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 * $DragonFly: src/sys/vm/vm_map.c,v 1.56 2007/04/29 18:25:41 dillon Exp $ 68 */ 69 70 /* 71 * Virtual memory mapping module. 72 */ 73 74 #include <sys/param.h> 75 #include <sys/systm.h> 76 #include <sys/kernel.h> 77 #include <sys/proc.h> 78 #include <sys/serialize.h> 79 #include <sys/lock.h> 80 #include <sys/vmmeter.h> 81 #include <sys/mman.h> 82 #include <sys/vnode.h> 83 #include <sys/resourcevar.h> 84 #include <sys/shm.h> 85 #include <sys/tree.h> 86 #include <sys/malloc.h> 87 88 #include <vm/vm.h> 89 #include <vm/vm_param.h> 90 #include <vm/pmap.h> 91 #include <vm/vm_map.h> 92 #include <vm/vm_page.h> 93 #include <vm/vm_object.h> 94 #include <vm/vm_pager.h> 95 #include <vm/vm_kern.h> 96 #include <vm/vm_extern.h> 97 #include <vm/swap_pager.h> 98 #include <vm/vm_zone.h> 99 100 #include <sys/thread2.h> 101 #include <sys/sysref2.h> 102 #include <sys/random.h> 103 #include <sys/sysctl.h> 104 105 /* 106 * Virtual memory maps provide for the mapping, protection, and sharing 107 * of virtual memory objects. In addition, this module provides for an 108 * efficient virtual copy of memory from one map to another. 109 * 110 * Synchronization is required prior to most operations. 111 * 112 * Maps consist of an ordered doubly-linked list of simple entries. 113 * A hint and a RB tree is used to speed-up lookups. 114 * 115 * Callers looking to modify maps specify start/end addresses which cause 116 * the related map entry to be clipped if necessary, and then later 117 * recombined if the pieces remained compatible. 118 * 119 * Virtual copy operations are performed by copying VM object references 120 * from one map to another, and then marking both regions as copy-on-write. 121 */ 122 static void vmspace_terminate(struct vmspace *vm); 123 static void vmspace_lock(struct vmspace *vm); 124 static void vmspace_unlock(struct vmspace *vm); 125 static void vmspace_dtor(void *obj, void *private); 126 127 MALLOC_DEFINE(M_VMSPACE, "vmspace", "vmspace objcache backingstore"); 128 129 struct sysref_class vmspace_sysref_class = { 130 .name = "vmspace", 131 .mtype = M_VMSPACE, 132 .proto = SYSREF_PROTO_VMSPACE, 133 .offset = offsetof(struct vmspace, vm_sysref), 134 .objsize = sizeof(struct vmspace), 135 .mag_capacity = 32, 136 .flags = SRC_MANAGEDINIT, 137 .dtor = vmspace_dtor, 138 .ops = { 139 .terminate = (sysref_terminate_func_t)vmspace_terminate, 140 .lock = (sysref_lock_func_t)vmspace_lock, 141 .unlock = (sysref_lock_func_t)vmspace_unlock 142 } 143 }; 144 145 /* 146 * per-cpu page table cross mappings are initialized in early boot 147 * and might require a considerable number of vm_map_entry structures. 148 */ 149 #define VMEPERCPU (MAXCPU+1) 150 151 static struct vm_zone mapentzone_store, mapzone_store; 152 static vm_zone_t mapentzone, mapzone; 153 static struct vm_object mapentobj, mapobj; 154 155 static struct vm_map_entry map_entry_init[MAX_MAPENT]; 156 static struct vm_map_entry cpu_map_entry_init[MAXCPU][VMEPERCPU]; 157 static struct vm_map map_init[MAX_KMAP]; 158 159 static int randomize_mmap; 160 SYSCTL_INT(_vm, OID_AUTO, randomize_mmap, CTLFLAG_RW, &randomize_mmap, 0, 161 "Randomize mmap offsets"); 162 163 static void vm_map_entry_shadow(vm_map_entry_t entry); 164 static vm_map_entry_t vm_map_entry_create(vm_map_t map, int *); 165 static void vm_map_entry_dispose (vm_map_t map, vm_map_entry_t entry, int *); 166 static void _vm_map_clip_end (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 167 static void _vm_map_clip_start (vm_map_t, vm_map_entry_t, vm_offset_t, int *); 168 static void vm_map_entry_delete (vm_map_t, vm_map_entry_t, int *); 169 static void vm_map_entry_unwire (vm_map_t, vm_map_entry_t); 170 static void vm_map_copy_entry (vm_map_t, vm_map_t, vm_map_entry_t, 171 vm_map_entry_t); 172 static void vm_map_split (vm_map_entry_t); 173 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); 174 175 /* 176 * Initialize the vm_map module. Must be called before any other vm_map 177 * routines. 178 * 179 * Map and entry structures are allocated from the general purpose 180 * memory pool with some exceptions: 181 * 182 * - The kernel map is allocated statically. 183 * - Initial kernel map entries are allocated out of a static pool. 184 * 185 * These restrictions are necessary since malloc() uses the 186 * maps and requires map entries. 187 * 188 * Called from the low level boot code only. 189 */ 190 void 191 vm_map_startup(void) 192 { 193 mapzone = &mapzone_store; 194 zbootinit(mapzone, "MAP", sizeof (struct vm_map), 195 map_init, MAX_KMAP); 196 mapentzone = &mapentzone_store; 197 zbootinit(mapentzone, "MAP ENTRY", sizeof (struct vm_map_entry), 198 map_entry_init, MAX_MAPENT); 199 } 200 201 /* 202 * Called prior to any vmspace allocations. 203 * 204 * Called from the low level boot code only. 205 */ 206 void 207 vm_init2(void) 208 { 209 zinitna(mapentzone, &mapentobj, NULL, 0, 0, 210 ZONE_USE_RESERVE | ZONE_SPECIAL, 1); 211 zinitna(mapzone, &mapobj, NULL, 0, 0, 0, 1); 212 pmap_init2(); 213 vm_object_init2(); 214 } 215 216 217 /* 218 * Red black tree functions 219 * 220 * The caller must hold the related map lock. 221 */ 222 static int rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b); 223 RB_GENERATE(vm_map_rb_tree, vm_map_entry, rb_entry, rb_vm_map_compare); 224 225 /* a->start is address, and the only field has to be initialized */ 226 static int 227 rb_vm_map_compare(vm_map_entry_t a, vm_map_entry_t b) 228 { 229 if (a->start < b->start) 230 return(-1); 231 else if (a->start > b->start) 232 return(1); 233 return(0); 234 } 235 236 /* 237 * Allocate a vmspace structure, including a vm_map and pmap. 238 * Initialize numerous fields. While the initial allocation is zerod, 239 * subsequence reuse from the objcache leaves elements of the structure 240 * intact (particularly the pmap), so portions must be zerod. 241 * 242 * The structure is not considered activated until we call sysref_activate(). 243 * 244 * No requirements. 245 */ 246 struct vmspace * 247 vmspace_alloc(vm_offset_t min, vm_offset_t max) 248 { 249 struct vmspace *vm; 250 251 lwkt_gettoken(&vmspace_token); 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); 256 pmap_pinit(vmspace_pmap(vm)); /* (some fields reused) */ 257 vm->vm_map.pmap = vmspace_pmap(vm); /* XXX */ 258 vm->vm_shm = NULL; 259 vm->vm_exitingcnt = 0; 260 cpu_vmspace_alloc(vm); 261 sysref_activate(&vm->vm_sysref); 262 lwkt_reltoken(&vmspace_token); 263 264 return (vm); 265 } 266 267 /* 268 * dtor function - Some elements of the pmap are retained in the 269 * free-cached vmspaces to improve performance. We have to clean them up 270 * here before returning the vmspace to the memory pool. 271 * 272 * No requirements. 273 */ 274 static void 275 vmspace_dtor(void *obj, void *private) 276 { 277 struct vmspace *vm = obj; 278 279 pmap_puninit(vmspace_pmap(vm)); 280 } 281 282 /* 283 * Called in two cases: 284 * 285 * (1) When the last sysref is dropped, but exitingcnt might still be 286 * non-zero. 287 * 288 * (2) When there are no sysrefs (i.e. refcnt is negative) left and the 289 * exitingcnt becomes zero 290 * 291 * sysref will not scrap the object until we call sysref_put() once more 292 * after the last ref has been dropped. 293 * 294 * Interlocked by the sysref API. 295 */ 296 static void 297 vmspace_terminate(struct vmspace *vm) 298 { 299 int count; 300 301 /* 302 * If exitingcnt is non-zero we can't get rid of the entire vmspace 303 * yet, but we can scrap user memory. 304 */ 305 lwkt_gettoken(&vmspace_token); 306 if (vm->vm_exitingcnt) { 307 shmexit(vm); 308 pmap_remove_pages(vmspace_pmap(vm), VM_MIN_USER_ADDRESS, 309 VM_MAX_USER_ADDRESS); 310 vm_map_remove(&vm->vm_map, VM_MIN_USER_ADDRESS, 311 VM_MAX_USER_ADDRESS); 312 lwkt_reltoken(&vmspace_token); 313 return; 314 } 315 cpu_vmspace_free(vm); 316 317 /* 318 * Make sure any SysV shm is freed, it might not have in 319 * exit1() 320 */ 321 shmexit(vm); 322 323 KKASSERT(vm->vm_upcalls == NULL); 324 325 /* 326 * Lock the map, to wait out all other references to it. 327 * Delete all of the mappings and pages they hold, then call 328 * the pmap module to reclaim anything left. 329 */ 330 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 331 vm_map_lock(&vm->vm_map); 332 vm_map_delete(&vm->vm_map, vm->vm_map.min_offset, 333 vm->vm_map.max_offset, &count); 334 vm_map_unlock(&vm->vm_map); 335 vm_map_entry_release(count); 336 337 pmap_release(vmspace_pmap(vm)); 338 sysref_put(&vm->vm_sysref); 339 lwkt_reltoken(&vmspace_token); 340 } 341 342 /* 343 * vmspaces are not currently locked. 344 */ 345 static void 346 vmspace_lock(struct vmspace *vm __unused) 347 { 348 } 349 350 static void 351 vmspace_unlock(struct vmspace *vm __unused) 352 { 353 } 354 355 /* 356 * This is called during exit indicating that the vmspace is no 357 * longer in used by an exiting process, but the process has not yet 358 * been cleaned up. 359 * 360 * No requirements. 361 */ 362 void 363 vmspace_exitbump(struct vmspace *vm) 364 { 365 lwkt_gettoken(&vmspace_token); 366 ++vm->vm_exitingcnt; 367 lwkt_reltoken(&vmspace_token); 368 } 369 370 /* 371 * This is called in the wait*() handling code. The vmspace can be terminated 372 * after the last wait is finished using it. 373 * 374 * No requirements. 375 */ 376 void 377 vmspace_exitfree(struct proc *p) 378 { 379 struct vmspace *vm; 380 381 lwkt_gettoken(&vmspace_token); 382 vm = p->p_vmspace; 383 p->p_vmspace = NULL; 384 385 if (--vm->vm_exitingcnt == 0 && sysref_isinactive(&vm->vm_sysref)) 386 vmspace_terminate(vm); 387 lwkt_reltoken(&vmspace_token); 388 } 389 390 /* 391 * Swap useage is determined by taking the proportional swap used by 392 * VM objects backing the VM map. To make up for fractional losses, 393 * if the VM object has any swap use at all the associated map entries 394 * count for at least 1 swap page. 395 * 396 * No requirements. 397 */ 398 int 399 vmspace_swap_count(struct vmspace *vmspace) 400 { 401 vm_map_t map = &vmspace->vm_map; 402 vm_map_entry_t cur; 403 vm_object_t object; 404 int count = 0; 405 int n; 406 407 lwkt_gettoken(&vmspace_token); 408 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 409 switch(cur->maptype) { 410 case VM_MAPTYPE_NORMAL: 411 case VM_MAPTYPE_VPAGETABLE: 412 if ((object = cur->object.vm_object) == NULL) 413 break; 414 if (object->swblock_count) { 415 n = (cur->end - cur->start) / PAGE_SIZE; 416 count += object->swblock_count * 417 SWAP_META_PAGES * n / object->size + 1; 418 } 419 break; 420 default: 421 break; 422 } 423 } 424 lwkt_reltoken(&vmspace_token); 425 return(count); 426 } 427 428 /* 429 * Calculate the approximate number of anonymous pages in use by 430 * this vmspace. To make up for fractional losses, we count each 431 * VM object as having at least 1 anonymous page. 432 * 433 * No requirements. 434 */ 435 int 436 vmspace_anonymous_count(struct vmspace *vmspace) 437 { 438 vm_map_t map = &vmspace->vm_map; 439 vm_map_entry_t cur; 440 vm_object_t object; 441 int count = 0; 442 443 lwkt_gettoken(&vmspace_token); 444 for (cur = map->header.next; cur != &map->header; cur = cur->next) { 445 switch(cur->maptype) { 446 case VM_MAPTYPE_NORMAL: 447 case VM_MAPTYPE_VPAGETABLE: 448 if ((object = cur->object.vm_object) == NULL) 449 break; 450 if (object->type != OBJT_DEFAULT && 451 object->type != OBJT_SWAP) { 452 break; 453 } 454 count += object->resident_page_count; 455 break; 456 default: 457 break; 458 } 459 } 460 lwkt_reltoken(&vmspace_token); 461 return(count); 462 } 463 464 /* 465 * Creates and returns a new empty VM map with the given physical map 466 * structure, and having the given lower and upper address bounds. 467 * 468 * No requirements. 469 */ 470 vm_map_t 471 vm_map_create(vm_map_t result, pmap_t pmap, vm_offset_t min, vm_offset_t max) 472 { 473 if (result == NULL) 474 result = zalloc(mapzone); 475 vm_map_init(result, min, max, pmap); 476 return (result); 477 } 478 479 /* 480 * Initialize an existing vm_map structure such as that in the vmspace 481 * structure. The pmap is initialized elsewhere. 482 * 483 * No requirements. 484 */ 485 void 486 vm_map_init(struct vm_map *map, vm_offset_t min, vm_offset_t max, pmap_t pmap) 487 { 488 map->header.next = map->header.prev = &map->header; 489 RB_INIT(&map->rb_root); 490 map->nentries = 0; 491 map->size = 0; 492 map->system_map = 0; 493 map->infork = 0; 494 map->min_offset = min; 495 map->max_offset = max; 496 map->pmap = pmap; 497 map->first_free = &map->header; 498 map->hint = &map->header; 499 map->timestamp = 0; 500 map->flags = 0; 501 lockinit(&map->lock, "thrd_sleep", 0, 0); 502 } 503 504 /* 505 * Shadow the vm_map_entry's object. This typically needs to be done when 506 * a write fault is taken on an entry which had previously been cloned by 507 * fork(). The shared object (which might be NULL) must become private so 508 * we add a shadow layer above it. 509 * 510 * Object allocation for anonymous mappings is defered as long as possible. 511 * When creating a shadow, however, the underlying object must be instantiated 512 * so it can be shared. 513 * 514 * If the map segment is governed by a virtual page table then it is 515 * possible to address offsets beyond the mapped area. Just allocate 516 * a maximally sized object for this case. 517 * 518 * The vm_map must be exclusively locked. 519 * No other requirements. 520 */ 521 static 522 void 523 vm_map_entry_shadow(vm_map_entry_t entry) 524 { 525 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 526 vm_object_shadow(&entry->object.vm_object, &entry->offset, 527 0x7FFFFFFF); /* XXX */ 528 } else { 529 vm_object_shadow(&entry->object.vm_object, &entry->offset, 530 atop(entry->end - entry->start)); 531 } 532 entry->eflags &= ~MAP_ENTRY_NEEDS_COPY; 533 } 534 535 /* 536 * Allocate an object for a vm_map_entry. 537 * 538 * Object allocation for anonymous mappings is defered as long as possible. 539 * This function is called when we can defer no longer, generally when a map 540 * entry might be split or forked or takes a page fault. 541 * 542 * If the map segment is governed by a virtual page table then it is 543 * possible to address offsets beyond the mapped area. Just allocate 544 * a maximally sized object for this case. 545 * 546 * The vm_map must be exclusively locked. 547 * No other requirements. 548 */ 549 void 550 vm_map_entry_allocate_object(vm_map_entry_t entry) 551 { 552 vm_object_t obj; 553 554 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 555 obj = vm_object_allocate(OBJT_DEFAULT, 0x7FFFFFFF); /* XXX */ 556 } else { 557 obj = vm_object_allocate(OBJT_DEFAULT, 558 atop(entry->end - entry->start)); 559 } 560 entry->object.vm_object = obj; 561 entry->offset = 0; 562 } 563 564 /* 565 * Set an initial negative count so the first attempt to reserve 566 * space preloads a bunch of vm_map_entry's for this cpu. Also 567 * pre-allocate 2 vm_map_entries which will be needed by zalloc() to 568 * map a new page for vm_map_entry structures. SMP systems are 569 * particularly sensitive. 570 * 571 * This routine is called in early boot so we cannot just call 572 * vm_map_entry_reserve(). 573 * 574 * Called from the low level boot code only (for each cpu) 575 */ 576 void 577 vm_map_entry_reserve_cpu_init(globaldata_t gd) 578 { 579 vm_map_entry_t entry; 580 int i; 581 582 gd->gd_vme_avail -= MAP_RESERVE_COUNT * 2; 583 entry = &cpu_map_entry_init[gd->gd_cpuid][0]; 584 for (i = 0; i < VMEPERCPU; ++i, ++entry) { 585 entry->next = gd->gd_vme_base; 586 gd->gd_vme_base = entry; 587 } 588 } 589 590 /* 591 * Reserves vm_map_entry structures so code later on can manipulate 592 * map_entry structures within a locked map without blocking trying 593 * to allocate a new vm_map_entry. 594 * 595 * No requirements. 596 */ 597 int 598 vm_map_entry_reserve(int count) 599 { 600 struct globaldata *gd = mycpu; 601 vm_map_entry_t entry; 602 603 /* 604 * Make sure we have enough structures in gd_vme_base to handle 605 * the reservation request. 606 */ 607 crit_enter(); 608 while (gd->gd_vme_avail < count) { 609 entry = zalloc(mapentzone); 610 entry->next = gd->gd_vme_base; 611 gd->gd_vme_base = entry; 612 ++gd->gd_vme_avail; 613 } 614 gd->gd_vme_avail -= count; 615 crit_exit(); 616 617 return(count); 618 } 619 620 /* 621 * Releases previously reserved vm_map_entry structures that were not 622 * used. If we have too much junk in our per-cpu cache clean some of 623 * it out. 624 * 625 * No requirements. 626 */ 627 void 628 vm_map_entry_release(int count) 629 { 630 struct globaldata *gd = mycpu; 631 vm_map_entry_t entry; 632 633 crit_enter(); 634 gd->gd_vme_avail += count; 635 while (gd->gd_vme_avail > MAP_RESERVE_SLOP) { 636 entry = gd->gd_vme_base; 637 KKASSERT(entry != NULL); 638 gd->gd_vme_base = entry->next; 639 --gd->gd_vme_avail; 640 crit_exit(); 641 zfree(mapentzone, entry); 642 crit_enter(); 643 } 644 crit_exit(); 645 } 646 647 /* 648 * Reserve map entry structures for use in kernel_map itself. These 649 * entries have *ALREADY* been reserved on a per-cpu basis when the map 650 * was inited. This function is used by zalloc() to avoid a recursion 651 * when zalloc() itself needs to allocate additional kernel memory. 652 * 653 * This function works like the normal reserve but does not load the 654 * vm_map_entry cache (because that would result in an infinite 655 * recursion). Note that gd_vme_avail may go negative. This is expected. 656 * 657 * Any caller of this function must be sure to renormalize after 658 * potentially eating entries to ensure that the reserve supply 659 * remains intact. 660 * 661 * No requirements. 662 */ 663 int 664 vm_map_entry_kreserve(int count) 665 { 666 struct globaldata *gd = mycpu; 667 668 crit_enter(); 669 gd->gd_vme_avail -= count; 670 crit_exit(); 671 KASSERT(gd->gd_vme_base != NULL, 672 ("no reserved entries left, gd_vme_avail = %d\n", 673 gd->gd_vme_avail)); 674 return(count); 675 } 676 677 /* 678 * Release previously reserved map entries for kernel_map. We do not 679 * attempt to clean up like the normal release function as this would 680 * cause an unnecessary (but probably not fatal) deep procedure call. 681 * 682 * No requirements. 683 */ 684 void 685 vm_map_entry_krelease(int count) 686 { 687 struct globaldata *gd = mycpu; 688 689 crit_enter(); 690 gd->gd_vme_avail += count; 691 crit_exit(); 692 } 693 694 /* 695 * Allocates a VM map entry for insertion. No entry fields are filled in. 696 * 697 * The entries should have previously been reserved. The reservation count 698 * is tracked in (*countp). 699 * 700 * No requirements. 701 */ 702 static vm_map_entry_t 703 vm_map_entry_create(vm_map_t map, int *countp) 704 { 705 struct globaldata *gd = mycpu; 706 vm_map_entry_t entry; 707 708 KKASSERT(*countp > 0); 709 --*countp; 710 crit_enter(); 711 entry = gd->gd_vme_base; 712 KASSERT(entry != NULL, ("gd_vme_base NULL! count %d", *countp)); 713 gd->gd_vme_base = entry->next; 714 crit_exit(); 715 716 return(entry); 717 } 718 719 /* 720 * Dispose of a vm_map_entry that is no longer being referenced. 721 * 722 * No requirements. 723 */ 724 static void 725 vm_map_entry_dispose(vm_map_t map, vm_map_entry_t entry, int *countp) 726 { 727 struct globaldata *gd = mycpu; 728 729 KKASSERT(map->hint != entry); 730 KKASSERT(map->first_free != entry); 731 732 ++*countp; 733 crit_enter(); 734 entry->next = gd->gd_vme_base; 735 gd->gd_vme_base = entry; 736 crit_exit(); 737 } 738 739 740 /* 741 * Insert/remove entries from maps. 742 * 743 * The related map must be exclusively locked. 744 * No other requirements. 745 * 746 * NOTE! We currently acquire the vmspace_token only to avoid races 747 * against the pageout daemon's calls to vmspace_*_count(), which 748 * are unable to safely lock the vm_map without potentially 749 * deadlocking. 750 */ 751 static __inline void 752 vm_map_entry_link(vm_map_t map, 753 vm_map_entry_t after_where, 754 vm_map_entry_t entry) 755 { 756 ASSERT_VM_MAP_LOCKED(map); 757 758 lwkt_gettoken(&vmspace_token); 759 map->nentries++; 760 entry->prev = after_where; 761 entry->next = after_where->next; 762 entry->next->prev = entry; 763 after_where->next = entry; 764 if (vm_map_rb_tree_RB_INSERT(&map->rb_root, entry)) 765 panic("vm_map_entry_link: dup addr map %p ent %p", map, entry); 766 lwkt_reltoken(&vmspace_token); 767 } 768 769 static __inline void 770 vm_map_entry_unlink(vm_map_t map, 771 vm_map_entry_t entry) 772 { 773 vm_map_entry_t prev; 774 vm_map_entry_t next; 775 776 ASSERT_VM_MAP_LOCKED(map); 777 778 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 779 panic("vm_map_entry_unlink: attempt to mess with " 780 "locked entry! %p", entry); 781 } 782 lwkt_gettoken(&vmspace_token); 783 prev = entry->prev; 784 next = entry->next; 785 next->prev = prev; 786 prev->next = next; 787 vm_map_rb_tree_RB_REMOVE(&map->rb_root, entry); 788 map->nentries--; 789 lwkt_reltoken(&vmspace_token); 790 } 791 792 /* 793 * Finds the map entry containing (or immediately preceding) the specified 794 * address in the given map. The entry is returned in (*entry). 795 * 796 * The boolean result indicates whether the address is actually contained 797 * in the map. 798 * 799 * The related map must be locked. 800 * No other requirements. 801 */ 802 boolean_t 803 vm_map_lookup_entry(vm_map_t map, vm_offset_t address, vm_map_entry_t *entry) 804 { 805 vm_map_entry_t tmp; 806 vm_map_entry_t last; 807 808 ASSERT_VM_MAP_LOCKED(map); 809 #if 0 810 /* 811 * XXX TEMPORARILY DISABLED. For some reason our attempt to revive 812 * the hint code with the red-black lookup meets with system crashes 813 * and lockups. We do not yet know why. 814 * 815 * It is possible that the problem is related to the setting 816 * of the hint during map_entry deletion, in the code specified 817 * at the GGG comment later on in this file. 818 */ 819 /* 820 * Quickly check the cached hint, there's a good chance of a match. 821 */ 822 if (map->hint != &map->header) { 823 tmp = map->hint; 824 if (address >= tmp->start && address < tmp->end) { 825 *entry = tmp; 826 return(TRUE); 827 } 828 } 829 #endif 830 831 /* 832 * Locate the record from the top of the tree. 'last' tracks the 833 * closest prior record and is returned if no match is found, which 834 * in binary tree terms means tracking the most recent right-branch 835 * taken. If there is no prior record, &map->header is returned. 836 */ 837 last = &map->header; 838 tmp = RB_ROOT(&map->rb_root); 839 840 while (tmp) { 841 if (address >= tmp->start) { 842 if (address < tmp->end) { 843 *entry = tmp; 844 map->hint = tmp; 845 return(TRUE); 846 } 847 last = tmp; 848 tmp = RB_RIGHT(tmp, rb_entry); 849 } else { 850 tmp = RB_LEFT(tmp, rb_entry); 851 } 852 } 853 *entry = last; 854 return (FALSE); 855 } 856 857 /* 858 * Inserts the given whole VM object into the target map at the specified 859 * address range. The object's size should match that of the address range. 860 * 861 * The map must be exclusively locked. 862 * The caller must have reserved sufficient vm_map_entry structures. 863 * 864 * If object is non-NULL, ref count must be bumped by caller 865 * prior to making call to account for the new entry. 866 */ 867 int 868 vm_map_insert(vm_map_t map, int *countp, 869 vm_object_t object, vm_ooffset_t offset, 870 vm_offset_t start, vm_offset_t end, 871 vm_maptype_t maptype, 872 vm_prot_t prot, vm_prot_t max, 873 int cow) 874 { 875 vm_map_entry_t new_entry; 876 vm_map_entry_t prev_entry; 877 vm_map_entry_t temp_entry; 878 vm_eflags_t protoeflags; 879 880 ASSERT_VM_MAP_LOCKED(map); 881 882 /* 883 * Check that the start and end points are not bogus. 884 */ 885 if ((start < map->min_offset) || (end > map->max_offset) || 886 (start >= end)) 887 return (KERN_INVALID_ADDRESS); 888 889 /* 890 * Find the entry prior to the proposed starting address; if it's part 891 * of an existing entry, this range is bogus. 892 */ 893 if (vm_map_lookup_entry(map, start, &temp_entry)) 894 return (KERN_NO_SPACE); 895 896 prev_entry = temp_entry; 897 898 /* 899 * Assert that the next entry doesn't overlap the end point. 900 */ 901 902 if ((prev_entry->next != &map->header) && 903 (prev_entry->next->start < end)) 904 return (KERN_NO_SPACE); 905 906 protoeflags = 0; 907 908 if (cow & MAP_COPY_ON_WRITE) 909 protoeflags |= MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY; 910 911 if (cow & MAP_NOFAULT) { 912 protoeflags |= MAP_ENTRY_NOFAULT; 913 914 KASSERT(object == NULL, 915 ("vm_map_insert: paradoxical MAP_NOFAULT request")); 916 } 917 if (cow & MAP_DISABLE_SYNCER) 918 protoeflags |= MAP_ENTRY_NOSYNC; 919 if (cow & MAP_DISABLE_COREDUMP) 920 protoeflags |= MAP_ENTRY_NOCOREDUMP; 921 if (cow & MAP_IS_STACK) 922 protoeflags |= MAP_ENTRY_STACK; 923 if (cow & MAP_IS_KSTACK) 924 protoeflags |= MAP_ENTRY_KSTACK; 925 926 lwkt_gettoken(&vm_token); 927 lwkt_gettoken(&vmobj_token); 928 929 if (object) { 930 /* 931 * When object is non-NULL, it could be shared with another 932 * process. We have to set or clear OBJ_ONEMAPPING 933 * appropriately. 934 */ 935 if ((object->ref_count > 1) || (object->shadow_count != 0)) { 936 vm_object_clear_flag(object, OBJ_ONEMAPPING); 937 } 938 } 939 else if ((prev_entry != &map->header) && 940 (prev_entry->eflags == protoeflags) && 941 (prev_entry->end == start) && 942 (prev_entry->wired_count == 0) && 943 prev_entry->maptype == maptype && 944 ((prev_entry->object.vm_object == NULL) || 945 vm_object_coalesce(prev_entry->object.vm_object, 946 OFF_TO_IDX(prev_entry->offset), 947 (vm_size_t)(prev_entry->end - prev_entry->start), 948 (vm_size_t)(end - prev_entry->end)))) { 949 /* 950 * We were able to extend the object. Determine if we 951 * can extend the previous map entry to include the 952 * new range as well. 953 */ 954 if ((prev_entry->inheritance == VM_INHERIT_DEFAULT) && 955 (prev_entry->protection == prot) && 956 (prev_entry->max_protection == max)) { 957 lwkt_reltoken(&vmobj_token); 958 lwkt_reltoken(&vm_token); 959 map->size += (end - prev_entry->end); 960 prev_entry->end = end; 961 vm_map_simplify_entry(map, prev_entry, countp); 962 return (KERN_SUCCESS); 963 } 964 965 /* 966 * If we can extend the object but cannot extend the 967 * map entry, we have to create a new map entry. We 968 * must bump the ref count on the extended object to 969 * account for it. object may be NULL. 970 */ 971 object = prev_entry->object.vm_object; 972 offset = prev_entry->offset + 973 (prev_entry->end - prev_entry->start); 974 vm_object_reference_locked(object); 975 } 976 977 lwkt_reltoken(&vmobj_token); 978 lwkt_reltoken(&vm_token); 979 980 /* 981 * NOTE: if conditionals fail, object can be NULL here. This occurs 982 * in things like the buffer map where we manage kva but do not manage 983 * backing objects. 984 */ 985 986 /* 987 * Create a new entry 988 */ 989 990 new_entry = vm_map_entry_create(map, countp); 991 new_entry->start = start; 992 new_entry->end = end; 993 994 new_entry->maptype = maptype; 995 new_entry->eflags = protoeflags; 996 new_entry->object.vm_object = object; 997 new_entry->offset = offset; 998 new_entry->aux.master_pde = 0; 999 1000 new_entry->inheritance = VM_INHERIT_DEFAULT; 1001 new_entry->protection = prot; 1002 new_entry->max_protection = max; 1003 new_entry->wired_count = 0; 1004 1005 /* 1006 * Insert the new entry into the list 1007 */ 1008 1009 vm_map_entry_link(map, prev_entry, new_entry); 1010 map->size += new_entry->end - new_entry->start; 1011 1012 /* 1013 * Update the free space hint. Entries cannot overlap. 1014 * An exact comparison is needed to avoid matching 1015 * against the map->header. 1016 */ 1017 if ((map->first_free == prev_entry) && 1018 (prev_entry->end == new_entry->start)) { 1019 map->first_free = new_entry; 1020 } 1021 1022 #if 0 1023 /* 1024 * Temporarily removed to avoid MAP_STACK panic, due to 1025 * MAP_STACK being a huge hack. Will be added back in 1026 * when MAP_STACK (and the user stack mapping) is fixed. 1027 */ 1028 /* 1029 * It may be possible to simplify the entry 1030 */ 1031 vm_map_simplify_entry(map, new_entry, countp); 1032 #endif 1033 1034 /* 1035 * Try to pre-populate the page table. Mappings governed by virtual 1036 * page tables cannot be prepopulated without a lot of work, so 1037 * don't try. 1038 */ 1039 if ((cow & (MAP_PREFAULT|MAP_PREFAULT_PARTIAL)) && 1040 maptype != VM_MAPTYPE_VPAGETABLE) { 1041 pmap_object_init_pt(map->pmap, start, prot, 1042 object, OFF_TO_IDX(offset), end - start, 1043 cow & MAP_PREFAULT_PARTIAL); 1044 } 1045 1046 return (KERN_SUCCESS); 1047 } 1048 1049 /* 1050 * Find sufficient space for `length' bytes in the given map, starting at 1051 * `start'. Returns 0 on success, 1 on no space. 1052 * 1053 * This function will returned an arbitrarily aligned pointer. If no 1054 * particular alignment is required you should pass align as 1. Note that 1055 * the map may return PAGE_SIZE aligned pointers if all the lengths used in 1056 * the map are a multiple of PAGE_SIZE, even if you pass a smaller align 1057 * argument. 1058 * 1059 * 'align' should be a power of 2 but is not required to be. 1060 * 1061 * The map must be exclusively locked. 1062 * No other requirements. 1063 */ 1064 int 1065 vm_map_findspace(vm_map_t map, vm_offset_t start, vm_size_t length, 1066 vm_size_t align, int flags, vm_offset_t *addr) 1067 { 1068 vm_map_entry_t entry, next; 1069 vm_offset_t end; 1070 vm_offset_t align_mask; 1071 1072 if (start < map->min_offset) 1073 start = map->min_offset; 1074 if (start > map->max_offset) 1075 return (1); 1076 1077 /* 1078 * If the alignment is not a power of 2 we will have to use 1079 * a mod/division, set align_mask to a special value. 1080 */ 1081 if ((align | (align - 1)) + 1 != (align << 1)) 1082 align_mask = (vm_offset_t)-1; 1083 else 1084 align_mask = align - 1; 1085 1086 /* 1087 * Look for the first possible address; if there's already something 1088 * at this address, we have to start after it. 1089 */ 1090 if (start == map->min_offset) { 1091 if ((entry = map->first_free) != &map->header) 1092 start = entry->end; 1093 } else { 1094 vm_map_entry_t tmp; 1095 1096 if (vm_map_lookup_entry(map, start, &tmp)) 1097 start = tmp->end; 1098 entry = tmp; 1099 } 1100 1101 /* 1102 * Look through the rest of the map, trying to fit a new region in the 1103 * gap between existing regions, or after the very last region. 1104 */ 1105 for (;; start = (entry = next)->end) { 1106 /* 1107 * Adjust the proposed start by the requested alignment, 1108 * be sure that we didn't wrap the address. 1109 */ 1110 if (align_mask == (vm_offset_t)-1) 1111 end = ((start + align - 1) / align) * align; 1112 else 1113 end = (start + align_mask) & ~align_mask; 1114 if (end < start) 1115 return (1); 1116 start = end; 1117 /* 1118 * Find the end of the proposed new region. Be sure we didn't 1119 * go beyond the end of the map, or wrap around the address. 1120 * Then check to see if this is the last entry or if the 1121 * proposed end fits in the gap between this and the next 1122 * entry. 1123 */ 1124 end = start + length; 1125 if (end > map->max_offset || end < start) 1126 return (1); 1127 next = entry->next; 1128 1129 /* 1130 * If the next entry's start address is beyond the desired 1131 * end address we may have found a good entry. 1132 * 1133 * If the next entry is a stack mapping we do not map into 1134 * the stack's reserved space. 1135 * 1136 * XXX continue to allow mapping into the stack's reserved 1137 * space if doing a MAP_STACK mapping inside a MAP_STACK 1138 * mapping, for backwards compatibility. But the caller 1139 * really should use MAP_STACK | MAP_TRYFIXED if they 1140 * want to do that. 1141 */ 1142 if (next == &map->header) 1143 break; 1144 if (next->start >= end) { 1145 if ((next->eflags & MAP_ENTRY_STACK) == 0) 1146 break; 1147 if (flags & MAP_STACK) 1148 break; 1149 if (next->start - next->aux.avail_ssize >= end) 1150 break; 1151 } 1152 } 1153 map->hint = entry; 1154 1155 /* 1156 * Grow the kernel_map if necessary. pmap_growkernel() will panic 1157 * if it fails. The kernel_map is locked and nothing can steal 1158 * our address space if pmap_growkernel() blocks. 1159 * 1160 * NOTE: This may be unconditionally called for kldload areas on 1161 * x86_64 because these do not bump kernel_vm_end (which would 1162 * fill 128G worth of page tables!). Therefore we must not 1163 * retry. 1164 */ 1165 if (map == &kernel_map) { 1166 vm_offset_t kstop; 1167 1168 kstop = round_page(start + length); 1169 if (kstop > kernel_vm_end) 1170 pmap_growkernel(start, kstop); 1171 } 1172 *addr = start; 1173 return (0); 1174 } 1175 1176 /* 1177 * vm_map_find finds an unallocated region in the target address map with 1178 * the given length. The search is defined to be first-fit from the 1179 * specified address; the region found is returned in the same parameter. 1180 * 1181 * If object is non-NULL, ref count must be bumped by caller 1182 * prior to making call to account for the new entry. 1183 * 1184 * No requirements. This function will lock the map temporarily. 1185 */ 1186 int 1187 vm_map_find(vm_map_t map, vm_object_t object, vm_ooffset_t offset, 1188 vm_offset_t *addr, vm_size_t length, vm_size_t align, 1189 boolean_t fitit, 1190 vm_maptype_t maptype, 1191 vm_prot_t prot, vm_prot_t max, 1192 int cow) 1193 { 1194 vm_offset_t start; 1195 int result; 1196 int count; 1197 1198 start = *addr; 1199 1200 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1201 vm_map_lock(map); 1202 if (fitit) { 1203 if (vm_map_findspace(map, start, length, align, 0, addr)) { 1204 vm_map_unlock(map); 1205 vm_map_entry_release(count); 1206 return (KERN_NO_SPACE); 1207 } 1208 start = *addr; 1209 } 1210 result = vm_map_insert(map, &count, object, offset, 1211 start, start + length, 1212 maptype, 1213 prot, max, 1214 cow); 1215 vm_map_unlock(map); 1216 vm_map_entry_release(count); 1217 1218 return (result); 1219 } 1220 1221 /* 1222 * Simplify the given map entry by merging with either neighbor. This 1223 * routine also has the ability to merge with both neighbors. 1224 * 1225 * This routine guarentees that the passed entry remains valid (though 1226 * possibly extended). When merging, this routine may delete one or 1227 * both neighbors. No action is taken on entries which have their 1228 * in-transition flag set. 1229 * 1230 * The map must be exclusively locked. 1231 */ 1232 void 1233 vm_map_simplify_entry(vm_map_t map, vm_map_entry_t entry, int *countp) 1234 { 1235 vm_map_entry_t next, prev; 1236 vm_size_t prevsize, esize; 1237 1238 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1239 ++mycpu->gd_cnt.v_intrans_coll; 1240 return; 1241 } 1242 1243 if (entry->maptype == VM_MAPTYPE_SUBMAP) 1244 return; 1245 1246 prev = entry->prev; 1247 if (prev != &map->header) { 1248 prevsize = prev->end - prev->start; 1249 if ( (prev->end == entry->start) && 1250 (prev->maptype == entry->maptype) && 1251 (prev->object.vm_object == entry->object.vm_object) && 1252 (!prev->object.vm_object || 1253 (prev->offset + prevsize == entry->offset)) && 1254 (prev->eflags == entry->eflags) && 1255 (prev->protection == entry->protection) && 1256 (prev->max_protection == entry->max_protection) && 1257 (prev->inheritance == entry->inheritance) && 1258 (prev->wired_count == entry->wired_count)) { 1259 if (map->first_free == prev) 1260 map->first_free = entry; 1261 if (map->hint == prev) 1262 map->hint = entry; 1263 vm_map_entry_unlink(map, prev); 1264 entry->start = prev->start; 1265 entry->offset = prev->offset; 1266 if (prev->object.vm_object) 1267 vm_object_deallocate(prev->object.vm_object); 1268 vm_map_entry_dispose(map, prev, countp); 1269 } 1270 } 1271 1272 next = entry->next; 1273 if (next != &map->header) { 1274 esize = entry->end - entry->start; 1275 if ((entry->end == next->start) && 1276 (next->maptype == entry->maptype) && 1277 (next->object.vm_object == entry->object.vm_object) && 1278 (!entry->object.vm_object || 1279 (entry->offset + esize == next->offset)) && 1280 (next->eflags == entry->eflags) && 1281 (next->protection == entry->protection) && 1282 (next->max_protection == entry->max_protection) && 1283 (next->inheritance == entry->inheritance) && 1284 (next->wired_count == entry->wired_count)) { 1285 if (map->first_free == next) 1286 map->first_free = entry; 1287 if (map->hint == next) 1288 map->hint = entry; 1289 vm_map_entry_unlink(map, next); 1290 entry->end = next->end; 1291 if (next->object.vm_object) 1292 vm_object_deallocate(next->object.vm_object); 1293 vm_map_entry_dispose(map, next, countp); 1294 } 1295 } 1296 } 1297 1298 /* 1299 * Asserts that the given entry begins at or after the specified address. 1300 * If necessary, it splits the entry into two. 1301 */ 1302 #define vm_map_clip_start(map, entry, startaddr, countp) \ 1303 { \ 1304 if (startaddr > entry->start) \ 1305 _vm_map_clip_start(map, entry, startaddr, countp); \ 1306 } 1307 1308 /* 1309 * This routine is called only when it is known that the entry must be split. 1310 * 1311 * The map must be exclusively locked. 1312 */ 1313 static void 1314 _vm_map_clip_start(vm_map_t map, vm_map_entry_t entry, vm_offset_t start, 1315 int *countp) 1316 { 1317 vm_map_entry_t new_entry; 1318 1319 /* 1320 * Split off the front portion -- note that we must insert the new 1321 * entry BEFORE this one, so that this entry has the specified 1322 * starting address. 1323 */ 1324 1325 vm_map_simplify_entry(map, entry, countp); 1326 1327 /* 1328 * If there is no object backing this entry, we might as well create 1329 * one now. If we defer it, an object can get created after the map 1330 * is clipped, and individual objects will be created for the split-up 1331 * map. This is a bit of a hack, but is also about the best place to 1332 * put this improvement. 1333 */ 1334 if (entry->object.vm_object == NULL && !map->system_map) { 1335 vm_map_entry_allocate_object(entry); 1336 } 1337 1338 new_entry = vm_map_entry_create(map, countp); 1339 *new_entry = *entry; 1340 1341 new_entry->end = start; 1342 entry->offset += (start - entry->start); 1343 entry->start = start; 1344 1345 vm_map_entry_link(map, entry->prev, new_entry); 1346 1347 switch(entry->maptype) { 1348 case VM_MAPTYPE_NORMAL: 1349 case VM_MAPTYPE_VPAGETABLE: 1350 vm_object_reference(new_entry->object.vm_object); 1351 break; 1352 default: 1353 break; 1354 } 1355 } 1356 1357 /* 1358 * Asserts that the given entry ends at or before the specified address. 1359 * If necessary, it splits the entry into two. 1360 * 1361 * The map must be exclusively locked. 1362 */ 1363 #define vm_map_clip_end(map, entry, endaddr, countp) \ 1364 { \ 1365 if (endaddr < entry->end) \ 1366 _vm_map_clip_end(map, entry, endaddr, countp); \ 1367 } 1368 1369 /* 1370 * This routine is called only when it is known that the entry must be split. 1371 * 1372 * The map must be exclusively locked. 1373 */ 1374 static void 1375 _vm_map_clip_end(vm_map_t map, vm_map_entry_t entry, vm_offset_t end, 1376 int *countp) 1377 { 1378 vm_map_entry_t new_entry; 1379 1380 /* 1381 * If there is no object backing this entry, we might as well create 1382 * one now. If we defer it, an object can get created after the map 1383 * is clipped, and individual objects will be created for the split-up 1384 * map. This is a bit of a hack, but is also about the best place to 1385 * put this improvement. 1386 */ 1387 1388 if (entry->object.vm_object == NULL && !map->system_map) { 1389 vm_map_entry_allocate_object(entry); 1390 } 1391 1392 /* 1393 * Create a new entry and insert it AFTER the specified entry 1394 */ 1395 1396 new_entry = vm_map_entry_create(map, countp); 1397 *new_entry = *entry; 1398 1399 new_entry->start = entry->end = end; 1400 new_entry->offset += (end - entry->start); 1401 1402 vm_map_entry_link(map, entry, new_entry); 1403 1404 switch(entry->maptype) { 1405 case VM_MAPTYPE_NORMAL: 1406 case VM_MAPTYPE_VPAGETABLE: 1407 vm_object_reference(new_entry->object.vm_object); 1408 break; 1409 default: 1410 break; 1411 } 1412 } 1413 1414 /* 1415 * Asserts that the starting and ending region addresses fall within the 1416 * valid range for the map. 1417 */ 1418 #define VM_MAP_RANGE_CHECK(map, start, end) \ 1419 { \ 1420 if (start < vm_map_min(map)) \ 1421 start = vm_map_min(map); \ 1422 if (end > vm_map_max(map)) \ 1423 end = vm_map_max(map); \ 1424 if (start > end) \ 1425 start = end; \ 1426 } 1427 1428 /* 1429 * Used to block when an in-transition collison occurs. The map 1430 * is unlocked for the sleep and relocked before the return. 1431 */ 1432 void 1433 vm_map_transition_wait(vm_map_t map) 1434 { 1435 tsleep_interlock(map, 0); 1436 vm_map_unlock(map); 1437 tsleep(map, PINTERLOCKED, "vment", 0); 1438 vm_map_lock(map); 1439 } 1440 1441 /* 1442 * When we do blocking operations with the map lock held it is 1443 * possible that a clip might have occured on our in-transit entry, 1444 * requiring an adjustment to the entry in our loop. These macros 1445 * help the pageable and clip_range code deal with the case. The 1446 * conditional costs virtually nothing if no clipping has occured. 1447 */ 1448 1449 #define CLIP_CHECK_BACK(entry, save_start) \ 1450 do { \ 1451 while (entry->start != save_start) { \ 1452 entry = entry->prev; \ 1453 KASSERT(entry != &map->header, ("bad entry clip")); \ 1454 } \ 1455 } while(0) 1456 1457 #define CLIP_CHECK_FWD(entry, save_end) \ 1458 do { \ 1459 while (entry->end != save_end) { \ 1460 entry = entry->next; \ 1461 KASSERT(entry != &map->header, ("bad entry clip")); \ 1462 } \ 1463 } while(0) 1464 1465 1466 /* 1467 * Clip the specified range and return the base entry. The 1468 * range may cover several entries starting at the returned base 1469 * and the first and last entry in the covering sequence will be 1470 * properly clipped to the requested start and end address. 1471 * 1472 * If no holes are allowed you should pass the MAP_CLIP_NO_HOLES 1473 * flag. 1474 * 1475 * The MAP_ENTRY_IN_TRANSITION flag will be set for the entries 1476 * covered by the requested range. 1477 * 1478 * The map must be exclusively locked on entry and will remain locked 1479 * on return. If no range exists or the range contains holes and you 1480 * specified that no holes were allowed, NULL will be returned. This 1481 * routine may temporarily unlock the map in order avoid a deadlock when 1482 * sleeping. 1483 */ 1484 static 1485 vm_map_entry_t 1486 vm_map_clip_range(vm_map_t map, vm_offset_t start, vm_offset_t end, 1487 int *countp, int flags) 1488 { 1489 vm_map_entry_t start_entry; 1490 vm_map_entry_t entry; 1491 1492 /* 1493 * Locate the entry and effect initial clipping. The in-transition 1494 * case does not occur very often so do not try to optimize it. 1495 */ 1496 again: 1497 if (vm_map_lookup_entry(map, start, &start_entry) == FALSE) 1498 return (NULL); 1499 entry = start_entry; 1500 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 1501 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1502 ++mycpu->gd_cnt.v_intrans_coll; 1503 ++mycpu->gd_cnt.v_intrans_wait; 1504 vm_map_transition_wait(map); 1505 /* 1506 * entry and/or start_entry may have been clipped while 1507 * we slept, or may have gone away entirely. We have 1508 * to restart from the lookup. 1509 */ 1510 goto again; 1511 } 1512 1513 /* 1514 * Since we hold an exclusive map lock we do not have to restart 1515 * after clipping, even though clipping may block in zalloc. 1516 */ 1517 vm_map_clip_start(map, entry, start, countp); 1518 vm_map_clip_end(map, entry, end, countp); 1519 entry->eflags |= MAP_ENTRY_IN_TRANSITION; 1520 1521 /* 1522 * Scan entries covered by the range. When working on the next 1523 * entry a restart need only re-loop on the current entry which 1524 * we have already locked, since 'next' may have changed. Also, 1525 * even though entry is safe, it may have been clipped so we 1526 * have to iterate forwards through the clip after sleeping. 1527 */ 1528 while (entry->next != &map->header && entry->next->start < end) { 1529 vm_map_entry_t next = entry->next; 1530 1531 if (flags & MAP_CLIP_NO_HOLES) { 1532 if (next->start > entry->end) { 1533 vm_map_unclip_range(map, start_entry, 1534 start, entry->end, countp, flags); 1535 return(NULL); 1536 } 1537 } 1538 1539 if (next->eflags & MAP_ENTRY_IN_TRANSITION) { 1540 vm_offset_t save_end = entry->end; 1541 next->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 1542 ++mycpu->gd_cnt.v_intrans_coll; 1543 ++mycpu->gd_cnt.v_intrans_wait; 1544 vm_map_transition_wait(map); 1545 1546 /* 1547 * clips might have occured while we blocked. 1548 */ 1549 CLIP_CHECK_FWD(entry, save_end); 1550 CLIP_CHECK_BACK(start_entry, start); 1551 continue; 1552 } 1553 /* 1554 * No restart necessary even though clip_end may block, we 1555 * are holding the map lock. 1556 */ 1557 vm_map_clip_end(map, next, end, countp); 1558 next->eflags |= MAP_ENTRY_IN_TRANSITION; 1559 entry = next; 1560 } 1561 if (flags & MAP_CLIP_NO_HOLES) { 1562 if (entry->end != end) { 1563 vm_map_unclip_range(map, start_entry, 1564 start, entry->end, countp, flags); 1565 return(NULL); 1566 } 1567 } 1568 return(start_entry); 1569 } 1570 1571 /* 1572 * Undo the effect of vm_map_clip_range(). You should pass the same 1573 * flags and the same range that you passed to vm_map_clip_range(). 1574 * This code will clear the in-transition flag on the entries and 1575 * wake up anyone waiting. This code will also simplify the sequence 1576 * and attempt to merge it with entries before and after the sequence. 1577 * 1578 * The map must be locked on entry and will remain locked on return. 1579 * 1580 * Note that you should also pass the start_entry returned by 1581 * vm_map_clip_range(). However, if you block between the two calls 1582 * with the map unlocked please be aware that the start_entry may 1583 * have been clipped and you may need to scan it backwards to find 1584 * the entry corresponding with the original start address. You are 1585 * responsible for this, vm_map_unclip_range() expects the correct 1586 * start_entry to be passed to it and will KASSERT otherwise. 1587 */ 1588 static 1589 void 1590 vm_map_unclip_range(vm_map_t map, vm_map_entry_t start_entry, 1591 vm_offset_t start, vm_offset_t end, 1592 int *countp, int flags) 1593 { 1594 vm_map_entry_t entry; 1595 1596 entry = start_entry; 1597 1598 KASSERT(entry->start == start, ("unclip_range: illegal base entry")); 1599 while (entry != &map->header && entry->start < end) { 1600 KASSERT(entry->eflags & MAP_ENTRY_IN_TRANSITION, 1601 ("in-transition flag not set during unclip on: %p", 1602 entry)); 1603 KASSERT(entry->end <= end, 1604 ("unclip_range: tail wasn't clipped")); 1605 entry->eflags &= ~MAP_ENTRY_IN_TRANSITION; 1606 if (entry->eflags & MAP_ENTRY_NEEDS_WAKEUP) { 1607 entry->eflags &= ~MAP_ENTRY_NEEDS_WAKEUP; 1608 wakeup(map); 1609 } 1610 entry = entry->next; 1611 } 1612 1613 /* 1614 * Simplification does not block so there is no restart case. 1615 */ 1616 entry = start_entry; 1617 while (entry != &map->header && entry->start < end) { 1618 vm_map_simplify_entry(map, entry, countp); 1619 entry = entry->next; 1620 } 1621 } 1622 1623 /* 1624 * Mark the given range as handled by a subordinate map. 1625 * 1626 * This range must have been created with vm_map_find(), and no other 1627 * operations may have been performed on this range prior to calling 1628 * vm_map_submap(). 1629 * 1630 * Submappings cannot be removed. 1631 * 1632 * No requirements. 1633 */ 1634 int 1635 vm_map_submap(vm_map_t map, vm_offset_t start, vm_offset_t end, vm_map_t submap) 1636 { 1637 vm_map_entry_t entry; 1638 int result = KERN_INVALID_ARGUMENT; 1639 int count; 1640 1641 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1642 vm_map_lock(map); 1643 1644 VM_MAP_RANGE_CHECK(map, start, end); 1645 1646 if (vm_map_lookup_entry(map, start, &entry)) { 1647 vm_map_clip_start(map, entry, start, &count); 1648 } else { 1649 entry = entry->next; 1650 } 1651 1652 vm_map_clip_end(map, entry, end, &count); 1653 1654 if ((entry->start == start) && (entry->end == end) && 1655 ((entry->eflags & MAP_ENTRY_COW) == 0) && 1656 (entry->object.vm_object == NULL)) { 1657 entry->object.sub_map = submap; 1658 entry->maptype = VM_MAPTYPE_SUBMAP; 1659 result = KERN_SUCCESS; 1660 } 1661 vm_map_unlock(map); 1662 vm_map_entry_release(count); 1663 1664 return (result); 1665 } 1666 1667 /* 1668 * Sets the protection of the specified address region in the target map. 1669 * If "set_max" is specified, the maximum protection is to be set; 1670 * otherwise, only the current protection is affected. 1671 * 1672 * The protection is not applicable to submaps, but is applicable to normal 1673 * maps and maps governed by virtual page tables. For example, when operating 1674 * on a virtual page table our protection basically controls how COW occurs 1675 * on the backing object, whereas the virtual page table abstraction itself 1676 * is an abstraction for userland. 1677 * 1678 * No requirements. 1679 */ 1680 int 1681 vm_map_protect(vm_map_t map, vm_offset_t start, vm_offset_t end, 1682 vm_prot_t new_prot, boolean_t set_max) 1683 { 1684 vm_map_entry_t current; 1685 vm_map_entry_t entry; 1686 int count; 1687 1688 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1689 vm_map_lock(map); 1690 1691 VM_MAP_RANGE_CHECK(map, start, end); 1692 1693 if (vm_map_lookup_entry(map, start, &entry)) { 1694 vm_map_clip_start(map, entry, start, &count); 1695 } else { 1696 entry = entry->next; 1697 } 1698 1699 /* 1700 * Make a first pass to check for protection violations. 1701 */ 1702 current = entry; 1703 while ((current != &map->header) && (current->start < end)) { 1704 if (current->maptype == VM_MAPTYPE_SUBMAP) { 1705 vm_map_unlock(map); 1706 vm_map_entry_release(count); 1707 return (KERN_INVALID_ARGUMENT); 1708 } 1709 if ((new_prot & current->max_protection) != new_prot) { 1710 vm_map_unlock(map); 1711 vm_map_entry_release(count); 1712 return (KERN_PROTECTION_FAILURE); 1713 } 1714 current = current->next; 1715 } 1716 1717 /* 1718 * Go back and fix up protections. [Note that clipping is not 1719 * necessary the second time.] 1720 */ 1721 current = entry; 1722 1723 while ((current != &map->header) && (current->start < end)) { 1724 vm_prot_t old_prot; 1725 1726 vm_map_clip_end(map, current, end, &count); 1727 1728 old_prot = current->protection; 1729 if (set_max) { 1730 current->protection = 1731 (current->max_protection = new_prot) & 1732 old_prot; 1733 } else { 1734 current->protection = new_prot; 1735 } 1736 1737 /* 1738 * Update physical map if necessary. Worry about copy-on-write 1739 * here -- CHECK THIS XXX 1740 */ 1741 1742 if (current->protection != old_prot) { 1743 #define MASK(entry) (((entry)->eflags & MAP_ENTRY_COW) ? ~VM_PROT_WRITE : \ 1744 VM_PROT_ALL) 1745 1746 pmap_protect(map->pmap, current->start, 1747 current->end, 1748 current->protection & MASK(current)); 1749 #undef MASK 1750 } 1751 1752 vm_map_simplify_entry(map, current, &count); 1753 1754 current = current->next; 1755 } 1756 1757 vm_map_unlock(map); 1758 vm_map_entry_release(count); 1759 return (KERN_SUCCESS); 1760 } 1761 1762 /* 1763 * This routine traverses a processes map handling the madvise 1764 * system call. Advisories are classified as either those effecting 1765 * the vm_map_entry structure, or those effecting the underlying 1766 * objects. 1767 * 1768 * The <value> argument is used for extended madvise calls. 1769 * 1770 * No requirements. 1771 */ 1772 int 1773 vm_map_madvise(vm_map_t map, vm_offset_t start, vm_offset_t end, 1774 int behav, off_t value) 1775 { 1776 vm_map_entry_t current, entry; 1777 int modify_map = 0; 1778 int error = 0; 1779 int count; 1780 1781 /* 1782 * Some madvise calls directly modify the vm_map_entry, in which case 1783 * we need to use an exclusive lock on the map and we need to perform 1784 * various clipping operations. Otherwise we only need a read-lock 1785 * on the map. 1786 */ 1787 1788 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1789 1790 switch(behav) { 1791 case MADV_NORMAL: 1792 case MADV_SEQUENTIAL: 1793 case MADV_RANDOM: 1794 case MADV_NOSYNC: 1795 case MADV_AUTOSYNC: 1796 case MADV_NOCORE: 1797 case MADV_CORE: 1798 case MADV_SETMAP: 1799 case MADV_INVAL: 1800 modify_map = 1; 1801 vm_map_lock(map); 1802 break; 1803 case MADV_WILLNEED: 1804 case MADV_DONTNEED: 1805 case MADV_FREE: 1806 vm_map_lock_read(map); 1807 break; 1808 default: 1809 vm_map_entry_release(count); 1810 return (EINVAL); 1811 } 1812 1813 /* 1814 * Locate starting entry and clip if necessary. 1815 */ 1816 1817 VM_MAP_RANGE_CHECK(map, start, end); 1818 1819 if (vm_map_lookup_entry(map, start, &entry)) { 1820 if (modify_map) 1821 vm_map_clip_start(map, entry, start, &count); 1822 } else { 1823 entry = entry->next; 1824 } 1825 1826 if (modify_map) { 1827 /* 1828 * madvise behaviors that are implemented in the vm_map_entry. 1829 * 1830 * We clip the vm_map_entry so that behavioral changes are 1831 * limited to the specified address range. 1832 */ 1833 for (current = entry; 1834 (current != &map->header) && (current->start < end); 1835 current = current->next 1836 ) { 1837 if (current->maptype == VM_MAPTYPE_SUBMAP) 1838 continue; 1839 1840 vm_map_clip_end(map, current, end, &count); 1841 1842 switch (behav) { 1843 case MADV_NORMAL: 1844 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_NORMAL); 1845 break; 1846 case MADV_SEQUENTIAL: 1847 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_SEQUENTIAL); 1848 break; 1849 case MADV_RANDOM: 1850 vm_map_entry_set_behavior(current, MAP_ENTRY_BEHAV_RANDOM); 1851 break; 1852 case MADV_NOSYNC: 1853 current->eflags |= MAP_ENTRY_NOSYNC; 1854 break; 1855 case MADV_AUTOSYNC: 1856 current->eflags &= ~MAP_ENTRY_NOSYNC; 1857 break; 1858 case MADV_NOCORE: 1859 current->eflags |= MAP_ENTRY_NOCOREDUMP; 1860 break; 1861 case MADV_CORE: 1862 current->eflags &= ~MAP_ENTRY_NOCOREDUMP; 1863 break; 1864 case MADV_INVAL: 1865 /* 1866 * Invalidate the related pmap entries, used 1867 * to flush portions of the real kernel's 1868 * pmap when the caller has removed or 1869 * modified existing mappings in a virtual 1870 * page table. 1871 */ 1872 pmap_remove(map->pmap, 1873 current->start, current->end); 1874 break; 1875 case MADV_SETMAP: 1876 /* 1877 * Set the page directory page for a map 1878 * governed by a virtual page table. Mark 1879 * the entry as being governed by a virtual 1880 * page table if it is not. 1881 * 1882 * XXX the page directory page is stored 1883 * in the avail_ssize field if the map_entry. 1884 * 1885 * XXX the map simplification code does not 1886 * compare this field so weird things may 1887 * happen if you do not apply this function 1888 * to the entire mapping governed by the 1889 * virtual page table. 1890 */ 1891 if (current->maptype != VM_MAPTYPE_VPAGETABLE) { 1892 error = EINVAL; 1893 break; 1894 } 1895 current->aux.master_pde = value; 1896 pmap_remove(map->pmap, 1897 current->start, current->end); 1898 break; 1899 default: 1900 error = EINVAL; 1901 break; 1902 } 1903 vm_map_simplify_entry(map, current, &count); 1904 } 1905 vm_map_unlock(map); 1906 } else { 1907 vm_pindex_t pindex; 1908 int count; 1909 1910 /* 1911 * madvise behaviors that are implemented in the underlying 1912 * vm_object. 1913 * 1914 * Since we don't clip the vm_map_entry, we have to clip 1915 * the vm_object pindex and count. 1916 * 1917 * NOTE! We currently do not support these functions on 1918 * virtual page tables. 1919 */ 1920 for (current = entry; 1921 (current != &map->header) && (current->start < end); 1922 current = current->next 1923 ) { 1924 vm_offset_t useStart; 1925 1926 if (current->maptype != VM_MAPTYPE_NORMAL) 1927 continue; 1928 1929 pindex = OFF_TO_IDX(current->offset); 1930 count = atop(current->end - current->start); 1931 useStart = current->start; 1932 1933 if (current->start < start) { 1934 pindex += atop(start - current->start); 1935 count -= atop(start - current->start); 1936 useStart = start; 1937 } 1938 if (current->end > end) 1939 count -= atop(current->end - end); 1940 1941 if (count <= 0) 1942 continue; 1943 1944 vm_object_madvise(current->object.vm_object, 1945 pindex, count, behav); 1946 1947 /* 1948 * Try to populate the page table. Mappings governed 1949 * by virtual page tables cannot be pre-populated 1950 * without a lot of work so don't try. 1951 */ 1952 if (behav == MADV_WILLNEED && 1953 current->maptype != VM_MAPTYPE_VPAGETABLE) { 1954 pmap_object_init_pt( 1955 map->pmap, 1956 useStart, 1957 current->protection, 1958 current->object.vm_object, 1959 pindex, 1960 (count << PAGE_SHIFT), 1961 MAP_PREFAULT_MADVISE 1962 ); 1963 } 1964 } 1965 vm_map_unlock_read(map); 1966 } 1967 vm_map_entry_release(count); 1968 return(error); 1969 } 1970 1971 1972 /* 1973 * Sets the inheritance of the specified address range in the target map. 1974 * Inheritance affects how the map will be shared with child maps at the 1975 * time of vm_map_fork. 1976 */ 1977 int 1978 vm_map_inherit(vm_map_t map, vm_offset_t start, vm_offset_t end, 1979 vm_inherit_t new_inheritance) 1980 { 1981 vm_map_entry_t entry; 1982 vm_map_entry_t temp_entry; 1983 int count; 1984 1985 switch (new_inheritance) { 1986 case VM_INHERIT_NONE: 1987 case VM_INHERIT_COPY: 1988 case VM_INHERIT_SHARE: 1989 break; 1990 default: 1991 return (KERN_INVALID_ARGUMENT); 1992 } 1993 1994 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 1995 vm_map_lock(map); 1996 1997 VM_MAP_RANGE_CHECK(map, start, end); 1998 1999 if (vm_map_lookup_entry(map, start, &temp_entry)) { 2000 entry = temp_entry; 2001 vm_map_clip_start(map, entry, start, &count); 2002 } else 2003 entry = temp_entry->next; 2004 2005 while ((entry != &map->header) && (entry->start < end)) { 2006 vm_map_clip_end(map, entry, end, &count); 2007 2008 entry->inheritance = new_inheritance; 2009 2010 vm_map_simplify_entry(map, entry, &count); 2011 2012 entry = entry->next; 2013 } 2014 vm_map_unlock(map); 2015 vm_map_entry_release(count); 2016 return (KERN_SUCCESS); 2017 } 2018 2019 /* 2020 * Implement the semantics of mlock 2021 */ 2022 int 2023 vm_map_unwire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, 2024 boolean_t new_pageable) 2025 { 2026 vm_map_entry_t entry; 2027 vm_map_entry_t start_entry; 2028 vm_offset_t end; 2029 int rv = KERN_SUCCESS; 2030 int count; 2031 2032 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2033 vm_map_lock(map); 2034 VM_MAP_RANGE_CHECK(map, start, real_end); 2035 end = real_end; 2036 2037 start_entry = vm_map_clip_range(map, start, end, &count, 2038 MAP_CLIP_NO_HOLES); 2039 if (start_entry == NULL) { 2040 vm_map_unlock(map); 2041 vm_map_entry_release(count); 2042 return (KERN_INVALID_ADDRESS); 2043 } 2044 2045 if (new_pageable == 0) { 2046 entry = start_entry; 2047 while ((entry != &map->header) && (entry->start < end)) { 2048 vm_offset_t save_start; 2049 vm_offset_t save_end; 2050 2051 /* 2052 * Already user wired or hard wired (trivial cases) 2053 */ 2054 if (entry->eflags & MAP_ENTRY_USER_WIRED) { 2055 entry = entry->next; 2056 continue; 2057 } 2058 if (entry->wired_count != 0) { 2059 entry->wired_count++; 2060 entry->eflags |= MAP_ENTRY_USER_WIRED; 2061 entry = entry->next; 2062 continue; 2063 } 2064 2065 /* 2066 * A new wiring requires instantiation of appropriate 2067 * management structures and the faulting in of the 2068 * page. 2069 */ 2070 if (entry->maptype != VM_MAPTYPE_SUBMAP) { 2071 int copyflag = entry->eflags & 2072 MAP_ENTRY_NEEDS_COPY; 2073 if (copyflag && ((entry->protection & 2074 VM_PROT_WRITE) != 0)) { 2075 vm_map_entry_shadow(entry); 2076 } else if (entry->object.vm_object == NULL && 2077 !map->system_map) { 2078 vm_map_entry_allocate_object(entry); 2079 } 2080 } 2081 entry->wired_count++; 2082 entry->eflags |= MAP_ENTRY_USER_WIRED; 2083 2084 /* 2085 * Now fault in the area. Note that vm_fault_wire() 2086 * may release the map lock temporarily, it will be 2087 * relocked on return. The in-transition 2088 * flag protects the entries. 2089 */ 2090 save_start = entry->start; 2091 save_end = entry->end; 2092 rv = vm_fault_wire(map, entry, TRUE); 2093 if (rv) { 2094 CLIP_CHECK_BACK(entry, save_start); 2095 for (;;) { 2096 KASSERT(entry->wired_count == 1, ("bad wired_count on entry")); 2097 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2098 entry->wired_count = 0; 2099 if (entry->end == save_end) 2100 break; 2101 entry = entry->next; 2102 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2103 } 2104 end = save_start; /* unwire the rest */ 2105 break; 2106 } 2107 /* 2108 * note that even though the entry might have been 2109 * clipped, the USER_WIRED flag we set prevents 2110 * duplication so we do not have to do a 2111 * clip check. 2112 */ 2113 entry = entry->next; 2114 } 2115 2116 /* 2117 * If we failed fall through to the unwiring section to 2118 * unwire what we had wired so far. 'end' has already 2119 * been adjusted. 2120 */ 2121 if (rv) 2122 new_pageable = 1; 2123 2124 /* 2125 * start_entry might have been clipped if we unlocked the 2126 * map and blocked. No matter how clipped it has gotten 2127 * there should be a fragment that is on our start boundary. 2128 */ 2129 CLIP_CHECK_BACK(start_entry, start); 2130 } 2131 2132 /* 2133 * Deal with the unwiring case. 2134 */ 2135 if (new_pageable) { 2136 /* 2137 * This is the unwiring case. We must first ensure that the 2138 * range to be unwired is really wired down. We know there 2139 * are no holes. 2140 */ 2141 entry = start_entry; 2142 while ((entry != &map->header) && (entry->start < end)) { 2143 if ((entry->eflags & MAP_ENTRY_USER_WIRED) == 0) { 2144 rv = KERN_INVALID_ARGUMENT; 2145 goto done; 2146 } 2147 KASSERT(entry->wired_count != 0, ("wired count was 0 with USER_WIRED set! %p", entry)); 2148 entry = entry->next; 2149 } 2150 2151 /* 2152 * Now decrement the wiring count for each region. If a region 2153 * becomes completely unwired, unwire its physical pages and 2154 * mappings. 2155 */ 2156 /* 2157 * The map entries are processed in a loop, checking to 2158 * make sure the entry is wired and asserting it has a wired 2159 * count. However, another loop was inserted more-or-less in 2160 * the middle of the unwiring path. This loop picks up the 2161 * "entry" loop variable from the first loop without first 2162 * setting it to start_entry. Naturally, the secound loop 2163 * is never entered and the pages backing the entries are 2164 * never unwired. This can lead to a leak of wired pages. 2165 */ 2166 entry = start_entry; 2167 while ((entry != &map->header) && (entry->start < end)) { 2168 KASSERT(entry->eflags & MAP_ENTRY_USER_WIRED, 2169 ("expected USER_WIRED on entry %p", entry)); 2170 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2171 entry->wired_count--; 2172 if (entry->wired_count == 0) 2173 vm_fault_unwire(map, entry); 2174 entry = entry->next; 2175 } 2176 } 2177 done: 2178 vm_map_unclip_range(map, start_entry, start, real_end, &count, 2179 MAP_CLIP_NO_HOLES); 2180 map->timestamp++; 2181 vm_map_unlock(map); 2182 vm_map_entry_release(count); 2183 return (rv); 2184 } 2185 2186 /* 2187 * Sets the pageability of the specified address range in the target map. 2188 * Regions specified as not pageable require locked-down physical 2189 * memory and physical page maps. 2190 * 2191 * The map must not be locked, but a reference must remain to the map 2192 * throughout the call. 2193 * 2194 * This function may be called via the zalloc path and must properly 2195 * reserve map entries for kernel_map. 2196 * 2197 * No requirements. 2198 */ 2199 int 2200 vm_map_wire(vm_map_t map, vm_offset_t start, vm_offset_t real_end, int kmflags) 2201 { 2202 vm_map_entry_t entry; 2203 vm_map_entry_t start_entry; 2204 vm_offset_t end; 2205 int rv = KERN_SUCCESS; 2206 int count; 2207 2208 if (kmflags & KM_KRESERVE) 2209 count = vm_map_entry_kreserve(MAP_RESERVE_COUNT); 2210 else 2211 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2212 vm_map_lock(map); 2213 VM_MAP_RANGE_CHECK(map, start, real_end); 2214 end = real_end; 2215 2216 start_entry = vm_map_clip_range(map, start, end, &count, 2217 MAP_CLIP_NO_HOLES); 2218 if (start_entry == NULL) { 2219 vm_map_unlock(map); 2220 rv = KERN_INVALID_ADDRESS; 2221 goto failure; 2222 } 2223 if ((kmflags & KM_PAGEABLE) == 0) { 2224 /* 2225 * Wiring. 2226 * 2227 * 1. Holding the write lock, we create any shadow or zero-fill 2228 * objects that need to be created. Then we clip each map 2229 * entry to the region to be wired and increment its wiring 2230 * count. We create objects before clipping the map entries 2231 * to avoid object proliferation. 2232 * 2233 * 2. We downgrade to a read lock, and call vm_fault_wire to 2234 * fault in the pages for any newly wired area (wired_count is 2235 * 1). 2236 * 2237 * Downgrading to a read lock for vm_fault_wire avoids a 2238 * possible deadlock with another process that may have faulted 2239 * on one of the pages to be wired (it would mark the page busy, 2240 * blocking us, then in turn block on the map lock that we 2241 * hold). Because of problems in the recursive lock package, 2242 * we cannot upgrade to a write lock in vm_map_lookup. Thus, 2243 * any actions that require the write lock must be done 2244 * beforehand. Because we keep the read lock on the map, the 2245 * copy-on-write status of the entries we modify here cannot 2246 * change. 2247 */ 2248 entry = start_entry; 2249 while ((entry != &map->header) && (entry->start < end)) { 2250 /* 2251 * Trivial case if the entry is already wired 2252 */ 2253 if (entry->wired_count) { 2254 entry->wired_count++; 2255 entry = entry->next; 2256 continue; 2257 } 2258 2259 /* 2260 * The entry is being newly wired, we have to setup 2261 * appropriate management structures. A shadow 2262 * object is required for a copy-on-write region, 2263 * or a normal object for a zero-fill region. We 2264 * do not have to do this for entries that point to sub 2265 * maps because we won't hold the lock on the sub map. 2266 */ 2267 if (entry->maptype != VM_MAPTYPE_SUBMAP) { 2268 int copyflag = entry->eflags & 2269 MAP_ENTRY_NEEDS_COPY; 2270 if (copyflag && ((entry->protection & 2271 VM_PROT_WRITE) != 0)) { 2272 vm_map_entry_shadow(entry); 2273 } else if (entry->object.vm_object == NULL && 2274 !map->system_map) { 2275 vm_map_entry_allocate_object(entry); 2276 } 2277 } 2278 2279 entry->wired_count++; 2280 entry = entry->next; 2281 } 2282 2283 /* 2284 * Pass 2. 2285 */ 2286 2287 /* 2288 * HACK HACK HACK HACK 2289 * 2290 * vm_fault_wire() temporarily unlocks the map to avoid 2291 * deadlocks. The in-transition flag from vm_map_clip_range 2292 * call should protect us from changes while the map is 2293 * unlocked. T 2294 * 2295 * NOTE: Previously this comment stated that clipping might 2296 * still occur while the entry is unlocked, but from 2297 * what I can tell it actually cannot. 2298 * 2299 * It is unclear whether the CLIP_CHECK_*() calls 2300 * are still needed but we keep them in anyway. 2301 * 2302 * HACK HACK HACK HACK 2303 */ 2304 2305 entry = start_entry; 2306 while (entry != &map->header && entry->start < end) { 2307 /* 2308 * If vm_fault_wire fails for any page we need to undo 2309 * what has been done. We decrement the wiring count 2310 * for those pages which have not yet been wired (now) 2311 * and unwire those that have (later). 2312 */ 2313 vm_offset_t save_start = entry->start; 2314 vm_offset_t save_end = entry->end; 2315 2316 if (entry->wired_count == 1) 2317 rv = vm_fault_wire(map, entry, FALSE); 2318 if (rv) { 2319 CLIP_CHECK_BACK(entry, save_start); 2320 for (;;) { 2321 KASSERT(entry->wired_count == 1, ("wired_count changed unexpectedly")); 2322 entry->wired_count = 0; 2323 if (entry->end == save_end) 2324 break; 2325 entry = entry->next; 2326 KASSERT(entry != &map->header, ("bad entry clip during backout")); 2327 } 2328 end = save_start; 2329 break; 2330 } 2331 CLIP_CHECK_FWD(entry, save_end); 2332 entry = entry->next; 2333 } 2334 2335 /* 2336 * If a failure occured undo everything by falling through 2337 * to the unwiring code. 'end' has already been adjusted 2338 * appropriately. 2339 */ 2340 if (rv) 2341 kmflags |= KM_PAGEABLE; 2342 2343 /* 2344 * start_entry is still IN_TRANSITION but may have been 2345 * clipped since vm_fault_wire() unlocks and relocks the 2346 * map. No matter how clipped it has gotten there should 2347 * be a fragment that is on our start boundary. 2348 */ 2349 CLIP_CHECK_BACK(start_entry, start); 2350 } 2351 2352 if (kmflags & KM_PAGEABLE) { 2353 /* 2354 * This is the unwiring case. We must first ensure that the 2355 * range to be unwired is really wired down. We know there 2356 * are no holes. 2357 */ 2358 entry = start_entry; 2359 while ((entry != &map->header) && (entry->start < end)) { 2360 if (entry->wired_count == 0) { 2361 rv = KERN_INVALID_ARGUMENT; 2362 goto done; 2363 } 2364 entry = entry->next; 2365 } 2366 2367 /* 2368 * Now decrement the wiring count for each region. If a region 2369 * becomes completely unwired, unwire its physical pages and 2370 * mappings. 2371 */ 2372 entry = start_entry; 2373 while ((entry != &map->header) && (entry->start < end)) { 2374 entry->wired_count--; 2375 if (entry->wired_count == 0) 2376 vm_fault_unwire(map, entry); 2377 entry = entry->next; 2378 } 2379 } 2380 done: 2381 vm_map_unclip_range(map, start_entry, start, real_end, 2382 &count, MAP_CLIP_NO_HOLES); 2383 map->timestamp++; 2384 vm_map_unlock(map); 2385 failure: 2386 if (kmflags & KM_KRESERVE) 2387 vm_map_entry_krelease(count); 2388 else 2389 vm_map_entry_release(count); 2390 return (rv); 2391 } 2392 2393 /* 2394 * Mark a newly allocated address range as wired but do not fault in 2395 * the pages. The caller is expected to load the pages into the object. 2396 * 2397 * The map must be locked on entry and will remain locked on return. 2398 * No other requirements. 2399 */ 2400 void 2401 vm_map_set_wired_quick(vm_map_t map, vm_offset_t addr, vm_size_t size, 2402 int *countp) 2403 { 2404 vm_map_entry_t scan; 2405 vm_map_entry_t entry; 2406 2407 entry = vm_map_clip_range(map, addr, addr + size, 2408 countp, MAP_CLIP_NO_HOLES); 2409 for (scan = entry; 2410 scan != &map->header && scan->start < addr + size; 2411 scan = scan->next) { 2412 KKASSERT(entry->wired_count == 0); 2413 entry->wired_count = 1; 2414 } 2415 vm_map_unclip_range(map, entry, addr, addr + size, 2416 countp, MAP_CLIP_NO_HOLES); 2417 } 2418 2419 /* 2420 * Push any dirty cached pages in the address range to their pager. 2421 * If syncio is TRUE, dirty pages are written synchronously. 2422 * If invalidate is TRUE, any cached pages are freed as well. 2423 * 2424 * This routine is called by sys_msync() 2425 * 2426 * Returns an error if any part of the specified range is not mapped. 2427 * 2428 * No requirements. 2429 */ 2430 int 2431 vm_map_clean(vm_map_t map, vm_offset_t start, vm_offset_t end, 2432 boolean_t syncio, boolean_t invalidate) 2433 { 2434 vm_map_entry_t current; 2435 vm_map_entry_t entry; 2436 vm_size_t size; 2437 vm_object_t object; 2438 vm_ooffset_t offset; 2439 2440 vm_map_lock_read(map); 2441 VM_MAP_RANGE_CHECK(map, start, end); 2442 if (!vm_map_lookup_entry(map, start, &entry)) { 2443 vm_map_unlock_read(map); 2444 return (KERN_INVALID_ADDRESS); 2445 } 2446 /* 2447 * Make a first pass to check for holes. 2448 */ 2449 for (current = entry; current->start < end; current = current->next) { 2450 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2451 vm_map_unlock_read(map); 2452 return (KERN_INVALID_ARGUMENT); 2453 } 2454 if (end > current->end && 2455 (current->next == &map->header || 2456 current->end != current->next->start)) { 2457 vm_map_unlock_read(map); 2458 return (KERN_INVALID_ADDRESS); 2459 } 2460 } 2461 2462 if (invalidate) 2463 pmap_remove(vm_map_pmap(map), start, end); 2464 2465 /* 2466 * Make a second pass, cleaning/uncaching pages from the indicated 2467 * objects as we go. 2468 * 2469 * Hold vm_token to avoid blocking in vm_object_reference() 2470 */ 2471 lwkt_gettoken(&vm_token); 2472 lwkt_gettoken(&vmobj_token); 2473 2474 for (current = entry; current->start < end; current = current->next) { 2475 offset = current->offset + (start - current->start); 2476 size = (end <= current->end ? end : current->end) - start; 2477 if (current->maptype == VM_MAPTYPE_SUBMAP) { 2478 vm_map_t smap; 2479 vm_map_entry_t tentry; 2480 vm_size_t tsize; 2481 2482 smap = current->object.sub_map; 2483 vm_map_lock_read(smap); 2484 vm_map_lookup_entry(smap, offset, &tentry); 2485 tsize = tentry->end - offset; 2486 if (tsize < size) 2487 size = tsize; 2488 object = tentry->object.vm_object; 2489 offset = tentry->offset + (offset - tentry->start); 2490 vm_map_unlock_read(smap); 2491 } else { 2492 object = current->object.vm_object; 2493 } 2494 /* 2495 * Note that there is absolutely no sense in writing out 2496 * anonymous objects, so we track down the vnode object 2497 * to write out. 2498 * We invalidate (remove) all pages from the address space 2499 * anyway, for semantic correctness. 2500 * 2501 * note: certain anonymous maps, such as MAP_NOSYNC maps, 2502 * may start out with a NULL object. 2503 */ 2504 while (object && object->backing_object) { 2505 offset += object->backing_object_offset; 2506 object = object->backing_object; 2507 if (object->size < OFF_TO_IDX( offset + size)) 2508 size = IDX_TO_OFF(object->size) - offset; 2509 } 2510 if (object && (object->type == OBJT_VNODE) && 2511 (current->protection & VM_PROT_WRITE) && 2512 (object->flags & OBJ_NOMSYNC) == 0) { 2513 /* 2514 * Flush pages if writing is allowed, invalidate them 2515 * if invalidation requested. Pages undergoing I/O 2516 * will be ignored by vm_object_page_remove(). 2517 * 2518 * We cannot lock the vnode and then wait for paging 2519 * to complete without deadlocking against vm_fault. 2520 * Instead we simply call vm_object_page_remove() and 2521 * allow it to block internally on a page-by-page 2522 * basis when it encounters pages undergoing async 2523 * I/O. 2524 */ 2525 int flags; 2526 2527 vm_object_reference_locked(object); 2528 vn_lock(object->handle, LK_EXCLUSIVE | LK_RETRY); 2529 flags = (syncio || invalidate) ? OBJPC_SYNC : 0; 2530 flags |= invalidate ? OBJPC_INVAL : 0; 2531 2532 /* 2533 * When operating on a virtual page table just 2534 * flush the whole object. XXX we probably ought 2535 * to 2536 */ 2537 switch(current->maptype) { 2538 case VM_MAPTYPE_NORMAL: 2539 vm_object_page_clean(object, 2540 OFF_TO_IDX(offset), 2541 OFF_TO_IDX(offset + size + PAGE_MASK), 2542 flags); 2543 break; 2544 case VM_MAPTYPE_VPAGETABLE: 2545 vm_object_page_clean(object, 0, 0, flags); 2546 break; 2547 } 2548 vn_unlock(((struct vnode *)object->handle)); 2549 vm_object_deallocate_locked(object); 2550 } 2551 if (object && invalidate && 2552 ((object->type == OBJT_VNODE) || 2553 (object->type == OBJT_DEVICE))) { 2554 int clean_only = 2555 (object->type == OBJT_DEVICE) ? FALSE : TRUE; 2556 vm_object_reference_locked(object); 2557 switch(current->maptype) { 2558 case VM_MAPTYPE_NORMAL: 2559 vm_object_page_remove(object, 2560 OFF_TO_IDX(offset), 2561 OFF_TO_IDX(offset + size + PAGE_MASK), 2562 clean_only); 2563 break; 2564 case VM_MAPTYPE_VPAGETABLE: 2565 vm_object_page_remove(object, 0, 0, clean_only); 2566 break; 2567 } 2568 vm_object_deallocate_locked(object); 2569 } 2570 start += size; 2571 } 2572 2573 lwkt_reltoken(&vmobj_token); 2574 lwkt_reltoken(&vm_token); 2575 vm_map_unlock_read(map); 2576 2577 return (KERN_SUCCESS); 2578 } 2579 2580 /* 2581 * Make the region specified by this entry pageable. 2582 * 2583 * The vm_map must be exclusively locked. 2584 */ 2585 static void 2586 vm_map_entry_unwire(vm_map_t map, vm_map_entry_t entry) 2587 { 2588 entry->eflags &= ~MAP_ENTRY_USER_WIRED; 2589 entry->wired_count = 0; 2590 vm_fault_unwire(map, entry); 2591 } 2592 2593 /* 2594 * Deallocate the given entry from the target map. 2595 * 2596 * The vm_map must be exclusively locked. 2597 */ 2598 static void 2599 vm_map_entry_delete(vm_map_t map, vm_map_entry_t entry, int *countp) 2600 { 2601 vm_map_entry_unlink(map, entry); 2602 map->size -= entry->end - entry->start; 2603 2604 switch(entry->maptype) { 2605 case VM_MAPTYPE_NORMAL: 2606 case VM_MAPTYPE_VPAGETABLE: 2607 vm_object_deallocate(entry->object.vm_object); 2608 break; 2609 default: 2610 break; 2611 } 2612 2613 vm_map_entry_dispose(map, entry, countp); 2614 } 2615 2616 /* 2617 * Deallocates the given address range from the target map. 2618 * 2619 * The vm_map must be exclusively locked. 2620 */ 2621 int 2622 vm_map_delete(vm_map_t map, vm_offset_t start, vm_offset_t end, int *countp) 2623 { 2624 vm_object_t object; 2625 vm_map_entry_t entry; 2626 vm_map_entry_t first_entry; 2627 2628 ASSERT_VM_MAP_LOCKED(map); 2629 again: 2630 /* 2631 * Find the start of the region, and clip it. Set entry to point 2632 * at the first record containing the requested address or, if no 2633 * such record exists, the next record with a greater address. The 2634 * loop will run from this point until a record beyond the termination 2635 * address is encountered. 2636 * 2637 * map->hint must be adjusted to not point to anything we delete, 2638 * so set it to the entry prior to the one being deleted. 2639 * 2640 * GGG see other GGG comment. 2641 */ 2642 if (vm_map_lookup_entry(map, start, &first_entry)) { 2643 entry = first_entry; 2644 vm_map_clip_start(map, entry, start, countp); 2645 map->hint = entry->prev; /* possible problem XXX */ 2646 } else { 2647 map->hint = first_entry; /* possible problem XXX */ 2648 entry = first_entry->next; 2649 } 2650 2651 /* 2652 * If a hole opens up prior to the current first_free then 2653 * adjust first_free. As with map->hint, map->first_free 2654 * cannot be left set to anything we might delete. 2655 */ 2656 if (entry == &map->header) { 2657 map->first_free = &map->header; 2658 } else if (map->first_free->start >= start) { 2659 map->first_free = entry->prev; 2660 } 2661 2662 /* 2663 * Step through all entries in this region 2664 */ 2665 while ((entry != &map->header) && (entry->start < end)) { 2666 vm_map_entry_t next; 2667 vm_offset_t s, e; 2668 vm_pindex_t offidxstart, offidxend, count; 2669 2670 /* 2671 * If we hit an in-transition entry we have to sleep and 2672 * retry. It's easier (and not really slower) to just retry 2673 * since this case occurs so rarely and the hint is already 2674 * pointing at the right place. We have to reset the 2675 * start offset so as not to accidently delete an entry 2676 * another process just created in vacated space. 2677 */ 2678 if (entry->eflags & MAP_ENTRY_IN_TRANSITION) { 2679 entry->eflags |= MAP_ENTRY_NEEDS_WAKEUP; 2680 start = entry->start; 2681 ++mycpu->gd_cnt.v_intrans_coll; 2682 ++mycpu->gd_cnt.v_intrans_wait; 2683 vm_map_transition_wait(map); 2684 goto again; 2685 } 2686 vm_map_clip_end(map, entry, end, countp); 2687 2688 s = entry->start; 2689 e = entry->end; 2690 next = entry->next; 2691 2692 offidxstart = OFF_TO_IDX(entry->offset); 2693 count = OFF_TO_IDX(e - s); 2694 object = entry->object.vm_object; 2695 2696 /* 2697 * Unwire before removing addresses from the pmap; otherwise, 2698 * unwiring will put the entries back in the pmap. 2699 */ 2700 if (entry->wired_count != 0) 2701 vm_map_entry_unwire(map, entry); 2702 2703 offidxend = offidxstart + count; 2704 2705 /* 2706 * Hold vm_token when manipulating vm_objects, 2707 * 2708 * Hold vmobj_token when potentially adding or removing 2709 * objects (collapse requires both). 2710 */ 2711 lwkt_gettoken(&vm_token); 2712 lwkt_gettoken(&vmobj_token); 2713 2714 if (object == &kernel_object) { 2715 vm_object_page_remove(object, offidxstart, 2716 offidxend, FALSE); 2717 } else { 2718 pmap_remove(map->pmap, s, e); 2719 2720 if (object != NULL && 2721 object->ref_count != 1 && 2722 (object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == 2723 OBJ_ONEMAPPING && 2724 (object->type == OBJT_DEFAULT || 2725 object->type == OBJT_SWAP)) { 2726 vm_object_collapse(object); 2727 vm_object_page_remove(object, offidxstart, 2728 offidxend, FALSE); 2729 if (object->type == OBJT_SWAP) { 2730 swap_pager_freespace(object, 2731 offidxstart, 2732 count); 2733 } 2734 if (offidxend >= object->size && 2735 offidxstart < object->size) { 2736 object->size = offidxstart; 2737 } 2738 } 2739 } 2740 lwkt_reltoken(&vmobj_token); 2741 lwkt_reltoken(&vm_token); 2742 2743 /* 2744 * Delete the entry (which may delete the object) only after 2745 * removing all pmap entries pointing to its pages. 2746 * (Otherwise, its page frames may be reallocated, and any 2747 * modify bits will be set in the wrong object!) 2748 */ 2749 vm_map_entry_delete(map, entry, countp); 2750 entry = next; 2751 } 2752 return (KERN_SUCCESS); 2753 } 2754 2755 /* 2756 * Remove the given address range from the target map. 2757 * This is the exported form of vm_map_delete. 2758 * 2759 * No requirements. 2760 */ 2761 int 2762 vm_map_remove(vm_map_t map, vm_offset_t start, vm_offset_t end) 2763 { 2764 int result; 2765 int count; 2766 2767 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 2768 vm_map_lock(map); 2769 VM_MAP_RANGE_CHECK(map, start, end); 2770 result = vm_map_delete(map, start, end, &count); 2771 vm_map_unlock(map); 2772 vm_map_entry_release(count); 2773 2774 return (result); 2775 } 2776 2777 /* 2778 * Assert that the target map allows the specified privilege on the 2779 * entire address region given. The entire region must be allocated. 2780 * 2781 * The caller must specify whether the vm_map is already locked or not. 2782 */ 2783 boolean_t 2784 vm_map_check_protection(vm_map_t map, vm_offset_t start, vm_offset_t end, 2785 vm_prot_t protection, boolean_t have_lock) 2786 { 2787 vm_map_entry_t entry; 2788 vm_map_entry_t tmp_entry; 2789 boolean_t result; 2790 2791 if (have_lock == FALSE) 2792 vm_map_lock_read(map); 2793 2794 if (!vm_map_lookup_entry(map, start, &tmp_entry)) { 2795 if (have_lock == FALSE) 2796 vm_map_unlock_read(map); 2797 return (FALSE); 2798 } 2799 entry = tmp_entry; 2800 2801 result = TRUE; 2802 while (start < end) { 2803 if (entry == &map->header) { 2804 result = FALSE; 2805 break; 2806 } 2807 /* 2808 * No holes allowed! 2809 */ 2810 2811 if (start < entry->start) { 2812 result = FALSE; 2813 break; 2814 } 2815 /* 2816 * Check protection associated with entry. 2817 */ 2818 2819 if ((entry->protection & protection) != protection) { 2820 result = FALSE; 2821 break; 2822 } 2823 /* go to next entry */ 2824 2825 start = entry->end; 2826 entry = entry->next; 2827 } 2828 if (have_lock == FALSE) 2829 vm_map_unlock_read(map); 2830 return (result); 2831 } 2832 2833 /* 2834 * Split the pages in a map entry into a new object. This affords 2835 * easier removal of unused pages, and keeps object inheritance from 2836 * being a negative impact on memory usage. 2837 * 2838 * The vm_map must be exclusively locked. 2839 */ 2840 static void 2841 vm_map_split(vm_map_entry_t entry) 2842 { 2843 vm_page_t m; 2844 vm_object_t orig_object, new_object, source; 2845 vm_offset_t s, e; 2846 vm_pindex_t offidxstart, offidxend, idx; 2847 vm_size_t size; 2848 vm_ooffset_t offset; 2849 2850 orig_object = entry->object.vm_object; 2851 if (orig_object->type != OBJT_DEFAULT && orig_object->type != OBJT_SWAP) 2852 return; 2853 if (orig_object->ref_count <= 1) 2854 return; 2855 2856 offset = entry->offset; 2857 s = entry->start; 2858 e = entry->end; 2859 2860 offidxstart = OFF_TO_IDX(offset); 2861 offidxend = offidxstart + OFF_TO_IDX(e - s); 2862 size = offidxend - offidxstart; 2863 2864 switch(orig_object->type) { 2865 case OBJT_DEFAULT: 2866 new_object = default_pager_alloc(NULL, IDX_TO_OFF(size), 2867 VM_PROT_ALL, 0); 2868 break; 2869 case OBJT_SWAP: 2870 new_object = swap_pager_alloc(NULL, IDX_TO_OFF(size), 2871 VM_PROT_ALL, 0); 2872 break; 2873 default: 2874 /* not reached */ 2875 new_object = NULL; 2876 KKASSERT(0); 2877 } 2878 if (new_object == NULL) 2879 return; 2880 2881 /* 2882 * vm_token required when manipulating vm_objects. 2883 */ 2884 lwkt_gettoken(&vm_token); 2885 lwkt_gettoken(&vmobj_token); 2886 2887 source = orig_object->backing_object; 2888 if (source != NULL) { 2889 /* Referenced by new_object */ 2890 vm_object_reference_locked(source); 2891 LIST_INSERT_HEAD(&source->shadow_head, 2892 new_object, shadow_list); 2893 vm_object_clear_flag(source, OBJ_ONEMAPPING); 2894 new_object->backing_object_offset = 2895 orig_object->backing_object_offset + 2896 IDX_TO_OFF(offidxstart); 2897 new_object->backing_object = source; 2898 source->shadow_count++; 2899 source->generation++; 2900 } 2901 2902 for (idx = 0; idx < size; idx++) { 2903 vm_page_t m; 2904 2905 retry: 2906 m = vm_page_lookup(orig_object, offidxstart + idx); 2907 if (m == NULL) 2908 continue; 2909 2910 /* 2911 * We must wait for pending I/O to complete before we can 2912 * rename the page. 2913 * 2914 * We do not have to VM_PROT_NONE the page as mappings should 2915 * not be changed by this operation. 2916 */ 2917 if (vm_page_sleep_busy(m, TRUE, "spltwt")) 2918 goto retry; 2919 vm_page_busy(m); 2920 vm_page_rename(m, new_object, idx); 2921 /* page automatically made dirty by rename and cache handled */ 2922 vm_page_busy(m); 2923 } 2924 2925 if (orig_object->type == OBJT_SWAP) { 2926 vm_object_pip_add(orig_object, 1); 2927 /* 2928 * copy orig_object pages into new_object 2929 * and destroy unneeded pages in 2930 * shadow object. 2931 */ 2932 swap_pager_copy(orig_object, new_object, offidxstart, 0); 2933 vm_object_pip_wakeup(orig_object); 2934 } 2935 2936 /* 2937 * Wakeup the pages we played with. No spl protection is needed 2938 * for a simple wakeup. 2939 */ 2940 for (idx = 0; idx < size; idx++) { 2941 m = vm_page_lookup(new_object, idx); 2942 if (m) 2943 vm_page_wakeup(m); 2944 } 2945 2946 entry->object.vm_object = new_object; 2947 entry->offset = 0LL; 2948 vm_object_deallocate_locked(orig_object); 2949 lwkt_reltoken(&vmobj_token); 2950 lwkt_reltoken(&vm_token); 2951 } 2952 2953 /* 2954 * Copies the contents of the source entry to the destination 2955 * entry. The entries *must* be aligned properly. 2956 * 2957 * The vm_map must be exclusively locked. 2958 * vm_token must be held 2959 */ 2960 static void 2961 vm_map_copy_entry(vm_map_t src_map, vm_map_t dst_map, 2962 vm_map_entry_t src_entry, vm_map_entry_t dst_entry) 2963 { 2964 vm_object_t src_object; 2965 2966 if (dst_entry->maptype == VM_MAPTYPE_SUBMAP) 2967 return; 2968 if (src_entry->maptype == VM_MAPTYPE_SUBMAP) 2969 return; 2970 2971 ASSERT_LWKT_TOKEN_HELD(&vm_token); 2972 lwkt_gettoken(&vmobj_token); /* required for collapse */ 2973 2974 if (src_entry->wired_count == 0) { 2975 /* 2976 * If the source entry is marked needs_copy, it is already 2977 * write-protected. 2978 */ 2979 if ((src_entry->eflags & MAP_ENTRY_NEEDS_COPY) == 0) { 2980 pmap_protect(src_map->pmap, 2981 src_entry->start, 2982 src_entry->end, 2983 src_entry->protection & ~VM_PROT_WRITE); 2984 } 2985 2986 /* 2987 * Make a copy of the object. 2988 */ 2989 if ((src_object = src_entry->object.vm_object) != NULL) { 2990 if ((src_object->handle == NULL) && 2991 (src_object->type == OBJT_DEFAULT || 2992 src_object->type == OBJT_SWAP)) { 2993 vm_object_collapse(src_object); 2994 if ((src_object->flags & (OBJ_NOSPLIT|OBJ_ONEMAPPING)) == OBJ_ONEMAPPING) { 2995 vm_map_split(src_entry); 2996 src_object = src_entry->object.vm_object; 2997 } 2998 } 2999 3000 vm_object_reference_locked(src_object); 3001 vm_object_clear_flag(src_object, OBJ_ONEMAPPING); 3002 dst_entry->object.vm_object = src_object; 3003 src_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY); 3004 dst_entry->eflags |= (MAP_ENTRY_COW|MAP_ENTRY_NEEDS_COPY); 3005 dst_entry->offset = src_entry->offset; 3006 } else { 3007 dst_entry->object.vm_object = NULL; 3008 dst_entry->offset = 0; 3009 } 3010 3011 pmap_copy(dst_map->pmap, src_map->pmap, dst_entry->start, 3012 dst_entry->end - dst_entry->start, src_entry->start); 3013 } else { 3014 /* 3015 * Of course, wired down pages can't be set copy-on-write. 3016 * Cause wired pages to be copied into the new map by 3017 * simulating faults (the new pages are pageable) 3018 */ 3019 vm_fault_copy_entry(dst_map, src_map, dst_entry, src_entry); 3020 } 3021 lwkt_reltoken(&vmobj_token); 3022 } 3023 3024 /* 3025 * vmspace_fork: 3026 * Create a new process vmspace structure and vm_map 3027 * based on those of an existing process. The new map 3028 * is based on the old map, according to the inheritance 3029 * values on the regions in that map. 3030 * 3031 * The source map must not be locked. 3032 * No requirements. 3033 */ 3034 struct vmspace * 3035 vmspace_fork(struct vmspace *vm1) 3036 { 3037 struct vmspace *vm2; 3038 vm_map_t old_map = &vm1->vm_map; 3039 vm_map_t new_map; 3040 vm_map_entry_t old_entry; 3041 vm_map_entry_t new_entry; 3042 vm_object_t object; 3043 int count; 3044 3045 lwkt_gettoken(&vm_token); 3046 lwkt_gettoken(&vmspace_token); 3047 lwkt_gettoken(&vmobj_token); 3048 vm_map_lock(old_map); 3049 old_map->infork = 1; 3050 3051 /* 3052 * XXX Note: upcalls are not copied. 3053 */ 3054 vm2 = vmspace_alloc(old_map->min_offset, old_map->max_offset); 3055 bcopy(&vm1->vm_startcopy, &vm2->vm_startcopy, 3056 (caddr_t)&vm1->vm_endcopy - (caddr_t)&vm1->vm_startcopy); 3057 new_map = &vm2->vm_map; /* XXX */ 3058 new_map->timestamp = 1; 3059 3060 vm_map_lock(new_map); 3061 3062 count = 0; 3063 old_entry = old_map->header.next; 3064 while (old_entry != &old_map->header) { 3065 ++count; 3066 old_entry = old_entry->next; 3067 } 3068 3069 count = vm_map_entry_reserve(count + MAP_RESERVE_COUNT); 3070 3071 old_entry = old_map->header.next; 3072 while (old_entry != &old_map->header) { 3073 if (old_entry->maptype == VM_MAPTYPE_SUBMAP) 3074 panic("vm_map_fork: encountered a submap"); 3075 3076 switch (old_entry->inheritance) { 3077 case VM_INHERIT_NONE: 3078 break; 3079 case VM_INHERIT_SHARE: 3080 /* 3081 * Clone the entry, creating the shared object if 3082 * necessary. 3083 */ 3084 object = old_entry->object.vm_object; 3085 if (object == NULL) { 3086 vm_map_entry_allocate_object(old_entry); 3087 object = old_entry->object.vm_object; 3088 } 3089 3090 /* 3091 * Add the reference before calling vm_map_entry_shadow 3092 * to insure that a shadow object is created. 3093 */ 3094 vm_object_reference_locked(object); 3095 if (old_entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3096 vm_map_entry_shadow(old_entry); 3097 /* Transfer the second reference too. */ 3098 vm_object_reference_locked( 3099 old_entry->object.vm_object); 3100 vm_object_deallocate_locked(object); 3101 object = old_entry->object.vm_object; 3102 } 3103 vm_object_clear_flag(object, OBJ_ONEMAPPING); 3104 3105 /* 3106 * Clone the entry, referencing the shared object. 3107 */ 3108 new_entry = vm_map_entry_create(new_map, &count); 3109 *new_entry = *old_entry; 3110 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3111 new_entry->wired_count = 0; 3112 3113 /* 3114 * Insert the entry into the new map -- we know we're 3115 * inserting at the end of the new map. 3116 */ 3117 3118 vm_map_entry_link(new_map, new_map->header.prev, 3119 new_entry); 3120 3121 /* 3122 * Update the physical map 3123 */ 3124 pmap_copy(new_map->pmap, old_map->pmap, 3125 new_entry->start, 3126 (old_entry->end - old_entry->start), 3127 old_entry->start); 3128 break; 3129 case VM_INHERIT_COPY: 3130 /* 3131 * Clone the entry and link into the map. 3132 */ 3133 new_entry = vm_map_entry_create(new_map, &count); 3134 *new_entry = *old_entry; 3135 new_entry->eflags &= ~MAP_ENTRY_USER_WIRED; 3136 new_entry->wired_count = 0; 3137 new_entry->object.vm_object = NULL; 3138 vm_map_entry_link(new_map, new_map->header.prev, 3139 new_entry); 3140 vm_map_copy_entry(old_map, new_map, old_entry, 3141 new_entry); 3142 break; 3143 } 3144 old_entry = old_entry->next; 3145 } 3146 3147 new_map->size = old_map->size; 3148 old_map->infork = 0; 3149 vm_map_unlock(old_map); 3150 vm_map_unlock(new_map); 3151 vm_map_entry_release(count); 3152 3153 lwkt_reltoken(&vmobj_token); 3154 lwkt_reltoken(&vmspace_token); 3155 lwkt_reltoken(&vm_token); 3156 3157 return (vm2); 3158 } 3159 3160 /* 3161 * Create an auto-grow stack entry 3162 * 3163 * No requirements. 3164 */ 3165 int 3166 vm_map_stack (vm_map_t map, vm_offset_t addrbos, vm_size_t max_ssize, 3167 int flags, vm_prot_t prot, vm_prot_t max, int cow) 3168 { 3169 vm_map_entry_t prev_entry; 3170 vm_map_entry_t new_stack_entry; 3171 vm_size_t init_ssize; 3172 int rv; 3173 int count; 3174 vm_offset_t tmpaddr; 3175 3176 cow |= MAP_IS_STACK; 3177 3178 if (max_ssize < sgrowsiz) 3179 init_ssize = max_ssize; 3180 else 3181 init_ssize = sgrowsiz; 3182 3183 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3184 vm_map_lock(map); 3185 3186 /* 3187 * Find space for the mapping 3188 */ 3189 if ((flags & (MAP_FIXED | MAP_TRYFIXED)) == 0) { 3190 if (vm_map_findspace(map, addrbos, max_ssize, 1, 3191 flags, &tmpaddr)) { 3192 vm_map_unlock(map); 3193 vm_map_entry_release(count); 3194 return (KERN_NO_SPACE); 3195 } 3196 addrbos = tmpaddr; 3197 } 3198 3199 /* If addr is already mapped, no go */ 3200 if (vm_map_lookup_entry(map, addrbos, &prev_entry)) { 3201 vm_map_unlock(map); 3202 vm_map_entry_release(count); 3203 return (KERN_NO_SPACE); 3204 } 3205 3206 #if 0 3207 /* XXX already handled by kern_mmap() */ 3208 /* If we would blow our VMEM resource limit, no go */ 3209 if (map->size + init_ssize > 3210 curproc->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3211 vm_map_unlock(map); 3212 vm_map_entry_release(count); 3213 return (KERN_NO_SPACE); 3214 } 3215 #endif 3216 3217 /* 3218 * If we can't accomodate max_ssize in the current mapping, 3219 * no go. However, we need to be aware that subsequent user 3220 * mappings might map into the space we have reserved for 3221 * stack, and currently this space is not protected. 3222 * 3223 * Hopefully we will at least detect this condition 3224 * when we try to grow the stack. 3225 */ 3226 if ((prev_entry->next != &map->header) && 3227 (prev_entry->next->start < addrbos + max_ssize)) { 3228 vm_map_unlock(map); 3229 vm_map_entry_release(count); 3230 return (KERN_NO_SPACE); 3231 } 3232 3233 /* 3234 * We initially map a stack of only init_ssize. We will 3235 * grow as needed later. Since this is to be a grow 3236 * down stack, we map at the top of the range. 3237 * 3238 * Note: we would normally expect prot and max to be 3239 * VM_PROT_ALL, and cow to be 0. Possibly we should 3240 * eliminate these as input parameters, and just 3241 * pass these values here in the insert call. 3242 */ 3243 rv = vm_map_insert(map, &count, 3244 NULL, 0, addrbos + max_ssize - init_ssize, 3245 addrbos + max_ssize, 3246 VM_MAPTYPE_NORMAL, 3247 prot, max, 3248 cow); 3249 3250 /* Now set the avail_ssize amount */ 3251 if (rv == KERN_SUCCESS) { 3252 if (prev_entry != &map->header) 3253 vm_map_clip_end(map, prev_entry, addrbos + max_ssize - init_ssize, &count); 3254 new_stack_entry = prev_entry->next; 3255 if (new_stack_entry->end != addrbos + max_ssize || 3256 new_stack_entry->start != addrbos + max_ssize - init_ssize) 3257 panic ("Bad entry start/end for new stack entry"); 3258 else 3259 new_stack_entry->aux.avail_ssize = max_ssize - init_ssize; 3260 } 3261 3262 vm_map_unlock(map); 3263 vm_map_entry_release(count); 3264 return (rv); 3265 } 3266 3267 /* 3268 * Attempts to grow a vm stack entry. Returns KERN_SUCCESS if the 3269 * desired address is already mapped, or if we successfully grow 3270 * the stack. Also returns KERN_SUCCESS if addr is outside the 3271 * stack range (this is strange, but preserves compatibility with 3272 * the grow function in vm_machdep.c). 3273 * 3274 * No requirements. 3275 */ 3276 int 3277 vm_map_growstack (struct proc *p, vm_offset_t addr) 3278 { 3279 vm_map_entry_t prev_entry; 3280 vm_map_entry_t stack_entry; 3281 vm_map_entry_t new_stack_entry; 3282 struct vmspace *vm = p->p_vmspace; 3283 vm_map_t map = &vm->vm_map; 3284 vm_offset_t end; 3285 int grow_amount; 3286 int rv = KERN_SUCCESS; 3287 int is_procstack; 3288 int use_read_lock = 1; 3289 int count; 3290 3291 count = vm_map_entry_reserve(MAP_RESERVE_COUNT); 3292 Retry: 3293 if (use_read_lock) 3294 vm_map_lock_read(map); 3295 else 3296 vm_map_lock(map); 3297 3298 /* If addr is already in the entry range, no need to grow.*/ 3299 if (vm_map_lookup_entry(map, addr, &prev_entry)) 3300 goto done; 3301 3302 if ((stack_entry = prev_entry->next) == &map->header) 3303 goto done; 3304 if (prev_entry == &map->header) 3305 end = stack_entry->start - stack_entry->aux.avail_ssize; 3306 else 3307 end = prev_entry->end; 3308 3309 /* 3310 * This next test mimics the old grow function in vm_machdep.c. 3311 * It really doesn't quite make sense, but we do it anyway 3312 * for compatibility. 3313 * 3314 * If not growable stack, return success. This signals the 3315 * caller to proceed as he would normally with normal vm. 3316 */ 3317 if (stack_entry->aux.avail_ssize < 1 || 3318 addr >= stack_entry->start || 3319 addr < stack_entry->start - stack_entry->aux.avail_ssize) { 3320 goto done; 3321 } 3322 3323 /* Find the minimum grow amount */ 3324 grow_amount = roundup (stack_entry->start - addr, PAGE_SIZE); 3325 if (grow_amount > stack_entry->aux.avail_ssize) { 3326 rv = KERN_NO_SPACE; 3327 goto done; 3328 } 3329 3330 /* 3331 * If there is no longer enough space between the entries 3332 * nogo, and adjust the available space. Note: this 3333 * should only happen if the user has mapped into the 3334 * stack area after the stack was created, and is 3335 * probably an error. 3336 * 3337 * This also effectively destroys any guard page the user 3338 * might have intended by limiting the stack size. 3339 */ 3340 if (grow_amount > stack_entry->start - end) { 3341 if (use_read_lock && vm_map_lock_upgrade(map)) { 3342 use_read_lock = 0; 3343 goto Retry; 3344 } 3345 use_read_lock = 0; 3346 stack_entry->aux.avail_ssize = stack_entry->start - end; 3347 rv = KERN_NO_SPACE; 3348 goto done; 3349 } 3350 3351 is_procstack = addr >= (vm_offset_t)vm->vm_maxsaddr; 3352 3353 /* If this is the main process stack, see if we're over the 3354 * stack limit. 3355 */ 3356 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3357 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3358 rv = KERN_NO_SPACE; 3359 goto done; 3360 } 3361 3362 /* Round up the grow amount modulo SGROWSIZ */ 3363 grow_amount = roundup (grow_amount, sgrowsiz); 3364 if (grow_amount > stack_entry->aux.avail_ssize) { 3365 grow_amount = stack_entry->aux.avail_ssize; 3366 } 3367 if (is_procstack && (ctob(vm->vm_ssize) + grow_amount > 3368 p->p_rlimit[RLIMIT_STACK].rlim_cur)) { 3369 grow_amount = p->p_rlimit[RLIMIT_STACK].rlim_cur - 3370 ctob(vm->vm_ssize); 3371 } 3372 3373 /* If we would blow our VMEM resource limit, no go */ 3374 if (map->size + grow_amount > p->p_rlimit[RLIMIT_VMEM].rlim_cur) { 3375 rv = KERN_NO_SPACE; 3376 goto done; 3377 } 3378 3379 if (use_read_lock && vm_map_lock_upgrade(map)) { 3380 use_read_lock = 0; 3381 goto Retry; 3382 } 3383 use_read_lock = 0; 3384 3385 /* Get the preliminary new entry start value */ 3386 addr = stack_entry->start - grow_amount; 3387 3388 /* If this puts us into the previous entry, cut back our growth 3389 * to the available space. Also, see the note above. 3390 */ 3391 if (addr < end) { 3392 stack_entry->aux.avail_ssize = stack_entry->start - end; 3393 addr = end; 3394 } 3395 3396 rv = vm_map_insert(map, &count, 3397 NULL, 0, addr, stack_entry->start, 3398 VM_MAPTYPE_NORMAL, 3399 VM_PROT_ALL, VM_PROT_ALL, 3400 0); 3401 3402 /* Adjust the available stack space by the amount we grew. */ 3403 if (rv == KERN_SUCCESS) { 3404 if (prev_entry != &map->header) 3405 vm_map_clip_end(map, prev_entry, addr, &count); 3406 new_stack_entry = prev_entry->next; 3407 if (new_stack_entry->end != stack_entry->start || 3408 new_stack_entry->start != addr) 3409 panic ("Bad stack grow start/end in new stack entry"); 3410 else { 3411 new_stack_entry->aux.avail_ssize = 3412 stack_entry->aux.avail_ssize - 3413 (new_stack_entry->end - new_stack_entry->start); 3414 if (is_procstack) 3415 vm->vm_ssize += btoc(new_stack_entry->end - 3416 new_stack_entry->start); 3417 } 3418 3419 if (map->flags & MAP_WIREFUTURE) 3420 vm_map_unwire(map, new_stack_entry->start, 3421 new_stack_entry->end, FALSE); 3422 } 3423 3424 done: 3425 if (use_read_lock) 3426 vm_map_unlock_read(map); 3427 else 3428 vm_map_unlock(map); 3429 vm_map_entry_release(count); 3430 return (rv); 3431 } 3432 3433 /* 3434 * Unshare the specified VM space for exec. If other processes are 3435 * mapped to it, then create a new one. The new vmspace is null. 3436 * 3437 * No requirements. 3438 */ 3439 void 3440 vmspace_exec(struct proc *p, struct vmspace *vmcopy) 3441 { 3442 struct vmspace *oldvmspace = p->p_vmspace; 3443 struct vmspace *newvmspace; 3444 vm_map_t map = &p->p_vmspace->vm_map; 3445 3446 /* 3447 * If we are execing a resident vmspace we fork it, otherwise 3448 * we create a new vmspace. Note that exitingcnt and upcalls 3449 * are not copied to the new vmspace. 3450 */ 3451 lwkt_gettoken(&vmspace_token); 3452 if (vmcopy) { 3453 newvmspace = vmspace_fork(vmcopy); 3454 } else { 3455 newvmspace = vmspace_alloc(map->min_offset, map->max_offset); 3456 bcopy(&oldvmspace->vm_startcopy, &newvmspace->vm_startcopy, 3457 (caddr_t)&oldvmspace->vm_endcopy - 3458 (caddr_t)&oldvmspace->vm_startcopy); 3459 } 3460 3461 /* 3462 * Finish initializing the vmspace before assigning it 3463 * to the process. The vmspace will become the current vmspace 3464 * if p == curproc. 3465 */ 3466 pmap_pinit2(vmspace_pmap(newvmspace)); 3467 pmap_replacevm(p, newvmspace, 0); 3468 sysref_put(&oldvmspace->vm_sysref); 3469 lwkt_reltoken(&vmspace_token); 3470 } 3471 3472 /* 3473 * Unshare the specified VM space for forcing COW. This 3474 * is called by rfork, for the (RFMEM|RFPROC) == 0 case. 3475 * 3476 * The exitingcnt test is not strictly necessary but has been 3477 * included for code sanity (to make the code a bit more deterministic). 3478 */ 3479 void 3480 vmspace_unshare(struct proc *p) 3481 { 3482 struct vmspace *oldvmspace = p->p_vmspace; 3483 struct vmspace *newvmspace; 3484 3485 lwkt_gettoken(&vmspace_token); 3486 if (oldvmspace->vm_sysref.refcnt == 1 && oldvmspace->vm_exitingcnt == 0) 3487 return; 3488 newvmspace = vmspace_fork(oldvmspace); 3489 pmap_pinit2(vmspace_pmap(newvmspace)); 3490 pmap_replacevm(p, newvmspace, 0); 3491 sysref_put(&oldvmspace->vm_sysref); 3492 lwkt_reltoken(&vmspace_token); 3493 } 3494 3495 /* 3496 * vm_map_hint: return the beginning of the best area suitable for 3497 * creating a new mapping with "prot" protection. 3498 * 3499 * No requirements. 3500 */ 3501 vm_offset_t 3502 vm_map_hint(struct proc *p, vm_offset_t addr, vm_prot_t prot) 3503 { 3504 struct vmspace *vms = p->p_vmspace; 3505 3506 if (!randomize_mmap) { 3507 /* 3508 * Set a reasonable start point for the hint if it was 3509 * not specified or if it falls within the heap space. 3510 * Hinted mmap()s do not allocate out of the heap space. 3511 */ 3512 if (addr == 0 || 3513 (addr >= round_page((vm_offset_t)vms->vm_taddr) && 3514 addr < round_page((vm_offset_t)vms->vm_daddr + maxdsiz))) { 3515 addr = round_page((vm_offset_t)vms->vm_daddr + maxdsiz); 3516 } 3517 3518 return addr; 3519 } 3520 3521 if (addr != 0 && addr >= (vm_offset_t)vms->vm_daddr) 3522 return addr; 3523 3524 #ifdef notyet 3525 #ifdef __i386__ 3526 /* 3527 * If executable skip first two pages, otherwise start 3528 * after data + heap region. 3529 */ 3530 if ((prot & VM_PROT_EXECUTE) && 3531 ((vm_offset_t)vms->vm_daddr >= I386_MAX_EXE_ADDR)) { 3532 addr = (PAGE_SIZE * 2) + 3533 (karc4random() & (I386_MAX_EXE_ADDR / 2 - 1)); 3534 return (round_page(addr)); 3535 } 3536 #endif /* __i386__ */ 3537 #endif /* notyet */ 3538 3539 addr = (vm_offset_t)vms->vm_daddr + MAXDSIZ; 3540 addr += karc4random() & (MIN((256 * 1024 * 1024), MAXDSIZ) - 1); 3541 3542 return (round_page(addr)); 3543 } 3544 3545 /* 3546 * Finds the VM object, offset, and protection for a given virtual address 3547 * in the specified map, assuming a page fault of the type specified. 3548 * 3549 * Leaves the map in question locked for read; return values are guaranteed 3550 * until a vm_map_lookup_done call is performed. Note that the map argument 3551 * is in/out; the returned map must be used in the call to vm_map_lookup_done. 3552 * 3553 * A handle (out_entry) is returned for use in vm_map_lookup_done, to make 3554 * that fast. 3555 * 3556 * If a lookup is requested with "write protection" specified, the map may 3557 * be changed to perform virtual copying operations, although the data 3558 * referenced will remain the same. 3559 * 3560 * No requirements. 3561 */ 3562 int 3563 vm_map_lookup(vm_map_t *var_map, /* IN/OUT */ 3564 vm_offset_t vaddr, 3565 vm_prot_t fault_typea, 3566 vm_map_entry_t *out_entry, /* OUT */ 3567 vm_object_t *object, /* OUT */ 3568 vm_pindex_t *pindex, /* OUT */ 3569 vm_prot_t *out_prot, /* OUT */ 3570 boolean_t *wired) /* OUT */ 3571 { 3572 vm_map_entry_t entry; 3573 vm_map_t map = *var_map; 3574 vm_prot_t prot; 3575 vm_prot_t fault_type = fault_typea; 3576 int use_read_lock = 1; 3577 int rv = KERN_SUCCESS; 3578 3579 RetryLookup: 3580 if (use_read_lock) 3581 vm_map_lock_read(map); 3582 else 3583 vm_map_lock(map); 3584 3585 /* 3586 * If the map has an interesting hint, try it before calling full 3587 * blown lookup routine. 3588 */ 3589 entry = map->hint; 3590 *out_entry = entry; 3591 3592 if ((entry == &map->header) || 3593 (vaddr < entry->start) || (vaddr >= entry->end)) { 3594 vm_map_entry_t tmp_entry; 3595 3596 /* 3597 * Entry was either not a valid hint, or the vaddr was not 3598 * contained in the entry, so do a full lookup. 3599 */ 3600 if (!vm_map_lookup_entry(map, vaddr, &tmp_entry)) { 3601 rv = KERN_INVALID_ADDRESS; 3602 goto done; 3603 } 3604 3605 entry = tmp_entry; 3606 *out_entry = entry; 3607 } 3608 3609 /* 3610 * Handle submaps. 3611 */ 3612 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 3613 vm_map_t old_map = map; 3614 3615 *var_map = map = entry->object.sub_map; 3616 if (use_read_lock) 3617 vm_map_unlock_read(old_map); 3618 else 3619 vm_map_unlock(old_map); 3620 use_read_lock = 1; 3621 goto RetryLookup; 3622 } 3623 3624 /* 3625 * Check whether this task is allowed to have this page. 3626 * Note the special case for MAP_ENTRY_COW 3627 * pages with an override. This is to implement a forced 3628 * COW for debuggers. 3629 */ 3630 3631 if (fault_type & VM_PROT_OVERRIDE_WRITE) 3632 prot = entry->max_protection; 3633 else 3634 prot = entry->protection; 3635 3636 fault_type &= (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE); 3637 if ((fault_type & prot) != fault_type) { 3638 rv = KERN_PROTECTION_FAILURE; 3639 goto done; 3640 } 3641 3642 if ((entry->eflags & MAP_ENTRY_USER_WIRED) && 3643 (entry->eflags & MAP_ENTRY_COW) && 3644 (fault_type & VM_PROT_WRITE) && 3645 (fault_typea & VM_PROT_OVERRIDE_WRITE) == 0) { 3646 rv = KERN_PROTECTION_FAILURE; 3647 goto done; 3648 } 3649 3650 /* 3651 * If this page is not pageable, we have to get it for all possible 3652 * accesses. 3653 */ 3654 *wired = (entry->wired_count != 0); 3655 if (*wired) 3656 prot = fault_type = entry->protection; 3657 3658 /* 3659 * Virtual page tables may need to update the accessed (A) bit 3660 * in a page table entry. Upgrade the fault to a write fault for 3661 * that case if the map will support it. If the map does not support 3662 * it the page table entry simply will not be updated. 3663 */ 3664 if (entry->maptype == VM_MAPTYPE_VPAGETABLE) { 3665 if (prot & VM_PROT_WRITE) 3666 fault_type |= VM_PROT_WRITE; 3667 } 3668 3669 /* 3670 * If the entry was copy-on-write, we either ... 3671 */ 3672 if (entry->eflags & MAP_ENTRY_NEEDS_COPY) { 3673 /* 3674 * If we want to write the page, we may as well handle that 3675 * now since we've got the map locked. 3676 * 3677 * If we don't need to write the page, we just demote the 3678 * permissions allowed. 3679 */ 3680 3681 if (fault_type & VM_PROT_WRITE) { 3682 /* 3683 * Make a new object, and place it in the object 3684 * chain. Note that no new references have appeared 3685 * -- one just moved from the map to the new 3686 * object. 3687 */ 3688 3689 if (use_read_lock && vm_map_lock_upgrade(map)) { 3690 use_read_lock = 0; 3691 goto RetryLookup; 3692 } 3693 use_read_lock = 0; 3694 3695 vm_map_entry_shadow(entry); 3696 } else { 3697 /* 3698 * We're attempting to read a copy-on-write page -- 3699 * don't allow writes. 3700 */ 3701 3702 prot &= ~VM_PROT_WRITE; 3703 } 3704 } 3705 3706 /* 3707 * Create an object if necessary. 3708 */ 3709 if (entry->object.vm_object == NULL && 3710 !map->system_map) { 3711 if (use_read_lock && vm_map_lock_upgrade(map)) { 3712 use_read_lock = 0; 3713 goto RetryLookup; 3714 } 3715 use_read_lock = 0; 3716 vm_map_entry_allocate_object(entry); 3717 } 3718 3719 /* 3720 * Return the object/offset from this entry. If the entry was 3721 * copy-on-write or empty, it has been fixed up. 3722 */ 3723 3724 *pindex = OFF_TO_IDX((vaddr - entry->start) + entry->offset); 3725 *object = entry->object.vm_object; 3726 3727 /* 3728 * Return whether this is the only map sharing this data. On 3729 * success we return with a read lock held on the map. On failure 3730 * we return with the map unlocked. 3731 */ 3732 *out_prot = prot; 3733 done: 3734 if (rv == KERN_SUCCESS) { 3735 if (use_read_lock == 0) 3736 vm_map_lock_downgrade(map); 3737 } else if (use_read_lock) { 3738 vm_map_unlock_read(map); 3739 } else { 3740 vm_map_unlock(map); 3741 } 3742 return (rv); 3743 } 3744 3745 /* 3746 * Releases locks acquired by a vm_map_lookup() 3747 * (according to the handle returned by that lookup). 3748 * 3749 * No other requirements. 3750 */ 3751 void 3752 vm_map_lookup_done(vm_map_t map, vm_map_entry_t entry, int count) 3753 { 3754 /* 3755 * Unlock the main-level map 3756 */ 3757 vm_map_unlock_read(map); 3758 if (count) 3759 vm_map_entry_release(count); 3760 } 3761 3762 #include "opt_ddb.h" 3763 #ifdef DDB 3764 #include <sys/kernel.h> 3765 3766 #include <ddb/ddb.h> 3767 3768 /* 3769 * Debugging only 3770 */ 3771 DB_SHOW_COMMAND(map, vm_map_print) 3772 { 3773 static int nlines; 3774 /* XXX convert args. */ 3775 vm_map_t map = (vm_map_t)addr; 3776 boolean_t full = have_addr; 3777 3778 vm_map_entry_t entry; 3779 3780 db_iprintf("Task map %p: pmap=%p, nentries=%d, version=%u\n", 3781 (void *)map, 3782 (void *)map->pmap, map->nentries, map->timestamp); 3783 nlines++; 3784 3785 if (!full && db_indent) 3786 return; 3787 3788 db_indent += 2; 3789 for (entry = map->header.next; entry != &map->header; 3790 entry = entry->next) { 3791 db_iprintf("map entry %p: start=%p, end=%p\n", 3792 (void *)entry, (void *)entry->start, (void *)entry->end); 3793 nlines++; 3794 { 3795 static char *inheritance_name[4] = 3796 {"share", "copy", "none", "donate_copy"}; 3797 3798 db_iprintf(" prot=%x/%x/%s", 3799 entry->protection, 3800 entry->max_protection, 3801 inheritance_name[(int)(unsigned char)entry->inheritance]); 3802 if (entry->wired_count != 0) 3803 db_printf(", wired"); 3804 } 3805 if (entry->maptype == VM_MAPTYPE_SUBMAP) { 3806 /* XXX no %qd in kernel. Truncate entry->offset. */ 3807 db_printf(", share=%p, offset=0x%lx\n", 3808 (void *)entry->object.sub_map, 3809 (long)entry->offset); 3810 nlines++; 3811 if ((entry->prev == &map->header) || 3812 (entry->prev->object.sub_map != 3813 entry->object.sub_map)) { 3814 db_indent += 2; 3815 vm_map_print((db_expr_t)(intptr_t) 3816 entry->object.sub_map, 3817 full, 0, NULL); 3818 db_indent -= 2; 3819 } 3820 } else { 3821 /* XXX no %qd in kernel. Truncate entry->offset. */ 3822 db_printf(", object=%p, offset=0x%lx", 3823 (void *)entry->object.vm_object, 3824 (long)entry->offset); 3825 if (entry->eflags & MAP_ENTRY_COW) 3826 db_printf(", copy (%s)", 3827 (entry->eflags & MAP_ENTRY_NEEDS_COPY) ? "needed" : "done"); 3828 db_printf("\n"); 3829 nlines++; 3830 3831 if ((entry->prev == &map->header) || 3832 (entry->prev->object.vm_object != 3833 entry->object.vm_object)) { 3834 db_indent += 2; 3835 vm_object_print((db_expr_t)(intptr_t) 3836 entry->object.vm_object, 3837 full, 0, NULL); 3838 nlines += 4; 3839 db_indent -= 2; 3840 } 3841 } 3842 } 3843 db_indent -= 2; 3844 if (db_indent == 0) 3845 nlines = 0; 3846 } 3847 3848 /* 3849 * Debugging only 3850 */ 3851 DB_SHOW_COMMAND(procvm, procvm) 3852 { 3853 struct proc *p; 3854 3855 if (have_addr) { 3856 p = (struct proc *) addr; 3857 } else { 3858 p = curproc; 3859 } 3860 3861 db_printf("p = %p, vmspace = %p, map = %p, pmap = %p\n", 3862 (void *)p, (void *)p->p_vmspace, (void *)&p->p_vmspace->vm_map, 3863 (void *)vmspace_pmap(p->p_vmspace)); 3864 3865 vm_map_print((db_expr_t)(intptr_t)&p->p_vmspace->vm_map, 1, 0, NULL); 3866 } 3867 3868 #endif /* DDB */ 3869