1 /* $NetBSD: uvm_map.c,v 1.295 2011/02/02 15:25:27 chuck Exp $ */ 2 3 /* 4 * Copyright (c) 1997 Charles D. Cranor and Washington University. 5 * Copyright (c) 1991, 1993, The Regents of the University of California. 6 * 7 * All rights reserved. 8 * 9 * This code is derived from software contributed to Berkeley by 10 * The Mach Operating System project at Carnegie-Mellon University. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. Neither the name of the University nor the names of its contributors 21 * may be used to endorse or promote products derived from this software 22 * without specific prior written permission. 23 * 24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 34 * SUCH DAMAGE. 35 * 36 * @(#)vm_map.c 8.3 (Berkeley) 1/12/94 37 * from: Id: uvm_map.c,v 1.1.2.27 1998/02/07 01:16:54 chs Exp 38 * 39 * 40 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 41 * All rights reserved. 42 * 43 * Permission to use, copy, modify and distribute this software and 44 * its documentation is hereby granted, provided that both the copyright 45 * notice and this permission notice appear in all copies of the 46 * software, derivative works or modified versions, and any portions 47 * thereof, and that both notices appear in supporting documentation. 48 * 49 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 50 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 51 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 52 * 53 * Carnegie Mellon requests users of this software to return to 54 * 55 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 56 * School of Computer Science 57 * Carnegie Mellon University 58 * Pittsburgh PA 15213-3890 59 * 60 * any improvements or extensions that they make and grant Carnegie the 61 * rights to redistribute these changes. 62 */ 63 64 /* 65 * uvm_map.c: uvm map operations 66 */ 67 68 #include <sys/cdefs.h> 69 __KERNEL_RCSID(0, "$NetBSD: uvm_map.c,v 1.295 2011/02/02 15:25:27 chuck Exp $"); 70 71 #include "opt_ddb.h" 72 #include "opt_uvmhist.h" 73 #include "opt_uvm.h" 74 #include "opt_sysv.h" 75 76 #include <sys/param.h> 77 #include <sys/systm.h> 78 #include <sys/mman.h> 79 #include <sys/proc.h> 80 #include <sys/malloc.h> 81 #include <sys/pool.h> 82 #include <sys/kernel.h> 83 #include <sys/mount.h> 84 #include <sys/vnode.h> 85 #include <sys/lockdebug.h> 86 #include <sys/atomic.h> 87 #ifndef __USER_VA0_IS_SAFE 88 #include <sys/sysctl.h> 89 #include <sys/kauth.h> 90 #include "opt_user_va0_disable_default.h" 91 #endif 92 93 #ifdef SYSVSHM 94 #include <sys/shm.h> 95 #endif 96 97 #include <uvm/uvm.h> 98 #include <uvm/uvm_readahead.h> 99 100 #if defined(DDB) || defined(DEBUGPRINT) 101 #include <uvm/uvm_ddb.h> 102 #endif 103 104 #if !defined(UVMMAP_COUNTERS) 105 106 #define UVMMAP_EVCNT_DEFINE(name) /* nothing */ 107 #define UVMMAP_EVCNT_INCR(ev) /* nothing */ 108 #define UVMMAP_EVCNT_DECR(ev) /* nothing */ 109 110 #else /* defined(UVMMAP_NOCOUNTERS) */ 111 112 #include <sys/evcnt.h> 113 #define UVMMAP_EVCNT_DEFINE(name) \ 114 struct evcnt uvmmap_evcnt_##name = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, \ 115 "uvmmap", #name); \ 116 EVCNT_ATTACH_STATIC(uvmmap_evcnt_##name); 117 #define UVMMAP_EVCNT_INCR(ev) uvmmap_evcnt_##ev.ev_count++ 118 #define UVMMAP_EVCNT_DECR(ev) uvmmap_evcnt_##ev.ev_count-- 119 120 #endif /* defined(UVMMAP_NOCOUNTERS) */ 121 122 UVMMAP_EVCNT_DEFINE(ubackmerge) 123 UVMMAP_EVCNT_DEFINE(uforwmerge) 124 UVMMAP_EVCNT_DEFINE(ubimerge) 125 UVMMAP_EVCNT_DEFINE(unomerge) 126 UVMMAP_EVCNT_DEFINE(kbackmerge) 127 UVMMAP_EVCNT_DEFINE(kforwmerge) 128 UVMMAP_EVCNT_DEFINE(kbimerge) 129 UVMMAP_EVCNT_DEFINE(knomerge) 130 UVMMAP_EVCNT_DEFINE(map_call) 131 UVMMAP_EVCNT_DEFINE(mlk_call) 132 UVMMAP_EVCNT_DEFINE(mlk_hint) 133 UVMMAP_EVCNT_DEFINE(mlk_list) 134 UVMMAP_EVCNT_DEFINE(mlk_tree) 135 UVMMAP_EVCNT_DEFINE(mlk_treeloop) 136 UVMMAP_EVCNT_DEFINE(mlk_listloop) 137 138 UVMMAP_EVCNT_DEFINE(uke_alloc) 139 UVMMAP_EVCNT_DEFINE(uke_free) 140 UVMMAP_EVCNT_DEFINE(ukh_alloc) 141 UVMMAP_EVCNT_DEFINE(ukh_free) 142 143 const char vmmapbsy[] = "vmmapbsy"; 144 145 /* 146 * cache for vmspace structures. 147 */ 148 149 static struct pool_cache uvm_vmspace_cache; 150 151 /* 152 * cache for dynamically-allocated map entries. 153 */ 154 155 static struct pool_cache uvm_map_entry_cache; 156 157 MALLOC_DEFINE(M_VMMAP, "VM map", "VM map structures"); 158 MALLOC_DEFINE(M_VMPMAP, "VM pmap", "VM pmap"); 159 160 #ifdef PMAP_GROWKERNEL 161 /* 162 * This global represents the end of the kernel virtual address 163 * space. If we want to exceed this, we must grow the kernel 164 * virtual address space dynamically. 165 * 166 * Note, this variable is locked by kernel_map's lock. 167 */ 168 vaddr_t uvm_maxkaddr; 169 #endif 170 171 #ifndef __USER_VA0_IS_SAFE 172 #ifndef __USER_VA0_DISABLE_DEFAULT 173 #define __USER_VA0_DISABLE_DEFAULT 1 174 #endif 175 #ifdef USER_VA0_DISABLE_DEFAULT /* kernel config option overrides */ 176 #undef __USER_VA0_DISABLE_DEFAULT 177 #define __USER_VA0_DISABLE_DEFAULT USER_VA0_DISABLE_DEFAULT 178 #endif 179 static int user_va0_disable = __USER_VA0_DISABLE_DEFAULT; 180 #endif 181 182 /* 183 * macros 184 */ 185 186 /* 187 * VM_MAP_USE_KMAPENT: determine if uvm_kmapent_alloc/free is used 188 * for the vm_map. 189 */ 190 extern struct vm_map *pager_map; /* XXX */ 191 #define VM_MAP_USE_KMAPENT_FLAGS(flags) \ 192 (((flags) & VM_MAP_INTRSAFE) != 0) 193 #define VM_MAP_USE_KMAPENT(map) \ 194 (VM_MAP_USE_KMAPENT_FLAGS((map)->flags) || (map) == kernel_map) 195 196 /* 197 * UVM_ET_ISCOMPATIBLE: check some requirements for map entry merging 198 */ 199 200 #define UVM_ET_ISCOMPATIBLE(ent, type, uobj, meflags, \ 201 prot, maxprot, inh, adv, wire) \ 202 ((ent)->etype == (type) && \ 203 (((ent)->flags ^ (meflags)) & (UVM_MAP_NOMERGE | UVM_MAP_QUANTUM)) \ 204 == 0 && \ 205 (ent)->object.uvm_obj == (uobj) && \ 206 (ent)->protection == (prot) && \ 207 (ent)->max_protection == (maxprot) && \ 208 (ent)->inheritance == (inh) && \ 209 (ent)->advice == (adv) && \ 210 (ent)->wired_count == (wire)) 211 212 /* 213 * uvm_map_entry_link: insert entry into a map 214 * 215 * => map must be locked 216 */ 217 #define uvm_map_entry_link(map, after_where, entry) do { \ 218 uvm_mapent_check(entry); \ 219 (map)->nentries++; \ 220 (entry)->prev = (after_where); \ 221 (entry)->next = (after_where)->next; \ 222 (entry)->prev->next = (entry); \ 223 (entry)->next->prev = (entry); \ 224 uvm_rb_insert((map), (entry)); \ 225 } while (/*CONSTCOND*/ 0) 226 227 /* 228 * uvm_map_entry_unlink: remove entry from a map 229 * 230 * => map must be locked 231 */ 232 #define uvm_map_entry_unlink(map, entry) do { \ 233 KASSERT((entry) != (map)->first_free); \ 234 KASSERT((entry) != (map)->hint); \ 235 uvm_mapent_check(entry); \ 236 (map)->nentries--; \ 237 (entry)->next->prev = (entry)->prev; \ 238 (entry)->prev->next = (entry)->next; \ 239 uvm_rb_remove((map), (entry)); \ 240 } while (/*CONSTCOND*/ 0) 241 242 /* 243 * SAVE_HINT: saves the specified entry as the hint for future lookups. 244 * 245 * => map need not be locked. 246 */ 247 #define SAVE_HINT(map, check, value) do { \ 248 if ((map)->hint == (check)) \ 249 (map)->hint = (value); \ 250 } while (/*CONSTCOND*/ 0) 251 252 /* 253 * clear_hints: ensure that hints don't point to the entry. 254 * 255 * => map must be write-locked. 256 */ 257 static void 258 clear_hints(struct vm_map *map, struct vm_map_entry *ent) 259 { 260 261 SAVE_HINT(map, ent, ent->prev); 262 if (map->first_free == ent) { 263 map->first_free = ent->prev; 264 } 265 } 266 267 /* 268 * VM_MAP_RANGE_CHECK: check and correct range 269 * 270 * => map must at least be read locked 271 */ 272 273 #define VM_MAP_RANGE_CHECK(map, start, end) do { \ 274 if (start < vm_map_min(map)) \ 275 start = vm_map_min(map); \ 276 if (end > vm_map_max(map)) \ 277 end = vm_map_max(map); \ 278 if (start > end) \ 279 start = end; \ 280 } while (/*CONSTCOND*/ 0) 281 282 /* 283 * local prototypes 284 */ 285 286 static struct vm_map_entry * 287 uvm_mapent_alloc(struct vm_map *, int); 288 static struct vm_map_entry * 289 uvm_mapent_alloc_split(struct vm_map *, 290 const struct vm_map_entry *, int, 291 struct uvm_mapent_reservation *); 292 static void uvm_mapent_copy(struct vm_map_entry *, struct vm_map_entry *); 293 static void uvm_mapent_free(struct vm_map_entry *); 294 #if defined(DEBUG) 295 static void _uvm_mapent_check(const struct vm_map_entry *, const char *, 296 int); 297 #define uvm_mapent_check(map) _uvm_mapent_check(map, __FILE__, __LINE__) 298 #else /* defined(DEBUG) */ 299 #define uvm_mapent_check(e) /* nothing */ 300 #endif /* defined(DEBUG) */ 301 static struct vm_map_entry * 302 uvm_kmapent_alloc(struct vm_map *, int); 303 static void uvm_kmapent_free(struct vm_map_entry *); 304 static vsize_t uvm_kmapent_overhead(vsize_t); 305 306 static void uvm_map_entry_unwire(struct vm_map *, struct vm_map_entry *); 307 static void uvm_map_reference_amap(struct vm_map_entry *, int); 308 static int uvm_map_space_avail(vaddr_t *, vsize_t, voff_t, vsize_t, int, 309 struct vm_map_entry *); 310 static void uvm_map_unreference_amap(struct vm_map_entry *, int); 311 312 int _uvm_map_sanity(struct vm_map *); 313 int _uvm_tree_sanity(struct vm_map *); 314 static vsize_t uvm_rb_maxgap(const struct vm_map_entry *); 315 316 #define ROOT_ENTRY(map) ((struct vm_map_entry *)(map)->rb_tree.rbt_root) 317 #define LEFT_ENTRY(entry) ((struct vm_map_entry *)(entry)->rb_node.rb_left) 318 #define RIGHT_ENTRY(entry) ((struct vm_map_entry *)(entry)->rb_node.rb_right) 319 #define PARENT_ENTRY(map, entry) \ 320 (ROOT_ENTRY(map) == (entry) \ 321 ? NULL : (struct vm_map_entry *)RB_FATHER(&(entry)->rb_node)) 322 323 static int 324 uvm_map_compare_nodes(void *ctx, const void *nparent, const void *nkey) 325 { 326 const struct vm_map_entry *eparent = nparent; 327 const struct vm_map_entry *ekey = nkey; 328 329 KASSERT(eparent->start < ekey->start || eparent->start >= ekey->end); 330 KASSERT(ekey->start < eparent->start || ekey->start >= eparent->end); 331 332 if (eparent->start < ekey->start) 333 return -1; 334 if (eparent->end >= ekey->start) 335 return 1; 336 return 0; 337 } 338 339 static int 340 uvm_map_compare_key(void *ctx, const void *nparent, const void *vkey) 341 { 342 const struct vm_map_entry *eparent = nparent; 343 const vaddr_t va = *(const vaddr_t *) vkey; 344 345 if (eparent->start < va) 346 return -1; 347 if (eparent->end >= va) 348 return 1; 349 return 0; 350 } 351 352 static const rb_tree_ops_t uvm_map_tree_ops = { 353 .rbto_compare_nodes = uvm_map_compare_nodes, 354 .rbto_compare_key = uvm_map_compare_key, 355 .rbto_node_offset = offsetof(struct vm_map_entry, rb_node), 356 .rbto_context = NULL 357 }; 358 359 /* 360 * uvm_rb_gap: return the gap size between our entry and next entry. 361 */ 362 static inline vsize_t 363 uvm_rb_gap(const struct vm_map_entry *entry) 364 { 365 366 KASSERT(entry->next != NULL); 367 return entry->next->start - entry->end; 368 } 369 370 static vsize_t 371 uvm_rb_maxgap(const struct vm_map_entry *entry) 372 { 373 struct vm_map_entry *child; 374 vsize_t maxgap = entry->gap; 375 376 /* 377 * We need maxgap to be the largest gap of us or any of our 378 * descendents. Since each of our children's maxgap is the 379 * cached value of their largest gap of themselves or their 380 * descendents, we can just use that value and avoid recursing 381 * down the tree to calculate it. 382 */ 383 if ((child = LEFT_ENTRY(entry)) != NULL && maxgap < child->maxgap) 384 maxgap = child->maxgap; 385 386 if ((child = RIGHT_ENTRY(entry)) != NULL && maxgap < child->maxgap) 387 maxgap = child->maxgap; 388 389 return maxgap; 390 } 391 392 static void 393 uvm_rb_fixup(struct vm_map *map, struct vm_map_entry *entry) 394 { 395 struct vm_map_entry *parent; 396 397 KASSERT(entry->gap == uvm_rb_gap(entry)); 398 entry->maxgap = uvm_rb_maxgap(entry); 399 400 while ((parent = PARENT_ENTRY(map, entry)) != NULL) { 401 struct vm_map_entry *brother; 402 vsize_t maxgap = parent->gap; 403 unsigned int which; 404 405 KDASSERT(parent->gap == uvm_rb_gap(parent)); 406 if (maxgap < entry->maxgap) 407 maxgap = entry->maxgap; 408 /* 409 * Since we work towards the root, we know entry's maxgap 410 * value is OK, but its brothers may now be out-of-date due 411 * to rebalancing. So refresh it. 412 */ 413 which = RB_POSITION(&entry->rb_node) ^ RB_DIR_OTHER; 414 brother = (struct vm_map_entry *)parent->rb_node.rb_nodes[which]; 415 if (brother != NULL) { 416 KDASSERT(brother->gap == uvm_rb_gap(brother)); 417 brother->maxgap = uvm_rb_maxgap(brother); 418 if (maxgap < brother->maxgap) 419 maxgap = brother->maxgap; 420 } 421 422 parent->maxgap = maxgap; 423 entry = parent; 424 } 425 } 426 427 static void 428 uvm_rb_insert(struct vm_map *map, struct vm_map_entry *entry) 429 { 430 struct vm_map_entry *ret; 431 432 entry->gap = entry->maxgap = uvm_rb_gap(entry); 433 if (entry->prev != &map->header) 434 entry->prev->gap = uvm_rb_gap(entry->prev); 435 436 ret = rb_tree_insert_node(&map->rb_tree, entry); 437 KASSERTMSG(ret == entry, 438 ("uvm_rb_insert: map %p: duplicate entry %p", map, ret) 439 ); 440 441 /* 442 * If the previous entry is not our immediate left child, then it's an 443 * ancestor and will be fixed up on the way to the root. We don't 444 * have to check entry->prev against &map->header since &map->header 445 * will never be in the tree. 446 */ 447 uvm_rb_fixup(map, 448 LEFT_ENTRY(entry) == entry->prev ? entry->prev : entry); 449 } 450 451 static void 452 uvm_rb_remove(struct vm_map *map, struct vm_map_entry *entry) 453 { 454 struct vm_map_entry *prev_parent = NULL, *next_parent = NULL; 455 456 /* 457 * If we are removing an interior node, then an adjacent node will 458 * be used to replace its position in the tree. Therefore we will 459 * need to fixup the tree starting at the parent of the replacement 460 * node. So record their parents for later use. 461 */ 462 if (entry->prev != &map->header) 463 prev_parent = PARENT_ENTRY(map, entry->prev); 464 if (entry->next != &map->header) 465 next_parent = PARENT_ENTRY(map, entry->next); 466 467 rb_tree_remove_node(&map->rb_tree, entry); 468 469 /* 470 * If the previous node has a new parent, fixup the tree starting 471 * at the previous node's old parent. 472 */ 473 if (entry->prev != &map->header) { 474 /* 475 * Update the previous entry's gap due to our absence. 476 */ 477 entry->prev->gap = uvm_rb_gap(entry->prev); 478 uvm_rb_fixup(map, entry->prev); 479 if (prev_parent != NULL 480 && prev_parent != entry 481 && prev_parent != PARENT_ENTRY(map, entry->prev)) 482 uvm_rb_fixup(map, prev_parent); 483 } 484 485 /* 486 * If the next node has a new parent, fixup the tree starting 487 * at the next node's old parent. 488 */ 489 if (entry->next != &map->header) { 490 uvm_rb_fixup(map, entry->next); 491 if (next_parent != NULL 492 && next_parent != entry 493 && next_parent != PARENT_ENTRY(map, entry->next)) 494 uvm_rb_fixup(map, next_parent); 495 } 496 } 497 498 #if defined(DEBUG) 499 int uvm_debug_check_map = 0; 500 int uvm_debug_check_rbtree = 0; 501 #define uvm_map_check(map, name) \ 502 _uvm_map_check((map), (name), __FILE__, __LINE__) 503 static void 504 _uvm_map_check(struct vm_map *map, const char *name, 505 const char *file, int line) 506 { 507 508 if ((uvm_debug_check_map && _uvm_map_sanity(map)) || 509 (uvm_debug_check_rbtree && _uvm_tree_sanity(map))) { 510 panic("uvm_map_check failed: \"%s\" map=%p (%s:%d)", 511 name, map, file, line); 512 } 513 } 514 #else /* defined(DEBUG) */ 515 #define uvm_map_check(map, name) /* nothing */ 516 #endif /* defined(DEBUG) */ 517 518 #if defined(DEBUG) || defined(DDB) 519 int 520 _uvm_map_sanity(struct vm_map *map) 521 { 522 bool first_free_found = false; 523 bool hint_found = false; 524 const struct vm_map_entry *e; 525 struct vm_map_entry *hint = map->hint; 526 527 e = &map->header; 528 for (;;) { 529 if (map->first_free == e) { 530 first_free_found = true; 531 } else if (!first_free_found && e->next->start > e->end) { 532 printf("first_free %p should be %p\n", 533 map->first_free, e); 534 return -1; 535 } 536 if (hint == e) { 537 hint_found = true; 538 } 539 540 e = e->next; 541 if (e == &map->header) { 542 break; 543 } 544 } 545 if (!first_free_found) { 546 printf("stale first_free\n"); 547 return -1; 548 } 549 if (!hint_found) { 550 printf("stale hint\n"); 551 return -1; 552 } 553 return 0; 554 } 555 556 int 557 _uvm_tree_sanity(struct vm_map *map) 558 { 559 struct vm_map_entry *tmp, *trtmp; 560 int n = 0, i = 1; 561 562 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 563 if (tmp->gap != uvm_rb_gap(tmp)) { 564 printf("%d/%d gap %lx != %lx %s\n", 565 n + 1, map->nentries, 566 (ulong)tmp->gap, (ulong)uvm_rb_gap(tmp), 567 tmp->next == &map->header ? "(last)" : ""); 568 goto error; 569 } 570 /* 571 * If any entries are out of order, tmp->gap will be unsigned 572 * and will likely exceed the size of the map. 573 */ 574 if (tmp->gap >= vm_map_max(map) - vm_map_min(map)) { 575 printf("too large gap %zu\n", (size_t)tmp->gap); 576 goto error; 577 } 578 n++; 579 } 580 581 if (n != map->nentries) { 582 printf("nentries: %d vs %d\n", n, map->nentries); 583 goto error; 584 } 585 586 trtmp = NULL; 587 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 588 if (tmp->maxgap != uvm_rb_maxgap(tmp)) { 589 printf("maxgap %lx != %lx\n", 590 (ulong)tmp->maxgap, 591 (ulong)uvm_rb_maxgap(tmp)); 592 goto error; 593 } 594 if (trtmp != NULL && trtmp->start >= tmp->start) { 595 printf("corrupt: 0x%"PRIxVADDR"x >= 0x%"PRIxVADDR"x\n", 596 trtmp->start, tmp->start); 597 goto error; 598 } 599 600 trtmp = tmp; 601 } 602 603 for (tmp = map->header.next; tmp != &map->header; 604 tmp = tmp->next, i++) { 605 trtmp = rb_tree_iterate(&map->rb_tree, tmp, RB_DIR_LEFT); 606 if (trtmp == NULL) 607 trtmp = &map->header; 608 if (tmp->prev != trtmp) { 609 printf("lookup: %d: %p->prev=%p: %p\n", 610 i, tmp, tmp->prev, trtmp); 611 goto error; 612 } 613 trtmp = rb_tree_iterate(&map->rb_tree, tmp, RB_DIR_RIGHT); 614 if (trtmp == NULL) 615 trtmp = &map->header; 616 if (tmp->next != trtmp) { 617 printf("lookup: %d: %p->next=%p: %p\n", 618 i, tmp, tmp->next, trtmp); 619 goto error; 620 } 621 trtmp = rb_tree_find_node(&map->rb_tree, &tmp->start); 622 if (trtmp != tmp) { 623 printf("lookup: %d: %p - %p: %p\n", i, tmp, trtmp, 624 PARENT_ENTRY(map, tmp)); 625 goto error; 626 } 627 } 628 629 return (0); 630 error: 631 return (-1); 632 } 633 #endif /* defined(DEBUG) || defined(DDB) */ 634 635 #ifdef DIAGNOSTIC 636 static struct vm_map *uvm_kmapent_map(struct vm_map_entry *); 637 #endif 638 639 /* 640 * vm_map_lock: acquire an exclusive (write) lock on a map. 641 * 642 * => Note that "intrsafe" maps use only exclusive, spin locks. 643 * 644 * => The locking protocol provides for guaranteed upgrade from shared -> 645 * exclusive by whichever thread currently has the map marked busy. 646 * See "LOCKING PROTOCOL NOTES" in uvm_map.h. This is horrible; among 647 * other problems, it defeats any fairness guarantees provided by RW 648 * locks. 649 */ 650 651 void 652 vm_map_lock(struct vm_map *map) 653 { 654 655 if ((map->flags & VM_MAP_INTRSAFE) != 0) { 656 mutex_spin_enter(&map->mutex); 657 return; 658 } 659 660 for (;;) { 661 rw_enter(&map->lock, RW_WRITER); 662 if (map->busy == NULL) 663 break; 664 if (map->busy == curlwp) 665 break; 666 mutex_enter(&map->misc_lock); 667 rw_exit(&map->lock); 668 if (map->busy != NULL) 669 cv_wait(&map->cv, &map->misc_lock); 670 mutex_exit(&map->misc_lock); 671 } 672 673 map->timestamp++; 674 } 675 676 /* 677 * vm_map_lock_try: try to lock a map, failing if it is already locked. 678 */ 679 680 bool 681 vm_map_lock_try(struct vm_map *map) 682 { 683 684 if ((map->flags & VM_MAP_INTRSAFE) != 0) 685 return mutex_tryenter(&map->mutex); 686 if (!rw_tryenter(&map->lock, RW_WRITER)) 687 return false; 688 if (map->busy != NULL) { 689 rw_exit(&map->lock); 690 return false; 691 } 692 693 map->timestamp++; 694 return true; 695 } 696 697 /* 698 * vm_map_unlock: release an exclusive lock on a map. 699 */ 700 701 void 702 vm_map_unlock(struct vm_map *map) 703 { 704 705 if ((map->flags & VM_MAP_INTRSAFE) != 0) 706 mutex_spin_exit(&map->mutex); 707 else { 708 KASSERT(rw_write_held(&map->lock)); 709 KASSERT(map->busy == NULL || map->busy == curlwp); 710 rw_exit(&map->lock); 711 } 712 } 713 714 /* 715 * vm_map_unbusy: mark the map as unbusy, and wake any waiters that 716 * want an exclusive lock. 717 */ 718 719 void 720 vm_map_unbusy(struct vm_map *map) 721 { 722 723 KASSERT(map->busy == curlwp); 724 725 /* 726 * Safe to clear 'busy' and 'waiters' with only a read lock held: 727 * 728 * o they can only be set with a write lock held 729 * o writers are blocked out with a read or write hold 730 * o at any time, only one thread owns the set of values 731 */ 732 mutex_enter(&map->misc_lock); 733 map->busy = NULL; 734 cv_broadcast(&map->cv); 735 mutex_exit(&map->misc_lock); 736 } 737 738 /* 739 * vm_map_lock_read: acquire a shared (read) lock on a map. 740 */ 741 742 void 743 vm_map_lock_read(struct vm_map *map) 744 { 745 746 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 747 748 rw_enter(&map->lock, RW_READER); 749 } 750 751 /* 752 * vm_map_unlock_read: release a shared lock on a map. 753 */ 754 755 void 756 vm_map_unlock_read(struct vm_map *map) 757 { 758 759 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 760 761 rw_exit(&map->lock); 762 } 763 764 /* 765 * vm_map_busy: mark a map as busy. 766 * 767 * => the caller must hold the map write locked 768 */ 769 770 void 771 vm_map_busy(struct vm_map *map) 772 { 773 774 KASSERT(rw_write_held(&map->lock)); 775 KASSERT(map->busy == NULL); 776 777 map->busy = curlwp; 778 } 779 780 /* 781 * vm_map_locked_p: return true if the map is write locked. 782 * 783 * => only for debug purposes like KASSERTs. 784 * => should not be used to verify that a map is not locked. 785 */ 786 787 bool 788 vm_map_locked_p(struct vm_map *map) 789 { 790 791 if ((map->flags & VM_MAP_INTRSAFE) != 0) { 792 return mutex_owned(&map->mutex); 793 } else { 794 return rw_write_held(&map->lock); 795 } 796 } 797 798 /* 799 * uvm_mapent_alloc: allocate a map entry 800 */ 801 802 static struct vm_map_entry * 803 uvm_mapent_alloc(struct vm_map *map, int flags) 804 { 805 struct vm_map_entry *me; 806 int pflags = (flags & UVM_FLAG_NOWAIT) ? PR_NOWAIT : PR_WAITOK; 807 UVMHIST_FUNC("uvm_mapent_alloc"); UVMHIST_CALLED(maphist); 808 809 if (VM_MAP_USE_KMAPENT(map)) { 810 me = uvm_kmapent_alloc(map, flags); 811 } else { 812 me = pool_cache_get(&uvm_map_entry_cache, pflags); 813 if (__predict_false(me == NULL)) 814 return NULL; 815 me->flags = 0; 816 } 817 818 UVMHIST_LOG(maphist, "<- new entry=0x%x [kentry=%d]", me, 819 ((map->flags & VM_MAP_INTRSAFE) != 0 || map == kernel_map), 0, 0); 820 return (me); 821 } 822 823 /* 824 * uvm_mapent_alloc_split: allocate a map entry for clipping. 825 * 826 * => map must be locked by caller if UVM_MAP_QUANTUM is set. 827 */ 828 829 static struct vm_map_entry * 830 uvm_mapent_alloc_split(struct vm_map *map, 831 const struct vm_map_entry *old_entry, int flags, 832 struct uvm_mapent_reservation *umr) 833 { 834 struct vm_map_entry *me; 835 836 KASSERT(!VM_MAP_USE_KMAPENT(map) || 837 (old_entry->flags & UVM_MAP_QUANTUM) || !UMR_EMPTY(umr)); 838 839 if (old_entry->flags & UVM_MAP_QUANTUM) { 840 struct vm_map_kernel *vmk = vm_map_to_kernel(map); 841 842 KASSERT(vm_map_locked_p(map)); 843 me = vmk->vmk_merged_entries; 844 KASSERT(me); 845 vmk->vmk_merged_entries = me->next; 846 KASSERT(me->flags & UVM_MAP_QUANTUM); 847 } else { 848 me = uvm_mapent_alloc(map, flags); 849 } 850 851 return me; 852 } 853 854 /* 855 * uvm_mapent_free: free map entry 856 */ 857 858 static void 859 uvm_mapent_free(struct vm_map_entry *me) 860 { 861 UVMHIST_FUNC("uvm_mapent_free"); UVMHIST_CALLED(maphist); 862 863 UVMHIST_LOG(maphist,"<- freeing map entry=0x%x [flags=%d]", 864 me, me->flags, 0, 0); 865 if (me->flags & UVM_MAP_KERNEL) { 866 uvm_kmapent_free(me); 867 } else { 868 pool_cache_put(&uvm_map_entry_cache, me); 869 } 870 } 871 872 /* 873 * uvm_mapent_free_merged: free merged map entry 874 * 875 * => keep the entry if needed. 876 * => caller shouldn't hold map locked if VM_MAP_USE_KMAPENT(map) is true. 877 * => map should be locked if UVM_MAP_QUANTUM is set. 878 */ 879 880 static void 881 uvm_mapent_free_merged(struct vm_map *map, struct vm_map_entry *me) 882 { 883 884 KASSERT(!(me->flags & UVM_MAP_KERNEL) || uvm_kmapent_map(me) == map); 885 886 if (me->flags & UVM_MAP_QUANTUM) { 887 /* 888 * keep this entry for later splitting. 889 */ 890 struct vm_map_kernel *vmk; 891 892 KASSERT(vm_map_locked_p(map)); 893 KASSERT(VM_MAP_IS_KERNEL(map)); 894 KASSERT(!VM_MAP_USE_KMAPENT(map) || 895 (me->flags & UVM_MAP_KERNEL)); 896 897 vmk = vm_map_to_kernel(map); 898 me->next = vmk->vmk_merged_entries; 899 vmk->vmk_merged_entries = me; 900 } else { 901 uvm_mapent_free(me); 902 } 903 } 904 905 /* 906 * uvm_mapent_copy: copy a map entry, preserving flags 907 */ 908 909 static inline void 910 uvm_mapent_copy(struct vm_map_entry *src, struct vm_map_entry *dst) 911 { 912 913 memcpy(dst, src, ((char *)&src->uvm_map_entry_stop_copy) - 914 ((char *)src)); 915 } 916 917 /* 918 * uvm_mapent_overhead: calculate maximum kva overhead necessary for 919 * map entries. 920 * 921 * => size and flags are the same as uvm_km_suballoc's ones. 922 */ 923 924 vsize_t 925 uvm_mapent_overhead(vsize_t size, int flags) 926 { 927 928 if (VM_MAP_USE_KMAPENT_FLAGS(flags)) { 929 return uvm_kmapent_overhead(size); 930 } 931 return 0; 932 } 933 934 #if defined(DEBUG) 935 static void 936 _uvm_mapent_check(const struct vm_map_entry *entry, const char *file, int line) 937 { 938 939 if (entry->start >= entry->end) { 940 goto bad; 941 } 942 if (UVM_ET_ISOBJ(entry)) { 943 if (entry->object.uvm_obj == NULL) { 944 goto bad; 945 } 946 } else if (UVM_ET_ISSUBMAP(entry)) { 947 if (entry->object.sub_map == NULL) { 948 goto bad; 949 } 950 } else { 951 if (entry->object.uvm_obj != NULL || 952 entry->object.sub_map != NULL) { 953 goto bad; 954 } 955 } 956 if (!UVM_ET_ISOBJ(entry)) { 957 if (entry->offset != 0) { 958 goto bad; 959 } 960 } 961 962 return; 963 964 bad: 965 panic("%s: bad entry %p (%s:%d)", __func__, entry, file, line); 966 } 967 #endif /* defined(DEBUG) */ 968 969 /* 970 * uvm_map_entry_unwire: unwire a map entry 971 * 972 * => map should be locked by caller 973 */ 974 975 static inline void 976 uvm_map_entry_unwire(struct vm_map *map, struct vm_map_entry *entry) 977 { 978 979 entry->wired_count = 0; 980 uvm_fault_unwire_locked(map, entry->start, entry->end); 981 } 982 983 984 /* 985 * wrapper for calling amap_ref() 986 */ 987 static inline void 988 uvm_map_reference_amap(struct vm_map_entry *entry, int flags) 989 { 990 991 amap_ref(entry->aref.ar_amap, entry->aref.ar_pageoff, 992 (entry->end - entry->start) >> PAGE_SHIFT, flags); 993 } 994 995 996 /* 997 * wrapper for calling amap_unref() 998 */ 999 static inline void 1000 uvm_map_unreference_amap(struct vm_map_entry *entry, int flags) 1001 { 1002 1003 amap_unref(entry->aref.ar_amap, entry->aref.ar_pageoff, 1004 (entry->end - entry->start) >> PAGE_SHIFT, flags); 1005 } 1006 1007 1008 /* 1009 * uvm_map_init: init mapping system at boot time. 1010 */ 1011 1012 void 1013 uvm_map_init(void) 1014 { 1015 #if defined(UVMHIST) 1016 static struct uvm_history_ent maphistbuf[100]; 1017 static struct uvm_history_ent pdhistbuf[100]; 1018 #endif 1019 1020 /* 1021 * first, init logging system. 1022 */ 1023 1024 UVMHIST_FUNC("uvm_map_init"); 1025 UVMHIST_INIT_STATIC(maphist, maphistbuf); 1026 UVMHIST_INIT_STATIC(pdhist, pdhistbuf); 1027 UVMHIST_CALLED(maphist); 1028 UVMHIST_LOG(maphist,"<starting uvm map system>", 0, 0, 0, 0); 1029 1030 /* 1031 * initialize the global lock for kernel map entry. 1032 */ 1033 1034 mutex_init(&uvm_kentry_lock, MUTEX_DRIVER, IPL_VM); 1035 1036 /* 1037 * initialize caches. 1038 */ 1039 1040 pool_cache_bootstrap(&uvm_map_entry_cache, sizeof(struct vm_map_entry), 1041 0, 0, 0, "vmmpepl", NULL, IPL_NONE, NULL, NULL, NULL); 1042 pool_cache_bootstrap(&uvm_vmspace_cache, sizeof(struct vmspace), 1043 0, 0, 0, "vmsppl", NULL, IPL_NONE, NULL, NULL, NULL); 1044 } 1045 1046 /* 1047 * clippers 1048 */ 1049 1050 /* 1051 * uvm_mapent_splitadj: adjust map entries for splitting, after uvm_mapent_copy. 1052 */ 1053 1054 static void 1055 uvm_mapent_splitadj(struct vm_map_entry *entry1, struct vm_map_entry *entry2, 1056 vaddr_t splitat) 1057 { 1058 vaddr_t adj; 1059 1060 KASSERT(entry1->start < splitat); 1061 KASSERT(splitat < entry1->end); 1062 1063 adj = splitat - entry1->start; 1064 entry1->end = entry2->start = splitat; 1065 1066 if (entry1->aref.ar_amap) { 1067 amap_splitref(&entry1->aref, &entry2->aref, adj); 1068 } 1069 if (UVM_ET_ISSUBMAP(entry1)) { 1070 /* ... unlikely to happen, but play it safe */ 1071 uvm_map_reference(entry1->object.sub_map); 1072 } else if (UVM_ET_ISOBJ(entry1)) { 1073 KASSERT(entry1->object.uvm_obj != NULL); /* suppress coverity */ 1074 entry2->offset += adj; 1075 if (entry1->object.uvm_obj->pgops && 1076 entry1->object.uvm_obj->pgops->pgo_reference) 1077 entry1->object.uvm_obj->pgops->pgo_reference( 1078 entry1->object.uvm_obj); 1079 } 1080 } 1081 1082 /* 1083 * uvm_map_clip_start: ensure that the entry begins at or after 1084 * the starting address, if it doesn't we split the entry. 1085 * 1086 * => caller should use UVM_MAP_CLIP_START macro rather than calling 1087 * this directly 1088 * => map must be locked by caller 1089 */ 1090 1091 void 1092 uvm_map_clip_start(struct vm_map *map, struct vm_map_entry *entry, 1093 vaddr_t start, struct uvm_mapent_reservation *umr) 1094 { 1095 struct vm_map_entry *new_entry; 1096 1097 /* uvm_map_simplify_entry(map, entry); */ /* XXX */ 1098 1099 uvm_map_check(map, "clip_start entry"); 1100 uvm_mapent_check(entry); 1101 1102 /* 1103 * Split off the front portion. note that we must insert the new 1104 * entry BEFORE this one, so that this entry has the specified 1105 * starting address. 1106 */ 1107 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 1108 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1109 uvm_mapent_splitadj(new_entry, entry, start); 1110 uvm_map_entry_link(map, entry->prev, new_entry); 1111 1112 uvm_map_check(map, "clip_start leave"); 1113 } 1114 1115 /* 1116 * uvm_map_clip_end: ensure that the entry ends at or before 1117 * the ending address, if it does't we split the reference 1118 * 1119 * => caller should use UVM_MAP_CLIP_END macro rather than calling 1120 * this directly 1121 * => map must be locked by caller 1122 */ 1123 1124 void 1125 uvm_map_clip_end(struct vm_map *map, struct vm_map_entry *entry, vaddr_t end, 1126 struct uvm_mapent_reservation *umr) 1127 { 1128 struct vm_map_entry *new_entry; 1129 1130 uvm_map_check(map, "clip_end entry"); 1131 uvm_mapent_check(entry); 1132 1133 /* 1134 * Create a new entry and insert it 1135 * AFTER the specified entry 1136 */ 1137 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 1138 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1139 uvm_mapent_splitadj(entry, new_entry, end); 1140 uvm_map_entry_link(map, entry, new_entry); 1141 1142 uvm_map_check(map, "clip_end leave"); 1143 } 1144 1145 static void 1146 vm_map_drain(struct vm_map *map, uvm_flag_t flags) 1147 { 1148 1149 if (!VM_MAP_IS_KERNEL(map)) { 1150 return; 1151 } 1152 1153 uvm_km_va_drain(map, flags); 1154 } 1155 1156 /* 1157 * M A P - m a i n e n t r y p o i n t 1158 */ 1159 /* 1160 * uvm_map: establish a valid mapping in a map 1161 * 1162 * => assume startp is page aligned. 1163 * => assume size is a multiple of PAGE_SIZE. 1164 * => assume sys_mmap provides enough of a "hint" to have us skip 1165 * over text/data/bss area. 1166 * => map must be unlocked (we will lock it) 1167 * => <uobj,uoffset> value meanings (4 cases): 1168 * [1] <NULL,uoffset> == uoffset is a hint for PMAP_PREFER 1169 * [2] <NULL,UVM_UNKNOWN_OFFSET> == don't PMAP_PREFER 1170 * [3] <uobj,uoffset> == normal mapping 1171 * [4] <uobj,UVM_UNKNOWN_OFFSET> == uvm_map finds offset based on VA 1172 * 1173 * case [4] is for kernel mappings where we don't know the offset until 1174 * we've found a virtual address. note that kernel object offsets are 1175 * always relative to vm_map_min(kernel_map). 1176 * 1177 * => if `align' is non-zero, we align the virtual address to the specified 1178 * alignment. 1179 * this is provided as a mechanism for large pages. 1180 * 1181 * => XXXCDC: need way to map in external amap? 1182 */ 1183 1184 int 1185 uvm_map(struct vm_map *map, vaddr_t *startp /* IN/OUT */, vsize_t size, 1186 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags) 1187 { 1188 struct uvm_map_args args; 1189 struct vm_map_entry *new_entry; 1190 int error; 1191 1192 KASSERT((flags & UVM_FLAG_QUANTUM) == 0 || VM_MAP_IS_KERNEL(map)); 1193 KASSERT((size & PAGE_MASK) == 0); 1194 1195 #ifndef __USER_VA0_IS_SAFE 1196 if ((flags & UVM_FLAG_FIXED) && *startp == 0 && 1197 !VM_MAP_IS_KERNEL(map) && user_va0_disable) 1198 return EACCES; 1199 #endif 1200 1201 /* 1202 * for pager_map, allocate the new entry first to avoid sleeping 1203 * for memory while we have the map locked. 1204 * 1205 * Also, because we allocate entries for in-kernel maps 1206 * a bit differently (cf. uvm_kmapent_alloc/free), we need to 1207 * allocate them before locking the map. 1208 */ 1209 1210 new_entry = NULL; 1211 if (VM_MAP_USE_KMAPENT(map) || (flags & UVM_FLAG_QUANTUM) || 1212 map == pager_map) { 1213 new_entry = uvm_mapent_alloc(map, (flags & UVM_FLAG_NOWAIT)); 1214 if (__predict_false(new_entry == NULL)) 1215 return ENOMEM; 1216 if (flags & UVM_FLAG_QUANTUM) 1217 new_entry->flags |= UVM_MAP_QUANTUM; 1218 } 1219 if (map == pager_map) 1220 flags |= UVM_FLAG_NOMERGE; 1221 1222 error = uvm_map_prepare(map, *startp, size, uobj, uoffset, align, 1223 flags, &args); 1224 if (!error) { 1225 error = uvm_map_enter(map, &args, new_entry); 1226 *startp = args.uma_start; 1227 } else if (new_entry) { 1228 uvm_mapent_free(new_entry); 1229 } 1230 1231 #if defined(DEBUG) 1232 if (!error && VM_MAP_IS_KERNEL(map)) { 1233 uvm_km_check_empty(map, *startp, *startp + size); 1234 } 1235 #endif /* defined(DEBUG) */ 1236 1237 return error; 1238 } 1239 1240 int 1241 uvm_map_prepare(struct vm_map *map, vaddr_t start, vsize_t size, 1242 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags, 1243 struct uvm_map_args *args) 1244 { 1245 struct vm_map_entry *prev_entry; 1246 vm_prot_t prot = UVM_PROTECTION(flags); 1247 vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1248 1249 UVMHIST_FUNC("uvm_map_prepare"); 1250 UVMHIST_CALLED(maphist); 1251 1252 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1253 map, start, size, flags); 1254 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1255 1256 /* 1257 * detect a popular device driver bug. 1258 */ 1259 1260 KASSERT(doing_shutdown || curlwp != NULL || 1261 (map->flags & VM_MAP_INTRSAFE)); 1262 1263 /* 1264 * zero-sized mapping doesn't make any sense. 1265 */ 1266 KASSERT(size > 0); 1267 1268 KASSERT((~flags & (UVM_FLAG_NOWAIT | UVM_FLAG_WAITVA)) != 0); 1269 1270 uvm_map_check(map, "map entry"); 1271 1272 /* 1273 * check sanity of protection code 1274 */ 1275 1276 if ((prot & maxprot) != prot) { 1277 UVMHIST_LOG(maphist, "<- prot. failure: prot=0x%x, max=0x%x", 1278 prot, maxprot,0,0); 1279 return EACCES; 1280 } 1281 1282 /* 1283 * figure out where to put new VM range 1284 */ 1285 1286 retry: 1287 if (vm_map_lock_try(map) == false) { 1288 if ((flags & UVM_FLAG_TRYLOCK) != 0 && 1289 (map->flags & VM_MAP_INTRSAFE) == 0) { 1290 return EAGAIN; 1291 } 1292 vm_map_lock(map); /* could sleep here */ 1293 } 1294 prev_entry = uvm_map_findspace(map, start, size, &start, 1295 uobj, uoffset, align, flags); 1296 if (prev_entry == NULL) { 1297 unsigned int timestamp; 1298 1299 timestamp = map->timestamp; 1300 UVMHIST_LOG(maphist,"waiting va timestamp=0x%x", 1301 timestamp,0,0,0); 1302 map->flags |= VM_MAP_WANTVA; 1303 vm_map_unlock(map); 1304 1305 /* 1306 * try to reclaim kva and wait until someone does unmap. 1307 * fragile locking here, so we awaken every second to 1308 * recheck the condition. 1309 */ 1310 1311 vm_map_drain(map, flags); 1312 1313 mutex_enter(&map->misc_lock); 1314 while ((map->flags & VM_MAP_WANTVA) != 0 && 1315 map->timestamp == timestamp) { 1316 if ((flags & UVM_FLAG_WAITVA) == 0) { 1317 mutex_exit(&map->misc_lock); 1318 UVMHIST_LOG(maphist, 1319 "<- uvm_map_findspace failed!", 0,0,0,0); 1320 return ENOMEM; 1321 } else { 1322 cv_timedwait(&map->cv, &map->misc_lock, hz); 1323 } 1324 } 1325 mutex_exit(&map->misc_lock); 1326 goto retry; 1327 } 1328 1329 #ifdef PMAP_GROWKERNEL 1330 /* 1331 * If the kernel pmap can't map the requested space, 1332 * then allocate more resources for it. 1333 */ 1334 if (map == kernel_map && uvm_maxkaddr < (start + size)) 1335 uvm_maxkaddr = pmap_growkernel(start + size); 1336 #endif 1337 1338 UVMMAP_EVCNT_INCR(map_call); 1339 1340 /* 1341 * if uobj is null, then uoffset is either a VAC hint for PMAP_PREFER 1342 * [typically from uvm_map_reserve] or it is UVM_UNKNOWN_OFFSET. in 1343 * either case we want to zero it before storing it in the map entry 1344 * (because it looks strange and confusing when debugging...) 1345 * 1346 * if uobj is not null 1347 * if uoffset is not UVM_UNKNOWN_OFFSET then we have a normal mapping 1348 * and we do not need to change uoffset. 1349 * if uoffset is UVM_UNKNOWN_OFFSET then we need to find the offset 1350 * now (based on the starting address of the map). this case is 1351 * for kernel object mappings where we don't know the offset until 1352 * the virtual address is found (with uvm_map_findspace). the 1353 * offset is the distance we are from the start of the map. 1354 */ 1355 1356 if (uobj == NULL) { 1357 uoffset = 0; 1358 } else { 1359 if (uoffset == UVM_UNKNOWN_OFFSET) { 1360 KASSERT(UVM_OBJ_IS_KERN_OBJECT(uobj)); 1361 uoffset = start - vm_map_min(kernel_map); 1362 } 1363 } 1364 1365 args->uma_flags = flags; 1366 args->uma_prev = prev_entry; 1367 args->uma_start = start; 1368 args->uma_size = size; 1369 args->uma_uobj = uobj; 1370 args->uma_uoffset = uoffset; 1371 1372 UVMHIST_LOG(maphist, "<- done!", 0,0,0,0); 1373 return 0; 1374 } 1375 1376 int 1377 uvm_map_enter(struct vm_map *map, const struct uvm_map_args *args, 1378 struct vm_map_entry *new_entry) 1379 { 1380 struct vm_map_entry *prev_entry = args->uma_prev; 1381 struct vm_map_entry *dead = NULL; 1382 1383 const uvm_flag_t flags = args->uma_flags; 1384 const vm_prot_t prot = UVM_PROTECTION(flags); 1385 const vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1386 const vm_inherit_t inherit = UVM_INHERIT(flags); 1387 const int amapwaitflag = (flags & UVM_FLAG_NOWAIT) ? 1388 AMAP_EXTEND_NOWAIT : 0; 1389 const int advice = UVM_ADVICE(flags); 1390 const int meflagval = (flags & UVM_FLAG_QUANTUM) ? 1391 UVM_MAP_QUANTUM : 0; 1392 1393 vaddr_t start = args->uma_start; 1394 vsize_t size = args->uma_size; 1395 struct uvm_object *uobj = args->uma_uobj; 1396 voff_t uoffset = args->uma_uoffset; 1397 1398 const int kmap = (vm_map_pmap(map) == pmap_kernel()); 1399 int merged = 0; 1400 int error; 1401 int newetype; 1402 1403 UVMHIST_FUNC("uvm_map_enter"); 1404 UVMHIST_CALLED(maphist); 1405 1406 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1407 map, start, size, flags); 1408 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1409 1410 KASSERT(map->hint == prev_entry); /* bimerge case assumes this */ 1411 1412 if (flags & UVM_FLAG_QUANTUM) { 1413 KASSERT(new_entry); 1414 KASSERT(new_entry->flags & UVM_MAP_QUANTUM); 1415 } 1416 1417 if (uobj) 1418 newetype = UVM_ET_OBJ; 1419 else 1420 newetype = 0; 1421 1422 if (flags & UVM_FLAG_COPYONW) { 1423 newetype |= UVM_ET_COPYONWRITE; 1424 if ((flags & UVM_FLAG_OVERLAY) == 0) 1425 newetype |= UVM_ET_NEEDSCOPY; 1426 } 1427 1428 /* 1429 * try and insert in map by extending previous entry, if possible. 1430 * XXX: we don't try and pull back the next entry. might be useful 1431 * for a stack, but we are currently allocating our stack in advance. 1432 */ 1433 1434 if (flags & UVM_FLAG_NOMERGE) 1435 goto nomerge; 1436 1437 if (prev_entry->end == start && 1438 prev_entry != &map->header && 1439 UVM_ET_ISCOMPATIBLE(prev_entry, newetype, uobj, meflagval, 1440 prot, maxprot, inherit, advice, 0)) { 1441 1442 if (uobj && prev_entry->offset + 1443 (prev_entry->end - prev_entry->start) != uoffset) 1444 goto forwardmerge; 1445 1446 /* 1447 * can't extend a shared amap. note: no need to lock amap to 1448 * look at refs since we don't care about its exact value. 1449 * if it is one (i.e. we have only reference) it will stay there 1450 */ 1451 1452 if (prev_entry->aref.ar_amap && 1453 amap_refs(prev_entry->aref.ar_amap) != 1) { 1454 goto forwardmerge; 1455 } 1456 1457 if (prev_entry->aref.ar_amap) { 1458 error = amap_extend(prev_entry, size, 1459 amapwaitflag | AMAP_EXTEND_FORWARDS); 1460 if (error) 1461 goto nomerge; 1462 } 1463 1464 if (kmap) { 1465 UVMMAP_EVCNT_INCR(kbackmerge); 1466 } else { 1467 UVMMAP_EVCNT_INCR(ubackmerge); 1468 } 1469 UVMHIST_LOG(maphist," starting back merge", 0, 0, 0, 0); 1470 1471 /* 1472 * drop our reference to uobj since we are extending a reference 1473 * that we already have (the ref count can not drop to zero). 1474 */ 1475 1476 if (uobj && uobj->pgops->pgo_detach) 1477 uobj->pgops->pgo_detach(uobj); 1478 1479 /* 1480 * Now that we've merged the entries, note that we've grown 1481 * and our gap has shrunk. Then fix the tree. 1482 */ 1483 prev_entry->end += size; 1484 prev_entry->gap -= size; 1485 uvm_rb_fixup(map, prev_entry); 1486 1487 uvm_map_check(map, "map backmerged"); 1488 1489 UVMHIST_LOG(maphist,"<- done (via backmerge)!", 0, 0, 0, 0); 1490 merged++; 1491 } 1492 1493 forwardmerge: 1494 if (prev_entry->next->start == (start + size) && 1495 prev_entry->next != &map->header && 1496 UVM_ET_ISCOMPATIBLE(prev_entry->next, newetype, uobj, meflagval, 1497 prot, maxprot, inherit, advice, 0)) { 1498 1499 if (uobj && prev_entry->next->offset != uoffset + size) 1500 goto nomerge; 1501 1502 /* 1503 * can't extend a shared amap. note: no need to lock amap to 1504 * look at refs since we don't care about its exact value. 1505 * if it is one (i.e. we have only reference) it will stay there. 1506 * 1507 * note that we also can't merge two amaps, so if we 1508 * merged with the previous entry which has an amap, 1509 * and the next entry also has an amap, we give up. 1510 * 1511 * Interesting cases: 1512 * amap, new, amap -> give up second merge (single fwd extend) 1513 * amap, new, none -> double forward extend (extend again here) 1514 * none, new, amap -> double backward extend (done here) 1515 * uobj, new, amap -> single backward extend (done here) 1516 * 1517 * XXX should we attempt to deal with someone refilling 1518 * the deallocated region between two entries that are 1519 * backed by the same amap (ie, arefs is 2, "prev" and 1520 * "next" refer to it, and adding this allocation will 1521 * close the hole, thus restoring arefs to 1 and 1522 * deallocating the "next" vm_map_entry)? -- @@@ 1523 */ 1524 1525 if (prev_entry->next->aref.ar_amap && 1526 (amap_refs(prev_entry->next->aref.ar_amap) != 1 || 1527 (merged && prev_entry->aref.ar_amap))) { 1528 goto nomerge; 1529 } 1530 1531 if (merged) { 1532 /* 1533 * Try to extend the amap of the previous entry to 1534 * cover the next entry as well. If it doesn't work 1535 * just skip on, don't actually give up, since we've 1536 * already completed the back merge. 1537 */ 1538 if (prev_entry->aref.ar_amap) { 1539 if (amap_extend(prev_entry, 1540 prev_entry->next->end - 1541 prev_entry->next->start, 1542 amapwaitflag | AMAP_EXTEND_FORWARDS)) 1543 goto nomerge; 1544 } 1545 1546 /* 1547 * Try to extend the amap of the *next* entry 1548 * back to cover the new allocation *and* the 1549 * previous entry as well (the previous merge 1550 * didn't have an amap already otherwise we 1551 * wouldn't be checking here for an amap). If 1552 * it doesn't work just skip on, again, don't 1553 * actually give up, since we've already 1554 * completed the back merge. 1555 */ 1556 else if (prev_entry->next->aref.ar_amap) { 1557 if (amap_extend(prev_entry->next, 1558 prev_entry->end - 1559 prev_entry->start, 1560 amapwaitflag | AMAP_EXTEND_BACKWARDS)) 1561 goto nomerge; 1562 } 1563 } else { 1564 /* 1565 * Pull the next entry's amap backwards to cover this 1566 * new allocation. 1567 */ 1568 if (prev_entry->next->aref.ar_amap) { 1569 error = amap_extend(prev_entry->next, size, 1570 amapwaitflag | AMAP_EXTEND_BACKWARDS); 1571 if (error) 1572 goto nomerge; 1573 } 1574 } 1575 1576 if (merged) { 1577 if (kmap) { 1578 UVMMAP_EVCNT_DECR(kbackmerge); 1579 UVMMAP_EVCNT_INCR(kbimerge); 1580 } else { 1581 UVMMAP_EVCNT_DECR(ubackmerge); 1582 UVMMAP_EVCNT_INCR(ubimerge); 1583 } 1584 } else { 1585 if (kmap) { 1586 UVMMAP_EVCNT_INCR(kforwmerge); 1587 } else { 1588 UVMMAP_EVCNT_INCR(uforwmerge); 1589 } 1590 } 1591 UVMHIST_LOG(maphist," starting forward merge", 0, 0, 0, 0); 1592 1593 /* 1594 * drop our reference to uobj since we are extending a reference 1595 * that we already have (the ref count can not drop to zero). 1596 * (if merged, we've already detached) 1597 */ 1598 if (uobj && uobj->pgops->pgo_detach && !merged) 1599 uobj->pgops->pgo_detach(uobj); 1600 1601 if (merged) { 1602 dead = prev_entry->next; 1603 prev_entry->end = dead->end; 1604 uvm_map_entry_unlink(map, dead); 1605 if (dead->aref.ar_amap != NULL) { 1606 prev_entry->aref = dead->aref; 1607 dead->aref.ar_amap = NULL; 1608 } 1609 } else { 1610 prev_entry->next->start -= size; 1611 if (prev_entry != &map->header) { 1612 prev_entry->gap -= size; 1613 KASSERT(prev_entry->gap == uvm_rb_gap(prev_entry)); 1614 uvm_rb_fixup(map, prev_entry); 1615 } 1616 if (uobj) 1617 prev_entry->next->offset = uoffset; 1618 } 1619 1620 uvm_map_check(map, "map forwardmerged"); 1621 1622 UVMHIST_LOG(maphist,"<- done forwardmerge", 0, 0, 0, 0); 1623 merged++; 1624 } 1625 1626 nomerge: 1627 if (!merged) { 1628 UVMHIST_LOG(maphist," allocating new map entry", 0, 0, 0, 0); 1629 if (kmap) { 1630 UVMMAP_EVCNT_INCR(knomerge); 1631 } else { 1632 UVMMAP_EVCNT_INCR(unomerge); 1633 } 1634 1635 /* 1636 * allocate new entry and link it in. 1637 */ 1638 1639 if (new_entry == NULL) { 1640 new_entry = uvm_mapent_alloc(map, 1641 (flags & UVM_FLAG_NOWAIT)); 1642 if (__predict_false(new_entry == NULL)) { 1643 error = ENOMEM; 1644 goto done; 1645 } 1646 } 1647 new_entry->start = start; 1648 new_entry->end = new_entry->start + size; 1649 new_entry->object.uvm_obj = uobj; 1650 new_entry->offset = uoffset; 1651 1652 new_entry->etype = newetype; 1653 1654 if (flags & UVM_FLAG_NOMERGE) { 1655 new_entry->flags |= UVM_MAP_NOMERGE; 1656 } 1657 1658 new_entry->protection = prot; 1659 new_entry->max_protection = maxprot; 1660 new_entry->inheritance = inherit; 1661 new_entry->wired_count = 0; 1662 new_entry->advice = advice; 1663 if (flags & UVM_FLAG_OVERLAY) { 1664 1665 /* 1666 * to_add: for BSS we overallocate a little since we 1667 * are likely to extend 1668 */ 1669 1670 vaddr_t to_add = (flags & UVM_FLAG_AMAPPAD) ? 1671 UVM_AMAP_CHUNK << PAGE_SHIFT : 0; 1672 struct vm_amap *amap = amap_alloc(size, to_add, 1673 (flags & UVM_FLAG_NOWAIT)); 1674 if (__predict_false(amap == NULL)) { 1675 error = ENOMEM; 1676 goto done; 1677 } 1678 new_entry->aref.ar_pageoff = 0; 1679 new_entry->aref.ar_amap = amap; 1680 } else { 1681 new_entry->aref.ar_pageoff = 0; 1682 new_entry->aref.ar_amap = NULL; 1683 } 1684 uvm_map_entry_link(map, prev_entry, new_entry); 1685 1686 /* 1687 * Update the free space hint 1688 */ 1689 1690 if ((map->first_free == prev_entry) && 1691 (prev_entry->end >= new_entry->start)) 1692 map->first_free = new_entry; 1693 1694 new_entry = NULL; 1695 } 1696 1697 map->size += size; 1698 1699 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 1700 1701 error = 0; 1702 done: 1703 if ((flags & UVM_FLAG_QUANTUM) == 0) { 1704 /* 1705 * vmk_merged_entries is locked by the map's lock. 1706 */ 1707 vm_map_unlock(map); 1708 } 1709 if (new_entry && error == 0) { 1710 KDASSERT(merged); 1711 uvm_mapent_free_merged(map, new_entry); 1712 new_entry = NULL; 1713 } 1714 if (dead) { 1715 KDASSERT(merged); 1716 uvm_mapent_free_merged(map, dead); 1717 } 1718 if ((flags & UVM_FLAG_QUANTUM) != 0) { 1719 vm_map_unlock(map); 1720 } 1721 if (new_entry != NULL) { 1722 uvm_mapent_free(new_entry); 1723 } 1724 return error; 1725 } 1726 1727 /* 1728 * uvm_map_lookup_entry_bytree: lookup an entry in tree 1729 */ 1730 1731 static inline bool 1732 uvm_map_lookup_entry_bytree(struct vm_map *map, vaddr_t address, 1733 struct vm_map_entry **entry /* OUT */) 1734 { 1735 struct vm_map_entry *prev = &map->header; 1736 struct vm_map_entry *cur = ROOT_ENTRY(map); 1737 1738 while (cur) { 1739 UVMMAP_EVCNT_INCR(mlk_treeloop); 1740 if (address >= cur->start) { 1741 if (address < cur->end) { 1742 *entry = cur; 1743 return true; 1744 } 1745 prev = cur; 1746 cur = RIGHT_ENTRY(cur); 1747 } else 1748 cur = LEFT_ENTRY(cur); 1749 } 1750 *entry = prev; 1751 return false; 1752 } 1753 1754 /* 1755 * uvm_map_lookup_entry: find map entry at or before an address 1756 * 1757 * => map must at least be read-locked by caller 1758 * => entry is returned in "entry" 1759 * => return value is true if address is in the returned entry 1760 */ 1761 1762 bool 1763 uvm_map_lookup_entry(struct vm_map *map, vaddr_t address, 1764 struct vm_map_entry **entry /* OUT */) 1765 { 1766 struct vm_map_entry *cur; 1767 bool use_tree = false; 1768 UVMHIST_FUNC("uvm_map_lookup_entry"); 1769 UVMHIST_CALLED(maphist); 1770 1771 UVMHIST_LOG(maphist,"(map=0x%x,addr=0x%x,ent=0x%x)", 1772 map, address, entry, 0); 1773 1774 /* 1775 * start looking either from the head of the 1776 * list, or from the hint. 1777 */ 1778 1779 cur = map->hint; 1780 1781 if (cur == &map->header) 1782 cur = cur->next; 1783 1784 UVMMAP_EVCNT_INCR(mlk_call); 1785 if (address >= cur->start) { 1786 1787 /* 1788 * go from hint to end of list. 1789 * 1790 * but first, make a quick check to see if 1791 * we are already looking at the entry we 1792 * want (which is usually the case). 1793 * note also that we don't need to save the hint 1794 * here... it is the same hint (unless we are 1795 * at the header, in which case the hint didn't 1796 * buy us anything anyway). 1797 */ 1798 1799 if (cur != &map->header && cur->end > address) { 1800 UVMMAP_EVCNT_INCR(mlk_hint); 1801 *entry = cur; 1802 UVMHIST_LOG(maphist,"<- got it via hint (0x%x)", 1803 cur, 0, 0, 0); 1804 uvm_mapent_check(*entry); 1805 return (true); 1806 } 1807 1808 if (map->nentries > 15) 1809 use_tree = true; 1810 } else { 1811 1812 /* 1813 * invalid hint. use tree. 1814 */ 1815 use_tree = true; 1816 } 1817 1818 uvm_map_check(map, __func__); 1819 1820 if (use_tree) { 1821 /* 1822 * Simple lookup in the tree. Happens when the hint is 1823 * invalid, or nentries reach a threshold. 1824 */ 1825 UVMMAP_EVCNT_INCR(mlk_tree); 1826 if (uvm_map_lookup_entry_bytree(map, address, entry)) { 1827 goto got; 1828 } else { 1829 goto failed; 1830 } 1831 } 1832 1833 /* 1834 * search linearly 1835 */ 1836 1837 UVMMAP_EVCNT_INCR(mlk_list); 1838 while (cur != &map->header) { 1839 UVMMAP_EVCNT_INCR(mlk_listloop); 1840 if (cur->end > address) { 1841 if (address >= cur->start) { 1842 /* 1843 * save this lookup for future 1844 * hints, and return 1845 */ 1846 1847 *entry = cur; 1848 got: 1849 SAVE_HINT(map, map->hint, *entry); 1850 UVMHIST_LOG(maphist,"<- search got it (0x%x)", 1851 cur, 0, 0, 0); 1852 KDASSERT((*entry)->start <= address); 1853 KDASSERT(address < (*entry)->end); 1854 uvm_mapent_check(*entry); 1855 return (true); 1856 } 1857 break; 1858 } 1859 cur = cur->next; 1860 } 1861 *entry = cur->prev; 1862 failed: 1863 SAVE_HINT(map, map->hint, *entry); 1864 UVMHIST_LOG(maphist,"<- failed!",0,0,0,0); 1865 KDASSERT((*entry) == &map->header || (*entry)->end <= address); 1866 KDASSERT((*entry)->next == &map->header || 1867 address < (*entry)->next->start); 1868 return (false); 1869 } 1870 1871 /* 1872 * See if the range between start and start + length fits in the gap 1873 * entry->next->start and entry->end. Returns 1 if fits, 0 if doesn't 1874 * fit, and -1 address wraps around. 1875 */ 1876 static int 1877 uvm_map_space_avail(vaddr_t *start, vsize_t length, voff_t uoffset, 1878 vsize_t align, int topdown, struct vm_map_entry *entry) 1879 { 1880 vaddr_t end; 1881 1882 #ifdef PMAP_PREFER 1883 /* 1884 * push start address forward as needed to avoid VAC alias problems. 1885 * we only do this if a valid offset is specified. 1886 */ 1887 1888 if (uoffset != UVM_UNKNOWN_OFFSET) 1889 PMAP_PREFER(uoffset, start, length, topdown); 1890 #endif 1891 if (align != 0) { 1892 if ((*start & (align - 1)) != 0) { 1893 if (topdown) 1894 *start &= ~(align - 1); 1895 else 1896 *start = roundup(*start, align); 1897 } 1898 /* 1899 * XXX Should we PMAP_PREFER() here again? 1900 * eh...i think we're okay 1901 */ 1902 } 1903 1904 /* 1905 * Find the end of the proposed new region. Be sure we didn't 1906 * wrap around the address; if so, we lose. Otherwise, if the 1907 * proposed new region fits before the next entry, we win. 1908 */ 1909 1910 end = *start + length; 1911 if (end < *start) 1912 return (-1); 1913 1914 if (entry->next->start >= end && *start >= entry->end) 1915 return (1); 1916 1917 return (0); 1918 } 1919 1920 /* 1921 * uvm_map_findspace: find "length" sized space in "map". 1922 * 1923 * => "hint" is a hint about where we want it, unless UVM_FLAG_FIXED is 1924 * set in "flags" (in which case we insist on using "hint"). 1925 * => "result" is VA returned 1926 * => uobj/uoffset are to be used to handle VAC alignment, if required 1927 * => if "align" is non-zero, we attempt to align to that value. 1928 * => caller must at least have read-locked map 1929 * => returns NULL on failure, or pointer to prev. map entry if success 1930 * => note this is a cross between the old vm_map_findspace and vm_map_find 1931 */ 1932 1933 struct vm_map_entry * 1934 uvm_map_findspace(struct vm_map *map, vaddr_t hint, vsize_t length, 1935 vaddr_t *result /* OUT */, struct uvm_object *uobj, voff_t uoffset, 1936 vsize_t align, int flags) 1937 { 1938 struct vm_map_entry *entry; 1939 struct vm_map_entry *child, *prev, *tmp; 1940 vaddr_t orig_hint; 1941 const int topdown = map->flags & VM_MAP_TOPDOWN; 1942 UVMHIST_FUNC("uvm_map_findspace"); 1943 UVMHIST_CALLED(maphist); 1944 1945 UVMHIST_LOG(maphist, "(map=0x%x, hint=0x%x, len=%d, flags=0x%x)", 1946 map, hint, length, flags); 1947 KASSERT((align & (align - 1)) == 0); 1948 KASSERT((flags & UVM_FLAG_FIXED) == 0 || align == 0); 1949 1950 uvm_map_check(map, "map_findspace entry"); 1951 1952 /* 1953 * remember the original hint. if we are aligning, then we 1954 * may have to try again with no alignment constraint if 1955 * we fail the first time. 1956 */ 1957 1958 orig_hint = hint; 1959 if (hint < vm_map_min(map)) { /* check ranges ... */ 1960 if (flags & UVM_FLAG_FIXED) { 1961 UVMHIST_LOG(maphist,"<- VA below map range",0,0,0,0); 1962 return (NULL); 1963 } 1964 hint = vm_map_min(map); 1965 } 1966 if (hint > vm_map_max(map)) { 1967 UVMHIST_LOG(maphist,"<- VA 0x%x > range [0x%x->0x%x]", 1968 hint, vm_map_min(map), vm_map_max(map), 0); 1969 return (NULL); 1970 } 1971 1972 /* 1973 * Look for the first possible address; if there's already 1974 * something at this address, we have to start after it. 1975 */ 1976 1977 /* 1978 * @@@: there are four, no, eight cases to consider. 1979 * 1980 * 0: found, fixed, bottom up -> fail 1981 * 1: found, fixed, top down -> fail 1982 * 2: found, not fixed, bottom up -> start after entry->end, 1983 * loop up 1984 * 3: found, not fixed, top down -> start before entry->start, 1985 * loop down 1986 * 4: not found, fixed, bottom up -> check entry->next->start, fail 1987 * 5: not found, fixed, top down -> check entry->next->start, fail 1988 * 6: not found, not fixed, bottom up -> check entry->next->start, 1989 * loop up 1990 * 7: not found, not fixed, top down -> check entry->next->start, 1991 * loop down 1992 * 1993 * as you can see, it reduces to roughly five cases, and that 1994 * adding top down mapping only adds one unique case (without 1995 * it, there would be four cases). 1996 */ 1997 1998 if ((flags & UVM_FLAG_FIXED) == 0 && hint == vm_map_min(map)) { 1999 entry = map->first_free; 2000 } else { 2001 if (uvm_map_lookup_entry(map, hint, &entry)) { 2002 /* "hint" address already in use ... */ 2003 if (flags & UVM_FLAG_FIXED) { 2004 UVMHIST_LOG(maphist, "<- fixed & VA in use", 2005 0, 0, 0, 0); 2006 return (NULL); 2007 } 2008 if (topdown) 2009 /* Start from lower gap. */ 2010 entry = entry->prev; 2011 } else if (flags & UVM_FLAG_FIXED) { 2012 if (entry->next->start >= hint + length && 2013 hint + length > hint) 2014 goto found; 2015 2016 /* "hint" address is gap but too small */ 2017 UVMHIST_LOG(maphist, "<- fixed mapping failed", 2018 0, 0, 0, 0); 2019 return (NULL); /* only one shot at it ... */ 2020 } else { 2021 /* 2022 * See if given hint fits in this gap. 2023 */ 2024 switch (uvm_map_space_avail(&hint, length, 2025 uoffset, align, topdown, entry)) { 2026 case 1: 2027 goto found; 2028 case -1: 2029 goto wraparound; 2030 } 2031 2032 if (topdown) { 2033 /* 2034 * Still there is a chance to fit 2035 * if hint > entry->end. 2036 */ 2037 } else { 2038 /* Start from higher gap. */ 2039 entry = entry->next; 2040 if (entry == &map->header) 2041 goto notfound; 2042 goto nextgap; 2043 } 2044 } 2045 } 2046 2047 /* 2048 * Note that all UVM_FLAGS_FIXED case is already handled. 2049 */ 2050 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2051 2052 /* Try to find the space in the red-black tree */ 2053 2054 /* Check slot before any entry */ 2055 hint = topdown ? entry->next->start - length : entry->end; 2056 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2057 topdown, entry)) { 2058 case 1: 2059 goto found; 2060 case -1: 2061 goto wraparound; 2062 } 2063 2064 nextgap: 2065 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2066 /* If there is not enough space in the whole tree, we fail */ 2067 tmp = ROOT_ENTRY(map); 2068 if (tmp == NULL || tmp->maxgap < length) 2069 goto notfound; 2070 2071 prev = NULL; /* previous candidate */ 2072 2073 /* Find an entry close to hint that has enough space */ 2074 for (; tmp;) { 2075 KASSERT(tmp->next->start == tmp->end + tmp->gap); 2076 if (topdown) { 2077 if (tmp->next->start < hint + length && 2078 (prev == NULL || tmp->end > prev->end)) { 2079 if (tmp->gap >= length) 2080 prev = tmp; 2081 else if ((child = LEFT_ENTRY(tmp)) != NULL 2082 && child->maxgap >= length) 2083 prev = tmp; 2084 } 2085 } else { 2086 if (tmp->end >= hint && 2087 (prev == NULL || tmp->end < prev->end)) { 2088 if (tmp->gap >= length) 2089 prev = tmp; 2090 else if ((child = RIGHT_ENTRY(tmp)) != NULL 2091 && child->maxgap >= length) 2092 prev = tmp; 2093 } 2094 } 2095 if (tmp->next->start < hint + length) 2096 child = RIGHT_ENTRY(tmp); 2097 else if (tmp->end > hint) 2098 child = LEFT_ENTRY(tmp); 2099 else { 2100 if (tmp->gap >= length) 2101 break; 2102 if (topdown) 2103 child = LEFT_ENTRY(tmp); 2104 else 2105 child = RIGHT_ENTRY(tmp); 2106 } 2107 if (child == NULL || child->maxgap < length) 2108 break; 2109 tmp = child; 2110 } 2111 2112 if (tmp != NULL && tmp->start < hint && hint < tmp->next->start) { 2113 /* 2114 * Check if the entry that we found satifies the 2115 * space requirement 2116 */ 2117 if (topdown) { 2118 if (hint > tmp->next->start - length) 2119 hint = tmp->next->start - length; 2120 } else { 2121 if (hint < tmp->end) 2122 hint = tmp->end; 2123 } 2124 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2125 topdown, tmp)) { 2126 case 1: 2127 entry = tmp; 2128 goto found; 2129 case -1: 2130 goto wraparound; 2131 } 2132 if (tmp->gap >= length) 2133 goto listsearch; 2134 } 2135 if (prev == NULL) 2136 goto notfound; 2137 2138 if (topdown) { 2139 KASSERT(orig_hint >= prev->next->start - length || 2140 prev->next->start - length > prev->next->start); 2141 hint = prev->next->start - length; 2142 } else { 2143 KASSERT(orig_hint <= prev->end); 2144 hint = prev->end; 2145 } 2146 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2147 topdown, prev)) { 2148 case 1: 2149 entry = prev; 2150 goto found; 2151 case -1: 2152 goto wraparound; 2153 } 2154 if (prev->gap >= length) 2155 goto listsearch; 2156 2157 if (topdown) 2158 tmp = LEFT_ENTRY(prev); 2159 else 2160 tmp = RIGHT_ENTRY(prev); 2161 for (;;) { 2162 KASSERT(tmp && tmp->maxgap >= length); 2163 if (topdown) 2164 child = RIGHT_ENTRY(tmp); 2165 else 2166 child = LEFT_ENTRY(tmp); 2167 if (child && child->maxgap >= length) { 2168 tmp = child; 2169 continue; 2170 } 2171 if (tmp->gap >= length) 2172 break; 2173 if (topdown) 2174 tmp = LEFT_ENTRY(tmp); 2175 else 2176 tmp = RIGHT_ENTRY(tmp); 2177 } 2178 2179 if (topdown) { 2180 KASSERT(orig_hint >= tmp->next->start - length || 2181 tmp->next->start - length > tmp->next->start); 2182 hint = tmp->next->start - length; 2183 } else { 2184 KASSERT(orig_hint <= tmp->end); 2185 hint = tmp->end; 2186 } 2187 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2188 topdown, tmp)) { 2189 case 1: 2190 entry = tmp; 2191 goto found; 2192 case -1: 2193 goto wraparound; 2194 } 2195 2196 /* 2197 * The tree fails to find an entry because of offset or alignment 2198 * restrictions. Search the list instead. 2199 */ 2200 listsearch: 2201 /* 2202 * Look through the rest of the map, trying to fit a new region in 2203 * the gap between existing regions, or after the very last region. 2204 * note: entry->end = base VA of current gap, 2205 * entry->next->start = VA of end of current gap 2206 */ 2207 2208 for (;;) { 2209 /* Update hint for current gap. */ 2210 hint = topdown ? entry->next->start - length : entry->end; 2211 2212 /* See if it fits. */ 2213 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2214 topdown, entry)) { 2215 case 1: 2216 goto found; 2217 case -1: 2218 goto wraparound; 2219 } 2220 2221 /* Advance to next/previous gap */ 2222 if (topdown) { 2223 if (entry == &map->header) { 2224 UVMHIST_LOG(maphist, "<- failed (off start)", 2225 0,0,0,0); 2226 goto notfound; 2227 } 2228 entry = entry->prev; 2229 } else { 2230 entry = entry->next; 2231 if (entry == &map->header) { 2232 UVMHIST_LOG(maphist, "<- failed (off end)", 2233 0,0,0,0); 2234 goto notfound; 2235 } 2236 } 2237 } 2238 2239 found: 2240 SAVE_HINT(map, map->hint, entry); 2241 *result = hint; 2242 UVMHIST_LOG(maphist,"<- got it! (result=0x%x)", hint, 0,0,0); 2243 KASSERT( topdown || hint >= orig_hint); 2244 KASSERT(!topdown || hint <= orig_hint); 2245 KASSERT(entry->end <= hint); 2246 KASSERT(hint + length <= entry->next->start); 2247 return (entry); 2248 2249 wraparound: 2250 UVMHIST_LOG(maphist, "<- failed (wrap around)", 0,0,0,0); 2251 2252 return (NULL); 2253 2254 notfound: 2255 UVMHIST_LOG(maphist, "<- failed (notfound)", 0,0,0,0); 2256 2257 return (NULL); 2258 } 2259 2260 /* 2261 * U N M A P - m a i n h e l p e r f u n c t i o n s 2262 */ 2263 2264 /* 2265 * uvm_unmap_remove: remove mappings from a vm_map (from "start" up to "stop") 2266 * 2267 * => caller must check alignment and size 2268 * => map must be locked by caller 2269 * => we return a list of map entries that we've remove from the map 2270 * in "entry_list" 2271 */ 2272 2273 void 2274 uvm_unmap_remove(struct vm_map *map, vaddr_t start, vaddr_t end, 2275 struct vm_map_entry **entry_list /* OUT */, 2276 struct uvm_mapent_reservation *umr, int flags) 2277 { 2278 struct vm_map_entry *entry, *first_entry, *next; 2279 vaddr_t len; 2280 UVMHIST_FUNC("uvm_unmap_remove"); UVMHIST_CALLED(maphist); 2281 2282 UVMHIST_LOG(maphist,"(map=0x%x, start=0x%x, end=0x%x)", 2283 map, start, end, 0); 2284 VM_MAP_RANGE_CHECK(map, start, end); 2285 2286 uvm_map_check(map, "unmap_remove entry"); 2287 2288 /* 2289 * find first entry 2290 */ 2291 2292 if (uvm_map_lookup_entry(map, start, &first_entry) == true) { 2293 /* clip and go... */ 2294 entry = first_entry; 2295 UVM_MAP_CLIP_START(map, entry, start, umr); 2296 /* critical! prevents stale hint */ 2297 SAVE_HINT(map, entry, entry->prev); 2298 } else { 2299 entry = first_entry->next; 2300 } 2301 2302 /* 2303 * Save the free space hint 2304 */ 2305 2306 if (map->first_free != &map->header && map->first_free->start >= start) 2307 map->first_free = entry->prev; 2308 2309 /* 2310 * note: we now re-use first_entry for a different task. we remove 2311 * a number of map entries from the map and save them in a linked 2312 * list headed by "first_entry". once we remove them from the map 2313 * the caller should unlock the map and drop the references to the 2314 * backing objects [c.f. uvm_unmap_detach]. the object is to 2315 * separate unmapping from reference dropping. why? 2316 * [1] the map has to be locked for unmapping 2317 * [2] the map need not be locked for reference dropping 2318 * [3] dropping references may trigger pager I/O, and if we hit 2319 * a pager that does synchronous I/O we may have to wait for it. 2320 * [4] we would like all waiting for I/O to occur with maps unlocked 2321 * so that we don't block other threads. 2322 */ 2323 2324 first_entry = NULL; 2325 *entry_list = NULL; 2326 2327 /* 2328 * break up the area into map entry sized regions and unmap. note 2329 * that all mappings have to be removed before we can even consider 2330 * dropping references to amaps or VM objects (otherwise we could end 2331 * up with a mapping to a page on the free list which would be very bad) 2332 */ 2333 2334 while ((entry != &map->header) && (entry->start < end)) { 2335 KASSERT((entry->flags & UVM_MAP_FIRST) == 0); 2336 2337 UVM_MAP_CLIP_END(map, entry, end, umr); 2338 next = entry->next; 2339 len = entry->end - entry->start; 2340 2341 /* 2342 * unwire before removing addresses from the pmap; otherwise 2343 * unwiring will put the entries back into the pmap (XXX). 2344 */ 2345 2346 if (VM_MAPENT_ISWIRED(entry)) { 2347 uvm_map_entry_unwire(map, entry); 2348 } 2349 if (flags & UVM_FLAG_VAONLY) { 2350 2351 /* nothing */ 2352 2353 } else if ((map->flags & VM_MAP_PAGEABLE) == 0) { 2354 2355 /* 2356 * if the map is non-pageable, any pages mapped there 2357 * must be wired and entered with pmap_kenter_pa(), 2358 * and we should free any such pages immediately. 2359 * this is mostly used for kmem_map. 2360 */ 2361 KASSERT(vm_map_pmap(map) == pmap_kernel()); 2362 2363 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2364 uvm_km_pgremove_intrsafe(map, entry->start, 2365 entry->end); 2366 pmap_kremove(entry->start, len); 2367 } 2368 } else if (UVM_ET_ISOBJ(entry) && 2369 UVM_OBJ_IS_KERN_OBJECT(entry->object.uvm_obj)) { 2370 KASSERT(vm_map_pmap(map) == pmap_kernel()); 2371 2372 /* 2373 * note: kernel object mappings are currently used in 2374 * two ways: 2375 * [1] "normal" mappings of pages in the kernel object 2376 * [2] uvm_km_valloc'd allocations in which we 2377 * pmap_enter in some non-kernel-object page 2378 * (e.g. vmapbuf). 2379 * 2380 * for case [1], we need to remove the mapping from 2381 * the pmap and then remove the page from the kernel 2382 * object (because, once pages in a kernel object are 2383 * unmapped they are no longer needed, unlike, say, 2384 * a vnode where you might want the data to persist 2385 * until flushed out of a queue). 2386 * 2387 * for case [2], we need to remove the mapping from 2388 * the pmap. there shouldn't be any pages at the 2389 * specified offset in the kernel object [but it 2390 * doesn't hurt to call uvm_km_pgremove just to be 2391 * safe?] 2392 * 2393 * uvm_km_pgremove currently does the following: 2394 * for pages in the kernel object in range: 2395 * - drops the swap slot 2396 * - uvm_pagefree the page 2397 */ 2398 2399 /* 2400 * remove mappings from pmap and drop the pages 2401 * from the object. offsets are always relative 2402 * to vm_map_min(kernel_map). 2403 */ 2404 2405 pmap_remove(pmap_kernel(), entry->start, 2406 entry->start + len); 2407 uvm_km_pgremove(entry->start, entry->end); 2408 2409 /* 2410 * null out kernel_object reference, we've just 2411 * dropped it 2412 */ 2413 2414 entry->etype &= ~UVM_ET_OBJ; 2415 entry->object.uvm_obj = NULL; 2416 } else if (UVM_ET_ISOBJ(entry) || entry->aref.ar_amap) { 2417 2418 /* 2419 * remove mappings the standard way. 2420 */ 2421 2422 pmap_remove(map->pmap, entry->start, entry->end); 2423 } 2424 2425 #if defined(DEBUG) 2426 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2427 2428 /* 2429 * check if there's remaining mapping, 2430 * which is a bug in caller. 2431 */ 2432 2433 vaddr_t va; 2434 for (va = entry->start; va < entry->end; 2435 va += PAGE_SIZE) { 2436 if (pmap_extract(vm_map_pmap(map), va, NULL)) { 2437 panic("uvm_unmap_remove: has mapping"); 2438 } 2439 } 2440 2441 if (VM_MAP_IS_KERNEL(map)) { 2442 uvm_km_check_empty(map, entry->start, 2443 entry->end); 2444 } 2445 } 2446 #endif /* defined(DEBUG) */ 2447 2448 /* 2449 * remove entry from map and put it on our list of entries 2450 * that we've nuked. then go to next entry. 2451 */ 2452 2453 UVMHIST_LOG(maphist, " removed map entry 0x%x", entry, 0, 0,0); 2454 2455 /* critical! prevents stale hint */ 2456 SAVE_HINT(map, entry, entry->prev); 2457 2458 uvm_map_entry_unlink(map, entry); 2459 KASSERT(map->size >= len); 2460 map->size -= len; 2461 entry->prev = NULL; 2462 entry->next = first_entry; 2463 first_entry = entry; 2464 entry = next; 2465 } 2466 2467 /* 2468 * Note: if map is dying, leave pmap_update() for pmap_destroy(), 2469 * which will be called later. 2470 */ 2471 if ((map->flags & VM_MAP_DYING) == 0) { 2472 pmap_update(vm_map_pmap(map)); 2473 } else { 2474 KASSERT(vm_map_pmap(map) != pmap_kernel()); 2475 } 2476 2477 uvm_map_check(map, "unmap_remove leave"); 2478 2479 /* 2480 * now we've cleaned up the map and are ready for the caller to drop 2481 * references to the mapped objects. 2482 */ 2483 2484 *entry_list = first_entry; 2485 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 2486 2487 if (map->flags & VM_MAP_WANTVA) { 2488 mutex_enter(&map->misc_lock); 2489 map->flags &= ~VM_MAP_WANTVA; 2490 cv_broadcast(&map->cv); 2491 mutex_exit(&map->misc_lock); 2492 } 2493 } 2494 2495 /* 2496 * uvm_unmap_detach: drop references in a chain of map entries 2497 * 2498 * => we will free the map entries as we traverse the list. 2499 */ 2500 2501 void 2502 uvm_unmap_detach(struct vm_map_entry *first_entry, int flags) 2503 { 2504 struct vm_map_entry *next_entry; 2505 UVMHIST_FUNC("uvm_unmap_detach"); UVMHIST_CALLED(maphist); 2506 2507 while (first_entry) { 2508 KASSERT(!VM_MAPENT_ISWIRED(first_entry)); 2509 UVMHIST_LOG(maphist, 2510 " detach 0x%x: amap=0x%x, obj=0x%x, submap?=%d", 2511 first_entry, first_entry->aref.ar_amap, 2512 first_entry->object.uvm_obj, 2513 UVM_ET_ISSUBMAP(first_entry)); 2514 2515 /* 2516 * drop reference to amap, if we've got one 2517 */ 2518 2519 if (first_entry->aref.ar_amap) 2520 uvm_map_unreference_amap(first_entry, flags); 2521 2522 /* 2523 * drop reference to our backing object, if we've got one 2524 */ 2525 2526 KASSERT(!UVM_ET_ISSUBMAP(first_entry)); 2527 if (UVM_ET_ISOBJ(first_entry) && 2528 first_entry->object.uvm_obj->pgops->pgo_detach) { 2529 (*first_entry->object.uvm_obj->pgops->pgo_detach) 2530 (first_entry->object.uvm_obj); 2531 } 2532 next_entry = first_entry->next; 2533 uvm_mapent_free(first_entry); 2534 first_entry = next_entry; 2535 } 2536 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 2537 } 2538 2539 /* 2540 * E X T R A C T I O N F U N C T I O N S 2541 */ 2542 2543 /* 2544 * uvm_map_reserve: reserve space in a vm_map for future use. 2545 * 2546 * => we reserve space in a map by putting a dummy map entry in the 2547 * map (dummy means obj=NULL, amap=NULL, prot=VM_PROT_NONE) 2548 * => map should be unlocked (we will write lock it) 2549 * => we return true if we were able to reserve space 2550 * => XXXCDC: should be inline? 2551 */ 2552 2553 int 2554 uvm_map_reserve(struct vm_map *map, vsize_t size, 2555 vaddr_t offset /* hint for pmap_prefer */, 2556 vsize_t align /* alignment */, 2557 vaddr_t *raddr /* IN:hint, OUT: reserved VA */, 2558 uvm_flag_t flags /* UVM_FLAG_FIXED or 0 */) 2559 { 2560 UVMHIST_FUNC("uvm_map_reserve"); UVMHIST_CALLED(maphist); 2561 2562 UVMHIST_LOG(maphist, "(map=0x%x, size=0x%x, offset=0x%x,addr=0x%x)", 2563 map,size,offset,raddr); 2564 2565 size = round_page(size); 2566 2567 /* 2568 * reserve some virtual space. 2569 */ 2570 2571 if (uvm_map(map, raddr, size, NULL, offset, align, 2572 UVM_MAPFLAG(UVM_PROT_NONE, UVM_PROT_NONE, UVM_INH_NONE, 2573 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE|flags)) != 0) { 2574 UVMHIST_LOG(maphist, "<- done (no VM)", 0,0,0,0); 2575 return (false); 2576 } 2577 2578 UVMHIST_LOG(maphist, "<- done (*raddr=0x%x)", *raddr,0,0,0); 2579 return (true); 2580 } 2581 2582 /* 2583 * uvm_map_replace: replace a reserved (blank) area of memory with 2584 * real mappings. 2585 * 2586 * => caller must WRITE-LOCK the map 2587 * => we return true if replacement was a success 2588 * => we expect the newents chain to have nnewents entrys on it and 2589 * we expect newents->prev to point to the last entry on the list 2590 * => note newents is allowed to be NULL 2591 */ 2592 2593 static int 2594 uvm_map_replace(struct vm_map *map, vaddr_t start, vaddr_t end, 2595 struct vm_map_entry *newents, int nnewents, vsize_t nsize, 2596 struct vm_map_entry **oldentryp) 2597 { 2598 struct vm_map_entry *oldent, *last; 2599 2600 uvm_map_check(map, "map_replace entry"); 2601 2602 /* 2603 * first find the blank map entry at the specified address 2604 */ 2605 2606 if (!uvm_map_lookup_entry(map, start, &oldent)) { 2607 return (false); 2608 } 2609 2610 /* 2611 * check to make sure we have a proper blank entry 2612 */ 2613 2614 if (end < oldent->end && !VM_MAP_USE_KMAPENT(map)) { 2615 UVM_MAP_CLIP_END(map, oldent, end, NULL); 2616 } 2617 if (oldent->start != start || oldent->end != end || 2618 oldent->object.uvm_obj != NULL || oldent->aref.ar_amap != NULL) { 2619 return (false); 2620 } 2621 2622 #ifdef DIAGNOSTIC 2623 2624 /* 2625 * sanity check the newents chain 2626 */ 2627 2628 { 2629 struct vm_map_entry *tmpent = newents; 2630 int nent = 0; 2631 vsize_t sz = 0; 2632 vaddr_t cur = start; 2633 2634 while (tmpent) { 2635 nent++; 2636 sz += tmpent->end - tmpent->start; 2637 if (tmpent->start < cur) 2638 panic("uvm_map_replace1"); 2639 if (tmpent->start >= tmpent->end || tmpent->end > end) { 2640 panic("uvm_map_replace2: " 2641 "tmpent->start=0x%"PRIxVADDR 2642 ", tmpent->end=0x%"PRIxVADDR 2643 ", end=0x%"PRIxVADDR, 2644 tmpent->start, tmpent->end, end); 2645 } 2646 cur = tmpent->end; 2647 if (tmpent->next) { 2648 if (tmpent->next->prev != tmpent) 2649 panic("uvm_map_replace3"); 2650 } else { 2651 if (newents->prev != tmpent) 2652 panic("uvm_map_replace4"); 2653 } 2654 tmpent = tmpent->next; 2655 } 2656 if (nent != nnewents) 2657 panic("uvm_map_replace5"); 2658 if (sz != nsize) 2659 panic("uvm_map_replace6"); 2660 } 2661 #endif 2662 2663 /* 2664 * map entry is a valid blank! replace it. (this does all the 2665 * work of map entry link/unlink...). 2666 */ 2667 2668 if (newents) { 2669 last = newents->prev; 2670 2671 /* critical: flush stale hints out of map */ 2672 SAVE_HINT(map, map->hint, newents); 2673 if (map->first_free == oldent) 2674 map->first_free = last; 2675 2676 last->next = oldent->next; 2677 last->next->prev = last; 2678 2679 /* Fix RB tree */ 2680 uvm_rb_remove(map, oldent); 2681 2682 newents->prev = oldent->prev; 2683 newents->prev->next = newents; 2684 map->nentries = map->nentries + (nnewents - 1); 2685 2686 /* Fixup the RB tree */ 2687 { 2688 int i; 2689 struct vm_map_entry *tmp; 2690 2691 tmp = newents; 2692 for (i = 0; i < nnewents && tmp; i++) { 2693 uvm_rb_insert(map, tmp); 2694 tmp = tmp->next; 2695 } 2696 } 2697 } else { 2698 /* NULL list of new entries: just remove the old one */ 2699 clear_hints(map, oldent); 2700 uvm_map_entry_unlink(map, oldent); 2701 } 2702 map->size -= end - start - nsize; 2703 2704 uvm_map_check(map, "map_replace leave"); 2705 2706 /* 2707 * now we can free the old blank entry and return. 2708 */ 2709 2710 *oldentryp = oldent; 2711 return (true); 2712 } 2713 2714 /* 2715 * uvm_map_extract: extract a mapping from a map and put it somewhere 2716 * (maybe removing the old mapping) 2717 * 2718 * => maps should be unlocked (we will write lock them) 2719 * => returns 0 on success, error code otherwise 2720 * => start must be page aligned 2721 * => len must be page sized 2722 * => flags: 2723 * UVM_EXTRACT_REMOVE: remove mappings from srcmap 2724 * UVM_EXTRACT_CONTIG: abort if unmapped area (advisory only) 2725 * UVM_EXTRACT_QREF: for a temporary extraction do quick obj refs 2726 * UVM_EXTRACT_FIXPROT: set prot to maxprot as we go 2727 * >>>NOTE: if you set REMOVE, you are not allowed to use CONTIG or QREF!<<< 2728 * >>>NOTE: QREF's must be unmapped via the QREF path, thus should only 2729 * be used from within the kernel in a kernel level map <<< 2730 */ 2731 2732 int 2733 uvm_map_extract(struct vm_map *srcmap, vaddr_t start, vsize_t len, 2734 struct vm_map *dstmap, vaddr_t *dstaddrp, int flags) 2735 { 2736 vaddr_t dstaddr, end, newend, oldoffset, fudge, orig_fudge; 2737 struct vm_map_entry *chain, *endchain, *entry, *orig_entry, *newentry, 2738 *deadentry, *oldentry; 2739 struct vm_map_entry *resentry = NULL; /* a dummy reservation entry */ 2740 vsize_t elen; 2741 int nchain, error, copy_ok; 2742 vsize_t nsize; 2743 UVMHIST_FUNC("uvm_map_extract"); UVMHIST_CALLED(maphist); 2744 2745 UVMHIST_LOG(maphist,"(srcmap=0x%x,start=0x%x, len=0x%x", srcmap, start, 2746 len,0); 2747 UVMHIST_LOG(maphist," ...,dstmap=0x%x, flags=0x%x)", dstmap,flags,0,0); 2748 2749 /* 2750 * step 0: sanity check: start must be on a page boundary, length 2751 * must be page sized. can't ask for CONTIG/QREF if you asked for 2752 * REMOVE. 2753 */ 2754 2755 KASSERT((start & PAGE_MASK) == 0 && (len & PAGE_MASK) == 0); 2756 KASSERT((flags & UVM_EXTRACT_REMOVE) == 0 || 2757 (flags & (UVM_EXTRACT_CONTIG|UVM_EXTRACT_QREF)) == 0); 2758 2759 /* 2760 * step 1: reserve space in the target map for the extracted area 2761 */ 2762 2763 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 2764 dstaddr = vm_map_min(dstmap); 2765 if (!uvm_map_reserve(dstmap, len, start, 0, &dstaddr, 0)) 2766 return (ENOMEM); 2767 *dstaddrp = dstaddr; /* pass address back to caller */ 2768 UVMHIST_LOG(maphist, " dstaddr=0x%x", dstaddr,0,0,0); 2769 } else { 2770 dstaddr = *dstaddrp; 2771 } 2772 2773 /* 2774 * step 2: setup for the extraction process loop by init'ing the 2775 * map entry chain, locking src map, and looking up the first useful 2776 * entry in the map. 2777 */ 2778 2779 end = start + len; 2780 newend = dstaddr + len; 2781 chain = endchain = NULL; 2782 nchain = 0; 2783 nsize = 0; 2784 vm_map_lock(srcmap); 2785 2786 if (uvm_map_lookup_entry(srcmap, start, &entry)) { 2787 2788 /* "start" is within an entry */ 2789 if (flags & UVM_EXTRACT_QREF) { 2790 2791 /* 2792 * for quick references we don't clip the entry, so 2793 * the entry may map space "before" the starting 2794 * virtual address... this is the "fudge" factor 2795 * (which can be non-zero only the first time 2796 * through the "while" loop in step 3). 2797 */ 2798 2799 fudge = start - entry->start; 2800 } else { 2801 2802 /* 2803 * normal reference: we clip the map to fit (thus 2804 * fudge is zero) 2805 */ 2806 2807 UVM_MAP_CLIP_START(srcmap, entry, start, NULL); 2808 SAVE_HINT(srcmap, srcmap->hint, entry->prev); 2809 fudge = 0; 2810 } 2811 } else { 2812 2813 /* "start" is not within an entry ... skip to next entry */ 2814 if (flags & UVM_EXTRACT_CONTIG) { 2815 error = EINVAL; 2816 goto bad; /* definite hole here ... */ 2817 } 2818 2819 entry = entry->next; 2820 fudge = 0; 2821 } 2822 2823 /* save values from srcmap for step 6 */ 2824 orig_entry = entry; 2825 orig_fudge = fudge; 2826 2827 /* 2828 * step 3: now start looping through the map entries, extracting 2829 * as we go. 2830 */ 2831 2832 while (entry->start < end && entry != &srcmap->header) { 2833 2834 /* if we are not doing a quick reference, clip it */ 2835 if ((flags & UVM_EXTRACT_QREF) == 0) 2836 UVM_MAP_CLIP_END(srcmap, entry, end, NULL); 2837 2838 /* clear needs_copy (allow chunking) */ 2839 if (UVM_ET_ISNEEDSCOPY(entry)) { 2840 amap_copy(srcmap, entry, 2841 AMAP_COPY_NOWAIT|AMAP_COPY_NOMERGE, start, end); 2842 if (UVM_ET_ISNEEDSCOPY(entry)) { /* failed? */ 2843 error = ENOMEM; 2844 goto bad; 2845 } 2846 2847 /* amap_copy could clip (during chunk)! update fudge */ 2848 if (fudge) { 2849 fudge = start - entry->start; 2850 orig_fudge = fudge; 2851 } 2852 } 2853 2854 /* calculate the offset of this from "start" */ 2855 oldoffset = (entry->start + fudge) - start; 2856 2857 /* allocate a new map entry */ 2858 newentry = uvm_mapent_alloc(dstmap, 0); 2859 if (newentry == NULL) { 2860 error = ENOMEM; 2861 goto bad; 2862 } 2863 2864 /* set up new map entry */ 2865 newentry->next = NULL; 2866 newentry->prev = endchain; 2867 newentry->start = dstaddr + oldoffset; 2868 newentry->end = 2869 newentry->start + (entry->end - (entry->start + fudge)); 2870 if (newentry->end > newend || newentry->end < newentry->start) 2871 newentry->end = newend; 2872 newentry->object.uvm_obj = entry->object.uvm_obj; 2873 if (newentry->object.uvm_obj) { 2874 if (newentry->object.uvm_obj->pgops->pgo_reference) 2875 newentry->object.uvm_obj->pgops-> 2876 pgo_reference(newentry->object.uvm_obj); 2877 newentry->offset = entry->offset + fudge; 2878 } else { 2879 newentry->offset = 0; 2880 } 2881 newentry->etype = entry->etype; 2882 newentry->protection = (flags & UVM_EXTRACT_FIXPROT) ? 2883 entry->max_protection : entry->protection; 2884 newentry->max_protection = entry->max_protection; 2885 newentry->inheritance = entry->inheritance; 2886 newentry->wired_count = 0; 2887 newentry->aref.ar_amap = entry->aref.ar_amap; 2888 if (newentry->aref.ar_amap) { 2889 newentry->aref.ar_pageoff = 2890 entry->aref.ar_pageoff + (fudge >> PAGE_SHIFT); 2891 uvm_map_reference_amap(newentry, AMAP_SHARED | 2892 ((flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0)); 2893 } else { 2894 newentry->aref.ar_pageoff = 0; 2895 } 2896 newentry->advice = entry->advice; 2897 if ((flags & UVM_EXTRACT_QREF) != 0) { 2898 newentry->flags |= UVM_MAP_NOMERGE; 2899 } 2900 2901 /* now link it on the chain */ 2902 nchain++; 2903 nsize += newentry->end - newentry->start; 2904 if (endchain == NULL) { 2905 chain = endchain = newentry; 2906 } else { 2907 endchain->next = newentry; 2908 endchain = newentry; 2909 } 2910 2911 /* end of 'while' loop! */ 2912 if ((flags & UVM_EXTRACT_CONTIG) && entry->end < end && 2913 (entry->next == &srcmap->header || 2914 entry->next->start != entry->end)) { 2915 error = EINVAL; 2916 goto bad; 2917 } 2918 entry = entry->next; 2919 fudge = 0; 2920 } 2921 2922 /* 2923 * step 4: close off chain (in format expected by uvm_map_replace) 2924 */ 2925 2926 if (chain) 2927 chain->prev = endchain; 2928 2929 /* 2930 * step 5: attempt to lock the dest map so we can pmap_copy. 2931 * note usage of copy_ok: 2932 * 1 => dstmap locked, pmap_copy ok, and we "replace" here (step 5) 2933 * 0 => dstmap unlocked, NO pmap_copy, and we will "replace" in step 7 2934 */ 2935 2936 if (srcmap == dstmap || vm_map_lock_try(dstmap) == true) { 2937 copy_ok = 1; 2938 if (!uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2939 nchain, nsize, &resentry)) { 2940 if (srcmap != dstmap) 2941 vm_map_unlock(dstmap); 2942 error = EIO; 2943 goto bad; 2944 } 2945 } else { 2946 copy_ok = 0; 2947 /* replace defered until step 7 */ 2948 } 2949 2950 /* 2951 * step 6: traverse the srcmap a second time to do the following: 2952 * - if we got a lock on the dstmap do pmap_copy 2953 * - if UVM_EXTRACT_REMOVE remove the entries 2954 * we make use of orig_entry and orig_fudge (saved in step 2) 2955 */ 2956 2957 if (copy_ok || (flags & UVM_EXTRACT_REMOVE)) { 2958 2959 /* purge possible stale hints from srcmap */ 2960 if (flags & UVM_EXTRACT_REMOVE) { 2961 SAVE_HINT(srcmap, srcmap->hint, orig_entry->prev); 2962 if (srcmap->first_free != &srcmap->header && 2963 srcmap->first_free->start >= start) 2964 srcmap->first_free = orig_entry->prev; 2965 } 2966 2967 entry = orig_entry; 2968 fudge = orig_fudge; 2969 deadentry = NULL; /* for UVM_EXTRACT_REMOVE */ 2970 2971 while (entry->start < end && entry != &srcmap->header) { 2972 if (copy_ok) { 2973 oldoffset = (entry->start + fudge) - start; 2974 elen = MIN(end, entry->end) - 2975 (entry->start + fudge); 2976 pmap_copy(dstmap->pmap, srcmap->pmap, 2977 dstaddr + oldoffset, elen, 2978 entry->start + fudge); 2979 } 2980 2981 /* we advance "entry" in the following if statement */ 2982 if (flags & UVM_EXTRACT_REMOVE) { 2983 pmap_remove(srcmap->pmap, entry->start, 2984 entry->end); 2985 oldentry = entry; /* save entry */ 2986 entry = entry->next; /* advance */ 2987 uvm_map_entry_unlink(srcmap, oldentry); 2988 /* add to dead list */ 2989 oldentry->next = deadentry; 2990 deadentry = oldentry; 2991 } else { 2992 entry = entry->next; /* advance */ 2993 } 2994 2995 /* end of 'while' loop */ 2996 fudge = 0; 2997 } 2998 pmap_update(srcmap->pmap); 2999 3000 /* 3001 * unlock dstmap. we will dispose of deadentry in 3002 * step 7 if needed 3003 */ 3004 3005 if (copy_ok && srcmap != dstmap) 3006 vm_map_unlock(dstmap); 3007 3008 } else { 3009 deadentry = NULL; 3010 } 3011 3012 /* 3013 * step 7: we are done with the source map, unlock. if copy_ok 3014 * is 0 then we have not replaced the dummy mapping in dstmap yet 3015 * and we need to do so now. 3016 */ 3017 3018 vm_map_unlock(srcmap); 3019 if ((flags & UVM_EXTRACT_REMOVE) && deadentry) 3020 uvm_unmap_detach(deadentry, 0); /* dispose of old entries */ 3021 3022 /* now do the replacement if we didn't do it in step 5 */ 3023 if (copy_ok == 0) { 3024 vm_map_lock(dstmap); 3025 error = uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 3026 nchain, nsize, &resentry); 3027 vm_map_unlock(dstmap); 3028 3029 if (error == false) { 3030 error = EIO; 3031 goto bad2; 3032 } 3033 } 3034 3035 if (resentry != NULL) 3036 uvm_mapent_free(resentry); 3037 3038 return (0); 3039 3040 /* 3041 * bad: failure recovery 3042 */ 3043 bad: 3044 vm_map_unlock(srcmap); 3045 bad2: /* src already unlocked */ 3046 if (chain) 3047 uvm_unmap_detach(chain, 3048 (flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0); 3049 3050 if (resentry != NULL) 3051 uvm_mapent_free(resentry); 3052 3053 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 3054 uvm_unmap(dstmap, dstaddr, dstaddr+len); /* ??? */ 3055 } 3056 return (error); 3057 } 3058 3059 /* end of extraction functions */ 3060 3061 /* 3062 * uvm_map_submap: punch down part of a map into a submap 3063 * 3064 * => only the kernel_map is allowed to be submapped 3065 * => the purpose of submapping is to break up the locking granularity 3066 * of a larger map 3067 * => the range specified must have been mapped previously with a uvm_map() 3068 * call [with uobj==NULL] to create a blank map entry in the main map. 3069 * [And it had better still be blank!] 3070 * => maps which contain submaps should never be copied or forked. 3071 * => to remove a submap, use uvm_unmap() on the main map 3072 * and then uvm_map_deallocate() the submap. 3073 * => main map must be unlocked. 3074 * => submap must have been init'd and have a zero reference count. 3075 * [need not be locked as we don't actually reference it] 3076 */ 3077 3078 int 3079 uvm_map_submap(struct vm_map *map, vaddr_t start, vaddr_t end, 3080 struct vm_map *submap) 3081 { 3082 struct vm_map_entry *entry; 3083 struct uvm_mapent_reservation umr; 3084 int error; 3085 3086 uvm_mapent_reserve(map, &umr, 2, 0); 3087 3088 vm_map_lock(map); 3089 VM_MAP_RANGE_CHECK(map, start, end); 3090 3091 if (uvm_map_lookup_entry(map, start, &entry)) { 3092 UVM_MAP_CLIP_START(map, entry, start, &umr); 3093 UVM_MAP_CLIP_END(map, entry, end, &umr); /* to be safe */ 3094 } else { 3095 entry = NULL; 3096 } 3097 3098 if (entry != NULL && 3099 entry->start == start && entry->end == end && 3100 entry->object.uvm_obj == NULL && entry->aref.ar_amap == NULL && 3101 !UVM_ET_ISCOPYONWRITE(entry) && !UVM_ET_ISNEEDSCOPY(entry)) { 3102 entry->etype |= UVM_ET_SUBMAP; 3103 entry->object.sub_map = submap; 3104 entry->offset = 0; 3105 uvm_map_reference(submap); 3106 error = 0; 3107 } else { 3108 error = EINVAL; 3109 } 3110 vm_map_unlock(map); 3111 3112 uvm_mapent_unreserve(map, &umr); 3113 3114 return error; 3115 } 3116 3117 /* 3118 * uvm_map_setup_kernel: init in-kernel map 3119 * 3120 * => map must not be in service yet. 3121 */ 3122 3123 void 3124 uvm_map_setup_kernel(struct vm_map_kernel *map, 3125 vaddr_t vmin, vaddr_t vmax, int flags) 3126 { 3127 3128 uvm_map_setup(&map->vmk_map, vmin, vmax, flags); 3129 callback_head_init(&map->vmk_reclaim_callback, IPL_VM); 3130 LIST_INIT(&map->vmk_kentry_free); 3131 map->vmk_merged_entries = NULL; 3132 } 3133 3134 3135 /* 3136 * uvm_map_protect: change map protection 3137 * 3138 * => set_max means set max_protection. 3139 * => map must be unlocked. 3140 */ 3141 3142 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 3143 ~VM_PROT_WRITE : VM_PROT_ALL) 3144 3145 int 3146 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end, 3147 vm_prot_t new_prot, bool set_max) 3148 { 3149 struct vm_map_entry *current, *entry; 3150 int error = 0; 3151 UVMHIST_FUNC("uvm_map_protect"); UVMHIST_CALLED(maphist); 3152 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_prot=0x%x)", 3153 map, start, end, new_prot); 3154 3155 vm_map_lock(map); 3156 VM_MAP_RANGE_CHECK(map, start, end); 3157 if (uvm_map_lookup_entry(map, start, &entry)) { 3158 UVM_MAP_CLIP_START(map, entry, start, NULL); 3159 } else { 3160 entry = entry->next; 3161 } 3162 3163 /* 3164 * make a first pass to check for protection violations. 3165 */ 3166 3167 current = entry; 3168 while ((current != &map->header) && (current->start < end)) { 3169 if (UVM_ET_ISSUBMAP(current)) { 3170 error = EINVAL; 3171 goto out; 3172 } 3173 if ((new_prot & current->max_protection) != new_prot) { 3174 error = EACCES; 3175 goto out; 3176 } 3177 /* 3178 * Don't allow VM_PROT_EXECUTE to be set on entries that 3179 * point to vnodes that are associated with a NOEXEC file 3180 * system. 3181 */ 3182 if (UVM_ET_ISOBJ(current) && 3183 UVM_OBJ_IS_VNODE(current->object.uvm_obj)) { 3184 struct vnode *vp = 3185 (struct vnode *) current->object.uvm_obj; 3186 3187 if ((new_prot & VM_PROT_EXECUTE) != 0 && 3188 (vp->v_mount->mnt_flag & MNT_NOEXEC) != 0) { 3189 error = EACCES; 3190 goto out; 3191 } 3192 } 3193 3194 current = current->next; 3195 } 3196 3197 /* go back and fix up protections (no need to clip this time). */ 3198 3199 current = entry; 3200 while ((current != &map->header) && (current->start < end)) { 3201 vm_prot_t old_prot; 3202 3203 UVM_MAP_CLIP_END(map, current, end, NULL); 3204 old_prot = current->protection; 3205 if (set_max) 3206 current->protection = 3207 (current->max_protection = new_prot) & old_prot; 3208 else 3209 current->protection = new_prot; 3210 3211 /* 3212 * update physical map if necessary. worry about copy-on-write 3213 * here -- CHECK THIS XXX 3214 */ 3215 3216 if (current->protection != old_prot) { 3217 /* update pmap! */ 3218 pmap_protect(map->pmap, current->start, current->end, 3219 current->protection & MASK(entry)); 3220 3221 /* 3222 * If this entry points at a vnode, and the 3223 * protection includes VM_PROT_EXECUTE, mark 3224 * the vnode as VEXECMAP. 3225 */ 3226 if (UVM_ET_ISOBJ(current)) { 3227 struct uvm_object *uobj = 3228 current->object.uvm_obj; 3229 3230 if (UVM_OBJ_IS_VNODE(uobj) && 3231 (current->protection & VM_PROT_EXECUTE)) { 3232 vn_markexec((struct vnode *) uobj); 3233 } 3234 } 3235 } 3236 3237 /* 3238 * If the map is configured to lock any future mappings, 3239 * wire this entry now if the old protection was VM_PROT_NONE 3240 * and the new protection is not VM_PROT_NONE. 3241 */ 3242 3243 if ((map->flags & VM_MAP_WIREFUTURE) != 0 && 3244 VM_MAPENT_ISWIRED(entry) == 0 && 3245 old_prot == VM_PROT_NONE && 3246 new_prot != VM_PROT_NONE) { 3247 if (uvm_map_pageable(map, entry->start, 3248 entry->end, false, 3249 UVM_LK_ENTER|UVM_LK_EXIT) != 0) { 3250 3251 /* 3252 * If locking the entry fails, remember the 3253 * error if it's the first one. Note we 3254 * still continue setting the protection in 3255 * the map, but will return the error 3256 * condition regardless. 3257 * 3258 * XXX Ignore what the actual error is, 3259 * XXX just call it a resource shortage 3260 * XXX so that it doesn't get confused 3261 * XXX what uvm_map_protect() itself would 3262 * XXX normally return. 3263 */ 3264 3265 error = ENOMEM; 3266 } 3267 } 3268 current = current->next; 3269 } 3270 pmap_update(map->pmap); 3271 3272 out: 3273 vm_map_unlock(map); 3274 3275 UVMHIST_LOG(maphist, "<- done, error=%d",error,0,0,0); 3276 return error; 3277 } 3278 3279 #undef MASK 3280 3281 /* 3282 * uvm_map_inherit: set inheritance code for range of addrs in map. 3283 * 3284 * => map must be unlocked 3285 * => note that the inherit code is used during a "fork". see fork 3286 * code for details. 3287 */ 3288 3289 int 3290 uvm_map_inherit(struct vm_map *map, vaddr_t start, vaddr_t end, 3291 vm_inherit_t new_inheritance) 3292 { 3293 struct vm_map_entry *entry, *temp_entry; 3294 UVMHIST_FUNC("uvm_map_inherit"); UVMHIST_CALLED(maphist); 3295 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_inh=0x%x)", 3296 map, start, end, new_inheritance); 3297 3298 switch (new_inheritance) { 3299 case MAP_INHERIT_NONE: 3300 case MAP_INHERIT_COPY: 3301 case MAP_INHERIT_SHARE: 3302 break; 3303 default: 3304 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3305 return EINVAL; 3306 } 3307 3308 vm_map_lock(map); 3309 VM_MAP_RANGE_CHECK(map, start, end); 3310 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3311 entry = temp_entry; 3312 UVM_MAP_CLIP_START(map, entry, start, NULL); 3313 } else { 3314 entry = temp_entry->next; 3315 } 3316 while ((entry != &map->header) && (entry->start < end)) { 3317 UVM_MAP_CLIP_END(map, entry, end, NULL); 3318 entry->inheritance = new_inheritance; 3319 entry = entry->next; 3320 } 3321 vm_map_unlock(map); 3322 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3323 return 0; 3324 } 3325 3326 /* 3327 * uvm_map_advice: set advice code for range of addrs in map. 3328 * 3329 * => map must be unlocked 3330 */ 3331 3332 int 3333 uvm_map_advice(struct vm_map *map, vaddr_t start, vaddr_t end, int new_advice) 3334 { 3335 struct vm_map_entry *entry, *temp_entry; 3336 UVMHIST_FUNC("uvm_map_advice"); UVMHIST_CALLED(maphist); 3337 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_adv=0x%x)", 3338 map, start, end, new_advice); 3339 3340 vm_map_lock(map); 3341 VM_MAP_RANGE_CHECK(map, start, end); 3342 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3343 entry = temp_entry; 3344 UVM_MAP_CLIP_START(map, entry, start, NULL); 3345 } else { 3346 entry = temp_entry->next; 3347 } 3348 3349 /* 3350 * XXXJRT: disallow holes? 3351 */ 3352 3353 while ((entry != &map->header) && (entry->start < end)) { 3354 UVM_MAP_CLIP_END(map, entry, end, NULL); 3355 3356 switch (new_advice) { 3357 case MADV_NORMAL: 3358 case MADV_RANDOM: 3359 case MADV_SEQUENTIAL: 3360 /* nothing special here */ 3361 break; 3362 3363 default: 3364 vm_map_unlock(map); 3365 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3366 return EINVAL; 3367 } 3368 entry->advice = new_advice; 3369 entry = entry->next; 3370 } 3371 3372 vm_map_unlock(map); 3373 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3374 return 0; 3375 } 3376 3377 /* 3378 * uvm_map_willneed: apply MADV_WILLNEED 3379 */ 3380 3381 int 3382 uvm_map_willneed(struct vm_map *map, vaddr_t start, vaddr_t end) 3383 { 3384 struct vm_map_entry *entry; 3385 UVMHIST_FUNC("uvm_map_willneed"); UVMHIST_CALLED(maphist); 3386 UVMHIST_LOG(maphist,"(map=0x%lx,start=0x%lx,end=0x%lx)", 3387 map, start, end, 0); 3388 3389 vm_map_lock_read(map); 3390 VM_MAP_RANGE_CHECK(map, start, end); 3391 if (!uvm_map_lookup_entry(map, start, &entry)) { 3392 entry = entry->next; 3393 } 3394 while (entry->start < end) { 3395 struct vm_amap * const amap = entry->aref.ar_amap; 3396 struct uvm_object * const uobj = entry->object.uvm_obj; 3397 3398 KASSERT(entry != &map->header); 3399 KASSERT(start < entry->end); 3400 /* 3401 * XXX IMPLEMENT ME. 3402 * Should invent a "weak" mode for uvm_fault() 3403 * which would only do the PGO_LOCKED pgo_get(). 3404 * 3405 * for now, we handle only the easy but common case. 3406 */ 3407 if (UVM_ET_ISOBJ(entry) && amap == NULL && uobj != NULL) { 3408 off_t offset; 3409 off_t size; 3410 3411 offset = entry->offset; 3412 if (start < entry->start) { 3413 offset += entry->start - start; 3414 } 3415 size = entry->offset + (entry->end - entry->start); 3416 if (entry->end < end) { 3417 size -= end - entry->end; 3418 } 3419 uvm_readahead(uobj, offset, size); 3420 } 3421 entry = entry->next; 3422 } 3423 vm_map_unlock_read(map); 3424 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3425 return 0; 3426 } 3427 3428 /* 3429 * uvm_map_pageable: sets the pageability of a range in a map. 3430 * 3431 * => wires map entries. should not be used for transient page locking. 3432 * for that, use uvm_fault_wire()/uvm_fault_unwire() (see uvm_vslock()). 3433 * => regions specified as not pageable require lock-down (wired) memory 3434 * and page tables. 3435 * => map must never be read-locked 3436 * => if islocked is true, map is already write-locked 3437 * => we always unlock the map, since we must downgrade to a read-lock 3438 * to call uvm_fault_wire() 3439 * => XXXCDC: check this and try and clean it up. 3440 */ 3441 3442 int 3443 uvm_map_pageable(struct vm_map *map, vaddr_t start, vaddr_t end, 3444 bool new_pageable, int lockflags) 3445 { 3446 struct vm_map_entry *entry, *start_entry, *failed_entry; 3447 int rv; 3448 #ifdef DIAGNOSTIC 3449 u_int timestamp_save; 3450 #endif 3451 UVMHIST_FUNC("uvm_map_pageable"); UVMHIST_CALLED(maphist); 3452 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_pageable=0x%x)", 3453 map, start, end, new_pageable); 3454 KASSERT(map->flags & VM_MAP_PAGEABLE); 3455 3456 if ((lockflags & UVM_LK_ENTER) == 0) 3457 vm_map_lock(map); 3458 VM_MAP_RANGE_CHECK(map, start, end); 3459 3460 /* 3461 * only one pageability change may take place at one time, since 3462 * uvm_fault_wire assumes it will be called only once for each 3463 * wiring/unwiring. therefore, we have to make sure we're actually 3464 * changing the pageability for the entire region. we do so before 3465 * making any changes. 3466 */ 3467 3468 if (uvm_map_lookup_entry(map, start, &start_entry) == false) { 3469 if ((lockflags & UVM_LK_EXIT) == 0) 3470 vm_map_unlock(map); 3471 3472 UVMHIST_LOG(maphist,"<- done (fault)",0,0,0,0); 3473 return EFAULT; 3474 } 3475 entry = start_entry; 3476 3477 /* 3478 * handle wiring and unwiring separately. 3479 */ 3480 3481 if (new_pageable) { /* unwire */ 3482 UVM_MAP_CLIP_START(map, entry, start, NULL); 3483 3484 /* 3485 * unwiring. first ensure that the range to be unwired is 3486 * really wired down and that there are no holes. 3487 */ 3488 3489 while ((entry != &map->header) && (entry->start < end)) { 3490 if (entry->wired_count == 0 || 3491 (entry->end < end && 3492 (entry->next == &map->header || 3493 entry->next->start > entry->end))) { 3494 if ((lockflags & UVM_LK_EXIT) == 0) 3495 vm_map_unlock(map); 3496 UVMHIST_LOG(maphist, "<- done (INVAL)",0,0,0,0); 3497 return EINVAL; 3498 } 3499 entry = entry->next; 3500 } 3501 3502 /* 3503 * POSIX 1003.1b - a single munlock call unlocks a region, 3504 * regardless of the number of mlock calls made on that 3505 * region. 3506 */ 3507 3508 entry = start_entry; 3509 while ((entry != &map->header) && (entry->start < end)) { 3510 UVM_MAP_CLIP_END(map, entry, end, NULL); 3511 if (VM_MAPENT_ISWIRED(entry)) 3512 uvm_map_entry_unwire(map, entry); 3513 entry = entry->next; 3514 } 3515 if ((lockflags & UVM_LK_EXIT) == 0) 3516 vm_map_unlock(map); 3517 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3518 return 0; 3519 } 3520 3521 /* 3522 * wire case: in two passes [XXXCDC: ugly block of code here] 3523 * 3524 * 1: holding the write lock, we create any anonymous maps that need 3525 * to be created. then we clip each map entry to the region to 3526 * be wired and increment its wiring count. 3527 * 3528 * 2: we downgrade to a read lock, and call uvm_fault_wire to fault 3529 * in the pages for any newly wired area (wired_count == 1). 3530 * 3531 * downgrading to a read lock for uvm_fault_wire avoids a possible 3532 * deadlock with another thread that may have faulted on one of 3533 * the pages to be wired (it would mark the page busy, blocking 3534 * us, then in turn block on the map lock that we hold). because 3535 * of problems in the recursive lock package, we cannot upgrade 3536 * to a write lock in vm_map_lookup. thus, any actions that 3537 * require the write lock must be done beforehand. because we 3538 * keep the read lock on the map, the copy-on-write status of the 3539 * entries we modify here cannot change. 3540 */ 3541 3542 while ((entry != &map->header) && (entry->start < end)) { 3543 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3544 3545 /* 3546 * perform actions of vm_map_lookup that need the 3547 * write lock on the map: create an anonymous map 3548 * for a copy-on-write region, or an anonymous map 3549 * for a zero-fill region. (XXXCDC: submap case 3550 * ok?) 3551 */ 3552 3553 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3554 if (UVM_ET_ISNEEDSCOPY(entry) && 3555 ((entry->max_protection & VM_PROT_WRITE) || 3556 (entry->object.uvm_obj == NULL))) { 3557 amap_copy(map, entry, 0, start, end); 3558 /* XXXCDC: wait OK? */ 3559 } 3560 } 3561 } 3562 UVM_MAP_CLIP_START(map, entry, start, NULL); 3563 UVM_MAP_CLIP_END(map, entry, end, NULL); 3564 entry->wired_count++; 3565 3566 /* 3567 * Check for holes 3568 */ 3569 3570 if (entry->protection == VM_PROT_NONE || 3571 (entry->end < end && 3572 (entry->next == &map->header || 3573 entry->next->start > entry->end))) { 3574 3575 /* 3576 * found one. amap creation actions do not need to 3577 * be undone, but the wired counts need to be restored. 3578 */ 3579 3580 while (entry != &map->header && entry->end > start) { 3581 entry->wired_count--; 3582 entry = entry->prev; 3583 } 3584 if ((lockflags & UVM_LK_EXIT) == 0) 3585 vm_map_unlock(map); 3586 UVMHIST_LOG(maphist,"<- done (INVALID WIRE)",0,0,0,0); 3587 return EINVAL; 3588 } 3589 entry = entry->next; 3590 } 3591 3592 /* 3593 * Pass 2. 3594 */ 3595 3596 #ifdef DIAGNOSTIC 3597 timestamp_save = map->timestamp; 3598 #endif 3599 vm_map_busy(map); 3600 vm_map_unlock(map); 3601 3602 rv = 0; 3603 entry = start_entry; 3604 while (entry != &map->header && entry->start < end) { 3605 if (entry->wired_count == 1) { 3606 rv = uvm_fault_wire(map, entry->start, entry->end, 3607 entry->max_protection, 1); 3608 if (rv) { 3609 3610 /* 3611 * wiring failed. break out of the loop. 3612 * we'll clean up the map below, once we 3613 * have a write lock again. 3614 */ 3615 3616 break; 3617 } 3618 } 3619 entry = entry->next; 3620 } 3621 3622 if (rv) { /* failed? */ 3623 3624 /* 3625 * Get back to an exclusive (write) lock. 3626 */ 3627 3628 vm_map_lock(map); 3629 vm_map_unbusy(map); 3630 3631 #ifdef DIAGNOSTIC 3632 if (timestamp_save + 1 != map->timestamp) 3633 panic("uvm_map_pageable: stale map"); 3634 #endif 3635 3636 /* 3637 * first drop the wiring count on all the entries 3638 * which haven't actually been wired yet. 3639 */ 3640 3641 failed_entry = entry; 3642 while (entry != &map->header && entry->start < end) { 3643 entry->wired_count--; 3644 entry = entry->next; 3645 } 3646 3647 /* 3648 * now, unwire all the entries that were successfully 3649 * wired above. 3650 */ 3651 3652 entry = start_entry; 3653 while (entry != failed_entry) { 3654 entry->wired_count--; 3655 if (VM_MAPENT_ISWIRED(entry) == 0) 3656 uvm_map_entry_unwire(map, entry); 3657 entry = entry->next; 3658 } 3659 if ((lockflags & UVM_LK_EXIT) == 0) 3660 vm_map_unlock(map); 3661 UVMHIST_LOG(maphist, "<- done (RV=%d)", rv,0,0,0); 3662 return (rv); 3663 } 3664 3665 if ((lockflags & UVM_LK_EXIT) == 0) { 3666 vm_map_unbusy(map); 3667 } else { 3668 3669 /* 3670 * Get back to an exclusive (write) lock. 3671 */ 3672 3673 vm_map_lock(map); 3674 vm_map_unbusy(map); 3675 } 3676 3677 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3678 return 0; 3679 } 3680 3681 /* 3682 * uvm_map_pageable_all: special case of uvm_map_pageable - affects 3683 * all mapped regions. 3684 * 3685 * => map must not be locked. 3686 * => if no flags are specified, all regions are unwired. 3687 * => XXXJRT: has some of the same problems as uvm_map_pageable() above. 3688 */ 3689 3690 int 3691 uvm_map_pageable_all(struct vm_map *map, int flags, vsize_t limit) 3692 { 3693 struct vm_map_entry *entry, *failed_entry; 3694 vsize_t size; 3695 int rv; 3696 #ifdef DIAGNOSTIC 3697 u_int timestamp_save; 3698 #endif 3699 UVMHIST_FUNC("uvm_map_pageable_all"); UVMHIST_CALLED(maphist); 3700 UVMHIST_LOG(maphist,"(map=0x%x,flags=0x%x)", map, flags, 0, 0); 3701 3702 KASSERT(map->flags & VM_MAP_PAGEABLE); 3703 3704 vm_map_lock(map); 3705 3706 /* 3707 * handle wiring and unwiring separately. 3708 */ 3709 3710 if (flags == 0) { /* unwire */ 3711 3712 /* 3713 * POSIX 1003.1b -- munlockall unlocks all regions, 3714 * regardless of how many times mlockall has been called. 3715 */ 3716 3717 for (entry = map->header.next; entry != &map->header; 3718 entry = entry->next) { 3719 if (VM_MAPENT_ISWIRED(entry)) 3720 uvm_map_entry_unwire(map, entry); 3721 } 3722 map->flags &= ~VM_MAP_WIREFUTURE; 3723 vm_map_unlock(map); 3724 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3725 return 0; 3726 } 3727 3728 if (flags & MCL_FUTURE) { 3729 3730 /* 3731 * must wire all future mappings; remember this. 3732 */ 3733 3734 map->flags |= VM_MAP_WIREFUTURE; 3735 } 3736 3737 if ((flags & MCL_CURRENT) == 0) { 3738 3739 /* 3740 * no more work to do! 3741 */ 3742 3743 UVMHIST_LOG(maphist,"<- done (OK no wire)",0,0,0,0); 3744 vm_map_unlock(map); 3745 return 0; 3746 } 3747 3748 /* 3749 * wire case: in three passes [XXXCDC: ugly block of code here] 3750 * 3751 * 1: holding the write lock, count all pages mapped by non-wired 3752 * entries. if this would cause us to go over our limit, we fail. 3753 * 3754 * 2: still holding the write lock, we create any anonymous maps that 3755 * need to be created. then we increment its wiring count. 3756 * 3757 * 3: we downgrade to a read lock, and call uvm_fault_wire to fault 3758 * in the pages for any newly wired area (wired_count == 1). 3759 * 3760 * downgrading to a read lock for uvm_fault_wire avoids a possible 3761 * deadlock with another thread that may have faulted on one of 3762 * the pages to be wired (it would mark the page busy, blocking 3763 * us, then in turn block on the map lock that we hold). because 3764 * of problems in the recursive lock package, we cannot upgrade 3765 * to a write lock in vm_map_lookup. thus, any actions that 3766 * require the write lock must be done beforehand. because we 3767 * keep the read lock on the map, the copy-on-write status of the 3768 * entries we modify here cannot change. 3769 */ 3770 3771 for (size = 0, entry = map->header.next; entry != &map->header; 3772 entry = entry->next) { 3773 if (entry->protection != VM_PROT_NONE && 3774 VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3775 size += entry->end - entry->start; 3776 } 3777 } 3778 3779 if (atop(size) + uvmexp.wired > uvmexp.wiredmax) { 3780 vm_map_unlock(map); 3781 return ENOMEM; 3782 } 3783 3784 if (limit != 0 && 3785 (size + ptoa(pmap_wired_count(vm_map_pmap(map))) > limit)) { 3786 vm_map_unlock(map); 3787 return ENOMEM; 3788 } 3789 3790 /* 3791 * Pass 2. 3792 */ 3793 3794 for (entry = map->header.next; entry != &map->header; 3795 entry = entry->next) { 3796 if (entry->protection == VM_PROT_NONE) 3797 continue; 3798 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3799 3800 /* 3801 * perform actions of vm_map_lookup that need the 3802 * write lock on the map: create an anonymous map 3803 * for a copy-on-write region, or an anonymous map 3804 * for a zero-fill region. (XXXCDC: submap case 3805 * ok?) 3806 */ 3807 3808 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3809 if (UVM_ET_ISNEEDSCOPY(entry) && 3810 ((entry->max_protection & VM_PROT_WRITE) || 3811 (entry->object.uvm_obj == NULL))) { 3812 amap_copy(map, entry, 0, entry->start, 3813 entry->end); 3814 /* XXXCDC: wait OK? */ 3815 } 3816 } 3817 } 3818 entry->wired_count++; 3819 } 3820 3821 /* 3822 * Pass 3. 3823 */ 3824 3825 #ifdef DIAGNOSTIC 3826 timestamp_save = map->timestamp; 3827 #endif 3828 vm_map_busy(map); 3829 vm_map_unlock(map); 3830 3831 rv = 0; 3832 for (entry = map->header.next; entry != &map->header; 3833 entry = entry->next) { 3834 if (entry->wired_count == 1) { 3835 rv = uvm_fault_wire(map, entry->start, entry->end, 3836 entry->max_protection, 1); 3837 if (rv) { 3838 3839 /* 3840 * wiring failed. break out of the loop. 3841 * we'll clean up the map below, once we 3842 * have a write lock again. 3843 */ 3844 3845 break; 3846 } 3847 } 3848 } 3849 3850 if (rv) { 3851 3852 /* 3853 * Get back an exclusive (write) lock. 3854 */ 3855 3856 vm_map_lock(map); 3857 vm_map_unbusy(map); 3858 3859 #ifdef DIAGNOSTIC 3860 if (timestamp_save + 1 != map->timestamp) 3861 panic("uvm_map_pageable_all: stale map"); 3862 #endif 3863 3864 /* 3865 * first drop the wiring count on all the entries 3866 * which haven't actually been wired yet. 3867 * 3868 * Skip VM_PROT_NONE entries like we did above. 3869 */ 3870 3871 failed_entry = entry; 3872 for (/* nothing */; entry != &map->header; 3873 entry = entry->next) { 3874 if (entry->protection == VM_PROT_NONE) 3875 continue; 3876 entry->wired_count--; 3877 } 3878 3879 /* 3880 * now, unwire all the entries that were successfully 3881 * wired above. 3882 * 3883 * Skip VM_PROT_NONE entries like we did above. 3884 */ 3885 3886 for (entry = map->header.next; entry != failed_entry; 3887 entry = entry->next) { 3888 if (entry->protection == VM_PROT_NONE) 3889 continue; 3890 entry->wired_count--; 3891 if (VM_MAPENT_ISWIRED(entry)) 3892 uvm_map_entry_unwire(map, entry); 3893 } 3894 vm_map_unlock(map); 3895 UVMHIST_LOG(maphist,"<- done (RV=%d)", rv,0,0,0); 3896 return (rv); 3897 } 3898 3899 vm_map_unbusy(map); 3900 3901 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3902 return 0; 3903 } 3904 3905 /* 3906 * uvm_map_clean: clean out a map range 3907 * 3908 * => valid flags: 3909 * if (flags & PGO_CLEANIT): dirty pages are cleaned first 3910 * if (flags & PGO_SYNCIO): dirty pages are written synchronously 3911 * if (flags & PGO_DEACTIVATE): any cached pages are deactivated after clean 3912 * if (flags & PGO_FREE): any cached pages are freed after clean 3913 * => returns an error if any part of the specified range isn't mapped 3914 * => never a need to flush amap layer since the anonymous memory has 3915 * no permanent home, but may deactivate pages there 3916 * => called from sys_msync() and sys_madvise() 3917 * => caller must not write-lock map (read OK). 3918 * => we may sleep while cleaning if SYNCIO [with map read-locked] 3919 */ 3920 3921 int 3922 uvm_map_clean(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 3923 { 3924 struct vm_map_entry *current, *entry; 3925 struct uvm_object *uobj; 3926 struct vm_amap *amap; 3927 struct vm_anon *anon; 3928 struct vm_page *pg; 3929 vaddr_t offset; 3930 vsize_t size; 3931 voff_t uoff; 3932 int error, refs; 3933 UVMHIST_FUNC("uvm_map_clean"); UVMHIST_CALLED(maphist); 3934 3935 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,flags=0x%x)", 3936 map, start, end, flags); 3937 KASSERT((flags & (PGO_FREE|PGO_DEACTIVATE)) != 3938 (PGO_FREE|PGO_DEACTIVATE)); 3939 3940 vm_map_lock_read(map); 3941 VM_MAP_RANGE_CHECK(map, start, end); 3942 if (uvm_map_lookup_entry(map, start, &entry) == false) { 3943 vm_map_unlock_read(map); 3944 return EFAULT; 3945 } 3946 3947 /* 3948 * Make a first pass to check for holes and wiring problems. 3949 */ 3950 3951 for (current = entry; current->start < end; current = current->next) { 3952 if (UVM_ET_ISSUBMAP(current)) { 3953 vm_map_unlock_read(map); 3954 return EINVAL; 3955 } 3956 if ((flags & PGO_FREE) != 0 && VM_MAPENT_ISWIRED(entry)) { 3957 vm_map_unlock_read(map); 3958 return EBUSY; 3959 } 3960 if (end <= current->end) { 3961 break; 3962 } 3963 if (current->end != current->next->start) { 3964 vm_map_unlock_read(map); 3965 return EFAULT; 3966 } 3967 } 3968 3969 error = 0; 3970 for (current = entry; start < end; current = current->next) { 3971 amap = current->aref.ar_amap; /* upper layer */ 3972 uobj = current->object.uvm_obj; /* lower layer */ 3973 KASSERT(start >= current->start); 3974 3975 /* 3976 * No amap cleaning necessary if: 3977 * 3978 * (1) There's no amap. 3979 * 3980 * (2) We're not deactivating or freeing pages. 3981 */ 3982 3983 if (amap == NULL || (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) 3984 goto flush_object; 3985 3986 amap_lock(amap); 3987 offset = start - current->start; 3988 size = MIN(end, current->end) - start; 3989 for ( ; size != 0; size -= PAGE_SIZE, offset += PAGE_SIZE) { 3990 anon = amap_lookup(¤t->aref, offset); 3991 if (anon == NULL) 3992 continue; 3993 3994 mutex_enter(&anon->an_lock); 3995 pg = anon->an_page; 3996 if (pg == NULL) { 3997 mutex_exit(&anon->an_lock); 3998 continue; 3999 } 4000 4001 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 4002 4003 /* 4004 * In these first 3 cases, we just deactivate the page. 4005 */ 4006 4007 case PGO_CLEANIT|PGO_FREE: 4008 case PGO_CLEANIT|PGO_DEACTIVATE: 4009 case PGO_DEACTIVATE: 4010 deactivate_it: 4011 /* 4012 * skip the page if it's loaned or wired, 4013 * since it shouldn't be on a paging queue 4014 * at all in these cases. 4015 */ 4016 4017 mutex_enter(&uvm_pageqlock); 4018 if (pg->loan_count != 0 || 4019 pg->wire_count != 0) { 4020 mutex_exit(&uvm_pageqlock); 4021 mutex_exit(&anon->an_lock); 4022 continue; 4023 } 4024 KASSERT(pg->uanon == anon); 4025 uvm_pagedeactivate(pg); 4026 mutex_exit(&uvm_pageqlock); 4027 mutex_exit(&anon->an_lock); 4028 continue; 4029 4030 case PGO_FREE: 4031 4032 /* 4033 * If there are multiple references to 4034 * the amap, just deactivate the page. 4035 */ 4036 4037 if (amap_refs(amap) > 1) 4038 goto deactivate_it; 4039 4040 /* skip the page if it's wired */ 4041 if (pg->wire_count != 0) { 4042 mutex_exit(&anon->an_lock); 4043 continue; 4044 } 4045 amap_unadd(¤t->aref, offset); 4046 refs = --anon->an_ref; 4047 mutex_exit(&anon->an_lock); 4048 if (refs == 0) 4049 uvm_anfree(anon); 4050 continue; 4051 } 4052 } 4053 amap_unlock(amap); 4054 4055 flush_object: 4056 /* 4057 * flush pages if we've got a valid backing object. 4058 * note that we must always clean object pages before 4059 * freeing them since otherwise we could reveal stale 4060 * data from files. 4061 */ 4062 4063 uoff = current->offset + (start - current->start); 4064 size = MIN(end, current->end) - start; 4065 if (uobj != NULL) { 4066 mutex_enter(&uobj->vmobjlock); 4067 if (uobj->pgops->pgo_put != NULL) 4068 error = (uobj->pgops->pgo_put)(uobj, uoff, 4069 uoff + size, flags | PGO_CLEANIT); 4070 else 4071 error = 0; 4072 } 4073 start += size; 4074 } 4075 vm_map_unlock_read(map); 4076 return (error); 4077 } 4078 4079 4080 /* 4081 * uvm_map_checkprot: check protection in map 4082 * 4083 * => must allow specified protection in a fully allocated region. 4084 * => map must be read or write locked by caller. 4085 */ 4086 4087 bool 4088 uvm_map_checkprot(struct vm_map *map, vaddr_t start, vaddr_t end, 4089 vm_prot_t protection) 4090 { 4091 struct vm_map_entry *entry; 4092 struct vm_map_entry *tmp_entry; 4093 4094 if (!uvm_map_lookup_entry(map, start, &tmp_entry)) { 4095 return (false); 4096 } 4097 entry = tmp_entry; 4098 while (start < end) { 4099 if (entry == &map->header) { 4100 return (false); 4101 } 4102 4103 /* 4104 * no holes allowed 4105 */ 4106 4107 if (start < entry->start) { 4108 return (false); 4109 } 4110 4111 /* 4112 * check protection associated with entry 4113 */ 4114 4115 if ((entry->protection & protection) != protection) { 4116 return (false); 4117 } 4118 start = entry->end; 4119 entry = entry->next; 4120 } 4121 return (true); 4122 } 4123 4124 /* 4125 * uvmspace_alloc: allocate a vmspace structure. 4126 * 4127 * - structure includes vm_map and pmap 4128 * - XXX: no locking on this structure 4129 * - refcnt set to 1, rest must be init'd by caller 4130 */ 4131 struct vmspace * 4132 uvmspace_alloc(vaddr_t vmin, vaddr_t vmax) 4133 { 4134 struct vmspace *vm; 4135 UVMHIST_FUNC("uvmspace_alloc"); UVMHIST_CALLED(maphist); 4136 4137 vm = pool_cache_get(&uvm_vmspace_cache, PR_WAITOK); 4138 uvmspace_init(vm, NULL, vmin, vmax); 4139 UVMHIST_LOG(maphist,"<- done (vm=0x%x)", vm,0,0,0); 4140 return (vm); 4141 } 4142 4143 /* 4144 * uvmspace_init: initialize a vmspace structure. 4145 * 4146 * - XXX: no locking on this structure 4147 * - refcnt set to 1, rest must be init'd by caller 4148 */ 4149 void 4150 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t vmin, vaddr_t vmax) 4151 { 4152 UVMHIST_FUNC("uvmspace_init"); UVMHIST_CALLED(maphist); 4153 4154 memset(vm, 0, sizeof(*vm)); 4155 uvm_map_setup(&vm->vm_map, vmin, vmax, VM_MAP_PAGEABLE 4156 #ifdef __USING_TOPDOWN_VM 4157 | VM_MAP_TOPDOWN 4158 #endif 4159 ); 4160 if (pmap) 4161 pmap_reference(pmap); 4162 else 4163 pmap = pmap_create(); 4164 vm->vm_map.pmap = pmap; 4165 vm->vm_refcnt = 1; 4166 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4167 } 4168 4169 /* 4170 * uvmspace_share: share a vmspace between two processes 4171 * 4172 * - used for vfork, threads(?) 4173 */ 4174 4175 void 4176 uvmspace_share(struct proc *p1, struct proc *p2) 4177 { 4178 4179 uvmspace_addref(p1->p_vmspace); 4180 p2->p_vmspace = p1->p_vmspace; 4181 } 4182 4183 #if 0 4184 4185 /* 4186 * uvmspace_unshare: ensure that process "p" has its own, unshared, vmspace 4187 * 4188 * - XXX: no locking on vmspace 4189 */ 4190 4191 void 4192 uvmspace_unshare(struct lwp *l) 4193 { 4194 struct proc *p = l->l_proc; 4195 struct vmspace *nvm, *ovm = p->p_vmspace; 4196 4197 if (ovm->vm_refcnt == 1) 4198 /* nothing to do: vmspace isn't shared in the first place */ 4199 return; 4200 4201 /* make a new vmspace, still holding old one */ 4202 nvm = uvmspace_fork(ovm); 4203 4204 kpreempt_disable(); 4205 pmap_deactivate(l); /* unbind old vmspace */ 4206 p->p_vmspace = nvm; 4207 pmap_activate(l); /* switch to new vmspace */ 4208 kpreempt_enable(); 4209 4210 uvmspace_free(ovm); /* drop reference to old vmspace */ 4211 } 4212 4213 #endif 4214 4215 /* 4216 * uvmspace_exec: the process wants to exec a new program 4217 */ 4218 4219 void 4220 uvmspace_exec(struct lwp *l, vaddr_t start, vaddr_t end) 4221 { 4222 struct proc *p = l->l_proc; 4223 struct vmspace *nvm, *ovm = p->p_vmspace; 4224 struct vm_map *map = &ovm->vm_map; 4225 4226 #ifdef __sparc__ 4227 /* XXX cgd 960926: the sparc #ifdef should be a MD hook */ 4228 kill_user_windows(l); /* before stack addresses go away */ 4229 #endif 4230 #ifdef __HAVE_CPU_VMSPACE_EXEC 4231 cpu_vmspace_exec(l, start, end); 4232 #endif 4233 4234 /* 4235 * see if more than one process is using this vmspace... 4236 */ 4237 4238 if (ovm->vm_refcnt == 1) { 4239 4240 /* 4241 * if p is the only process using its vmspace then we can safely 4242 * recycle that vmspace for the program that is being exec'd. 4243 */ 4244 4245 #ifdef SYSVSHM 4246 /* 4247 * SYSV SHM semantics require us to kill all segments on an exec 4248 */ 4249 4250 if (ovm->vm_shm) 4251 shmexit(ovm); 4252 #endif 4253 4254 /* 4255 * POSIX 1003.1b -- "lock future mappings" is revoked 4256 * when a process execs another program image. 4257 */ 4258 4259 map->flags &= ~VM_MAP_WIREFUTURE; 4260 4261 /* 4262 * now unmap the old program 4263 */ 4264 4265 pmap_remove_all(map->pmap); 4266 uvm_unmap(map, vm_map_min(map), vm_map_max(map)); 4267 KASSERT(map->header.prev == &map->header); 4268 KASSERT(map->nentries == 0); 4269 4270 /* 4271 * resize the map 4272 */ 4273 4274 vm_map_setmin(map, start); 4275 vm_map_setmax(map, end); 4276 } else { 4277 4278 /* 4279 * p's vmspace is being shared, so we can't reuse it for p since 4280 * it is still being used for others. allocate a new vmspace 4281 * for p 4282 */ 4283 4284 nvm = uvmspace_alloc(start, end); 4285 4286 /* 4287 * install new vmspace and drop our ref to the old one. 4288 */ 4289 4290 kpreempt_disable(); 4291 pmap_deactivate(l); 4292 p->p_vmspace = nvm; 4293 pmap_activate(l); 4294 kpreempt_enable(); 4295 4296 uvmspace_free(ovm); 4297 } 4298 } 4299 4300 /* 4301 * uvmspace_addref: add a referece to a vmspace. 4302 */ 4303 4304 void 4305 uvmspace_addref(struct vmspace *vm) 4306 { 4307 struct vm_map *map = &vm->vm_map; 4308 4309 KASSERT((map->flags & VM_MAP_DYING) == 0); 4310 4311 mutex_enter(&map->misc_lock); 4312 KASSERT(vm->vm_refcnt > 0); 4313 vm->vm_refcnt++; 4314 mutex_exit(&map->misc_lock); 4315 } 4316 4317 /* 4318 * uvmspace_free: free a vmspace data structure 4319 */ 4320 4321 void 4322 uvmspace_free(struct vmspace *vm) 4323 { 4324 struct vm_map_entry *dead_entries; 4325 struct vm_map *map = &vm->vm_map; 4326 int n; 4327 4328 UVMHIST_FUNC("uvmspace_free"); UVMHIST_CALLED(maphist); 4329 4330 UVMHIST_LOG(maphist,"(vm=0x%x) ref=%d", vm, vm->vm_refcnt,0,0); 4331 mutex_enter(&map->misc_lock); 4332 n = --vm->vm_refcnt; 4333 mutex_exit(&map->misc_lock); 4334 if (n > 0) 4335 return; 4336 4337 /* 4338 * at this point, there should be no other references to the map. 4339 * delete all of the mappings, then destroy the pmap. 4340 */ 4341 4342 map->flags |= VM_MAP_DYING; 4343 pmap_remove_all(map->pmap); 4344 #ifdef SYSVSHM 4345 /* Get rid of any SYSV shared memory segments. */ 4346 if (vm->vm_shm != NULL) 4347 shmexit(vm); 4348 #endif 4349 if (map->nentries) { 4350 uvm_unmap_remove(map, vm_map_min(map), vm_map_max(map), 4351 &dead_entries, NULL, 0); 4352 if (dead_entries != NULL) 4353 uvm_unmap_detach(dead_entries, 0); 4354 } 4355 KASSERT(map->nentries == 0); 4356 KASSERT(map->size == 0); 4357 mutex_destroy(&map->misc_lock); 4358 mutex_destroy(&map->mutex); 4359 rw_destroy(&map->lock); 4360 cv_destroy(&map->cv); 4361 pmap_destroy(map->pmap); 4362 pool_cache_put(&uvm_vmspace_cache, vm); 4363 } 4364 4365 /* 4366 * F O R K - m a i n e n t r y p o i n t 4367 */ 4368 /* 4369 * uvmspace_fork: fork a process' main map 4370 * 4371 * => create a new vmspace for child process from parent. 4372 * => parent's map must not be locked. 4373 */ 4374 4375 struct vmspace * 4376 uvmspace_fork(struct vmspace *vm1) 4377 { 4378 struct vmspace *vm2; 4379 struct vm_map *old_map = &vm1->vm_map; 4380 struct vm_map *new_map; 4381 struct vm_map_entry *old_entry; 4382 struct vm_map_entry *new_entry; 4383 UVMHIST_FUNC("uvmspace_fork"); UVMHIST_CALLED(maphist); 4384 4385 vm_map_lock(old_map); 4386 4387 vm2 = uvmspace_alloc(vm_map_min(old_map), vm_map_max(old_map)); 4388 memcpy(&vm2->vm_startcopy, &vm1->vm_startcopy, 4389 (char *) (vm1 + 1) - (char *) &vm1->vm_startcopy); 4390 new_map = &vm2->vm_map; /* XXX */ 4391 4392 old_entry = old_map->header.next; 4393 new_map->size = old_map->size; 4394 4395 /* 4396 * go entry-by-entry 4397 */ 4398 4399 while (old_entry != &old_map->header) { 4400 4401 /* 4402 * first, some sanity checks on the old entry 4403 */ 4404 4405 KASSERT(!UVM_ET_ISSUBMAP(old_entry)); 4406 KASSERT(UVM_ET_ISCOPYONWRITE(old_entry) || 4407 !UVM_ET_ISNEEDSCOPY(old_entry)); 4408 4409 switch (old_entry->inheritance) { 4410 case MAP_INHERIT_NONE: 4411 4412 /* 4413 * drop the mapping, modify size 4414 */ 4415 new_map->size -= old_entry->end - old_entry->start; 4416 break; 4417 4418 case MAP_INHERIT_SHARE: 4419 4420 /* 4421 * share the mapping: this means we want the old and 4422 * new entries to share amaps and backing objects. 4423 */ 4424 /* 4425 * if the old_entry needs a new amap (due to prev fork) 4426 * then we need to allocate it now so that we have 4427 * something we own to share with the new_entry. [in 4428 * other words, we need to clear needs_copy] 4429 */ 4430 4431 if (UVM_ET_ISNEEDSCOPY(old_entry)) { 4432 /* get our own amap, clears needs_copy */ 4433 amap_copy(old_map, old_entry, AMAP_COPY_NOCHUNK, 4434 0, 0); 4435 /* XXXCDC: WAITOK??? */ 4436 } 4437 4438 new_entry = uvm_mapent_alloc(new_map, 0); 4439 /* old_entry -> new_entry */ 4440 uvm_mapent_copy(old_entry, new_entry); 4441 4442 /* new pmap has nothing wired in it */ 4443 new_entry->wired_count = 0; 4444 4445 /* 4446 * gain reference to object backing the map (can't 4447 * be a submap, already checked this case). 4448 */ 4449 4450 if (new_entry->aref.ar_amap) 4451 uvm_map_reference_amap(new_entry, AMAP_SHARED); 4452 4453 if (new_entry->object.uvm_obj && 4454 new_entry->object.uvm_obj->pgops->pgo_reference) 4455 new_entry->object.uvm_obj-> 4456 pgops->pgo_reference( 4457 new_entry->object.uvm_obj); 4458 4459 /* insert entry at end of new_map's entry list */ 4460 uvm_map_entry_link(new_map, new_map->header.prev, 4461 new_entry); 4462 4463 break; 4464 4465 case MAP_INHERIT_COPY: 4466 4467 /* 4468 * copy-on-write the mapping (using mmap's 4469 * MAP_PRIVATE semantics) 4470 * 4471 * allocate new_entry, adjust reference counts. 4472 * (note that new references are read-only). 4473 */ 4474 4475 new_entry = uvm_mapent_alloc(new_map, 0); 4476 /* old_entry -> new_entry */ 4477 uvm_mapent_copy(old_entry, new_entry); 4478 4479 if (new_entry->aref.ar_amap) 4480 uvm_map_reference_amap(new_entry, 0); 4481 4482 if (new_entry->object.uvm_obj && 4483 new_entry->object.uvm_obj->pgops->pgo_reference) 4484 new_entry->object.uvm_obj->pgops->pgo_reference 4485 (new_entry->object.uvm_obj); 4486 4487 /* new pmap has nothing wired in it */ 4488 new_entry->wired_count = 0; 4489 4490 new_entry->etype |= 4491 (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY); 4492 uvm_map_entry_link(new_map, new_map->header.prev, 4493 new_entry); 4494 4495 /* 4496 * the new entry will need an amap. it will either 4497 * need to be copied from the old entry or created 4498 * from scratch (if the old entry does not have an 4499 * amap). can we defer this process until later 4500 * (by setting "needs_copy") or do we need to copy 4501 * the amap now? 4502 * 4503 * we must copy the amap now if any of the following 4504 * conditions hold: 4505 * 1. the old entry has an amap and that amap is 4506 * being shared. this means that the old (parent) 4507 * process is sharing the amap with another 4508 * process. if we do not clear needs_copy here 4509 * we will end up in a situation where both the 4510 * parent and child process are refering to the 4511 * same amap with "needs_copy" set. if the 4512 * parent write-faults, the fault routine will 4513 * clear "needs_copy" in the parent by allocating 4514 * a new amap. this is wrong because the 4515 * parent is supposed to be sharing the old amap 4516 * and the new amap will break that. 4517 * 4518 * 2. if the old entry has an amap and a non-zero 4519 * wire count then we are going to have to call 4520 * amap_cow_now to avoid page faults in the 4521 * parent process. since amap_cow_now requires 4522 * "needs_copy" to be clear we might as well 4523 * clear it here as well. 4524 * 4525 */ 4526 4527 if (old_entry->aref.ar_amap != NULL) { 4528 if ((amap_flags(old_entry->aref.ar_amap) & 4529 AMAP_SHARED) != 0 || 4530 VM_MAPENT_ISWIRED(old_entry)) { 4531 4532 amap_copy(new_map, new_entry, 4533 AMAP_COPY_NOCHUNK, 0, 0); 4534 /* XXXCDC: M_WAITOK ... ok? */ 4535 } 4536 } 4537 4538 /* 4539 * if the parent's entry is wired down, then the 4540 * parent process does not want page faults on 4541 * access to that memory. this means that we 4542 * cannot do copy-on-write because we can't write 4543 * protect the old entry. in this case we 4544 * resolve all copy-on-write faults now, using 4545 * amap_cow_now. note that we have already 4546 * allocated any needed amap (above). 4547 */ 4548 4549 if (VM_MAPENT_ISWIRED(old_entry)) { 4550 4551 /* 4552 * resolve all copy-on-write faults now 4553 * (note that there is nothing to do if 4554 * the old mapping does not have an amap). 4555 */ 4556 if (old_entry->aref.ar_amap) 4557 amap_cow_now(new_map, new_entry); 4558 4559 } else { 4560 4561 /* 4562 * setup mappings to trigger copy-on-write faults 4563 * we must write-protect the parent if it has 4564 * an amap and it is not already "needs_copy"... 4565 * if it is already "needs_copy" then the parent 4566 * has already been write-protected by a previous 4567 * fork operation. 4568 */ 4569 4570 if (old_entry->aref.ar_amap && 4571 !UVM_ET_ISNEEDSCOPY(old_entry)) { 4572 if (old_entry->max_protection & VM_PROT_WRITE) { 4573 pmap_protect(old_map->pmap, 4574 old_entry->start, 4575 old_entry->end, 4576 old_entry->protection & 4577 ~VM_PROT_WRITE); 4578 } 4579 old_entry->etype |= UVM_ET_NEEDSCOPY; 4580 } 4581 } 4582 break; 4583 } /* end of switch statement */ 4584 old_entry = old_entry->next; 4585 } 4586 4587 pmap_update(old_map->pmap); 4588 vm_map_unlock(old_map); 4589 4590 #ifdef SYSVSHM 4591 if (vm1->vm_shm) 4592 shmfork(vm1, vm2); 4593 #endif 4594 4595 #ifdef PMAP_FORK 4596 pmap_fork(vm1->vm_map.pmap, vm2->vm_map.pmap); 4597 #endif 4598 4599 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4600 return (vm2); 4601 } 4602 4603 4604 /* 4605 * in-kernel map entry allocation. 4606 */ 4607 4608 struct uvm_kmapent_hdr { 4609 LIST_ENTRY(uvm_kmapent_hdr) ukh_listq; 4610 int ukh_nused; 4611 struct vm_map_entry *ukh_freelist; 4612 struct vm_map *ukh_map; 4613 struct vm_map_entry ukh_entries[0]; 4614 }; 4615 4616 #define UVM_KMAPENT_CHUNK \ 4617 ((PAGE_SIZE - sizeof(struct uvm_kmapent_hdr)) \ 4618 / sizeof(struct vm_map_entry)) 4619 4620 #define UVM_KHDR_FIND(entry) \ 4621 ((struct uvm_kmapent_hdr *)(((vaddr_t)entry) & ~PAGE_MASK)) 4622 4623 4624 #ifdef DIAGNOSTIC 4625 static struct vm_map * 4626 uvm_kmapent_map(struct vm_map_entry *entry) 4627 { 4628 const struct uvm_kmapent_hdr *ukh; 4629 4630 ukh = UVM_KHDR_FIND(entry); 4631 return ukh->ukh_map; 4632 } 4633 #endif 4634 4635 static inline struct vm_map_entry * 4636 uvm_kmapent_get(struct uvm_kmapent_hdr *ukh) 4637 { 4638 struct vm_map_entry *entry; 4639 4640 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4641 KASSERT(ukh->ukh_nused >= 0); 4642 4643 entry = ukh->ukh_freelist; 4644 if (entry) { 4645 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4646 == UVM_MAP_KERNEL); 4647 ukh->ukh_freelist = entry->next; 4648 ukh->ukh_nused++; 4649 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4650 } else { 4651 KASSERT(ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4652 } 4653 4654 return entry; 4655 } 4656 4657 static inline void 4658 uvm_kmapent_put(struct uvm_kmapent_hdr *ukh, struct vm_map_entry *entry) 4659 { 4660 4661 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4662 == UVM_MAP_KERNEL); 4663 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4664 KASSERT(ukh->ukh_nused > 0); 4665 KASSERT(ukh->ukh_freelist != NULL || 4666 ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4667 KASSERT(ukh->ukh_freelist == NULL || 4668 ukh->ukh_nused < UVM_KMAPENT_CHUNK); 4669 4670 ukh->ukh_nused--; 4671 entry->next = ukh->ukh_freelist; 4672 ukh->ukh_freelist = entry; 4673 } 4674 4675 /* 4676 * uvm_kmapent_alloc: allocate a map entry for in-kernel map 4677 */ 4678 4679 static struct vm_map_entry * 4680 uvm_kmapent_alloc(struct vm_map *map, int flags) 4681 { 4682 struct vm_page *pg; 4683 struct uvm_kmapent_hdr *ukh; 4684 struct vm_map_entry *entry; 4685 #ifndef PMAP_MAP_POOLPAGE 4686 struct uvm_map_args args; 4687 uvm_flag_t mapflags = UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, 4688 UVM_INH_NONE, UVM_ADV_RANDOM, flags | UVM_FLAG_NOMERGE); 4689 int error; 4690 #endif 4691 vaddr_t va; 4692 int i; 4693 4694 KDASSERT(UVM_KMAPENT_CHUNK > 2); 4695 KDASSERT(kernel_map != NULL); 4696 KASSERT(vm_map_pmap(map) == pmap_kernel()); 4697 4698 UVMMAP_EVCNT_INCR(uke_alloc); 4699 entry = NULL; 4700 again: 4701 /* 4702 * try to grab an entry from freelist. 4703 */ 4704 mutex_spin_enter(&uvm_kentry_lock); 4705 ukh = LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free); 4706 if (ukh) { 4707 entry = uvm_kmapent_get(ukh); 4708 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK) 4709 LIST_REMOVE(ukh, ukh_listq); 4710 } 4711 mutex_spin_exit(&uvm_kentry_lock); 4712 4713 if (entry) 4714 return entry; 4715 4716 /* 4717 * there's no free entry for this vm_map. 4718 * now we need to allocate some vm_map_entry. 4719 * for simplicity, always allocate one page chunk of them at once. 4720 */ 4721 4722 pg = uvm_pagealloc(NULL, 0, NULL, 4723 (flags & UVM_KMF_NOWAIT) != 0 ? UVM_PGA_USERESERVE : 0); 4724 if (__predict_false(pg == NULL)) { 4725 if (flags & UVM_FLAG_NOWAIT) 4726 return NULL; 4727 uvm_wait("kme_alloc"); 4728 goto again; 4729 } 4730 4731 #ifdef PMAP_MAP_POOLPAGE 4732 va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg)); 4733 KASSERT(va != 0); 4734 #else 4735 error = uvm_map_prepare(map, 0, PAGE_SIZE, NULL, UVM_UNKNOWN_OFFSET, 4736 0, mapflags, &args); 4737 if (error) { 4738 uvm_pagefree(pg); 4739 return NULL; 4740 } 4741 4742 va = args.uma_start; 4743 4744 pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), 4745 VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE); 4746 pmap_update(vm_map_pmap(map)); 4747 4748 #endif 4749 ukh = (void *)va; 4750 4751 /* 4752 * use the last entry for ukh itsself. 4753 */ 4754 4755 i = UVM_KMAPENT_CHUNK - 1; 4756 #ifndef PMAP_MAP_POOLPAGE 4757 entry = &ukh->ukh_entries[i--]; 4758 entry->flags = UVM_MAP_KERNEL | UVM_MAP_KMAPENT; 4759 error = uvm_map_enter(map, &args, entry); 4760 KASSERT(error == 0); 4761 #endif 4762 4763 ukh->ukh_nused = UVM_KMAPENT_CHUNK; 4764 ukh->ukh_map = map; 4765 ukh->ukh_freelist = NULL; 4766 for (; i >= 1; i--) { 4767 struct vm_map_entry *xentry = &ukh->ukh_entries[i]; 4768 4769 xentry->flags = UVM_MAP_KERNEL; 4770 uvm_kmapent_put(ukh, xentry); 4771 } 4772 #ifdef PMAP_MAP_POOLPAGE 4773 KASSERT(ukh->ukh_nused == 1); 4774 #else 4775 KASSERT(ukh->ukh_nused == 2); 4776 #endif 4777 4778 mutex_spin_enter(&uvm_kentry_lock); 4779 LIST_INSERT_HEAD(&vm_map_to_kernel(map)->vmk_kentry_free, 4780 ukh, ukh_listq); 4781 mutex_spin_exit(&uvm_kentry_lock); 4782 4783 /* 4784 * return first entry. 4785 */ 4786 4787 entry = &ukh->ukh_entries[0]; 4788 entry->flags = UVM_MAP_KERNEL; 4789 UVMMAP_EVCNT_INCR(ukh_alloc); 4790 4791 return entry; 4792 } 4793 4794 /* 4795 * uvm_mapent_free: free map entry for in-kernel map 4796 */ 4797 4798 static void 4799 uvm_kmapent_free(struct vm_map_entry *entry) 4800 { 4801 struct uvm_kmapent_hdr *ukh; 4802 struct vm_page *pg; 4803 struct vm_map *map; 4804 #ifndef PMAP_UNMAP_POOLPAGE 4805 struct pmap *pmap; 4806 struct vm_map_entry *deadentry; 4807 #endif 4808 vaddr_t va; 4809 paddr_t pa; 4810 4811 UVMMAP_EVCNT_INCR(uke_free); 4812 ukh = UVM_KHDR_FIND(entry); 4813 map = ukh->ukh_map; 4814 4815 mutex_spin_enter(&uvm_kentry_lock); 4816 uvm_kmapent_put(ukh, entry); 4817 #ifdef PMAP_UNMAP_POOLPAGE 4818 if (ukh->ukh_nused > 0) { 4819 #else 4820 if (ukh->ukh_nused > 1) { 4821 #endif 4822 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK - 1) 4823 LIST_INSERT_HEAD( 4824 &vm_map_to_kernel(map)->vmk_kentry_free, 4825 ukh, ukh_listq); 4826 mutex_spin_exit(&uvm_kentry_lock); 4827 return; 4828 } 4829 4830 /* 4831 * now we can free this ukh. 4832 * 4833 * however, keep an empty ukh to avoid ping-pong. 4834 */ 4835 4836 if (LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free) == ukh && 4837 LIST_NEXT(ukh, ukh_listq) == NULL) { 4838 mutex_spin_exit(&uvm_kentry_lock); 4839 return; 4840 } 4841 LIST_REMOVE(ukh, ukh_listq); 4842 mutex_spin_exit(&uvm_kentry_lock); 4843 4844 va = (vaddr_t)ukh; 4845 4846 #ifdef PMAP_UNMAP_POOLPAGE 4847 KASSERT(ukh->ukh_nused == 0); 4848 pa = PMAP_UNMAP_POOLPAGE(va); 4849 KASSERT(pa != 0); 4850 #else 4851 KASSERT(ukh->ukh_nused == 1); 4852 4853 /* 4854 * remove map entry for ukh itsself. 4855 */ 4856 4857 KASSERT((va & PAGE_MASK) == 0); 4858 vm_map_lock(map); 4859 uvm_unmap_remove(map, va, va + PAGE_SIZE, &deadentry, NULL, 0); 4860 KASSERT(deadentry->flags & UVM_MAP_KERNEL); 4861 KASSERT(deadentry->flags & UVM_MAP_KMAPENT); 4862 KASSERT(deadentry->next == NULL); 4863 KASSERT(deadentry == &ukh->ukh_entries[UVM_KMAPENT_CHUNK - 1]); 4864 4865 /* 4866 * unmap the page from pmap and free it. 4867 */ 4868 4869 pmap = vm_map_pmap(map); 4870 KASSERT(pmap == pmap_kernel()); 4871 if (!pmap_extract(pmap, va, &pa)) 4872 panic("%s: no mapping", __func__); 4873 pmap_kremove(va, PAGE_SIZE); 4874 pmap_update(vm_map_pmap(map)); 4875 vm_map_unlock(map); 4876 #endif /* !PMAP_UNMAP_POOLPAGE */ 4877 pg = PHYS_TO_VM_PAGE(pa); 4878 uvm_pagefree(pg); 4879 UVMMAP_EVCNT_INCR(ukh_free); 4880 } 4881 4882 static vsize_t 4883 uvm_kmapent_overhead(vsize_t size) 4884 { 4885 4886 /* 4887 * - the max number of unmerged entries is howmany(size, PAGE_SIZE) 4888 * as the min allocation unit is PAGE_SIZE. 4889 * - UVM_KMAPENT_CHUNK "kmapent"s are allocated from a page. 4890 * one of them are used to map the page itself. 4891 */ 4892 4893 return howmany(howmany(size, PAGE_SIZE), (UVM_KMAPENT_CHUNK - 1)) * 4894 PAGE_SIZE; 4895 } 4896 4897 /* 4898 * map entry reservation 4899 */ 4900 4901 /* 4902 * uvm_mapent_reserve: reserve map entries for clipping before locking map. 4903 * 4904 * => needed when unmapping entries allocated without UVM_FLAG_QUANTUM. 4905 * => caller shouldn't hold map locked. 4906 */ 4907 int 4908 uvm_mapent_reserve(struct vm_map *map, struct uvm_mapent_reservation *umr, 4909 int nentries, int flags) 4910 { 4911 4912 umr->umr_nentries = 0; 4913 4914 if ((flags & UVM_FLAG_QUANTUM) != 0) 4915 return 0; 4916 4917 if (!VM_MAP_USE_KMAPENT(map)) 4918 return 0; 4919 4920 while (nentries--) { 4921 struct vm_map_entry *ent; 4922 ent = uvm_kmapent_alloc(map, flags); 4923 if (!ent) { 4924 uvm_mapent_unreserve(map, umr); 4925 return ENOMEM; 4926 } 4927 UMR_PUTENTRY(umr, ent); 4928 } 4929 4930 return 0; 4931 } 4932 4933 /* 4934 * uvm_mapent_unreserve: 4935 * 4936 * => caller shouldn't hold map locked. 4937 * => never fail or sleep. 4938 */ 4939 void 4940 uvm_mapent_unreserve(struct vm_map *map, struct uvm_mapent_reservation *umr) 4941 { 4942 4943 while (!UMR_EMPTY(umr)) 4944 uvm_kmapent_free(UMR_GETENTRY(umr)); 4945 } 4946 4947 /* 4948 * uvm_mapent_trymerge: try to merge an entry with its neighbors. 4949 * 4950 * => called with map locked. 4951 * => return non zero if successfully merged. 4952 */ 4953 4954 int 4955 uvm_mapent_trymerge(struct vm_map *map, struct vm_map_entry *entry, int flags) 4956 { 4957 struct uvm_object *uobj; 4958 struct vm_map_entry *next; 4959 struct vm_map_entry *prev; 4960 vsize_t size; 4961 int merged = 0; 4962 bool copying; 4963 int newetype; 4964 4965 if (VM_MAP_USE_KMAPENT(map)) { 4966 return 0; 4967 } 4968 if (entry->aref.ar_amap != NULL) { 4969 return 0; 4970 } 4971 if ((entry->flags & UVM_MAP_NOMERGE) != 0) { 4972 return 0; 4973 } 4974 4975 uobj = entry->object.uvm_obj; 4976 size = entry->end - entry->start; 4977 copying = (flags & UVM_MERGE_COPYING) != 0; 4978 newetype = copying ? (entry->etype & ~UVM_ET_NEEDSCOPY) : entry->etype; 4979 4980 next = entry->next; 4981 if (next != &map->header && 4982 next->start == entry->end && 4983 ((copying && next->aref.ar_amap != NULL && 4984 amap_refs(next->aref.ar_amap) == 1) || 4985 (!copying && next->aref.ar_amap == NULL)) && 4986 UVM_ET_ISCOMPATIBLE(next, newetype, 4987 uobj, entry->flags, entry->protection, 4988 entry->max_protection, entry->inheritance, entry->advice, 4989 entry->wired_count) && 4990 (uobj == NULL || entry->offset + size == next->offset)) { 4991 int error; 4992 4993 if (copying) { 4994 error = amap_extend(next, size, 4995 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_BACKWARDS); 4996 } else { 4997 error = 0; 4998 } 4999 if (error == 0) { 5000 if (uobj) { 5001 if (uobj->pgops->pgo_detach) { 5002 uobj->pgops->pgo_detach(uobj); 5003 } 5004 } 5005 5006 entry->end = next->end; 5007 clear_hints(map, next); 5008 uvm_map_entry_unlink(map, next); 5009 if (copying) { 5010 entry->aref = next->aref; 5011 entry->etype &= ~UVM_ET_NEEDSCOPY; 5012 } 5013 uvm_map_check(map, "trymerge forwardmerge"); 5014 uvm_mapent_free_merged(map, next); 5015 merged++; 5016 } 5017 } 5018 5019 prev = entry->prev; 5020 if (prev != &map->header && 5021 prev->end == entry->start && 5022 ((copying && !merged && prev->aref.ar_amap != NULL && 5023 amap_refs(prev->aref.ar_amap) == 1) || 5024 (!copying && prev->aref.ar_amap == NULL)) && 5025 UVM_ET_ISCOMPATIBLE(prev, newetype, 5026 uobj, entry->flags, entry->protection, 5027 entry->max_protection, entry->inheritance, entry->advice, 5028 entry->wired_count) && 5029 (uobj == NULL || 5030 prev->offset + prev->end - prev->start == entry->offset)) { 5031 int error; 5032 5033 if (copying) { 5034 error = amap_extend(prev, size, 5035 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_FORWARDS); 5036 } else { 5037 error = 0; 5038 } 5039 if (error == 0) { 5040 if (uobj) { 5041 if (uobj->pgops->pgo_detach) { 5042 uobj->pgops->pgo_detach(uobj); 5043 } 5044 entry->offset = prev->offset; 5045 } 5046 5047 entry->start = prev->start; 5048 clear_hints(map, prev); 5049 uvm_map_entry_unlink(map, prev); 5050 if (copying) { 5051 entry->aref = prev->aref; 5052 entry->etype &= ~UVM_ET_NEEDSCOPY; 5053 } 5054 uvm_map_check(map, "trymerge backmerge"); 5055 uvm_mapent_free_merged(map, prev); 5056 merged++; 5057 } 5058 } 5059 5060 return merged; 5061 } 5062 5063 /* 5064 * uvm_map_create: create map 5065 */ 5066 5067 struct vm_map * 5068 uvm_map_create(pmap_t pmap, vaddr_t vmin, vaddr_t vmax, int flags) 5069 { 5070 struct vm_map *result; 5071 5072 result = malloc(sizeof(struct vm_map), M_VMMAP, M_WAITOK); 5073 uvm_map_setup(result, vmin, vmax, flags); 5074 result->pmap = pmap; 5075 return(result); 5076 } 5077 5078 /* 5079 * uvm_map_setup: init map 5080 * 5081 * => map must not be in service yet. 5082 */ 5083 5084 void 5085 uvm_map_setup(struct vm_map *map, vaddr_t vmin, vaddr_t vmax, int flags) 5086 { 5087 int ipl; 5088 5089 rb_tree_init(&map->rb_tree, &uvm_map_tree_ops); 5090 map->header.next = map->header.prev = &map->header; 5091 map->nentries = 0; 5092 map->size = 0; 5093 map->ref_count = 1; 5094 vm_map_setmin(map, vmin); 5095 vm_map_setmax(map, vmax); 5096 map->flags = flags; 5097 map->first_free = &map->header; 5098 map->hint = &map->header; 5099 map->timestamp = 0; 5100 map->busy = NULL; 5101 5102 if ((flags & VM_MAP_INTRSAFE) != 0) { 5103 ipl = IPL_VM; 5104 } else { 5105 ipl = IPL_NONE; 5106 } 5107 5108 rw_init(&map->lock); 5109 cv_init(&map->cv, "vm_map"); 5110 mutex_init(&map->misc_lock, MUTEX_DRIVER, ipl); 5111 mutex_init(&map->mutex, MUTEX_DRIVER, ipl); 5112 } 5113 5114 5115 /* 5116 * U N M A P - m a i n e n t r y p o i n t 5117 */ 5118 5119 /* 5120 * uvm_unmap1: remove mappings from a vm_map (from "start" up to "stop") 5121 * 5122 * => caller must check alignment and size 5123 * => map must be unlocked (we will lock it) 5124 * => flags is UVM_FLAG_QUANTUM or 0. 5125 */ 5126 5127 void 5128 uvm_unmap1(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 5129 { 5130 struct vm_map_entry *dead_entries; 5131 struct uvm_mapent_reservation umr; 5132 UVMHIST_FUNC("uvm_unmap"); UVMHIST_CALLED(maphist); 5133 5134 UVMHIST_LOG(maphist, " (map=0x%x, start=0x%x, end=0x%x)", 5135 map, start, end, 0); 5136 if (map == kernel_map) { 5137 LOCKDEBUG_MEM_CHECK((void *)start, end - start); 5138 } 5139 /* 5140 * work now done by helper functions. wipe the pmap's and then 5141 * detach from the dead entries... 5142 */ 5143 uvm_mapent_reserve(map, &umr, 2, flags); 5144 vm_map_lock(map); 5145 uvm_unmap_remove(map, start, end, &dead_entries, &umr, flags); 5146 vm_map_unlock(map); 5147 uvm_mapent_unreserve(map, &umr); 5148 5149 if (dead_entries != NULL) 5150 uvm_unmap_detach(dead_entries, 0); 5151 5152 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 5153 } 5154 5155 5156 /* 5157 * uvm_map_reference: add reference to a map 5158 * 5159 * => map need not be locked (we use misc_lock). 5160 */ 5161 5162 void 5163 uvm_map_reference(struct vm_map *map) 5164 { 5165 mutex_enter(&map->misc_lock); 5166 map->ref_count++; 5167 mutex_exit(&map->misc_lock); 5168 } 5169 5170 struct vm_map_kernel * 5171 vm_map_to_kernel(struct vm_map *map) 5172 { 5173 5174 KASSERT(VM_MAP_IS_KERNEL(map)); 5175 5176 return (struct vm_map_kernel *)map; 5177 } 5178 5179 bool 5180 vm_map_starved_p(struct vm_map *map) 5181 { 5182 5183 if ((map->flags & VM_MAP_WANTVA) != 0) { 5184 return true; 5185 } 5186 /* XXX */ 5187 if ((vm_map_max(map) - vm_map_min(map)) / 16 * 15 < map->size) { 5188 return true; 5189 } 5190 return false; 5191 } 5192 5193 #if defined(DDB) || defined(DEBUGPRINT) 5194 5195 /* 5196 * uvm_map_printit: actually prints the map 5197 */ 5198 5199 void 5200 uvm_map_printit(struct vm_map *map, bool full, 5201 void (*pr)(const char *, ...)) 5202 { 5203 struct vm_map_entry *entry; 5204 5205 (*pr)("MAP %p: [0x%lx->0x%lx]\n", map, vm_map_min(map), 5206 vm_map_max(map)); 5207 (*pr)("\t#ent=%d, sz=%d, ref=%d, version=%d, flags=0x%x\n", 5208 map->nentries, map->size, map->ref_count, map->timestamp, 5209 map->flags); 5210 (*pr)("\tpmap=%p(resident=%ld, wired=%ld)\n", map->pmap, 5211 pmap_resident_count(map->pmap), pmap_wired_count(map->pmap)); 5212 if (!full) 5213 return; 5214 for (entry = map->header.next; entry != &map->header; 5215 entry = entry->next) { 5216 (*pr)(" - %p: 0x%lx->0x%lx: obj=%p/0x%llx, amap=%p/%d\n", 5217 entry, entry->start, entry->end, entry->object.uvm_obj, 5218 (long long)entry->offset, entry->aref.ar_amap, 5219 entry->aref.ar_pageoff); 5220 (*pr)( 5221 "\tsubmap=%c, cow=%c, nc=%c, prot(max)=%d/%d, inh=%d, " 5222 "wc=%d, adv=%d\n", 5223 (entry->etype & UVM_ET_SUBMAP) ? 'T' : 'F', 5224 (entry->etype & UVM_ET_COPYONWRITE) ? 'T' : 'F', 5225 (entry->etype & UVM_ET_NEEDSCOPY) ? 'T' : 'F', 5226 entry->protection, entry->max_protection, 5227 entry->inheritance, entry->wired_count, entry->advice); 5228 } 5229 } 5230 5231 void 5232 uvm_whatis(uintptr_t addr, void (*pr)(const char *, ...)) 5233 { 5234 struct vm_map *map; 5235 5236 for (map = kernel_map;;) { 5237 struct vm_map_entry *entry; 5238 5239 if (!uvm_map_lookup_entry_bytree(map, (vaddr_t)addr, &entry)) { 5240 break; 5241 } 5242 (*pr)("%p is %p+%zu from VMMAP %p\n", 5243 (void *)addr, (void *)entry->start, 5244 (size_t)(addr - (uintptr_t)entry->start), map); 5245 if (!UVM_ET_ISSUBMAP(entry)) { 5246 break; 5247 } 5248 map = entry->object.sub_map; 5249 } 5250 } 5251 5252 #endif /* DDB || DEBUGPRINT */ 5253 5254 #ifndef __USER_VA0_IS_SAFE 5255 static int 5256 sysctl_user_va0_disable(SYSCTLFN_ARGS) 5257 { 5258 struct sysctlnode node; 5259 int t, error; 5260 5261 node = *rnode; 5262 node.sysctl_data = &t; 5263 t = user_va0_disable; 5264 error = sysctl_lookup(SYSCTLFN_CALL(&node)); 5265 if (error || newp == NULL) 5266 return (error); 5267 5268 /* lower only at securelevel < 1 */ 5269 if (!t && user_va0_disable && 5270 kauth_authorize_system(l->l_cred, 5271 KAUTH_SYSTEM_CHSYSFLAGS /* XXX */, 0, 5272 NULL, NULL, NULL)) 5273 return EPERM; 5274 5275 user_va0_disable = !!t; 5276 return 0; 5277 } 5278 5279 SYSCTL_SETUP(sysctl_uvmmap_setup, "sysctl uvmmap setup") 5280 { 5281 5282 sysctl_createv(clog, 0, NULL, NULL, 5283 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 5284 CTLTYPE_INT, "user_va0_disable", 5285 SYSCTL_DESCR("Disable VA 0"), 5286 sysctl_user_va0_disable, 0, &user_va0_disable, 0, 5287 CTL_VM, CTL_CREATE, CTL_EOL); 5288 } 5289 #endif 5290