 2936 if ((entry->protection & protection) != protection) { 2937 result = FALSE; 2938 break; 2939 } 2940 /* go to next entry */ 2941 2942 start = entry->end; 2943 entry = entry->next; 2944 } 2945 if (have_lock == FALSE) 2946 vm_map_unlock_read(map); 2947 return (result); 2948 } 2949 2950 /* 2951 * If appropriate this function shadows the original object with a new object 2952 * and moves the VM pages from the original object to the new object. 2953 * The original object will also be collapsed, if possible. 2954 * 2955 * We can only do this for normal memory objects with a single mapping, and 2956 * it only makes sense to do it if there are 2 or more refs on the original 2957 * object. i.e. typically a memory object that has been extended into 2958 * multiple vm_map_entry's with non-overlapping ranges. 2959 * 2960 * This makes it easier to remove unused pages and keeps object inheritance 2961 * from being a negative impact on memory usage. 2962 * 2963 * On return the (possibly new) entry->object.vm_object will have an 2964 * additional ref on it for the caller to dispose of (usually by cloning 2965 * the vm_map_entry). The additional ref had to be done in this routine 2966 * to avoid racing a collapse. The object's ONEMAPPING flag will also be 2967 * cleared. 2968 * 2969 * The vm_map must be locked and its token held. 2970 */ 2971 static void 2972 vm_map_split(vm_map_entry_t entry) 2973 { 2974 #if 0 2975 /* UNOPTIMIZED */ 2976 vm_object_t oobject; 2977 2978 oobject = entry->object.vm_object; 2979 vm_object_hold(oobject); 2980 vm_object_chain_wait(oobject); 2981 vm_object_reference_locked(oobject); 2982 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 2983 vm_object_drop(oobject); 2984 #else 2985 /* OPTIMIZED */ 2986 vm_object_t oobject, nobject, bobject; 2987 vm_offset_t s, e; 2988 vm_page_t m; 2989 vm_pindex_t offidxstart, offidxend, idx; 2990 vm_size_t size; 2991 vm_ooffset_t offset; 2992 2993 /* 2994 * Setup. Chain lock the original object throughout the entire 2995 * routine to prevent new page faults from occuring. 2996 * 2997 * XXX can madvise WILLNEED interfere with us too? 2998 */ 2999 oobject = entry->object.vm_object; 3000 vm_object_hold(oobject); 3001 vm_object_chain_acquire(oobject); 3002 3003 /* 3004 * Original object cannot be split? 3005 */ 3006 if (oobject->handle == NULL || (oobject->type != OBJT_DEFAULT && 3007 oobject->type != OBJT_SWAP)) { 3008 vm_object_chain_release(oobject); 3009 vm_object_reference_locked(oobject); 3010 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3011 vm_object_drop(oobject); 3012 return; 3013 } 3014 3015 /* 3016 * Collapse original object with its backing store as an 3017 * optimization to reduce chain lengths when possible. 3018 * 3019 * If ref_count <= 1 there aren't other non-overlapping vm_map_entry's 3020 * for oobject, so there's no point collapsing it. 3021 * 3022 * Then re-check whether the object can be split. 3023 */ 3024 vm_object_collapse(oobject, NULL); 3025 3026 if (oobject->ref_count <= 1 || 3027 (oobject->type != OBJT_DEFAULT && oobject->type != OBJT_SWAP) || 3028 (oobject->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) != OBJ_ONEMAPPING) { 3029 vm_object_chain_release(oobject); 3030 vm_object_reference_locked(oobject); 3031 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3032 vm_object_drop(oobject); 3033 return; 3034 } 3035 3036 /* 3037 * Acquire the chain lock on the backing object. 3038 * 3039 * Give bobject an additional ref count for when it will be shadowed 3040 * by nobject. 3041 */ 3042 if ((bobject = oobject->backing_object) != NULL) { 3043 vm_object_hold(bobject); 3044 vm_object_chain_wait(bobject); 3045 vm_object_reference_locked(bobject); 3046 vm_object_chain_acquire(bobject); 3047 KKASSERT(bobject->backing_object == bobject); 3048 KKASSERT((bobject->flags & OBJ_DEAD) == 0); 3049 } 3050 3051 /* 3052 * Calculate the object page range and allocate the new object. 3053 */ 3054 offset = entry->offset; 3055 s = entry->start; 3056 e = entry->end; 3057 3058 offidxstart = OFF_TO_IDX(offset); 3059 offidxend = offidxstart + OFF_TO_IDX(e - s); 3060 size = offidxend - offidxstart; 3061 3062 switch(oobject->type) { 3063 case OBJT_DEFAULT: 3064 nobject = default_pager_alloc(NULL, IDX_TO_OFF(size), 3065 VM_PROT_ALL, 0); 3066 break; 3067 case OBJT_SWAP: 3068 nobject = swap_pager_alloc(NULL, IDX_TO_OFF(size), 3069 VM_PROT_ALL, 0); 3070 break; 3071 default: 3072 /* not reached */ 3073 nobject = NULL; 3074 KKASSERT(0); 3075 } 3076 3077 if (nobject == NULL) { 3078 if (bobject) { 3079 vm_object_chain_release(bobject); 3080 vm_object_deallocate(bobject); 3081 vm_object_drop(bobject); 3082 } 3083 vm_object_chain_release(oobject); 3084 vm_object_reference_locked(oobject); 3085 vm_object_clear_flag(oobject, OBJ_ONEMAPPING); 3086 vm_object_drop(oobject); 3087 return; 3088 } 3089 3090 /* 3091 * The new object will replace entry->object.vm_object so it needs 3092 * a second reference (the caller expects an additional ref). 3093 */ 3094 vm_object_hold(nobject); 3095 vm_object_reference_locked(nobject); 3096 vm_object_chain_acquire(nobject); 3097 3098 /* 3099 * nobject shadows bobject (oobject already shadows bobject). 3100 */ 3101 if (bobject) { 3102 nobject->backing_object_offset = 3103 oobject->backing_object_offset + IDX_TO_OFF(offidxstart); 3104 nobject->backing_object = bobject; 3105 bobject->shadow_count++; 3106 bobject->generation++; 3107 LIST_INSERT_HEAD(&bobject->shadow_head, nobject, shadow_list); 3108 vm_object_clear_flag(bobject, OBJ_ONEMAPPING); /* XXX? */ 3109 vm_object_chain_release(bobject); 3110 vm_object_drop(bobject); 3111 } 3112 3113 /* 3114 * Move the VM pages from oobject to nobject 3115 */ 3116 for (idx = 0; idx < size; idx++) { 3117 vm_page_t m; 3118 3119 m = vm_page_lookup_busy_wait(oobject, offidxstart + idx, 3120 TRUE, "vmpg"); 3121 if (m == NULL) 3122 continue; 3123 3124 /* 3125 * We must wait for pending I/O to complete before we can 3126 * rename the page. 3127 * 3128 * We do not have to VM_PROT_NONE the page as mappings should 3129 * not be changed by this operation. 3130 * 3131 * NOTE: The act of renaming a page updates chaingen for both 3132 * objects. 3133 */ 3134 vm_page_rename(m, nobject, idx); 3135 /* page automatically made dirty by rename and cache handled */ 3136 /* page remains busy */ 3137 } 3138 3139 if (oobject->type == OBJT_SWAP) { 3140 vm_object_pip_add(oobject, 1); 3141 /* 3142 * copy oobject pages into nobject and destroy unneeded 3143 * pages in shadow object. 3144 */ 3145 swap_pager_copy(oobject, nobject, offidxstart, 0); 3146 vm_object_pip_wakeup(oobject); 3147 } 3148 3149 /* 3150 * Wakeup the pages we played with. No spl protection is needed 3151 * for a simple wakeup. 3152 */ 3153 for (idx = 0; idx < size; idx++) { 3154 m = vm_page_lookup(nobject, idx); 3155 if (m) { 3156 KKASSERT(m->flags & PG_BUSY); 3157 vm_page_wakeup(m); 3158 } 3159 } 3160 entry->object.vm_object = nobject; 3161 entry->offset = 0LL; 3162 3163 /* 3164 * Cleanup 3165 * 3166 * NOTE: There is no need to remove OBJ_ONEMAPPING from oobject, the 3167 * related pages were moved and are no longer applicable to the 3168 * original object. 3169 * 3170 * NOTE: Deallocate oobject (due to its entry->object.vm_object being 3171 * replaced by nobject). 3172 */ 3173 vm_object_chain_release(nobject); 3174 vm_object_drop(nobject); 3175 if (bobject) { 3176 vm_object_chain_release(bobject); 3177 vm_object_drop(bobject); 3178 } 3179 vm_object_chain_release(oobject); 3180 /*vm_object_clear_flag(oobject, OBJ_ONEMAPPING);*/ 3181 vm_object_deallocate_locked(oobject); 3182 vm_object_drop(oobject); 3183 #endif 3184 } 3185 3186 /* 3187 * Copies the contents of the source entry to the destination 3188 * entry. The entries *must* be aligned properly. 3189 * 3190 * The vm_maps must be exclusively locked. 3191 * The vm_map's token must be held. 3192 * 3193 * Because the maps are locked no faults can be in progress during the 3194 * operation. 3195 */ 3196 static void 3197 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 3198 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 3199 { 3200 vm_object_t src_object; 3201 3202 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP) 3203 return; 3204 if (src_entry->maptype == VM_MAPTYPE_SUBMAP) 3205 return; 3206 3207 if (src_entry->wired_count == 0) { 3208 /* 3209 * If the source entry is marked needs_copy, it is already 3210 * write-protected. 3211 */ 3212 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 3213 pmap_protect(src_map->pmap, 3214 src_entry->start, 3215 src_entry->end, 3216 src_entry->protection & ~VM_PROT_WRITE); 3217 } 3218 3219 /* 3220 * Make a copy of the object. 3221 * 3222 * The object must be locked prior to checking the object type 3223 * and for the call to vm_object_collapse() and vm_map_split(). 3224 * We cannot use *_hold() here because the split code will 3225 * probably try to destroy the object. The lock is a pool 3226 * token and doesn't care. 3227 * 3228 * We must bump src_map->timestamp when setting 3229 * MAP_ENTRY_NEEDS_COPY to force any concurrent fault 3230 * to retry, otherwise the concurrent fault might improperly 3231 * install a RW pte when its supposed to be a RO(COW) pte. 3232 * This race can occur because a vnode-backed fault may have 3233 * to temporarily release the map lock. 3234 */ 3235 if (src_entry->object.vm_object != NULL) { 3236 vm_map_split(src_entry); 3237 src_object = src_entry->object.vm_object; 3238 dst_entry->object.vm_object = src_object; 3239 src_entry->eflags |= (MAP_ENTRY_COW | 3240 MAP_ENTRY_NEEDS_COPY); 3241 dst_entry->eflags |= (MAP_ENTRY_COW | 3242 MAP_ENTRY_NEEDS_COPY); 3243 dst_entry->offset = src_entry->offset; 3244 ++src_map->timestamp; 3245 } else { 3246 dst_entry->object.vm_object = NULL; 3247 dst_entry->offset = 0; 3248 } 3249 3250 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, 3251 dst_entry->end - dst_entry->start, src_entry->start); 3252 } else { 3253 /* 3254 * Of course, wired down pages can't be set copy-on-write. 3255 * Cause wired pages to be copied into the new map by 3256 * simulating faults (the new pages are pageable) 3257 */ 3258 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 3259 } 3260 } 3261 3262 /* 3263 * vmspace_fork: 3264 * Create a new process vmspace structure and vm_map 3265 * based on those of an existing process. The new map 3266 * is based on the old map, according to the inheritance 3267 * values on the regions in that map. 3268 * 3269 * The source map must not be locked. 3270 * No requirements. 3271 */ 3272 struct vmspace * 3273 vmspace_fork(struct vmspace *vm1) 3274 { 3275 struct vmspace *vm2; 3276 vm_map_t old_map = &vm1->vm_map; 3277 vm_map_t new_map; 3278 vm_map_entry_t old_entry; 3279 vm_map_entry_t new_entry; 3280 vm_object_t object; 3281 int count; 3282 3283 lwkt_gettoken(&vm1->vm_map.token); 3284 vm_map_lock(old_map); 3285 3286 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset); 3287 lwkt_gettoken(&vm2->vm_map.token); 3288 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 3289 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 3290 new_map = &vm2->vm_map; /* XXX */ 3291 new_map->timestamp = 1; 3292 3293 vm_map_lock(new_map); 3294 3295 count = 0; 3296 old_entry = old_map->header.next; 3297 while (old_entry != &old_map->header) { 3298 ++count; 3299 old_entry = old_entry->next; 3300 } 3301 3302 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 3303 3304 old_entry = old_map->header.next; 3305 while (old_entry != &old_map->header) { 3306 if (old_entry->maptype == VM_MAPTYPE_SUBMAP) 3307 panic("vm_map_fork: encountered a submap"); 3308 3309 switch (old_entry->inheritance) { 3310 case VM_INHERIT_NONE: 3311 break; 3312 case VM_INHERIT_SHARE: 3313 /* 3314 * Clone the entry, creating the shared object if 3315 * necessary. 3316 */ 3317 if (old_entry->object.vm_object == NULL) 3318 vm_map_entry_allocate_object(old_entry); 3319 3320 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3321 /* 3322 * Shadow a map_entry which needs a copy, 3323 * replacing its object with a new object 3324 * that points to the old one. Ask the 3325 * shadow code to automatically add an 3326 * additional ref. We can't do it afterwords 3327 * because we might race a collapse. The call 3328 * to vm_map_entry_shadow() will also clear 3329 * OBJ_ONEMAPPING. 3330 */ 3331 vm_map_entry_shadow(old_entry, 1); 3332 } else { 3333 /* 3334 * We will make a shared copy of the object, 3335 * and must clear OBJ_ONEMAPPING. 3336 * 3337 * XXX assert that object.vm_object != NULL 3338 * since we allocate it above. 3339 */ 3340 if (old_entry->object.vm_object) { 3341 object = old_entry->object.vm_object; 3342 vm_object_hold(object); 3343 vm_object_chain_wait(object); 3344 vm_object_reference_locked(object); 3345 vm_object_clear_flag(object, 3346 OBJ_ONEMAPPING); 3347 vm_object_drop(object); 3348 } 3349 } 3350 3351 /* 3352 * Clone the entry. We've already bumped the ref on 3353 * any vm_object. 3354 */ 3355 new_entry = vm_map_entry_create(new_map, &count); 3356 *new_entry = *old_entry; 3357 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3358 new_entry->wired_count = 0; 3359 3360 /* 3361 * Insert the entry into the new map -- we know we're 3362 * inserting at the end of the new map. 3363 */ 3364 3365 vm_map_entry_link(new_map, new_map->header.prev, 3366 new_entry); 3367 3368 /* 3369 * Update the physical map 3370 */ 3371 pmap_copy(new_map->pmap, old_map->pmap, 3372 new_entry->start, 3373 (old_entry->end - old_entry->start), 3374 old_entry->start); 3375 break; 3376 case VM_INHERIT_COPY: 3377 /* 3378 * Clone the entry and link into the map. 3379 */ 3380 new_entry = vm_map_entry_create(new_map, &count); 3381 *new_entry = *old_entry; 3382 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3383 new_entry->wired_count = 0; 3384 new_entry->object.vm_object = NULL; 3385 vm_map_entry_link(new_map, new_map->header.prev, 3386 new_entry); 3387 vm_map_copy_entry(old_map, new_map, old_entry, 3388 new_entry); 3389 break; 3390 } 3391 old_entry = old_entry->next; 3392 } 3393 3394 new_map->size = old_map->size; 3395 vm_map_unlock(old_map); 3396 vm_map_unlock(new_map); 3397 vm_map_entry_release(count); 3398 3399 lwkt_reltoken(&vm2->vm_map.token); 3400 lwkt_reltoken(&vm1->vm_map.token); 3401 3402 return (vm2); 3403 } 3404 3405 /* 3406 * Create an auto-grow stack entry 3407 * 3408 * No requirements. 3409 */ 3410 int 3411 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 3412 int flags, vm_prot_t prot, vm_prot_t max, int cow) 3413 { 3414 vm_map_entry_t prev_entry; 3415 vm_map_entry_t new_stack_entry; 3416 vm_size_t init_ssize; 3417 int rv; 3418 int count; 3419 vm_offset_t tmpaddr; 3420 3421 cow |= MAP_IS_STACK; 3422 3423 if (max_ssize < sgrowsiz) 3424 init_ssize = max_ssize; 3425 else 3426 init_ssize = sgrowsiz; 3427 3428 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3429 vm_map_lock(map); 3430 3431 /* 3432 * Find space for the mapping 3433 */ 3434 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) { 3435 if (vm_map_findspace(map, addrbos, max_ssize, 1, 3436 flags, &tmpaddr)) { 3437 vm_map_unlock(map); 3438 vm_map_entry_release(count); 3439 return (KERN_NO_SPACE); 3440 } 3441 addrbos = tmpaddr; 3442 } 3443 3444 /* If addr is already mapped, no go */ 3445 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) { 3446 vm_map_unlock(map); 3447 vm_map_entry_release(count); 3448 return (KERN_NO_SPACE); 3449 } 3450 3451 #if 0 3452 /* XXX already handled by kern_mmap() */ 3453 /* If we would blow our VMEM resource limit, no go */ 3454 if (map->size + init_ssize > 3455 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3456 vm_map_unlock(map); 3457 vm_map_entry_release(count); 3458 return (KERN_NO_SPACE); 3459 } 3460 #endif 3461 3462 /* 3463 * If we can't accomodate max_ssize in the current mapping, 3464 * no go. However, we need to be aware that subsequent user 3465 * mappings might map into the space we have reserved for 3466 * stack, and currently this space is not protected. 3467 * 3468 * Hopefully we will at least detect this condition 3469 * when we try to grow the stack. 3470 */ 3471 if ((prev_entry->next != &map->header) && 3472 (prev_entry->next->start < addrbos + max_ssize)) { 3473 vm_map_unlock(map); 3474 vm_map_entry_release(count); 3475 return (KERN_NO_SPACE); 3476 } 3477 3478 /* 3479 * We initially map a stack of only init_ssize. We will 3480 * grow as needed later. Since this is to be a grow 3481 * down stack, we map at the top of the range. 3482 * 3483 * Note: we would normally expect prot and max to be 3484 * VM_PROT_ALL, and cow to be 0. Possibly we should 3485 * eliminate these as input parameters, and just 3486 * pass these values here in the insert call. 3487 */ 3488 rv = vm_map_insert(map, &count, 3489 NULL, 0, addrbos + max_ssize - init_ssize, 3490 addrbos + max_ssize, 3491 VM_MAPTYPE_NORMAL, 3492 prot, max, 3493 cow); 3494 3495 /* Now set the avail_ssize amount */ 3496 if (rv == KERN_SUCCESS) { 3497 if (prev_entry != &map->header) 3498 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count); 3499 new_stack_entry = prev_entry->next; 3500 if (new_stack_entry->end != addrbos + max_ssize || 3501 new_stack_entry->start != addrbos + max_ssize - init_ssize) 3502 panic ("Bad entry start/end for new stack entry"); 3503 else 3504 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize; 3505 } 3506 3507 vm_map_unlock(map); 3508 vm_map_entry_release(count); 3509 return (rv); 3510 } 3511 3512 /* 3513 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 3514 * desired address is already mapped, or if we successfully grow 3515 * the stack. Also returns KERN_SUCCESS if addr is outside the 3516 * stack range (this is strange, but preserves compatibility with 3517 * the grow function in vm_machdep.c). 3518 * 3519 * No requirements. 3520 */ 3521 int 3522 vm_map_growstack (struct proc *p, vm_offset_t addr) 3523 { 3524 vm_map_entry_t prev_entry; 3525 vm_map_entry_t stack_entry; 3526 vm_map_entry_t new_stack_entry; 3527 struct vmspace *vm = p->p_vmspace; 3528 vm_map_t map = &vm->vm_map; 3529 vm_offset_t end; 3530 int grow_amount; 3531 int rv = KERN_SUCCESS; 3532 int is_procstack; 3533 int use_read_lock = 1; 3534 int count; 3535 3536 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3537 Retry: 3538 if (use_read_lock) 3539 vm_map_lock_read(map); 3540 else 3541 vm_map_lock(map); 3542 3543 /* If addr is already in the entry range, no need to grow.*/ 3544 if (vm_map_lookup_entry(map, addr, &prev_entry)) 3545 goto done; 3546 3547 if ((stack_entry = prev_entry->next) == &map->header) 3548 goto done; 3549 if (prev_entry == &map->header) 3550 end = stack_entry->start - stack_entry->aux.avail_ssize; 3551 else 3552 end = prev_entry->end; 3553 3554 /* 3555 * This next test mimics the old grow function in vm_machdep.c. 3556 * It really doesn't quite make sense, but we do it anyway 3557 * for compatibility. 3558 * 3559 * If not growable stack, return success. This signals the 3560 * caller to proceed as he would normally with normal vm. 3561 */ 3562 if (stack_entry->aux.avail_ssize < 1 || 3563 addr >= stack_entry->start || 3564 addr < stack_entry->start - stack_entry->aux.avail_ssize) { 3565 goto done; 3566 } 3567 3568 /* Find the minimum grow amount */ 3569 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE); 3570 if (grow_amount > stack_entry->aux.avail_ssize) { 3571 rv = KERN_NO_SPACE; 3572 goto done; 3573 } 3574 3575 /* 3576 * If there is no longer enough space between the entries 3577 * nogo, and adjust the available space. Note: this 3578 * should only happen if the user has mapped into the 3579 * stack area after the stack was created, and is 3580 * probably an error. 3581 * 3582 * This also effectively destroys any guard page the user 3583 * might have intended by limiting the stack size. 3584 */ 3585 if (grow_amount > stack_entry->start - end) { 3586 if (use_read_lock && vm_map_lock_upgrade(map)) { 3587 /* lost lock */ 3588 use_read_lock = 0; 3589 goto Retry; 3590 } 3591 use_read_lock = 0; 3592 stack_entry->aux.avail_ssize = stack_entry->start - end; 3593 rv = KERN_NO_SPACE; 3594 goto done; 3595 } 3596 3597 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 3598 3599 /* If this is the main process stack, see if we're over the 3600 * stack limit. 3601 */ 3602 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3603 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3604 rv = KERN_NO_SPACE; 3605 goto done; 3606 } 3607 3608 /* Round up the grow amount modulo SGROWSIZ */ 3609 grow_amount = roundup (grow_amount, sgrowsiz); 3610 if (grow_amount > stack_entry->aux.avail_ssize) { 3611 grow_amount = stack_entry->aux.avail_ssize; 3612 } 3613 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3614 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3615 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - 3616 ctob(vm->vm_ssize); 3617 } 3618 3619 /* If we would blow our VMEM resource limit, no go */ 3620 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3621 rv = KERN_NO_SPACE; 3622 goto done; 3623 } 3624 3625 if (use_read_lock && vm_map_lock_upgrade(map)) { 3626 /* lost lock */ 3627 use_read_lock = 0; 3628 goto Retry; 3629 } 3630 use_read_lock = 0; 3631 3632 /* Get the preliminary new entry start value */ 3633 addr = stack_entry->start - grow_amount; 3634 3635 /* If this puts us into the previous entry, cut back our growth 3636 * to the available space. Also, see the note above. 3637 */ 3638 if (addr < end) { 3639 stack_entry->aux.avail_ssize = stack_entry->start - end; 3640 addr = end; 3641 } 3642 3643 rv = vm_map_insert(map, &count, 3644 NULL, 0, addr, stack_entry->start, 3645 VM_MAPTYPE_NORMAL, 3646 VM_PROT_ALL, VM_PROT_ALL, 3647 0); 3648 3649 /* Adjust the available stack space by the amount we grew. */ 3650 if (rv == KERN_SUCCESS) { 3651 if (prev_entry != &map->header) 3652 vm_map_clip_end(map, prev_entry, addr, &count); 3653 new_stack_entry = prev_entry->next; 3654 if (new_stack_entry->end != stack_entry->start || 3655 new_stack_entry->start != addr) 3656 panic ("Bad stack grow start/end in new stack entry"); 3657 else { 3658 new_stack_entry->aux.avail_ssize = 3659 stack_entry->aux.avail_ssize - 3660 (new_stack_entry->end - new_stack_entry->start); 3661 if (is_procstack) 3662 vm->vm_ssize += btoc(new_stack_entry->end - 3663 new_stack_entry->start); 3664 } 3665 3666 if (map->flags & MAP_WIREFUTURE) 3667 vm_map_unwire(map, new_stack_entry->start, 3668 new_stack_entry->end, FALSE); 3669 } 3670 3671 done: 3672 if (use_read_lock) 3673 vm_map_unlock_read(map); 3674 else 3675 vm_map_unlock(map); 3676 vm_map_entry_release(count); 3677 return (rv); 3678 } 3679 3680 /* 3681 * Unshare the specified VM space for exec. If other processes are 3682 * mapped to it, then create a new one. The new vmspace is null. 3683 * 3684 * No requirements. 3685 */ 3686 void 3687 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 3688 { 3689 struct vmspace *oldvmspace = p->p_vmspace; 3690 struct vmspace *newvmspace; 3691 vm_map_t map = &p->p_vmspace->vm_map; 3692 3693 /* 3694 * If we are execing a resident vmspace we fork it, otherwise 3695 * we create a new vmspace. Note that exitingcnt is not 3696 * copied to the new vmspace. 3697 */ 3698 lwkt_gettoken(&oldvmspace->vm_map.token); 3699 if (vmcopy) { 3700 newvmspace = vmspace_fork(vmcopy); 3701 lwkt_gettoken(&newvmspace->vm_map.token); 3702 } else { 3703 newvmspace = vmspace_alloc(map->min_offset, map->max_offset); 3704 lwkt_gettoken(&newvmspace->vm_map.token); 3705 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 3706 (caddr_t)&oldvmspace->vm_endcopy - 3707 (caddr_t)&oldvmspace->vm_startcopy); 3708 } 3709 3710 /* 3711 * Finish initializing the vmspace before assigning it 3712 * to the process. The vmspace will become the current vmspace 3713 * if p == curproc. 3714 */ 3715 pmap_pinit2(vmspace_pmap(newvmspace)); 3716 pmap_replacevm(p, newvmspace, 0); 3717 lwkt_reltoken(&newvmspace->vm_map.token); 3718 lwkt_reltoken(&oldvmspace->vm_map.token); 3719 vmspace_free(oldvmspace); 3720 } 3721 3722 /* 3723 * Unshare the specified VM space for forcing COW. This 3724 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 3725 */ 3726 void 3727 vmspace_unshare(struct proc *p) 3728 { 3729 struct vmspace *oldvmspace = p->p_vmspace; 3730 struct vmspace *newvmspace; 3731 3732 lwkt_gettoken(&oldvmspace->vm_map.token); 3733 if (oldvmspace->vm_sysref.refcnt == 1) { 3734 lwkt_reltoken(&oldvmspace->vm_map.token); 3735 return; 3736 } 3737 newvmspace = vmspace_fork(oldvmspace); 3738 lwkt_gettoken(&newvmspace->vm_map.token); 3739 pmap_pinit2(vmspace_pmap(newvmspace)); 3740 pmap_replacevm(p, newvmspace, 0); 3741 lwkt_reltoken(&newvmspace->vm_map.token); 3742 lwkt_reltoken(&oldvmspace->vm_map.token); 3743 vmspace_free(oldvmspace); 3744 } 3745 3746 /* 3747 * vm_map_hint: return the beginning of the best area suitable for 3748 * creating a new mapping with "prot" protection. 3749 * 3750 * No requirements. 3751 */ 3752 vm_offset_t 3753 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot) 3754 { 3755 struct vmspace *vms = p->p_vmspace; 3756 3757 if (!randomize_mmap) { 3758 /* 3759 * Set a reasonable start point for the hint if it was 3760 * not specified or if it falls within the heap space. 3761 * Hinted mmap()s do not allocate out of the heap space. 3762 */ 3763 if (addr == 0 || 3764 (addr >= round_page((vm_offset_t)vms->vm_taddr) && 3765 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) { 3766 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz); 3767 } 3768 3769 return addr; 3770 } 3771 3772 if (addr != 0 && addr >= (vm_offset_t)vms->vm_daddr) 3773 return addr; 3774 3775 #ifdef notyet 3776 #ifdef __i386__ 3777 /* 3778 * If executable skip first two pages, otherwise start 3779 * after data + heap region. 3780 */ 3781 if ((prot & VM_PROT_EXECUTE) && 3782 ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) { 3783 addr = (PAGE_SIZE * 2) + 3784 (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1)); 3785 return (round_page(addr)); 3786 } 3787 #endif /* __i386__ */ 3788 #endif /* notyet */ 3789 3790 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ; 3791 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1); 3792 3793 return (round_page(addr)); 3794 } 3795 3796 /* 3797 * Finds the VM object, offset, and protection for a given virtual address 3798 * in the specified map, assuming a page fault of the type specified. 3799 * 3800 * Leaves the map in question locked for read; return values are guaranteed 3801 * until a vm_map_lookup_done call is performed. Note that the map argument 3802 * is in/out; the returned map must be used in the call to vm_map_lookup_done. 3803 * 3804 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make 3805 * that fast. 3806 * 3807 * If a lookup is requested with "write protection" specified, the map may 3808 * be changed to perform virtual copying operations, although the data 3809 * referenced will remain the same. 3810 * 3811 * No requirements. 3812 */ 3813 int 3814 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 3815 vm_offset_t vaddr, 3816 vm_prot_t fault_typea, 3817 vm_map_entry_t *out_entry, /* OUT */ 3818 vm_object_t *object, /* OUT */ 3819 vm_pindex_t *pindex, /* OUT */ 3820 vm_prot_t *out_prot, /* OUT */ 3821 boolean_t *wired) /* OUT */ 3822 { 3823 vm_map_entry_t entry; 3824 vm_map_t map = *var_map; 3825 vm_prot_t prot; 3826 vm_prot_t fault_type = fault_typea; 3827 int use_read_lock = 1; 3828 int rv = KERN_SUCCESS; 3829 3830 RetryLookup: 3831 if (use_read_lock) 3832 vm_map_lock_read(map); 3833 else 3834 vm_map_lock(map); 3835 3836 /* 3837 * If the map has an interesting hint, try it before calling full 3838 * blown lookup routine. 3839 */ 3840 entry = map->hint; 3841 cpu_ccfence(); 3842 *out_entry = entry; 3843 *object = NULL; 3844 3845 if ((entry == &map->header) || 3846 (vaddr < entry->start) || (vaddr >= entry->end)) { 3847 vm_map_entry_t tmp_entry; 3848 3849 /* 3850 * Entry was either not a valid hint, or the vaddr was not 3851 * contained in the entry, so do a full lookup. 3852 */ 3853 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 3854 rv = KERN_INVALID_ADDRESS; 3855 goto done; 3856 } 3857 3858 entry = tmp_entry; 3859 *out_entry = entry; 3860 } 3861 3862 /* 3863 * Handle submaps. 3864 */ 3865 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 3866 vm_map_t old_map = map; 3867 3868 *var_map = map = entry->object.sub_map; 3869 if (use_read_lock) 3870 vm_map_unlock_read(old_map); 3871 else 3872 vm_map_unlock(old_map); 3873 use_read_lock = 1; 3874 goto RetryLookup; 3875 } 3876 3877 /* 3878 * Check whether this task is allowed to have this page. 3879 * Note the special case for MAP_ENTRY_COW 3880 * pages with an override. This is to implement a forced 3881 * COW for debuggers. 3882 */ 3883 3884 if (fault_type & VM_PROT_OVERRIDE_WRITE) 3885 prot = entry->max_protection; 3886 else 3887 prot = entry->protection; 3888 3889 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 3890 if ((fault_type & prot) != fault_type) { 3891 rv = KERN_PROTECTION_FAILURE; 3892 goto done; 3893 } 3894 3895 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 3896 (entry->eflags & MAP_ENTRY_COW) && 3897 (fault_type & VM_PROT_WRITE) && 3898 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 3899 rv = KERN_PROTECTION_FAILURE; 3900 goto done; 3901 } 3902 3903 /* 3904 * If this page is not pageable, we have to get it for all possible 3905 * accesses. 3906 */ 3907 *wired = (entry->wired_count != 0); 3908 if (*wired) 3909 prot = fault_type = entry->protection; 3910 3911 /* 3912 * Virtual page tables may need to update the accessed (A) bit 3913 * in a page table entry. Upgrade the fault to a write fault for 3914 * that case if the map will support it. If the map does not support 3915 * it the page table entry simply will not be updated. 3916 */ 3917 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 3918 if (prot & VM_PROT_WRITE) 3919 fault_type |= VM_PROT_WRITE; 3920 } 3921 3922 /* 3923 * If the entry was copy-on-write, we either ... 3924 */ 3925 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3926 /* 3927 * If we want to write the page, we may as well handle that 3928 * now since we've got the map locked. 3929 * 3930 * If we don't need to write the page, we just demote the 3931 * permissions allowed. 3932 */ 3933 3934 if (fault_type & VM_PROT_WRITE) { 3935 /* 3936 * Make a new object, and place it in the object 3937 * chain. Note that no new references have appeared 3938 * -- one just moved from the map to the new 3939 * object. 3940 */ 3941 3942 if (use_read_lock && vm_map_lock_upgrade(map)) { 3943 /* lost lock */ 3944 use_read_lock = 0; 3945 goto RetryLookup; 3946 } 3947 use_read_lock = 0; 3948 3949 vm_map_entry_shadow(entry, 0); 3950 } else { 3951 /* 3952 * We're attempting to read a copy-on-write page -- 3953 * don't allow writes. 3954 */ 3955 3956 prot &= ~VM_PROT_WRITE; 3957 } 3958 } 3959 3960 /* 3961 * Create an object if necessary. 3962 */ 3963 if (entry->object.vm_object == NULL && !map->system_map) { 3964 if (use_read_lock && vm_map_lock_upgrade(map)) { 3965 /* lost lock */ 3966 use_read_lock = 0; 3967 goto RetryLookup; 3968 } 3969 use_read_lock = 0; 3970 vm_map_entry_allocate_object(entry); 3971 } 3972 3973 /* 3974 * Return the object/offset from this entry. If the entry was 3975 * copy-on-write or empty, it has been fixed up. 3976 */ 3977 3978 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 3979 *object = entry->object.vm_object; 3980 3981 /* 3982 * Return whether this is the only map sharing this data. On 3983 * success we return with a read lock held on the map. On failure 3984 * we return with the map unlocked. 3985 */ 3986 *out_prot = prot; 3987 done: 3988 if (rv == KERN_SUCCESS) { 3989 if (use_read_lock == 0) 3990 vm_map_lock_downgrade(map); 3991 } else if (use_read_lock) { 3992 vm_map_unlock_read(map); 3993 } else { 3994 vm_map_unlock(map); 3995 } 3996 return (rv); 3997 } 3998 3999 /* 4000 * Releases locks acquired by a vm_map_lookup() 4001 * (according to the handle returned by that lookup). 4002 * 4003 * No other requirements. 4004 */ 4005 void 4006 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 4007 { 4008 /* 4009 * Unlock the main-level map 4010 */ 4011 vm_map_unlock_read(map); 4012 if (count) 4013 vm_map_entry_release(count); 4014 } 4015 4016 #include "opt_ddb.h" 4017 #ifdef DDB 4018 #include <sys/kernel.h> 4019 4020 #include <ddb/ddb.h> 4021 4022 /* 4023 * Debugging only 4024 */ 4025 DB_SHOW_COMMAND(map, vm_map_print) 4026 { 4027 static int nlines; 4028 /* XXX convert args. */ 4029 vm_map_t map = (vm_map_t)addr; 4030 boolean_t full = have_addr; 4031 4032 vm_map_entry_t entry; 4033 4034 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 4035 (void *)map, 4036 (void *)map->pmap, map->nentries, map->timestamp); 4037 nlines++; 4038 4039 if (!full && db_indent) 4040 return; 4041 4042 db_indent += 2; 4043 for (entry = map->header.next; entry != &map->header; 4044 entry = entry->next) { 4045 db_iprintf("map entry %p: start=%p, end=%p\n", 4046 (void *)entry, (void *)entry->start, (void *)entry->end); 4047 nlines++; 4048 { 4049 static char *inheritance_name[4] = 4050 {"share", "copy", "none", "donate_copy"}; 4051 4052 db_iprintf(" prot=%x/%x/%s", 4053 entry->protection, 4054 entry->max_protection, 4055 inheritance_name[(int)(unsigned char)entry->inheritance]); 4056 if (entry->wired_count != 0) 4057 db_printf(", wired"); 4058 } 4059 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 4060 /* XXX no %qd in kernel. Truncate entry->offset. */ 4061 db_printf(", share=%p, offset=0x%lx\n", 4062 (void *)entry->object.sub_map, 4063 (long)entry->offset); 4064 nlines++; 4065 if ((entry->prev == &map->header) || 4066 (entry->prev->object.sub_map != 4067 entry->object.sub_map)) { 4068 db_indent += 2; 4069 vm_map_print((db_expr_t)(intptr_t) 4070 entry->object.sub_map, 4071 full, 0, NULL); 4072 db_indent -= 2; 4073 } 4074 } else { 4075 /* XXX no %qd in kernel. Truncate entry->offset. */ 4076 db_printf(", object=%p, offset=0x%lx", 4077 (void *)entry->object.vm_object, 4078 (long)entry->offset); 4079 if (entry->eflags & MAP_ENTRY_COW) 4080 db_printf(", copy (%s)", 4081 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 4082 db_printf("\n"); 4083 nlines++; 4084 4085 if ((entry->prev == &map->header) || 4086 (entry->prev->object.vm_object != 4087 entry->object.vm_object)) { 4088 db_indent += 2; 4089 vm_object_print((db_expr_t)(intptr_t) 4090 entry->object.vm_object, 4091 full, 0, NULL); 4092 nlines += 4; 4093 db_indent -= 2; 4094 } 4095 } 4096 } 4097 db_indent -= 2; 4098 if (db_indent == 0) 4099 nlines = 0; 4100 } 4101 4102 /* 4103 * Debugging only 4104 */ 4105 DB_SHOW_COMMAND(procvm, procvm) 4106 { 4107 struct proc *p; 4108 4109 if (have_addr) { 4110 p = (struct proc *) addr; 4111 } else { 4112 p = curproc; 4113 } 4114 4115 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 4116 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 4117 (void *)vmspace_pmap(p->p_vmspace)); 4118 4119 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 4120 } 4121 4122 #endif /* DDB */ 4123