1 /* $NetBSD: uvm_map.c,v 1.305 2011/09/27 01:02:39 jym 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 1.305 2011/09/27 01:02:39 jym 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 int, 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 * If the previous entry is not our immediate left child, then it's an 442 * ancestor and will be fixed up on the way to the root. We don't 443 * have to check entry->prev against &map->header since &map->header 444 * will never be in the tree. 445 */ 446 uvm_rb_fixup(map, 447 LEFT_ENTRY(entry) == entry->prev ? entry->prev : entry); 448 } 449 450 static void 451 uvm_rb_remove(struct vm_map *map, struct vm_map_entry *entry) 452 { 453 struct vm_map_entry *prev_parent = NULL, *next_parent = NULL; 454 455 /* 456 * If we are removing an interior node, then an adjacent node will 457 * be used to replace its position in the tree. Therefore we will 458 * need to fixup the tree starting at the parent of the replacement 459 * node. So record their parents for later use. 460 */ 461 if (entry->prev != &map->header) 462 prev_parent = PARENT_ENTRY(map, entry->prev); 463 if (entry->next != &map->header) 464 next_parent = PARENT_ENTRY(map, entry->next); 465 466 rb_tree_remove_node(&map->rb_tree, entry); 467 468 /* 469 * If the previous node has a new parent, fixup the tree starting 470 * at the previous node's old parent. 471 */ 472 if (entry->prev != &map->header) { 473 /* 474 * Update the previous entry's gap due to our absence. 475 */ 476 entry->prev->gap = uvm_rb_gap(entry->prev); 477 uvm_rb_fixup(map, entry->prev); 478 if (prev_parent != NULL 479 && prev_parent != entry 480 && prev_parent != PARENT_ENTRY(map, entry->prev)) 481 uvm_rb_fixup(map, prev_parent); 482 } 483 484 /* 485 * If the next node has a new parent, fixup the tree starting 486 * at the next node's old parent. 487 */ 488 if (entry->next != &map->header) { 489 uvm_rb_fixup(map, entry->next); 490 if (next_parent != NULL 491 && next_parent != entry 492 && next_parent != PARENT_ENTRY(map, entry->next)) 493 uvm_rb_fixup(map, next_parent); 494 } 495 } 496 497 #if defined(DEBUG) 498 int uvm_debug_check_map = 0; 499 int uvm_debug_check_rbtree = 0; 500 #define uvm_map_check(map, name) \ 501 _uvm_map_check((map), (name), __FILE__, __LINE__) 502 static void 503 _uvm_map_check(struct vm_map *map, const char *name, 504 const char *file, int line) 505 { 506 507 if ((uvm_debug_check_map && _uvm_map_sanity(map)) || 508 (uvm_debug_check_rbtree && _uvm_tree_sanity(map))) { 509 panic("uvm_map_check failed: \"%s\" map=%p (%s:%d)", 510 name, map, file, line); 511 } 512 } 513 #else /* defined(DEBUG) */ 514 #define uvm_map_check(map, name) /* nothing */ 515 #endif /* defined(DEBUG) */ 516 517 #if defined(DEBUG) || defined(DDB) 518 int 519 _uvm_map_sanity(struct vm_map *map) 520 { 521 bool first_free_found = false; 522 bool hint_found = false; 523 const struct vm_map_entry *e; 524 struct vm_map_entry *hint = map->hint; 525 526 e = &map->header; 527 for (;;) { 528 if (map->first_free == e) { 529 first_free_found = true; 530 } else if (!first_free_found && e->next->start > e->end) { 531 printf("first_free %p should be %p\n", 532 map->first_free, e); 533 return -1; 534 } 535 if (hint == e) { 536 hint_found = true; 537 } 538 539 e = e->next; 540 if (e == &map->header) { 541 break; 542 } 543 } 544 if (!first_free_found) { 545 printf("stale first_free\n"); 546 return -1; 547 } 548 if (!hint_found) { 549 printf("stale hint\n"); 550 return -1; 551 } 552 return 0; 553 } 554 555 int 556 _uvm_tree_sanity(struct vm_map *map) 557 { 558 struct vm_map_entry *tmp, *trtmp; 559 int n = 0, i = 1; 560 561 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 562 if (tmp->gap != uvm_rb_gap(tmp)) { 563 printf("%d/%d gap %lx != %lx %s\n", 564 n + 1, map->nentries, 565 (ulong)tmp->gap, (ulong)uvm_rb_gap(tmp), 566 tmp->next == &map->header ? "(last)" : ""); 567 goto error; 568 } 569 /* 570 * If any entries are out of order, tmp->gap will be unsigned 571 * and will likely exceed the size of the map. 572 */ 573 if (tmp->gap >= vm_map_max(map) - vm_map_min(map)) { 574 printf("too large gap %zu\n", (size_t)tmp->gap); 575 goto error; 576 } 577 n++; 578 } 579 580 if (n != map->nentries) { 581 printf("nentries: %d vs %d\n", n, map->nentries); 582 goto error; 583 } 584 585 trtmp = NULL; 586 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 587 if (tmp->maxgap != uvm_rb_maxgap(tmp)) { 588 printf("maxgap %lx != %lx\n", 589 (ulong)tmp->maxgap, 590 (ulong)uvm_rb_maxgap(tmp)); 591 goto error; 592 } 593 if (trtmp != NULL && trtmp->start >= tmp->start) { 594 printf("corrupt: 0x%"PRIxVADDR"x >= 0x%"PRIxVADDR"x\n", 595 trtmp->start, tmp->start); 596 goto error; 597 } 598 599 trtmp = tmp; 600 } 601 602 for (tmp = map->header.next; tmp != &map->header; 603 tmp = tmp->next, i++) { 604 trtmp = rb_tree_iterate(&map->rb_tree, tmp, RB_DIR_LEFT); 605 if (trtmp == NULL) 606 trtmp = &map->header; 607 if (tmp->prev != trtmp) { 608 printf("lookup: %d: %p->prev=%p: %p\n", 609 i, tmp, tmp->prev, trtmp); 610 goto error; 611 } 612 trtmp = rb_tree_iterate(&map->rb_tree, tmp, RB_DIR_RIGHT); 613 if (trtmp == NULL) 614 trtmp = &map->header; 615 if (tmp->next != trtmp) { 616 printf("lookup: %d: %p->next=%p: %p\n", 617 i, tmp, tmp->next, trtmp); 618 goto error; 619 } 620 trtmp = rb_tree_find_node(&map->rb_tree, &tmp->start); 621 if (trtmp != tmp) { 622 printf("lookup: %d: %p - %p: %p\n", i, tmp, trtmp, 623 PARENT_ENTRY(map, tmp)); 624 goto error; 625 } 626 } 627 628 return (0); 629 error: 630 return (-1); 631 } 632 #endif /* defined(DEBUG) || defined(DDB) */ 633 634 #ifdef DIAGNOSTIC 635 static struct vm_map *uvm_kmapent_map(struct vm_map_entry *); 636 #endif 637 638 /* 639 * vm_map_lock: acquire an exclusive (write) lock on a map. 640 * 641 * => Note that "intrsafe" maps use only exclusive, spin locks. 642 * 643 * => The locking protocol provides for guaranteed upgrade from shared -> 644 * exclusive by whichever thread currently has the map marked busy. 645 * See "LOCKING PROTOCOL NOTES" in uvm_map.h. This is horrible; among 646 * other problems, it defeats any fairness guarantees provided by RW 647 * locks. 648 */ 649 650 void 651 vm_map_lock(struct vm_map *map) 652 { 653 654 if ((map->flags & VM_MAP_INTRSAFE) != 0) { 655 mutex_spin_enter(&map->mutex); 656 return; 657 } 658 659 for (;;) { 660 rw_enter(&map->lock, RW_WRITER); 661 if (map->busy == NULL) 662 break; 663 if (map->busy == curlwp) 664 break; 665 mutex_enter(&map->misc_lock); 666 rw_exit(&map->lock); 667 if (map->busy != NULL) 668 cv_wait(&map->cv, &map->misc_lock); 669 mutex_exit(&map->misc_lock); 670 } 671 672 map->timestamp++; 673 } 674 675 /* 676 * vm_map_lock_try: try to lock a map, failing if it is already locked. 677 */ 678 679 bool 680 vm_map_lock_try(struct vm_map *map) 681 { 682 683 if ((map->flags & VM_MAP_INTRSAFE) != 0) 684 return mutex_tryenter(&map->mutex); 685 if (!rw_tryenter(&map->lock, RW_WRITER)) 686 return false; 687 if (map->busy != NULL) { 688 rw_exit(&map->lock); 689 return false; 690 } 691 692 map->timestamp++; 693 return true; 694 } 695 696 /* 697 * vm_map_unlock: release an exclusive lock on a map. 698 */ 699 700 void 701 vm_map_unlock(struct vm_map *map) 702 { 703 704 if ((map->flags & VM_MAP_INTRSAFE) != 0) 705 mutex_spin_exit(&map->mutex); 706 else { 707 KASSERT(rw_write_held(&map->lock)); 708 KASSERT(map->busy == NULL || map->busy == curlwp); 709 rw_exit(&map->lock); 710 } 711 } 712 713 /* 714 * vm_map_unbusy: mark the map as unbusy, and wake any waiters that 715 * want an exclusive lock. 716 */ 717 718 void 719 vm_map_unbusy(struct vm_map *map) 720 { 721 722 KASSERT(map->busy == curlwp); 723 724 /* 725 * Safe to clear 'busy' and 'waiters' with only a read lock held: 726 * 727 * o they can only be set with a write lock held 728 * o writers are blocked out with a read or write hold 729 * o at any time, only one thread owns the set of values 730 */ 731 mutex_enter(&map->misc_lock); 732 map->busy = NULL; 733 cv_broadcast(&map->cv); 734 mutex_exit(&map->misc_lock); 735 } 736 737 /* 738 * vm_map_lock_read: acquire a shared (read) lock on a map. 739 */ 740 741 void 742 vm_map_lock_read(struct vm_map *map) 743 { 744 745 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 746 747 rw_enter(&map->lock, RW_READER); 748 } 749 750 /* 751 * vm_map_unlock_read: release a shared lock on a map. 752 */ 753 754 void 755 vm_map_unlock_read(struct vm_map *map) 756 { 757 758 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 759 760 rw_exit(&map->lock); 761 } 762 763 /* 764 * vm_map_busy: mark a map as busy. 765 * 766 * => the caller must hold the map write locked 767 */ 768 769 void 770 vm_map_busy(struct vm_map *map) 771 { 772 773 KASSERT(rw_write_held(&map->lock)); 774 KASSERT(map->busy == NULL); 775 776 map->busy = curlwp; 777 } 778 779 /* 780 * vm_map_locked_p: return true if the map is write locked. 781 * 782 * => only for debug purposes like KASSERTs. 783 * => should not be used to verify that a map is not locked. 784 */ 785 786 bool 787 vm_map_locked_p(struct vm_map *map) 788 { 789 790 if ((map->flags & VM_MAP_INTRSAFE) != 0) { 791 return mutex_owned(&map->mutex); 792 } else { 793 return rw_write_held(&map->lock); 794 } 795 } 796 797 /* 798 * uvm_mapent_alloc: allocate a map entry 799 */ 800 801 static struct vm_map_entry * 802 uvm_mapent_alloc(struct vm_map *map, int flags) 803 { 804 struct vm_map_entry *me; 805 int pflags = (flags & UVM_FLAG_NOWAIT) ? PR_NOWAIT : PR_WAITOK; 806 UVMHIST_FUNC("uvm_mapent_alloc"); UVMHIST_CALLED(maphist); 807 808 if (VM_MAP_USE_KMAPENT(map)) { 809 me = uvm_kmapent_alloc(map, flags); 810 } else { 811 me = pool_cache_get(&uvm_map_entry_cache, pflags); 812 if (__predict_false(me == NULL)) 813 return NULL; 814 me->flags = 0; 815 } 816 817 UVMHIST_LOG(maphist, "<- new entry=0x%x [kentry=%d]", me, 818 ((map->flags & VM_MAP_INTRSAFE) != 0 || map == kernel_map), 0, 0); 819 return (me); 820 } 821 822 /* 823 * uvm_mapent_alloc_split: allocate a map entry for clipping. 824 * 825 * => map must be locked by caller if UVM_MAP_QUANTUM is set. 826 */ 827 828 static struct vm_map_entry * 829 uvm_mapent_alloc_split(struct vm_map *map, 830 const struct vm_map_entry *old_entry, int flags, 831 struct uvm_mapent_reservation *umr) 832 { 833 struct vm_map_entry *me; 834 835 KASSERT(!VM_MAP_USE_KMAPENT(map) || 836 (old_entry->flags & UVM_MAP_QUANTUM) || !UMR_EMPTY(umr)); 837 838 if (old_entry->flags & UVM_MAP_QUANTUM) { 839 struct vm_map_kernel *vmk = vm_map_to_kernel(map); 840 841 KASSERT(vm_map_locked_p(map)); 842 me = vmk->vmk_merged_entries; 843 KASSERT(me); 844 vmk->vmk_merged_entries = me->next; 845 KASSERT(me->flags & UVM_MAP_QUANTUM); 846 } else { 847 me = uvm_mapent_alloc(map, flags); 848 } 849 850 return me; 851 } 852 853 /* 854 * uvm_mapent_free: free map entry 855 */ 856 857 static void 858 uvm_mapent_free(struct vm_map_entry *me) 859 { 860 UVMHIST_FUNC("uvm_mapent_free"); UVMHIST_CALLED(maphist); 861 862 UVMHIST_LOG(maphist,"<- freeing map entry=0x%x [flags=%d]", 863 me, me->flags, 0, 0); 864 if (me->flags & UVM_MAP_KERNEL) { 865 uvm_kmapent_free(me); 866 } else { 867 pool_cache_put(&uvm_map_entry_cache, me); 868 } 869 } 870 871 /* 872 * uvm_mapent_free_merged: free merged map entry 873 * 874 * => keep the entry if needed. 875 * => caller shouldn't hold map locked if VM_MAP_USE_KMAPENT(map) is true. 876 * => map should be locked if UVM_MAP_QUANTUM is set. 877 */ 878 879 static void 880 uvm_mapent_free_merged(struct vm_map *map, struct vm_map_entry *me) 881 { 882 883 KASSERT(!(me->flags & UVM_MAP_KERNEL) || uvm_kmapent_map(me) == map); 884 885 if (me->flags & UVM_MAP_QUANTUM) { 886 /* 887 * keep this entry for later splitting. 888 */ 889 struct vm_map_kernel *vmk; 890 891 KASSERT(vm_map_locked_p(map)); 892 KASSERT(VM_MAP_IS_KERNEL(map)); 893 KASSERT(!VM_MAP_USE_KMAPENT(map) || 894 (me->flags & UVM_MAP_KERNEL)); 895 896 vmk = vm_map_to_kernel(map); 897 me->next = vmk->vmk_merged_entries; 898 vmk->vmk_merged_entries = me; 899 } else { 900 uvm_mapent_free(me); 901 } 902 } 903 904 /* 905 * uvm_mapent_copy: copy a map entry, preserving flags 906 */ 907 908 static inline void 909 uvm_mapent_copy(struct vm_map_entry *src, struct vm_map_entry *dst) 910 { 911 912 memcpy(dst, src, ((char *)&src->uvm_map_entry_stop_copy) - 913 ((char *)src)); 914 } 915 916 /* 917 * uvm_mapent_overhead: calculate maximum kva overhead necessary for 918 * map entries. 919 * 920 * => size and flags are the same as uvm_km_suballoc's ones. 921 */ 922 923 vsize_t 924 uvm_mapent_overhead(vsize_t size, int flags) 925 { 926 927 if (VM_MAP_USE_KMAPENT_FLAGS(flags)) { 928 return uvm_kmapent_overhead(size); 929 } 930 return 0; 931 } 932 933 #if defined(DEBUG) 934 static void 935 _uvm_mapent_check(const struct vm_map_entry *entry, const char *file, int line) 936 { 937 938 if (entry->start >= entry->end) { 939 goto bad; 940 } 941 if (UVM_ET_ISOBJ(entry)) { 942 if (entry->object.uvm_obj == NULL) { 943 goto bad; 944 } 945 } else if (UVM_ET_ISSUBMAP(entry)) { 946 if (entry->object.sub_map == NULL) { 947 goto bad; 948 } 949 } else { 950 if (entry->object.uvm_obj != NULL || 951 entry->object.sub_map != NULL) { 952 goto bad; 953 } 954 } 955 if (!UVM_ET_ISOBJ(entry)) { 956 if (entry->offset != 0) { 957 goto bad; 958 } 959 } 960 961 return; 962 963 bad: 964 panic("%s: bad entry %p (%s:%d)", __func__, entry, file, line); 965 } 966 #endif /* defined(DEBUG) */ 967 968 /* 969 * uvm_map_entry_unwire: unwire a map entry 970 * 971 * => map should be locked by caller 972 */ 973 974 static inline void 975 uvm_map_entry_unwire(struct vm_map *map, struct vm_map_entry *entry) 976 { 977 978 entry->wired_count = 0; 979 uvm_fault_unwire_locked(map, entry->start, entry->end); 980 } 981 982 983 /* 984 * wrapper for calling amap_ref() 985 */ 986 static inline void 987 uvm_map_reference_amap(struct vm_map_entry *entry, int flags) 988 { 989 990 amap_ref(entry->aref.ar_amap, entry->aref.ar_pageoff, 991 (entry->end - entry->start) >> PAGE_SHIFT, flags); 992 } 993 994 995 /* 996 * wrapper for calling amap_unref() 997 */ 998 static inline void 999 uvm_map_unreference_amap(struct vm_map_entry *entry, int flags) 1000 { 1001 1002 amap_unref(entry->aref.ar_amap, entry->aref.ar_pageoff, 1003 (entry->end - entry->start) >> PAGE_SHIFT, flags); 1004 } 1005 1006 1007 /* 1008 * uvm_map_init: init mapping system at boot time. 1009 */ 1010 1011 void 1012 uvm_map_init(void) 1013 { 1014 #if defined(UVMHIST) 1015 static struct kern_history_ent maphistbuf[100]; 1016 static struct kern_history_ent pdhistbuf[100]; 1017 #endif 1018 1019 /* 1020 * first, init logging system. 1021 */ 1022 1023 UVMHIST_FUNC("uvm_map_init"); 1024 UVMHIST_INIT_STATIC(maphist, maphistbuf); 1025 UVMHIST_INIT_STATIC(pdhist, pdhistbuf); 1026 UVMHIST_CALLED(maphist); 1027 UVMHIST_LOG(maphist,"<starting uvm map system>", 0, 0, 0, 0); 1028 1029 /* 1030 * initialize the global lock for kernel map entry. 1031 */ 1032 1033 mutex_init(&uvm_kentry_lock, MUTEX_DRIVER, IPL_VM); 1034 1035 /* 1036 * initialize caches. 1037 */ 1038 1039 pool_cache_bootstrap(&uvm_map_entry_cache, sizeof(struct vm_map_entry), 1040 0, 0, 0, "vmmpepl", NULL, IPL_NONE, NULL, NULL, NULL); 1041 pool_cache_bootstrap(&uvm_vmspace_cache, sizeof(struct vmspace), 1042 0, 0, 0, "vmsppl", NULL, IPL_NONE, NULL, NULL, NULL); 1043 } 1044 1045 /* 1046 * clippers 1047 */ 1048 1049 /* 1050 * uvm_mapent_splitadj: adjust map entries for splitting, after uvm_mapent_copy. 1051 */ 1052 1053 static void 1054 uvm_mapent_splitadj(struct vm_map_entry *entry1, struct vm_map_entry *entry2, 1055 vaddr_t splitat) 1056 { 1057 vaddr_t adj; 1058 1059 KASSERT(entry1->start < splitat); 1060 KASSERT(splitat < entry1->end); 1061 1062 adj = splitat - entry1->start; 1063 entry1->end = entry2->start = splitat; 1064 1065 if (entry1->aref.ar_amap) { 1066 amap_splitref(&entry1->aref, &entry2->aref, adj); 1067 } 1068 if (UVM_ET_ISSUBMAP(entry1)) { 1069 /* ... unlikely to happen, but play it safe */ 1070 uvm_map_reference(entry1->object.sub_map); 1071 } else if (UVM_ET_ISOBJ(entry1)) { 1072 KASSERT(entry1->object.uvm_obj != NULL); /* suppress coverity */ 1073 entry2->offset += adj; 1074 if (entry1->object.uvm_obj->pgops && 1075 entry1->object.uvm_obj->pgops->pgo_reference) 1076 entry1->object.uvm_obj->pgops->pgo_reference( 1077 entry1->object.uvm_obj); 1078 } 1079 } 1080 1081 /* 1082 * uvm_map_clip_start: ensure that the entry begins at or after 1083 * the starting address, if it doesn't we split the entry. 1084 * 1085 * => caller should use UVM_MAP_CLIP_START macro rather than calling 1086 * this directly 1087 * => map must be locked by caller 1088 */ 1089 1090 void 1091 uvm_map_clip_start(struct vm_map *map, struct vm_map_entry *entry, 1092 vaddr_t start, struct uvm_mapent_reservation *umr) 1093 { 1094 struct vm_map_entry *new_entry; 1095 1096 /* uvm_map_simplify_entry(map, entry); */ /* XXX */ 1097 1098 uvm_map_check(map, "clip_start entry"); 1099 uvm_mapent_check(entry); 1100 1101 /* 1102 * Split off the front portion. note that we must insert the new 1103 * entry BEFORE this one, so that this entry has the specified 1104 * starting address. 1105 */ 1106 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 1107 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1108 uvm_mapent_splitadj(new_entry, entry, start); 1109 uvm_map_entry_link(map, entry->prev, new_entry); 1110 1111 uvm_map_check(map, "clip_start leave"); 1112 } 1113 1114 /* 1115 * uvm_map_clip_end: ensure that the entry ends at or before 1116 * the ending address, if it does't we split the reference 1117 * 1118 * => caller should use UVM_MAP_CLIP_END macro rather than calling 1119 * this directly 1120 * => map must be locked by caller 1121 */ 1122 1123 void 1124 uvm_map_clip_end(struct vm_map *map, struct vm_map_entry *entry, vaddr_t end, 1125 struct uvm_mapent_reservation *umr) 1126 { 1127 struct vm_map_entry *new_entry; 1128 1129 uvm_map_check(map, "clip_end entry"); 1130 uvm_mapent_check(entry); 1131 1132 /* 1133 * Create a new entry and insert it 1134 * AFTER the specified entry 1135 */ 1136 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 1137 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1138 uvm_mapent_splitadj(entry, new_entry, end); 1139 uvm_map_entry_link(map, entry, new_entry); 1140 1141 uvm_map_check(map, "clip_end leave"); 1142 } 1143 1144 static void 1145 vm_map_drain(struct vm_map *map, uvm_flag_t flags) 1146 { 1147 1148 if (!VM_MAP_IS_KERNEL(map)) { 1149 return; 1150 } 1151 1152 uvm_km_va_drain(map, flags); 1153 } 1154 1155 /* 1156 * M A P - m a i n e n t r y p o i n t 1157 */ 1158 /* 1159 * uvm_map: establish a valid mapping in a map 1160 * 1161 * => assume startp is page aligned. 1162 * => assume size is a multiple of PAGE_SIZE. 1163 * => assume sys_mmap provides enough of a "hint" to have us skip 1164 * over text/data/bss area. 1165 * => map must be unlocked (we will lock it) 1166 * => <uobj,uoffset> value meanings (4 cases): 1167 * [1] <NULL,uoffset> == uoffset is a hint for PMAP_PREFER 1168 * [2] <NULL,UVM_UNKNOWN_OFFSET> == don't PMAP_PREFER 1169 * [3] <uobj,uoffset> == normal mapping 1170 * [4] <uobj,UVM_UNKNOWN_OFFSET> == uvm_map finds offset based on VA 1171 * 1172 * case [4] is for kernel mappings where we don't know the offset until 1173 * we've found a virtual address. note that kernel object offsets are 1174 * always relative to vm_map_min(kernel_map). 1175 * 1176 * => if `align' is non-zero, we align the virtual address to the specified 1177 * alignment. 1178 * this is provided as a mechanism for large pages. 1179 * 1180 * => XXXCDC: need way to map in external amap? 1181 */ 1182 1183 int 1184 uvm_map(struct vm_map *map, vaddr_t *startp /* IN/OUT */, vsize_t size, 1185 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags) 1186 { 1187 struct uvm_map_args args; 1188 struct vm_map_entry *new_entry; 1189 int error; 1190 1191 KASSERT((flags & UVM_FLAG_QUANTUM) == 0 || VM_MAP_IS_KERNEL(map)); 1192 KASSERT((size & PAGE_MASK) == 0); 1193 1194 #ifndef __USER_VA0_IS_SAFE 1195 if ((flags & UVM_FLAG_FIXED) && *startp == 0 && 1196 !VM_MAP_IS_KERNEL(map) && user_va0_disable) 1197 return EACCES; 1198 #endif 1199 1200 /* 1201 * for pager_map, allocate the new entry first to avoid sleeping 1202 * for memory while we have the map locked. 1203 * 1204 * Also, because we allocate entries for in-kernel maps 1205 * a bit differently (cf. uvm_kmapent_alloc/free), we need to 1206 * allocate them before locking the map. 1207 */ 1208 1209 new_entry = NULL; 1210 if (VM_MAP_USE_KMAPENT(map) || (flags & UVM_FLAG_QUANTUM) || 1211 map == pager_map) { 1212 new_entry = uvm_mapent_alloc(map, (flags & UVM_FLAG_NOWAIT)); 1213 if (__predict_false(new_entry == NULL)) 1214 return ENOMEM; 1215 if (flags & UVM_FLAG_QUANTUM) 1216 new_entry->flags |= UVM_MAP_QUANTUM; 1217 } 1218 if (map == pager_map) 1219 flags |= UVM_FLAG_NOMERGE; 1220 1221 error = uvm_map_prepare(map, *startp, size, uobj, uoffset, align, 1222 flags, &args); 1223 if (!error) { 1224 error = uvm_map_enter(map, &args, new_entry); 1225 *startp = args.uma_start; 1226 } else if (new_entry) { 1227 uvm_mapent_free(new_entry); 1228 } 1229 1230 #if defined(DEBUG) 1231 if (!error && VM_MAP_IS_KERNEL(map)) { 1232 uvm_km_check_empty(map, *startp, *startp + size); 1233 } 1234 #endif /* defined(DEBUG) */ 1235 1236 return error; 1237 } 1238 1239 int 1240 uvm_map_prepare(struct vm_map *map, vaddr_t start, vsize_t size, 1241 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags, 1242 struct uvm_map_args *args) 1243 { 1244 struct vm_map_entry *prev_entry; 1245 vm_prot_t prot = UVM_PROTECTION(flags); 1246 vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1247 1248 UVMHIST_FUNC("uvm_map_prepare"); 1249 UVMHIST_CALLED(maphist); 1250 1251 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1252 map, start, size, flags); 1253 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1254 1255 /* 1256 * detect a popular device driver bug. 1257 */ 1258 1259 KASSERT(doing_shutdown || curlwp != NULL || 1260 (map->flags & VM_MAP_INTRSAFE)); 1261 1262 /* 1263 * zero-sized mapping doesn't make any sense. 1264 */ 1265 KASSERT(size > 0); 1266 1267 KASSERT((~flags & (UVM_FLAG_NOWAIT | UVM_FLAG_WAITVA)) != 0); 1268 1269 uvm_map_check(map, "map entry"); 1270 1271 /* 1272 * check sanity of protection code 1273 */ 1274 1275 if ((prot & maxprot) != prot) { 1276 UVMHIST_LOG(maphist, "<- prot. failure: prot=0x%x, max=0x%x", 1277 prot, maxprot,0,0); 1278 return EACCES; 1279 } 1280 1281 /* 1282 * figure out where to put new VM range 1283 */ 1284 1285 retry: 1286 if (vm_map_lock_try(map) == false) { 1287 if ((flags & UVM_FLAG_TRYLOCK) != 0 && 1288 (map->flags & VM_MAP_INTRSAFE) == 0) { 1289 return EAGAIN; 1290 } 1291 vm_map_lock(map); /* could sleep here */ 1292 } 1293 prev_entry = uvm_map_findspace(map, start, size, &start, 1294 uobj, uoffset, align, flags); 1295 if (prev_entry == NULL) { 1296 unsigned int timestamp; 1297 1298 timestamp = map->timestamp; 1299 UVMHIST_LOG(maphist,"waiting va timestamp=0x%x", 1300 timestamp,0,0,0); 1301 map->flags |= VM_MAP_WANTVA; 1302 vm_map_unlock(map); 1303 1304 /* 1305 * try to reclaim kva and wait until someone does unmap. 1306 * fragile locking here, so we awaken every second to 1307 * recheck the condition. 1308 */ 1309 1310 vm_map_drain(map, flags); 1311 1312 mutex_enter(&map->misc_lock); 1313 while ((map->flags & VM_MAP_WANTVA) != 0 && 1314 map->timestamp == timestamp) { 1315 if ((flags & UVM_FLAG_WAITVA) == 0) { 1316 mutex_exit(&map->misc_lock); 1317 UVMHIST_LOG(maphist, 1318 "<- uvm_map_findspace failed!", 0,0,0,0); 1319 return ENOMEM; 1320 } else { 1321 cv_timedwait(&map->cv, &map->misc_lock, hz); 1322 } 1323 } 1324 mutex_exit(&map->misc_lock); 1325 goto retry; 1326 } 1327 1328 #ifdef PMAP_GROWKERNEL 1329 /* 1330 * If the kernel pmap can't map the requested space, 1331 * then allocate more resources for it. 1332 */ 1333 if (map == kernel_map && uvm_maxkaddr < (start + size)) 1334 uvm_maxkaddr = pmap_growkernel(start + size); 1335 #endif 1336 1337 UVMMAP_EVCNT_INCR(map_call); 1338 1339 /* 1340 * if uobj is null, then uoffset is either a VAC hint for PMAP_PREFER 1341 * [typically from uvm_map_reserve] or it is UVM_UNKNOWN_OFFSET. in 1342 * either case we want to zero it before storing it in the map entry 1343 * (because it looks strange and confusing when debugging...) 1344 * 1345 * if uobj is not null 1346 * if uoffset is not UVM_UNKNOWN_OFFSET then we have a normal mapping 1347 * and we do not need to change uoffset. 1348 * if uoffset is UVM_UNKNOWN_OFFSET then we need to find the offset 1349 * now (based on the starting address of the map). this case is 1350 * for kernel object mappings where we don't know the offset until 1351 * the virtual address is found (with uvm_map_findspace). the 1352 * offset is the distance we are from the start of the map. 1353 */ 1354 1355 if (uobj == NULL) { 1356 uoffset = 0; 1357 } else { 1358 if (uoffset == UVM_UNKNOWN_OFFSET) { 1359 KASSERT(UVM_OBJ_IS_KERN_OBJECT(uobj)); 1360 uoffset = start - vm_map_min(kernel_map); 1361 } 1362 } 1363 1364 args->uma_flags = flags; 1365 args->uma_prev = prev_entry; 1366 args->uma_start = start; 1367 args->uma_size = size; 1368 args->uma_uobj = uobj; 1369 args->uma_uoffset = uoffset; 1370 1371 UVMHIST_LOG(maphist, "<- done!", 0,0,0,0); 1372 return 0; 1373 } 1374 1375 int 1376 uvm_map_enter(struct vm_map *map, const struct uvm_map_args *args, 1377 struct vm_map_entry *new_entry) 1378 { 1379 struct vm_map_entry *prev_entry = args->uma_prev; 1380 struct vm_map_entry *dead = NULL; 1381 1382 const uvm_flag_t flags = args->uma_flags; 1383 const vm_prot_t prot = UVM_PROTECTION(flags); 1384 const vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1385 const vm_inherit_t inherit = UVM_INHERIT(flags); 1386 const int amapwaitflag = (flags & UVM_FLAG_NOWAIT) ? 1387 AMAP_EXTEND_NOWAIT : 0; 1388 const int advice = UVM_ADVICE(flags); 1389 const int meflagval = (flags & UVM_FLAG_QUANTUM) ? 1390 UVM_MAP_QUANTUM : 0; 1391 1392 vaddr_t start = args->uma_start; 1393 vsize_t size = args->uma_size; 1394 struct uvm_object *uobj = args->uma_uobj; 1395 voff_t uoffset = args->uma_uoffset; 1396 1397 const int kmap = (vm_map_pmap(map) == pmap_kernel()); 1398 int merged = 0; 1399 int error; 1400 int newetype; 1401 1402 UVMHIST_FUNC("uvm_map_enter"); 1403 UVMHIST_CALLED(maphist); 1404 1405 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1406 map, start, size, flags); 1407 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1408 1409 KASSERT(map->hint == prev_entry); /* bimerge case assumes this */ 1410 1411 if (flags & UVM_FLAG_QUANTUM) { 1412 KASSERT(new_entry); 1413 KASSERT(new_entry->flags & UVM_MAP_QUANTUM); 1414 } 1415 1416 if (uobj) 1417 newetype = UVM_ET_OBJ; 1418 else 1419 newetype = 0; 1420 1421 if (flags & UVM_FLAG_COPYONW) { 1422 newetype |= UVM_ET_COPYONWRITE; 1423 if ((flags & UVM_FLAG_OVERLAY) == 0) 1424 newetype |= UVM_ET_NEEDSCOPY; 1425 } 1426 1427 /* 1428 * try and insert in map by extending previous entry, if possible. 1429 * XXX: we don't try and pull back the next entry. might be useful 1430 * for a stack, but we are currently allocating our stack in advance. 1431 */ 1432 1433 if (flags & UVM_FLAG_NOMERGE) 1434 goto nomerge; 1435 1436 if (prev_entry->end == start && 1437 prev_entry != &map->header && 1438 UVM_ET_ISCOMPATIBLE(prev_entry, newetype, uobj, meflagval, 1439 prot, maxprot, inherit, advice, 0)) { 1440 1441 if (uobj && prev_entry->offset + 1442 (prev_entry->end - prev_entry->start) != uoffset) 1443 goto forwardmerge; 1444 1445 /* 1446 * can't extend a shared amap. note: no need to lock amap to 1447 * look at refs since we don't care about its exact value. 1448 * if it is one (i.e. we have only reference) it will stay there 1449 */ 1450 1451 if (prev_entry->aref.ar_amap && 1452 amap_refs(prev_entry->aref.ar_amap) != 1) { 1453 goto forwardmerge; 1454 } 1455 1456 if (prev_entry->aref.ar_amap) { 1457 error = amap_extend(prev_entry, size, 1458 amapwaitflag | AMAP_EXTEND_FORWARDS); 1459 if (error) 1460 goto nomerge; 1461 } 1462 1463 if (kmap) { 1464 UVMMAP_EVCNT_INCR(kbackmerge); 1465 } else { 1466 UVMMAP_EVCNT_INCR(ubackmerge); 1467 } 1468 UVMHIST_LOG(maphist," starting back merge", 0, 0, 0, 0); 1469 1470 /* 1471 * drop our reference to uobj since we are extending a reference 1472 * that we already have (the ref count can not drop to zero). 1473 */ 1474 1475 if (uobj && uobj->pgops->pgo_detach) 1476 uobj->pgops->pgo_detach(uobj); 1477 1478 /* 1479 * Now that we've merged the entries, note that we've grown 1480 * and our gap has shrunk. Then fix the tree. 1481 */ 1482 prev_entry->end += size; 1483 prev_entry->gap -= size; 1484 uvm_rb_fixup(map, prev_entry); 1485 1486 uvm_map_check(map, "map backmerged"); 1487 1488 UVMHIST_LOG(maphist,"<- done (via backmerge)!", 0, 0, 0, 0); 1489 merged++; 1490 } 1491 1492 forwardmerge: 1493 if (prev_entry->next->start == (start + size) && 1494 prev_entry->next != &map->header && 1495 UVM_ET_ISCOMPATIBLE(prev_entry->next, newetype, uobj, meflagval, 1496 prot, maxprot, inherit, advice, 0)) { 1497 1498 if (uobj && prev_entry->next->offset != uoffset + size) 1499 goto nomerge; 1500 1501 /* 1502 * can't extend a shared amap. note: no need to lock amap to 1503 * look at refs since we don't care about its exact value. 1504 * if it is one (i.e. we have only reference) it will stay there. 1505 * 1506 * note that we also can't merge two amaps, so if we 1507 * merged with the previous entry which has an amap, 1508 * and the next entry also has an amap, we give up. 1509 * 1510 * Interesting cases: 1511 * amap, new, amap -> give up second merge (single fwd extend) 1512 * amap, new, none -> double forward extend (extend again here) 1513 * none, new, amap -> double backward extend (done here) 1514 * uobj, new, amap -> single backward extend (done here) 1515 * 1516 * XXX should we attempt to deal with someone refilling 1517 * the deallocated region between two entries that are 1518 * backed by the same amap (ie, arefs is 2, "prev" and 1519 * "next" refer to it, and adding this allocation will 1520 * close the hole, thus restoring arefs to 1 and 1521 * deallocating the "next" vm_map_entry)? -- @@@ 1522 */ 1523 1524 if (prev_entry->next->aref.ar_amap && 1525 (amap_refs(prev_entry->next->aref.ar_amap) != 1 || 1526 (merged && prev_entry->aref.ar_amap))) { 1527 goto nomerge; 1528 } 1529 1530 if (merged) { 1531 /* 1532 * Try to extend the amap of the previous entry to 1533 * cover the next entry as well. If it doesn't work 1534 * just skip on, don't actually give up, since we've 1535 * already completed the back merge. 1536 */ 1537 if (prev_entry->aref.ar_amap) { 1538 if (amap_extend(prev_entry, 1539 prev_entry->next->end - 1540 prev_entry->next->start, 1541 amapwaitflag | AMAP_EXTEND_FORWARDS)) 1542 goto nomerge; 1543 } 1544 1545 /* 1546 * Try to extend the amap of the *next* entry 1547 * back to cover the new allocation *and* the 1548 * previous entry as well (the previous merge 1549 * didn't have an amap already otherwise we 1550 * wouldn't be checking here for an amap). If 1551 * it doesn't work just skip on, again, don't 1552 * actually give up, since we've already 1553 * completed the back merge. 1554 */ 1555 else if (prev_entry->next->aref.ar_amap) { 1556 if (amap_extend(prev_entry->next, 1557 prev_entry->end - 1558 prev_entry->start, 1559 amapwaitflag | AMAP_EXTEND_BACKWARDS)) 1560 goto nomerge; 1561 } 1562 } else { 1563 /* 1564 * Pull the next entry's amap backwards to cover this 1565 * new allocation. 1566 */ 1567 if (prev_entry->next->aref.ar_amap) { 1568 error = amap_extend(prev_entry->next, size, 1569 amapwaitflag | AMAP_EXTEND_BACKWARDS); 1570 if (error) 1571 goto nomerge; 1572 } 1573 } 1574 1575 if (merged) { 1576 if (kmap) { 1577 UVMMAP_EVCNT_DECR(kbackmerge); 1578 UVMMAP_EVCNT_INCR(kbimerge); 1579 } else { 1580 UVMMAP_EVCNT_DECR(ubackmerge); 1581 UVMMAP_EVCNT_INCR(ubimerge); 1582 } 1583 } else { 1584 if (kmap) { 1585 UVMMAP_EVCNT_INCR(kforwmerge); 1586 } else { 1587 UVMMAP_EVCNT_INCR(uforwmerge); 1588 } 1589 } 1590 UVMHIST_LOG(maphist," starting forward merge", 0, 0, 0, 0); 1591 1592 /* 1593 * drop our reference to uobj since we are extending a reference 1594 * that we already have (the ref count can not drop to zero). 1595 * (if merged, we've already detached) 1596 */ 1597 if (uobj && uobj->pgops->pgo_detach && !merged) 1598 uobj->pgops->pgo_detach(uobj); 1599 1600 if (merged) { 1601 dead = prev_entry->next; 1602 prev_entry->end = dead->end; 1603 uvm_map_entry_unlink(map, dead); 1604 if (dead->aref.ar_amap != NULL) { 1605 prev_entry->aref = dead->aref; 1606 dead->aref.ar_amap = NULL; 1607 } 1608 } else { 1609 prev_entry->next->start -= size; 1610 if (prev_entry != &map->header) { 1611 prev_entry->gap -= size; 1612 KASSERT(prev_entry->gap == uvm_rb_gap(prev_entry)); 1613 uvm_rb_fixup(map, prev_entry); 1614 } 1615 if (uobj) 1616 prev_entry->next->offset = uoffset; 1617 } 1618 1619 uvm_map_check(map, "map forwardmerged"); 1620 1621 UVMHIST_LOG(maphist,"<- done forwardmerge", 0, 0, 0, 0); 1622 merged++; 1623 } 1624 1625 nomerge: 1626 if (!merged) { 1627 UVMHIST_LOG(maphist," allocating new map entry", 0, 0, 0, 0); 1628 if (kmap) { 1629 UVMMAP_EVCNT_INCR(knomerge); 1630 } else { 1631 UVMMAP_EVCNT_INCR(unomerge); 1632 } 1633 1634 /* 1635 * allocate new entry and link it in. 1636 */ 1637 1638 if (new_entry == NULL) { 1639 new_entry = uvm_mapent_alloc(map, 1640 (flags & UVM_FLAG_NOWAIT)); 1641 if (__predict_false(new_entry == NULL)) { 1642 error = ENOMEM; 1643 goto done; 1644 } 1645 } 1646 new_entry->start = start; 1647 new_entry->end = new_entry->start + size; 1648 new_entry->object.uvm_obj = uobj; 1649 new_entry->offset = uoffset; 1650 1651 new_entry->etype = newetype; 1652 1653 if (flags & UVM_FLAG_NOMERGE) { 1654 new_entry->flags |= UVM_MAP_NOMERGE; 1655 } 1656 1657 new_entry->protection = prot; 1658 new_entry->max_protection = maxprot; 1659 new_entry->inheritance = inherit; 1660 new_entry->wired_count = 0; 1661 new_entry->advice = advice; 1662 if (flags & UVM_FLAG_OVERLAY) { 1663 1664 /* 1665 * to_add: for BSS we overallocate a little since we 1666 * are likely to extend 1667 */ 1668 1669 vaddr_t to_add = (flags & UVM_FLAG_AMAPPAD) ? 1670 UVM_AMAP_CHUNK << PAGE_SHIFT : 0; 1671 struct vm_amap *amap = amap_alloc(size, to_add, 1672 (flags & UVM_FLAG_NOWAIT)); 1673 if (__predict_false(amap == NULL)) { 1674 error = ENOMEM; 1675 goto done; 1676 } 1677 new_entry->aref.ar_pageoff = 0; 1678 new_entry->aref.ar_amap = amap; 1679 } else { 1680 new_entry->aref.ar_pageoff = 0; 1681 new_entry->aref.ar_amap = NULL; 1682 } 1683 uvm_map_entry_link(map, prev_entry, new_entry); 1684 1685 /* 1686 * Update the free space hint 1687 */ 1688 1689 if ((map->first_free == prev_entry) && 1690 (prev_entry->end >= new_entry->start)) 1691 map->first_free = new_entry; 1692 1693 new_entry = NULL; 1694 } 1695 1696 map->size += size; 1697 1698 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 1699 1700 error = 0; 1701 done: 1702 if ((flags & UVM_FLAG_QUANTUM) == 0) { 1703 /* 1704 * vmk_merged_entries is locked by the map's lock. 1705 */ 1706 vm_map_unlock(map); 1707 } 1708 if (new_entry && error == 0) { 1709 KDASSERT(merged); 1710 uvm_mapent_free_merged(map, new_entry); 1711 new_entry = NULL; 1712 } 1713 if (dead) { 1714 KDASSERT(merged); 1715 uvm_mapent_free_merged(map, dead); 1716 } 1717 if ((flags & UVM_FLAG_QUANTUM) != 0) { 1718 vm_map_unlock(map); 1719 } 1720 if (new_entry != NULL) { 1721 uvm_mapent_free(new_entry); 1722 } 1723 return error; 1724 } 1725 1726 /* 1727 * uvm_map_lookup_entry_bytree: lookup an entry in tree 1728 */ 1729 1730 static inline bool 1731 uvm_map_lookup_entry_bytree(struct vm_map *map, vaddr_t address, 1732 struct vm_map_entry **entry /* OUT */) 1733 { 1734 struct vm_map_entry *prev = &map->header; 1735 struct vm_map_entry *cur = ROOT_ENTRY(map); 1736 1737 while (cur) { 1738 UVMMAP_EVCNT_INCR(mlk_treeloop); 1739 if (address >= cur->start) { 1740 if (address < cur->end) { 1741 *entry = cur; 1742 return true; 1743 } 1744 prev = cur; 1745 cur = RIGHT_ENTRY(cur); 1746 } else 1747 cur = LEFT_ENTRY(cur); 1748 } 1749 *entry = prev; 1750 return false; 1751 } 1752 1753 /* 1754 * uvm_map_lookup_entry: find map entry at or before an address 1755 * 1756 * => map must at least be read-locked by caller 1757 * => entry is returned in "entry" 1758 * => return value is true if address is in the returned entry 1759 */ 1760 1761 bool 1762 uvm_map_lookup_entry(struct vm_map *map, vaddr_t address, 1763 struct vm_map_entry **entry /* OUT */) 1764 { 1765 struct vm_map_entry *cur; 1766 bool use_tree = false; 1767 UVMHIST_FUNC("uvm_map_lookup_entry"); 1768 UVMHIST_CALLED(maphist); 1769 1770 UVMHIST_LOG(maphist,"(map=0x%x,addr=0x%x,ent=0x%x)", 1771 map, address, entry, 0); 1772 1773 /* 1774 * start looking either from the head of the 1775 * list, or from the hint. 1776 */ 1777 1778 cur = map->hint; 1779 1780 if (cur == &map->header) 1781 cur = cur->next; 1782 1783 UVMMAP_EVCNT_INCR(mlk_call); 1784 if (address >= cur->start) { 1785 1786 /* 1787 * go from hint to end of list. 1788 * 1789 * but first, make a quick check to see if 1790 * we are already looking at the entry we 1791 * want (which is usually the case). 1792 * note also that we don't need to save the hint 1793 * here... it is the same hint (unless we are 1794 * at the header, in which case the hint didn't 1795 * buy us anything anyway). 1796 */ 1797 1798 if (cur != &map->header && cur->end > address) { 1799 UVMMAP_EVCNT_INCR(mlk_hint); 1800 *entry = cur; 1801 UVMHIST_LOG(maphist,"<- got it via hint (0x%x)", 1802 cur, 0, 0, 0); 1803 uvm_mapent_check(*entry); 1804 return (true); 1805 } 1806 1807 if (map->nentries > 15) 1808 use_tree = true; 1809 } else { 1810 1811 /* 1812 * invalid hint. use tree. 1813 */ 1814 use_tree = true; 1815 } 1816 1817 uvm_map_check(map, __func__); 1818 1819 if (use_tree) { 1820 /* 1821 * Simple lookup in the tree. Happens when the hint is 1822 * invalid, or nentries reach a threshold. 1823 */ 1824 UVMMAP_EVCNT_INCR(mlk_tree); 1825 if (uvm_map_lookup_entry_bytree(map, address, entry)) { 1826 goto got; 1827 } else { 1828 goto failed; 1829 } 1830 } 1831 1832 /* 1833 * search linearly 1834 */ 1835 1836 UVMMAP_EVCNT_INCR(mlk_list); 1837 while (cur != &map->header) { 1838 UVMMAP_EVCNT_INCR(mlk_listloop); 1839 if (cur->end > address) { 1840 if (address >= cur->start) { 1841 /* 1842 * save this lookup for future 1843 * hints, and return 1844 */ 1845 1846 *entry = cur; 1847 got: 1848 SAVE_HINT(map, map->hint, *entry); 1849 UVMHIST_LOG(maphist,"<- search got it (0x%x)", 1850 cur, 0, 0, 0); 1851 KDASSERT((*entry)->start <= address); 1852 KDASSERT(address < (*entry)->end); 1853 uvm_mapent_check(*entry); 1854 return (true); 1855 } 1856 break; 1857 } 1858 cur = cur->next; 1859 } 1860 *entry = cur->prev; 1861 failed: 1862 SAVE_HINT(map, map->hint, *entry); 1863 UVMHIST_LOG(maphist,"<- failed!",0,0,0,0); 1864 KDASSERT((*entry) == &map->header || (*entry)->end <= address); 1865 KDASSERT((*entry)->next == &map->header || 1866 address < (*entry)->next->start); 1867 return (false); 1868 } 1869 1870 /* 1871 * See if the range between start and start + length fits in the gap 1872 * entry->next->start and entry->end. Returns 1 if fits, 0 if doesn't 1873 * fit, and -1 address wraps around. 1874 */ 1875 static int 1876 uvm_map_space_avail(vaddr_t *start, vsize_t length, voff_t uoffset, 1877 vsize_t align, int flags, int topdown, struct vm_map_entry *entry) 1878 { 1879 vaddr_t end; 1880 1881 #ifdef PMAP_PREFER 1882 /* 1883 * push start address forward as needed to avoid VAC alias problems. 1884 * we only do this if a valid offset is specified. 1885 */ 1886 1887 if (uoffset != UVM_UNKNOWN_OFFSET) 1888 PMAP_PREFER(uoffset, start, length, topdown); 1889 #endif 1890 if ((flags & UVM_FLAG_COLORMATCH) != 0) { 1891 KASSERT(align < uvmexp.ncolors); 1892 if (uvmexp.ncolors > 1) { 1893 const u_int colormask = uvmexp.colormask; 1894 const u_int colorsize = colormask + 1; 1895 vaddr_t hint = atop(*start); 1896 const u_int color = hint & colormask; 1897 if (color != align) { 1898 hint -= color; /* adjust to color boundary */ 1899 KASSERT((hint & colormask) == 0); 1900 if (topdown) { 1901 if (align > color) 1902 hint -= colorsize; 1903 } else { 1904 if (align < color) 1905 hint += colorsize; 1906 } 1907 *start = ptoa(hint + align); /* adjust to color */ 1908 } 1909 } 1910 } else if (align != 0) { 1911 if ((*start & (align - 1)) != 0) { 1912 if (topdown) 1913 *start &= ~(align - 1); 1914 else 1915 *start = roundup(*start, align); 1916 } 1917 /* 1918 * XXX Should we PMAP_PREFER() here again? 1919 * eh...i think we're okay 1920 */ 1921 } 1922 1923 /* 1924 * Find the end of the proposed new region. Be sure we didn't 1925 * wrap around the address; if so, we lose. Otherwise, if the 1926 * proposed new region fits before the next entry, we win. 1927 */ 1928 1929 end = *start + length; 1930 if (end < *start) 1931 return (-1); 1932 1933 if (entry->next->start >= end && *start >= entry->end) 1934 return (1); 1935 1936 return (0); 1937 } 1938 1939 /* 1940 * uvm_map_findspace: find "length" sized space in "map". 1941 * 1942 * => "hint" is a hint about where we want it, unless UVM_FLAG_FIXED is 1943 * set in "flags" (in which case we insist on using "hint"). 1944 * => "result" is VA returned 1945 * => uobj/uoffset are to be used to handle VAC alignment, if required 1946 * => if "align" is non-zero, we attempt to align to that value. 1947 * => caller must at least have read-locked map 1948 * => returns NULL on failure, or pointer to prev. map entry if success 1949 * => note this is a cross between the old vm_map_findspace and vm_map_find 1950 */ 1951 1952 struct vm_map_entry * 1953 uvm_map_findspace(struct vm_map *map, vaddr_t hint, vsize_t length, 1954 vaddr_t *result /* OUT */, struct uvm_object *uobj, voff_t uoffset, 1955 vsize_t align, int flags) 1956 { 1957 struct vm_map_entry *entry; 1958 struct vm_map_entry *child, *prev, *tmp; 1959 vaddr_t orig_hint; 1960 const int topdown = map->flags & VM_MAP_TOPDOWN; 1961 UVMHIST_FUNC("uvm_map_findspace"); 1962 UVMHIST_CALLED(maphist); 1963 1964 UVMHIST_LOG(maphist, "(map=0x%x, hint=0x%x, len=%d, flags=0x%x)", 1965 map, hint, length, flags); 1966 KASSERT((flags & UVM_FLAG_COLORMATCH) != 0 || (align & (align - 1)) == 0); 1967 KASSERT((flags & UVM_FLAG_COLORMATCH) == 0 || align < uvmexp.ncolors); 1968 KASSERT((flags & UVM_FLAG_FIXED) == 0 || align == 0); 1969 1970 uvm_map_check(map, "map_findspace entry"); 1971 1972 /* 1973 * remember the original hint. if we are aligning, then we 1974 * may have to try again with no alignment constraint if 1975 * we fail the first time. 1976 */ 1977 1978 orig_hint = hint; 1979 if (hint < vm_map_min(map)) { /* check ranges ... */ 1980 if (flags & UVM_FLAG_FIXED) { 1981 UVMHIST_LOG(maphist,"<- VA below map range",0,0,0,0); 1982 return (NULL); 1983 } 1984 hint = vm_map_min(map); 1985 } 1986 if (hint > vm_map_max(map)) { 1987 UVMHIST_LOG(maphist,"<- VA 0x%x > range [0x%x->0x%x]", 1988 hint, vm_map_min(map), vm_map_max(map), 0); 1989 return (NULL); 1990 } 1991 1992 /* 1993 * Look for the first possible address; if there's already 1994 * something at this address, we have to start after it. 1995 */ 1996 1997 /* 1998 * @@@: there are four, no, eight cases to consider. 1999 * 2000 * 0: found, fixed, bottom up -> fail 2001 * 1: found, fixed, top down -> fail 2002 * 2: found, not fixed, bottom up -> start after entry->end, 2003 * loop up 2004 * 3: found, not fixed, top down -> start before entry->start, 2005 * loop down 2006 * 4: not found, fixed, bottom up -> check entry->next->start, fail 2007 * 5: not found, fixed, top down -> check entry->next->start, fail 2008 * 6: not found, not fixed, bottom up -> check entry->next->start, 2009 * loop up 2010 * 7: not found, not fixed, top down -> check entry->next->start, 2011 * loop down 2012 * 2013 * as you can see, it reduces to roughly five cases, and that 2014 * adding top down mapping only adds one unique case (without 2015 * it, there would be four cases). 2016 */ 2017 2018 if ((flags & UVM_FLAG_FIXED) == 0 && hint == vm_map_min(map)) { 2019 entry = map->first_free; 2020 } else { 2021 if (uvm_map_lookup_entry(map, hint, &entry)) { 2022 /* "hint" address already in use ... */ 2023 if (flags & UVM_FLAG_FIXED) { 2024 UVMHIST_LOG(maphist, "<- fixed & VA in use", 2025 0, 0, 0, 0); 2026 return (NULL); 2027 } 2028 if (topdown) 2029 /* Start from lower gap. */ 2030 entry = entry->prev; 2031 } else if (flags & UVM_FLAG_FIXED) { 2032 if (entry->next->start >= hint + length && 2033 hint + length > hint) 2034 goto found; 2035 2036 /* "hint" address is gap but too small */ 2037 UVMHIST_LOG(maphist, "<- fixed mapping failed", 2038 0, 0, 0, 0); 2039 return (NULL); /* only one shot at it ... */ 2040 } else { 2041 /* 2042 * See if given hint fits in this gap. 2043 */ 2044 switch (uvm_map_space_avail(&hint, length, 2045 uoffset, align, flags, topdown, entry)) { 2046 case 1: 2047 goto found; 2048 case -1: 2049 goto wraparound; 2050 } 2051 2052 if (topdown) { 2053 /* 2054 * Still there is a chance to fit 2055 * if hint > entry->end. 2056 */ 2057 } else { 2058 /* Start from higher gap. */ 2059 entry = entry->next; 2060 if (entry == &map->header) 2061 goto notfound; 2062 goto nextgap; 2063 } 2064 } 2065 } 2066 2067 /* 2068 * Note that all UVM_FLAGS_FIXED case is already handled. 2069 */ 2070 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2071 2072 /* Try to find the space in the red-black tree */ 2073 2074 /* Check slot before any entry */ 2075 hint = topdown ? entry->next->start - length : entry->end; 2076 switch (uvm_map_space_avail(&hint, length, uoffset, align, flags, 2077 topdown, entry)) { 2078 case 1: 2079 goto found; 2080 case -1: 2081 goto wraparound; 2082 } 2083 2084 nextgap: 2085 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2086 /* If there is not enough space in the whole tree, we fail */ 2087 tmp = ROOT_ENTRY(map); 2088 if (tmp == NULL || tmp->maxgap < length) 2089 goto notfound; 2090 2091 prev = NULL; /* previous candidate */ 2092 2093 /* Find an entry close to hint that has enough space */ 2094 for (; tmp;) { 2095 KASSERT(tmp->next->start == tmp->end + tmp->gap); 2096 if (topdown) { 2097 if (tmp->next->start < hint + length && 2098 (prev == NULL || tmp->end > prev->end)) { 2099 if (tmp->gap >= length) 2100 prev = tmp; 2101 else if ((child = LEFT_ENTRY(tmp)) != NULL 2102 && child->maxgap >= length) 2103 prev = tmp; 2104 } 2105 } else { 2106 if (tmp->end >= hint && 2107 (prev == NULL || tmp->end < prev->end)) { 2108 if (tmp->gap >= length) 2109 prev = tmp; 2110 else if ((child = RIGHT_ENTRY(tmp)) != NULL 2111 && child->maxgap >= length) 2112 prev = tmp; 2113 } 2114 } 2115 if (tmp->next->start < hint + length) 2116 child = RIGHT_ENTRY(tmp); 2117 else if (tmp->end > hint) 2118 child = LEFT_ENTRY(tmp); 2119 else { 2120 if (tmp->gap >= length) 2121 break; 2122 if (topdown) 2123 child = LEFT_ENTRY(tmp); 2124 else 2125 child = RIGHT_ENTRY(tmp); 2126 } 2127 if (child == NULL || child->maxgap < length) 2128 break; 2129 tmp = child; 2130 } 2131 2132 if (tmp != NULL && tmp->start < hint && hint < tmp->next->start) { 2133 /* 2134 * Check if the entry that we found satifies the 2135 * space requirement 2136 */ 2137 if (topdown) { 2138 if (hint > tmp->next->start - length) 2139 hint = tmp->next->start - length; 2140 } else { 2141 if (hint < tmp->end) 2142 hint = tmp->end; 2143 } 2144 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2145 flags, topdown, tmp)) { 2146 case 1: 2147 entry = tmp; 2148 goto found; 2149 case -1: 2150 goto wraparound; 2151 } 2152 if (tmp->gap >= length) 2153 goto listsearch; 2154 } 2155 if (prev == NULL) 2156 goto notfound; 2157 2158 if (topdown) { 2159 KASSERT(orig_hint >= prev->next->start - length || 2160 prev->next->start - length > prev->next->start); 2161 hint = prev->next->start - length; 2162 } else { 2163 KASSERT(orig_hint <= prev->end); 2164 hint = prev->end; 2165 } 2166 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2167 flags, topdown, prev)) { 2168 case 1: 2169 entry = prev; 2170 goto found; 2171 case -1: 2172 goto wraparound; 2173 } 2174 if (prev->gap >= length) 2175 goto listsearch; 2176 2177 if (topdown) 2178 tmp = LEFT_ENTRY(prev); 2179 else 2180 tmp = RIGHT_ENTRY(prev); 2181 for (;;) { 2182 KASSERT(tmp && tmp->maxgap >= length); 2183 if (topdown) 2184 child = RIGHT_ENTRY(tmp); 2185 else 2186 child = LEFT_ENTRY(tmp); 2187 if (child && child->maxgap >= length) { 2188 tmp = child; 2189 continue; 2190 } 2191 if (tmp->gap >= length) 2192 break; 2193 if (topdown) 2194 tmp = LEFT_ENTRY(tmp); 2195 else 2196 tmp = RIGHT_ENTRY(tmp); 2197 } 2198 2199 if (topdown) { 2200 KASSERT(orig_hint >= tmp->next->start - length || 2201 tmp->next->start - length > tmp->next->start); 2202 hint = tmp->next->start - length; 2203 } else { 2204 KASSERT(orig_hint <= tmp->end); 2205 hint = tmp->end; 2206 } 2207 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2208 flags, topdown, tmp)) { 2209 case 1: 2210 entry = tmp; 2211 goto found; 2212 case -1: 2213 goto wraparound; 2214 } 2215 2216 /* 2217 * The tree fails to find an entry because of offset or alignment 2218 * restrictions. Search the list instead. 2219 */ 2220 listsearch: 2221 /* 2222 * Look through the rest of the map, trying to fit a new region in 2223 * the gap between existing regions, or after the very last region. 2224 * note: entry->end = base VA of current gap, 2225 * entry->next->start = VA of end of current gap 2226 */ 2227 2228 for (;;) { 2229 /* Update hint for current gap. */ 2230 hint = topdown ? entry->next->start - length : entry->end; 2231 2232 /* See if it fits. */ 2233 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2234 flags, topdown, entry)) { 2235 case 1: 2236 goto found; 2237 case -1: 2238 goto wraparound; 2239 } 2240 2241 /* Advance to next/previous gap */ 2242 if (topdown) { 2243 if (entry == &map->header) { 2244 UVMHIST_LOG(maphist, "<- failed (off start)", 2245 0,0,0,0); 2246 goto notfound; 2247 } 2248 entry = entry->prev; 2249 } else { 2250 entry = entry->next; 2251 if (entry == &map->header) { 2252 UVMHIST_LOG(maphist, "<- failed (off end)", 2253 0,0,0,0); 2254 goto notfound; 2255 } 2256 } 2257 } 2258 2259 found: 2260 SAVE_HINT(map, map->hint, entry); 2261 *result = hint; 2262 UVMHIST_LOG(maphist,"<- got it! (result=0x%x)", hint, 0,0,0); 2263 KASSERT( topdown || hint >= orig_hint); 2264 KASSERT(!topdown || hint <= orig_hint); 2265 KASSERT(entry->end <= hint); 2266 KASSERT(hint + length <= entry->next->start); 2267 return (entry); 2268 2269 wraparound: 2270 UVMHIST_LOG(maphist, "<- failed (wrap around)", 0,0,0,0); 2271 2272 return (NULL); 2273 2274 notfound: 2275 UVMHIST_LOG(maphist, "<- failed (notfound)", 0,0,0,0); 2276 2277 return (NULL); 2278 } 2279 2280 /* 2281 * U N M A P - m a i n h e l p e r f u n c t i o n s 2282 */ 2283 2284 /* 2285 * uvm_unmap_remove: remove mappings from a vm_map (from "start" up to "stop") 2286 * 2287 * => caller must check alignment and size 2288 * => map must be locked by caller 2289 * => we return a list of map entries that we've remove from the map 2290 * in "entry_list" 2291 */ 2292 2293 void 2294 uvm_unmap_remove(struct vm_map *map, vaddr_t start, vaddr_t end, 2295 struct vm_map_entry **entry_list /* OUT */, 2296 struct uvm_mapent_reservation *umr, int flags) 2297 { 2298 struct vm_map_entry *entry, *first_entry, *next; 2299 vaddr_t len; 2300 UVMHIST_FUNC("uvm_unmap_remove"); UVMHIST_CALLED(maphist); 2301 2302 UVMHIST_LOG(maphist,"(map=0x%x, start=0x%x, end=0x%x)", 2303 map, start, end, 0); 2304 VM_MAP_RANGE_CHECK(map, start, end); 2305 2306 uvm_map_check(map, "unmap_remove entry"); 2307 2308 /* 2309 * find first entry 2310 */ 2311 2312 if (uvm_map_lookup_entry(map, start, &first_entry) == true) { 2313 /* clip and go... */ 2314 entry = first_entry; 2315 UVM_MAP_CLIP_START(map, entry, start, umr); 2316 /* critical! prevents stale hint */ 2317 SAVE_HINT(map, entry, entry->prev); 2318 } else { 2319 entry = first_entry->next; 2320 } 2321 2322 /* 2323 * Save the free space hint 2324 */ 2325 2326 if (map->first_free != &map->header && map->first_free->start >= start) 2327 map->first_free = entry->prev; 2328 2329 /* 2330 * note: we now re-use first_entry for a different task. we remove 2331 * a number of map entries from the map and save them in a linked 2332 * list headed by "first_entry". once we remove them from the map 2333 * the caller should unlock the map and drop the references to the 2334 * backing objects [c.f. uvm_unmap_detach]. the object is to 2335 * separate unmapping from reference dropping. why? 2336 * [1] the map has to be locked for unmapping 2337 * [2] the map need not be locked for reference dropping 2338 * [3] dropping references may trigger pager I/O, and if we hit 2339 * a pager that does synchronous I/O we may have to wait for it. 2340 * [4] we would like all waiting for I/O to occur with maps unlocked 2341 * so that we don't block other threads. 2342 */ 2343 2344 first_entry = NULL; 2345 *entry_list = NULL; 2346 2347 /* 2348 * break up the area into map entry sized regions and unmap. note 2349 * that all mappings have to be removed before we can even consider 2350 * dropping references to amaps or VM objects (otherwise we could end 2351 * up with a mapping to a page on the free list which would be very bad) 2352 */ 2353 2354 while ((entry != &map->header) && (entry->start < end)) { 2355 KASSERT((entry->flags & UVM_MAP_FIRST) == 0); 2356 2357 UVM_MAP_CLIP_END(map, entry, end, umr); 2358 next = entry->next; 2359 len = entry->end - entry->start; 2360 2361 /* 2362 * unwire before removing addresses from the pmap; otherwise 2363 * unwiring will put the entries back into the pmap (XXX). 2364 */ 2365 2366 if (VM_MAPENT_ISWIRED(entry)) { 2367 uvm_map_entry_unwire(map, entry); 2368 } 2369 if (flags & UVM_FLAG_VAONLY) { 2370 2371 /* nothing */ 2372 2373 } else if ((map->flags & VM_MAP_PAGEABLE) == 0) { 2374 2375 /* 2376 * if the map is non-pageable, any pages mapped there 2377 * must be wired and entered with pmap_kenter_pa(), 2378 * and we should free any such pages immediately. 2379 * this is mostly used for kmem_map. 2380 */ 2381 KASSERT(vm_map_pmap(map) == pmap_kernel()); 2382 2383 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2384 uvm_km_pgremove_intrsafe(map, entry->start, 2385 entry->end); 2386 pmap_kremove(entry->start, len); 2387 } 2388 } else if (UVM_ET_ISOBJ(entry) && 2389 UVM_OBJ_IS_KERN_OBJECT(entry->object.uvm_obj)) { 2390 panic("%s: kernel object %p %p\n", 2391 __func__, map, entry); 2392 } else if (UVM_ET_ISOBJ(entry) || entry->aref.ar_amap) { 2393 /* 2394 * remove mappings the standard way. lock object 2395 * and/or amap to ensure vm_page state does not 2396 * change while in pmap_remove(). 2397 */ 2398 2399 uvm_map_lock_entry(entry); 2400 pmap_remove(map->pmap, entry->start, entry->end); 2401 uvm_map_unlock_entry(entry); 2402 } 2403 2404 #if defined(DEBUG) 2405 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2406 2407 /* 2408 * check if there's remaining mapping, 2409 * which is a bug in caller. 2410 */ 2411 2412 vaddr_t va; 2413 for (va = entry->start; va < entry->end; 2414 va += PAGE_SIZE) { 2415 if (pmap_extract(vm_map_pmap(map), va, NULL)) { 2416 panic("%s: %#"PRIxVADDR" has mapping", 2417 __func__, va); 2418 } 2419 } 2420 2421 if (VM_MAP_IS_KERNEL(map)) { 2422 uvm_km_check_empty(map, entry->start, 2423 entry->end); 2424 } 2425 } 2426 #endif /* defined(DEBUG) */ 2427 2428 /* 2429 * remove entry from map and put it on our list of entries 2430 * that we've nuked. then go to next entry. 2431 */ 2432 2433 UVMHIST_LOG(maphist, " removed map entry 0x%x", entry, 0, 0,0); 2434 2435 /* critical! prevents stale hint */ 2436 SAVE_HINT(map, entry, entry->prev); 2437 2438 uvm_map_entry_unlink(map, entry); 2439 KASSERT(map->size >= len); 2440 map->size -= len; 2441 entry->prev = NULL; 2442 entry->next = first_entry; 2443 first_entry = entry; 2444 entry = next; 2445 } 2446 2447 /* 2448 * Note: if map is dying, leave pmap_update() for pmap_destroy(), 2449 * which will be called later. 2450 */ 2451 if ((map->flags & VM_MAP_DYING) == 0) { 2452 pmap_update(vm_map_pmap(map)); 2453 } else { 2454 KASSERT(vm_map_pmap(map) != pmap_kernel()); 2455 } 2456 2457 uvm_map_check(map, "unmap_remove leave"); 2458 2459 /* 2460 * now we've cleaned up the map and are ready for the caller to drop 2461 * references to the mapped objects. 2462 */ 2463 2464 *entry_list = first_entry; 2465 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 2466 2467 if (map->flags & VM_MAP_WANTVA) { 2468 mutex_enter(&map->misc_lock); 2469 map->flags &= ~VM_MAP_WANTVA; 2470 cv_broadcast(&map->cv); 2471 mutex_exit(&map->misc_lock); 2472 } 2473 } 2474 2475 /* 2476 * uvm_unmap_detach: drop references in a chain of map entries 2477 * 2478 * => we will free the map entries as we traverse the list. 2479 */ 2480 2481 void 2482 uvm_unmap_detach(struct vm_map_entry *first_entry, int flags) 2483 { 2484 struct vm_map_entry *next_entry; 2485 UVMHIST_FUNC("uvm_unmap_detach"); UVMHIST_CALLED(maphist); 2486 2487 while (first_entry) { 2488 KASSERT(!VM_MAPENT_ISWIRED(first_entry)); 2489 UVMHIST_LOG(maphist, 2490 " detach 0x%x: amap=0x%x, obj=0x%x, submap?=%d", 2491 first_entry, first_entry->aref.ar_amap, 2492 first_entry->object.uvm_obj, 2493 UVM_ET_ISSUBMAP(first_entry)); 2494 2495 /* 2496 * drop reference to amap, if we've got one 2497 */ 2498 2499 if (first_entry->aref.ar_amap) 2500 uvm_map_unreference_amap(first_entry, flags); 2501 2502 /* 2503 * drop reference to our backing object, if we've got one 2504 */ 2505 2506 KASSERT(!UVM_ET_ISSUBMAP(first_entry)); 2507 if (UVM_ET_ISOBJ(first_entry) && 2508 first_entry->object.uvm_obj->pgops->pgo_detach) { 2509 (*first_entry->object.uvm_obj->pgops->pgo_detach) 2510 (first_entry->object.uvm_obj); 2511 } 2512 next_entry = first_entry->next; 2513 uvm_mapent_free(first_entry); 2514 first_entry = next_entry; 2515 } 2516 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 2517 } 2518 2519 /* 2520 * E X T R A C T I O N F U N C T I O N S 2521 */ 2522 2523 /* 2524 * uvm_map_reserve: reserve space in a vm_map for future use. 2525 * 2526 * => we reserve space in a map by putting a dummy map entry in the 2527 * map (dummy means obj=NULL, amap=NULL, prot=VM_PROT_NONE) 2528 * => map should be unlocked (we will write lock it) 2529 * => we return true if we were able to reserve space 2530 * => XXXCDC: should be inline? 2531 */ 2532 2533 int 2534 uvm_map_reserve(struct vm_map *map, vsize_t size, 2535 vaddr_t offset /* hint for pmap_prefer */, 2536 vsize_t align /* alignment */, 2537 vaddr_t *raddr /* IN:hint, OUT: reserved VA */, 2538 uvm_flag_t flags /* UVM_FLAG_FIXED or 0 */) 2539 { 2540 UVMHIST_FUNC("uvm_map_reserve"); UVMHIST_CALLED(maphist); 2541 2542 UVMHIST_LOG(maphist, "(map=0x%x, size=0x%x, offset=0x%x,addr=0x%x)", 2543 map,size,offset,raddr); 2544 2545 size = round_page(size); 2546 2547 /* 2548 * reserve some virtual space. 2549 */ 2550 2551 if (uvm_map(map, raddr, size, NULL, offset, align, 2552 UVM_MAPFLAG(UVM_PROT_NONE, UVM_PROT_NONE, UVM_INH_NONE, 2553 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE|flags)) != 0) { 2554 UVMHIST_LOG(maphist, "<- done (no VM)", 0,0,0,0); 2555 return (false); 2556 } 2557 2558 UVMHIST_LOG(maphist, "<- done (*raddr=0x%x)", *raddr,0,0,0); 2559 return (true); 2560 } 2561 2562 /* 2563 * uvm_map_replace: replace a reserved (blank) area of memory with 2564 * real mappings. 2565 * 2566 * => caller must WRITE-LOCK the map 2567 * => we return true if replacement was a success 2568 * => we expect the newents chain to have nnewents entrys on it and 2569 * we expect newents->prev to point to the last entry on the list 2570 * => note newents is allowed to be NULL 2571 */ 2572 2573 static int 2574 uvm_map_replace(struct vm_map *map, vaddr_t start, vaddr_t end, 2575 struct vm_map_entry *newents, int nnewents, vsize_t nsize, 2576 struct vm_map_entry **oldentryp) 2577 { 2578 struct vm_map_entry *oldent, *last; 2579 2580 uvm_map_check(map, "map_replace entry"); 2581 2582 /* 2583 * first find the blank map entry at the specified address 2584 */ 2585 2586 if (!uvm_map_lookup_entry(map, start, &oldent)) { 2587 return (false); 2588 } 2589 2590 /* 2591 * check to make sure we have a proper blank entry 2592 */ 2593 2594 if (end < oldent->end && !VM_MAP_USE_KMAPENT(map)) { 2595 UVM_MAP_CLIP_END(map, oldent, end, NULL); 2596 } 2597 if (oldent->start != start || oldent->end != end || 2598 oldent->object.uvm_obj != NULL || oldent->aref.ar_amap != NULL) { 2599 return (false); 2600 } 2601 2602 #ifdef DIAGNOSTIC 2603 2604 /* 2605 * sanity check the newents chain 2606 */ 2607 2608 { 2609 struct vm_map_entry *tmpent = newents; 2610 int nent = 0; 2611 vsize_t sz = 0; 2612 vaddr_t cur = start; 2613 2614 while (tmpent) { 2615 nent++; 2616 sz += tmpent->end - tmpent->start; 2617 if (tmpent->start < cur) 2618 panic("uvm_map_replace1"); 2619 if (tmpent->start >= tmpent->end || tmpent->end > end) { 2620 panic("uvm_map_replace2: " 2621 "tmpent->start=0x%"PRIxVADDR 2622 ", tmpent->end=0x%"PRIxVADDR 2623 ", end=0x%"PRIxVADDR, 2624 tmpent->start, tmpent->end, end); 2625 } 2626 cur = tmpent->end; 2627 if (tmpent->next) { 2628 if (tmpent->next->prev != tmpent) 2629 panic("uvm_map_replace3"); 2630 } else { 2631 if (newents->prev != tmpent) 2632 panic("uvm_map_replace4"); 2633 } 2634 tmpent = tmpent->next; 2635 } 2636 if (nent != nnewents) 2637 panic("uvm_map_replace5"); 2638 if (sz != nsize) 2639 panic("uvm_map_replace6"); 2640 } 2641 #endif 2642 2643 /* 2644 * map entry is a valid blank! replace it. (this does all the 2645 * work of map entry link/unlink...). 2646 */ 2647 2648 if (newents) { 2649 last = newents->prev; 2650 2651 /* critical: flush stale hints out of map */ 2652 SAVE_HINT(map, map->hint, newents); 2653 if (map->first_free == oldent) 2654 map->first_free = last; 2655 2656 last->next = oldent->next; 2657 last->next->prev = last; 2658 2659 /* Fix RB tree */ 2660 uvm_rb_remove(map, oldent); 2661 2662 newents->prev = oldent->prev; 2663 newents->prev->next = newents; 2664 map->nentries = map->nentries + (nnewents - 1); 2665 2666 /* Fixup the RB tree */ 2667 { 2668 int i; 2669 struct vm_map_entry *tmp; 2670 2671 tmp = newents; 2672 for (i = 0; i < nnewents && tmp; i++) { 2673 uvm_rb_insert(map, tmp); 2674 tmp = tmp->next; 2675 } 2676 } 2677 } else { 2678 /* NULL list of new entries: just remove the old one */ 2679 clear_hints(map, oldent); 2680 uvm_map_entry_unlink(map, oldent); 2681 } 2682 map->size -= end - start - nsize; 2683 2684 uvm_map_check(map, "map_replace leave"); 2685 2686 /* 2687 * now we can free the old blank entry and return. 2688 */ 2689 2690 *oldentryp = oldent; 2691 return (true); 2692 } 2693 2694 /* 2695 * uvm_map_extract: extract a mapping from a map and put it somewhere 2696 * (maybe removing the old mapping) 2697 * 2698 * => maps should be unlocked (we will write lock them) 2699 * => returns 0 on success, error code otherwise 2700 * => start must be page aligned 2701 * => len must be page sized 2702 * => flags: 2703 * UVM_EXTRACT_REMOVE: remove mappings from srcmap 2704 * UVM_EXTRACT_CONTIG: abort if unmapped area (advisory only) 2705 * UVM_EXTRACT_QREF: for a temporary extraction do quick obj refs 2706 * UVM_EXTRACT_FIXPROT: set prot to maxprot as we go 2707 * >>>NOTE: if you set REMOVE, you are not allowed to use CONTIG or QREF!<<< 2708 * >>>NOTE: QREF's must be unmapped via the QREF path, thus should only 2709 * be used from within the kernel in a kernel level map <<< 2710 */ 2711 2712 int 2713 uvm_map_extract(struct vm_map *srcmap, vaddr_t start, vsize_t len, 2714 struct vm_map *dstmap, vaddr_t *dstaddrp, int flags) 2715 { 2716 vaddr_t dstaddr, end, newend, oldoffset, fudge, orig_fudge; 2717 struct vm_map_entry *chain, *endchain, *entry, *orig_entry, *newentry, 2718 *deadentry, *oldentry; 2719 struct vm_map_entry *resentry = NULL; /* a dummy reservation entry */ 2720 vsize_t elen; 2721 int nchain, error, copy_ok; 2722 vsize_t nsize; 2723 UVMHIST_FUNC("uvm_map_extract"); UVMHIST_CALLED(maphist); 2724 2725 UVMHIST_LOG(maphist,"(srcmap=0x%x,start=0x%x, len=0x%x", srcmap, start, 2726 len,0); 2727 UVMHIST_LOG(maphist," ...,dstmap=0x%x, flags=0x%x)", dstmap,flags,0,0); 2728 2729 /* 2730 * step 0: sanity check: start must be on a page boundary, length 2731 * must be page sized. can't ask for CONTIG/QREF if you asked for 2732 * REMOVE. 2733 */ 2734 2735 KASSERT((start & PAGE_MASK) == 0 && (len & PAGE_MASK) == 0); 2736 KASSERT((flags & UVM_EXTRACT_REMOVE) == 0 || 2737 (flags & (UVM_EXTRACT_CONTIG|UVM_EXTRACT_QREF)) == 0); 2738 2739 /* 2740 * step 1: reserve space in the target map for the extracted area 2741 */ 2742 2743 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 2744 dstaddr = vm_map_min(dstmap); 2745 if (!uvm_map_reserve(dstmap, len, start, 0, &dstaddr, 0)) 2746 return (ENOMEM); 2747 *dstaddrp = dstaddr; /* pass address back to caller */ 2748 UVMHIST_LOG(maphist, " dstaddr=0x%x", dstaddr,0,0,0); 2749 } else { 2750 dstaddr = *dstaddrp; 2751 } 2752 2753 /* 2754 * step 2: setup for the extraction process loop by init'ing the 2755 * map entry chain, locking src map, and looking up the first useful 2756 * entry in the map. 2757 */ 2758 2759 end = start + len; 2760 newend = dstaddr + len; 2761 chain = endchain = NULL; 2762 nchain = 0; 2763 nsize = 0; 2764 vm_map_lock(srcmap); 2765 2766 if (uvm_map_lookup_entry(srcmap, start, &entry)) { 2767 2768 /* "start" is within an entry */ 2769 if (flags & UVM_EXTRACT_QREF) { 2770 2771 /* 2772 * for quick references we don't clip the entry, so 2773 * the entry may map space "before" the starting 2774 * virtual address... this is the "fudge" factor 2775 * (which can be non-zero only the first time 2776 * through the "while" loop in step 3). 2777 */ 2778 2779 fudge = start - entry->start; 2780 } else { 2781 2782 /* 2783 * normal reference: we clip the map to fit (thus 2784 * fudge is zero) 2785 */ 2786 2787 UVM_MAP_CLIP_START(srcmap, entry, start, NULL); 2788 SAVE_HINT(srcmap, srcmap->hint, entry->prev); 2789 fudge = 0; 2790 } 2791 } else { 2792 2793 /* "start" is not within an entry ... skip to next entry */ 2794 if (flags & UVM_EXTRACT_CONTIG) { 2795 error = EINVAL; 2796 goto bad; /* definite hole here ... */ 2797 } 2798 2799 entry = entry->next; 2800 fudge = 0; 2801 } 2802 2803 /* save values from srcmap for step 6 */ 2804 orig_entry = entry; 2805 orig_fudge = fudge; 2806 2807 /* 2808 * step 3: now start looping through the map entries, extracting 2809 * as we go. 2810 */ 2811 2812 while (entry->start < end && entry != &srcmap->header) { 2813 2814 /* if we are not doing a quick reference, clip it */ 2815 if ((flags & UVM_EXTRACT_QREF) == 0) 2816 UVM_MAP_CLIP_END(srcmap, entry, end, NULL); 2817 2818 /* clear needs_copy (allow chunking) */ 2819 if (UVM_ET_ISNEEDSCOPY(entry)) { 2820 amap_copy(srcmap, entry, 2821 AMAP_COPY_NOWAIT|AMAP_COPY_NOMERGE, start, end); 2822 if (UVM_ET_ISNEEDSCOPY(entry)) { /* failed? */ 2823 error = ENOMEM; 2824 goto bad; 2825 } 2826 2827 /* amap_copy could clip (during chunk)! update fudge */ 2828 if (fudge) { 2829 fudge = start - entry->start; 2830 orig_fudge = fudge; 2831 } 2832 } 2833 2834 /* calculate the offset of this from "start" */ 2835 oldoffset = (entry->start + fudge) - start; 2836 2837 /* allocate a new map entry */ 2838 newentry = uvm_mapent_alloc(dstmap, 0); 2839 if (newentry == NULL) { 2840 error = ENOMEM; 2841 goto bad; 2842 } 2843 2844 /* set up new map entry */ 2845 newentry->next = NULL; 2846 newentry->prev = endchain; 2847 newentry->start = dstaddr + oldoffset; 2848 newentry->end = 2849 newentry->start + (entry->end - (entry->start + fudge)); 2850 if (newentry->end > newend || newentry->end < newentry->start) 2851 newentry->end = newend; 2852 newentry->object.uvm_obj = entry->object.uvm_obj; 2853 if (newentry->object.uvm_obj) { 2854 if (newentry->object.uvm_obj->pgops->pgo_reference) 2855 newentry->object.uvm_obj->pgops-> 2856 pgo_reference(newentry->object.uvm_obj); 2857 newentry->offset = entry->offset + fudge; 2858 } else { 2859 newentry->offset = 0; 2860 } 2861 newentry->etype = entry->etype; 2862 newentry->protection = (flags & UVM_EXTRACT_FIXPROT) ? 2863 entry->max_protection : entry->protection; 2864 newentry->max_protection = entry->max_protection; 2865 newentry->inheritance = entry->inheritance; 2866 newentry->wired_count = 0; 2867 newentry->aref.ar_amap = entry->aref.ar_amap; 2868 if (newentry->aref.ar_amap) { 2869 newentry->aref.ar_pageoff = 2870 entry->aref.ar_pageoff + (fudge >> PAGE_SHIFT); 2871 uvm_map_reference_amap(newentry, AMAP_SHARED | 2872 ((flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0)); 2873 } else { 2874 newentry->aref.ar_pageoff = 0; 2875 } 2876 newentry->advice = entry->advice; 2877 if ((flags & UVM_EXTRACT_QREF) != 0) { 2878 newentry->flags |= UVM_MAP_NOMERGE; 2879 } 2880 2881 /* now link it on the chain */ 2882 nchain++; 2883 nsize += newentry->end - newentry->start; 2884 if (endchain == NULL) { 2885 chain = endchain = newentry; 2886 } else { 2887 endchain->next = newentry; 2888 endchain = newentry; 2889 } 2890 2891 /* end of 'while' loop! */ 2892 if ((flags & UVM_EXTRACT_CONTIG) && entry->end < end && 2893 (entry->next == &srcmap->header || 2894 entry->next->start != entry->end)) { 2895 error = EINVAL; 2896 goto bad; 2897 } 2898 entry = entry->next; 2899 fudge = 0; 2900 } 2901 2902 /* 2903 * step 4: close off chain (in format expected by uvm_map_replace) 2904 */ 2905 2906 if (chain) 2907 chain->prev = endchain; 2908 2909 /* 2910 * step 5: attempt to lock the dest map so we can pmap_copy. 2911 * note usage of copy_ok: 2912 * 1 => dstmap locked, pmap_copy ok, and we "replace" here (step 5) 2913 * 0 => dstmap unlocked, NO pmap_copy, and we will "replace" in step 7 2914 */ 2915 2916 if (srcmap == dstmap || vm_map_lock_try(dstmap) == true) { 2917 copy_ok = 1; 2918 if (!uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2919 nchain, nsize, &resentry)) { 2920 if (srcmap != dstmap) 2921 vm_map_unlock(dstmap); 2922 error = EIO; 2923 goto bad; 2924 } 2925 } else { 2926 copy_ok = 0; 2927 /* replace defered until step 7 */ 2928 } 2929 2930 /* 2931 * step 6: traverse the srcmap a second time to do the following: 2932 * - if we got a lock on the dstmap do pmap_copy 2933 * - if UVM_EXTRACT_REMOVE remove the entries 2934 * we make use of orig_entry and orig_fudge (saved in step 2) 2935 */ 2936 2937 if (copy_ok || (flags & UVM_EXTRACT_REMOVE)) { 2938 2939 /* purge possible stale hints from srcmap */ 2940 if (flags & UVM_EXTRACT_REMOVE) { 2941 SAVE_HINT(srcmap, srcmap->hint, orig_entry->prev); 2942 if (srcmap->first_free != &srcmap->header && 2943 srcmap->first_free->start >= start) 2944 srcmap->first_free = orig_entry->prev; 2945 } 2946 2947 entry = orig_entry; 2948 fudge = orig_fudge; 2949 deadentry = NULL; /* for UVM_EXTRACT_REMOVE */ 2950 2951 while (entry->start < end && entry != &srcmap->header) { 2952 if (copy_ok) { 2953 oldoffset = (entry->start + fudge) - start; 2954 elen = MIN(end, entry->end) - 2955 (entry->start + fudge); 2956 pmap_copy(dstmap->pmap, srcmap->pmap, 2957 dstaddr + oldoffset, elen, 2958 entry->start + fudge); 2959 } 2960 2961 /* we advance "entry" in the following if statement */ 2962 if (flags & UVM_EXTRACT_REMOVE) { 2963 uvm_map_lock_entry(entry); 2964 pmap_remove(srcmap->pmap, entry->start, 2965 entry->end); 2966 uvm_map_unlock_entry(entry); 2967 oldentry = entry; /* save entry */ 2968 entry = entry->next; /* advance */ 2969 uvm_map_entry_unlink(srcmap, oldentry); 2970 /* add to dead list */ 2971 oldentry->next = deadentry; 2972 deadentry = oldentry; 2973 } else { 2974 entry = entry->next; /* advance */ 2975 } 2976 2977 /* end of 'while' loop */ 2978 fudge = 0; 2979 } 2980 pmap_update(srcmap->pmap); 2981 2982 /* 2983 * unlock dstmap. we will dispose of deadentry in 2984 * step 7 if needed 2985 */ 2986 2987 if (copy_ok && srcmap != dstmap) 2988 vm_map_unlock(dstmap); 2989 2990 } else { 2991 deadentry = NULL; 2992 } 2993 2994 /* 2995 * step 7: we are done with the source map, unlock. if copy_ok 2996 * is 0 then we have not replaced the dummy mapping in dstmap yet 2997 * and we need to do so now. 2998 */ 2999 3000 vm_map_unlock(srcmap); 3001 if ((flags & UVM_EXTRACT_REMOVE) && deadentry) 3002 uvm_unmap_detach(deadentry, 0); /* dispose of old entries */ 3003 3004 /* now do the replacement if we didn't do it in step 5 */ 3005 if (copy_ok == 0) { 3006 vm_map_lock(dstmap); 3007 error = uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 3008 nchain, nsize, &resentry); 3009 vm_map_unlock(dstmap); 3010 3011 if (error == false) { 3012 error = EIO; 3013 goto bad2; 3014 } 3015 } 3016 3017 if (resentry != NULL) 3018 uvm_mapent_free(resentry); 3019 3020 return (0); 3021 3022 /* 3023 * bad: failure recovery 3024 */ 3025 bad: 3026 vm_map_unlock(srcmap); 3027 bad2: /* src already unlocked */ 3028 if (chain) 3029 uvm_unmap_detach(chain, 3030 (flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0); 3031 3032 if (resentry != NULL) 3033 uvm_mapent_free(resentry); 3034 3035 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 3036 uvm_unmap(dstmap, dstaddr, dstaddr+len); /* ??? */ 3037 } 3038 return (error); 3039 } 3040 3041 /* end of extraction functions */ 3042 3043 /* 3044 * uvm_map_submap: punch down part of a map into a submap 3045 * 3046 * => only the kernel_map is allowed to be submapped 3047 * => the purpose of submapping is to break up the locking granularity 3048 * of a larger map 3049 * => the range specified must have been mapped previously with a uvm_map() 3050 * call [with uobj==NULL] to create a blank map entry in the main map. 3051 * [And it had better still be blank!] 3052 * => maps which contain submaps should never be copied or forked. 3053 * => to remove a submap, use uvm_unmap() on the main map 3054 * and then uvm_map_deallocate() the submap. 3055 * => main map must be unlocked. 3056 * => submap must have been init'd and have a zero reference count. 3057 * [need not be locked as we don't actually reference it] 3058 */ 3059 3060 int 3061 uvm_map_submap(struct vm_map *map, vaddr_t start, vaddr_t end, 3062 struct vm_map *submap) 3063 { 3064 struct vm_map_entry *entry; 3065 struct uvm_mapent_reservation umr; 3066 int error; 3067 3068 uvm_mapent_reserve(map, &umr, 2, 0); 3069 3070 vm_map_lock(map); 3071 VM_MAP_RANGE_CHECK(map, start, end); 3072 3073 if (uvm_map_lookup_entry(map, start, &entry)) { 3074 UVM_MAP_CLIP_START(map, entry, start, &umr); 3075 UVM_MAP_CLIP_END(map, entry, end, &umr); /* to be safe */ 3076 } else { 3077 entry = NULL; 3078 } 3079 3080 if (entry != NULL && 3081 entry->start == start && entry->end == end && 3082 entry->object.uvm_obj == NULL && entry->aref.ar_amap == NULL && 3083 !UVM_ET_ISCOPYONWRITE(entry) && !UVM_ET_ISNEEDSCOPY(entry)) { 3084 entry->etype |= UVM_ET_SUBMAP; 3085 entry->object.sub_map = submap; 3086 entry->offset = 0; 3087 uvm_map_reference(submap); 3088 error = 0; 3089 } else { 3090 error = EINVAL; 3091 } 3092 vm_map_unlock(map); 3093 3094 uvm_mapent_unreserve(map, &umr); 3095 3096 return error; 3097 } 3098 3099 /* 3100 * uvm_map_setup_kernel: init in-kernel map 3101 * 3102 * => map must not be in service yet. 3103 */ 3104 3105 void 3106 uvm_map_setup_kernel(struct vm_map_kernel *map, 3107 vaddr_t vmin, vaddr_t vmax, int flags) 3108 { 3109 3110 uvm_map_setup(&map->vmk_map, vmin, vmax, flags); 3111 callback_head_init(&map->vmk_reclaim_callback, IPL_VM); 3112 LIST_INIT(&map->vmk_kentry_free); 3113 map->vmk_merged_entries = NULL; 3114 } 3115 3116 3117 /* 3118 * uvm_map_protect: change map protection 3119 * 3120 * => set_max means set max_protection. 3121 * => map must be unlocked. 3122 */ 3123 3124 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 3125 ~VM_PROT_WRITE : VM_PROT_ALL) 3126 3127 int 3128 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end, 3129 vm_prot_t new_prot, bool set_max) 3130 { 3131 struct vm_map_entry *current, *entry; 3132 int error = 0; 3133 UVMHIST_FUNC("uvm_map_protect"); UVMHIST_CALLED(maphist); 3134 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_prot=0x%x)", 3135 map, start, end, new_prot); 3136 3137 vm_map_lock(map); 3138 VM_MAP_RANGE_CHECK(map, start, end); 3139 if (uvm_map_lookup_entry(map, start, &entry)) { 3140 UVM_MAP_CLIP_START(map, entry, start, NULL); 3141 } else { 3142 entry = entry->next; 3143 } 3144 3145 /* 3146 * make a first pass to check for protection violations. 3147 */ 3148 3149 current = entry; 3150 while ((current != &map->header) && (current->start < end)) { 3151 if (UVM_ET_ISSUBMAP(current)) { 3152 error = EINVAL; 3153 goto out; 3154 } 3155 if ((new_prot & current->max_protection) != new_prot) { 3156 error = EACCES; 3157 goto out; 3158 } 3159 /* 3160 * Don't allow VM_PROT_EXECUTE to be set on entries that 3161 * point to vnodes that are associated with a NOEXEC file 3162 * system. 3163 */ 3164 if (UVM_ET_ISOBJ(current) && 3165 UVM_OBJ_IS_VNODE(current->object.uvm_obj)) { 3166 struct vnode *vp = 3167 (struct vnode *) current->object.uvm_obj; 3168 3169 if ((new_prot & VM_PROT_EXECUTE) != 0 && 3170 (vp->v_mount->mnt_flag & MNT_NOEXEC) != 0) { 3171 error = EACCES; 3172 goto out; 3173 } 3174 } 3175 3176 current = current->next; 3177 } 3178 3179 /* go back and fix up protections (no need to clip this time). */ 3180 3181 current = entry; 3182 while ((current != &map->header) && (current->start < end)) { 3183 vm_prot_t old_prot; 3184 3185 UVM_MAP_CLIP_END(map, current, end, NULL); 3186 old_prot = current->protection; 3187 if (set_max) 3188 current->protection = 3189 (current->max_protection = new_prot) & old_prot; 3190 else 3191 current->protection = new_prot; 3192 3193 /* 3194 * update physical map if necessary. worry about copy-on-write 3195 * here -- CHECK THIS XXX 3196 */ 3197 3198 if (current->protection != old_prot) { 3199 /* update pmap! */ 3200 uvm_map_lock_entry(current); 3201 pmap_protect(map->pmap, current->start, current->end, 3202 current->protection & MASK(entry)); 3203 uvm_map_unlock_entry(current); 3204 3205 /* 3206 * If this entry points at a vnode, and the 3207 * protection includes VM_PROT_EXECUTE, mark 3208 * the vnode as VEXECMAP. 3209 */ 3210 if (UVM_ET_ISOBJ(current)) { 3211 struct uvm_object *uobj = 3212 current->object.uvm_obj; 3213 3214 if (UVM_OBJ_IS_VNODE(uobj) && 3215 (current->protection & VM_PROT_EXECUTE)) { 3216 vn_markexec((struct vnode *) uobj); 3217 } 3218 } 3219 } 3220 3221 /* 3222 * If the map is configured to lock any future mappings, 3223 * wire this entry now if the old protection was VM_PROT_NONE 3224 * and the new protection is not VM_PROT_NONE. 3225 */ 3226 3227 if ((map->flags & VM_MAP_WIREFUTURE) != 0 && 3228 VM_MAPENT_ISWIRED(entry) == 0 && 3229 old_prot == VM_PROT_NONE && 3230 new_prot != VM_PROT_NONE) { 3231 if (uvm_map_pageable(map, entry->start, 3232 entry->end, false, 3233 UVM_LK_ENTER|UVM_LK_EXIT) != 0) { 3234 3235 /* 3236 * If locking the entry fails, remember the 3237 * error if it's the first one. Note we 3238 * still continue setting the protection in 3239 * the map, but will return the error 3240 * condition regardless. 3241 * 3242 * XXX Ignore what the actual error is, 3243 * XXX just call it a resource shortage 3244 * XXX so that it doesn't get confused 3245 * XXX what uvm_map_protect() itself would 3246 * XXX normally return. 3247 */ 3248 3249 error = ENOMEM; 3250 } 3251 } 3252 current = current->next; 3253 } 3254 pmap_update(map->pmap); 3255 3256 out: 3257 vm_map_unlock(map); 3258 3259 UVMHIST_LOG(maphist, "<- done, error=%d",error,0,0,0); 3260 return error; 3261 } 3262 3263 #undef MASK 3264 3265 /* 3266 * uvm_map_inherit: set inheritance code for range of addrs in map. 3267 * 3268 * => map must be unlocked 3269 * => note that the inherit code is used during a "fork". see fork 3270 * code for details. 3271 */ 3272 3273 int 3274 uvm_map_inherit(struct vm_map *map, vaddr_t start, vaddr_t end, 3275 vm_inherit_t new_inheritance) 3276 { 3277 struct vm_map_entry *entry, *temp_entry; 3278 UVMHIST_FUNC("uvm_map_inherit"); UVMHIST_CALLED(maphist); 3279 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_inh=0x%x)", 3280 map, start, end, new_inheritance); 3281 3282 switch (new_inheritance) { 3283 case MAP_INHERIT_NONE: 3284 case MAP_INHERIT_COPY: 3285 case MAP_INHERIT_SHARE: 3286 break; 3287 default: 3288 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3289 return EINVAL; 3290 } 3291 3292 vm_map_lock(map); 3293 VM_MAP_RANGE_CHECK(map, start, end); 3294 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3295 entry = temp_entry; 3296 UVM_MAP_CLIP_START(map, entry, start, NULL); 3297 } else { 3298 entry = temp_entry->next; 3299 } 3300 while ((entry != &map->header) && (entry->start < end)) { 3301 UVM_MAP_CLIP_END(map, entry, end, NULL); 3302 entry->inheritance = new_inheritance; 3303 entry = entry->next; 3304 } 3305 vm_map_unlock(map); 3306 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3307 return 0; 3308 } 3309 3310 /* 3311 * uvm_map_advice: set advice code for range of addrs in map. 3312 * 3313 * => map must be unlocked 3314 */ 3315 3316 int 3317 uvm_map_advice(struct vm_map *map, vaddr_t start, vaddr_t end, int new_advice) 3318 { 3319 struct vm_map_entry *entry, *temp_entry; 3320 UVMHIST_FUNC("uvm_map_advice"); UVMHIST_CALLED(maphist); 3321 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_adv=0x%x)", 3322 map, start, end, new_advice); 3323 3324 vm_map_lock(map); 3325 VM_MAP_RANGE_CHECK(map, start, end); 3326 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3327 entry = temp_entry; 3328 UVM_MAP_CLIP_START(map, entry, start, NULL); 3329 } else { 3330 entry = temp_entry->next; 3331 } 3332 3333 /* 3334 * XXXJRT: disallow holes? 3335 */ 3336 3337 while ((entry != &map->header) && (entry->start < end)) { 3338 UVM_MAP_CLIP_END(map, entry, end, NULL); 3339 3340 switch (new_advice) { 3341 case MADV_NORMAL: 3342 case MADV_RANDOM: 3343 case MADV_SEQUENTIAL: 3344 /* nothing special here */ 3345 break; 3346 3347 default: 3348 vm_map_unlock(map); 3349 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3350 return EINVAL; 3351 } 3352 entry->advice = new_advice; 3353 entry = entry->next; 3354 } 3355 3356 vm_map_unlock(map); 3357 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3358 return 0; 3359 } 3360 3361 /* 3362 * uvm_map_willneed: apply MADV_WILLNEED 3363 */ 3364 3365 int 3366 uvm_map_willneed(struct vm_map *map, vaddr_t start, vaddr_t end) 3367 { 3368 struct vm_map_entry *entry; 3369 UVMHIST_FUNC("uvm_map_willneed"); UVMHIST_CALLED(maphist); 3370 UVMHIST_LOG(maphist,"(map=0x%lx,start=0x%lx,end=0x%lx)", 3371 map, start, end, 0); 3372 3373 vm_map_lock_read(map); 3374 VM_MAP_RANGE_CHECK(map, start, end); 3375 if (!uvm_map_lookup_entry(map, start, &entry)) { 3376 entry = entry->next; 3377 } 3378 while (entry->start < end) { 3379 struct vm_amap * const amap = entry->aref.ar_amap; 3380 struct uvm_object * const uobj = entry->object.uvm_obj; 3381 3382 KASSERT(entry != &map->header); 3383 KASSERT(start < entry->end); 3384 /* 3385 * For now, we handle only the easy but commonly-requested case. 3386 * ie. start prefetching of backing uobj pages. 3387 * 3388 * XXX It might be useful to pmap_enter() the already-in-core 3389 * pages by inventing a "weak" mode for uvm_fault() which would 3390 * only do the PGO_LOCKED pgo_get(). 3391 */ 3392 if (UVM_ET_ISOBJ(entry) && amap == NULL && uobj != NULL) { 3393 off_t offset; 3394 off_t size; 3395 3396 offset = entry->offset; 3397 if (start < entry->start) { 3398 offset += entry->start - start; 3399 } 3400 size = entry->offset + (entry->end - entry->start); 3401 if (entry->end < end) { 3402 size -= end - entry->end; 3403 } 3404 uvm_readahead(uobj, offset, size); 3405 } 3406 entry = entry->next; 3407 } 3408 vm_map_unlock_read(map); 3409 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3410 return 0; 3411 } 3412 3413 /* 3414 * uvm_map_pageable: sets the pageability of a range in a map. 3415 * 3416 * => wires map entries. should not be used for transient page locking. 3417 * for that, use uvm_fault_wire()/uvm_fault_unwire() (see uvm_vslock()). 3418 * => regions specified as not pageable require lock-down (wired) memory 3419 * and page tables. 3420 * => map must never be read-locked 3421 * => if islocked is true, map is already write-locked 3422 * => we always unlock the map, since we must downgrade to a read-lock 3423 * to call uvm_fault_wire() 3424 * => XXXCDC: check this and try and clean it up. 3425 */ 3426 3427 int 3428 uvm_map_pageable(struct vm_map *map, vaddr_t start, vaddr_t end, 3429 bool new_pageable, int lockflags) 3430 { 3431 struct vm_map_entry *entry, *start_entry, *failed_entry; 3432 int rv; 3433 #ifdef DIAGNOSTIC 3434 u_int timestamp_save; 3435 #endif 3436 UVMHIST_FUNC("uvm_map_pageable"); UVMHIST_CALLED(maphist); 3437 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_pageable=0x%x)", 3438 map, start, end, new_pageable); 3439 KASSERT(map->flags & VM_MAP_PAGEABLE); 3440 3441 if ((lockflags & UVM_LK_ENTER) == 0) 3442 vm_map_lock(map); 3443 VM_MAP_RANGE_CHECK(map, start, end); 3444 3445 /* 3446 * only one pageability change may take place at one time, since 3447 * uvm_fault_wire assumes it will be called only once for each 3448 * wiring/unwiring. therefore, we have to make sure we're actually 3449 * changing the pageability for the entire region. we do so before 3450 * making any changes. 3451 */ 3452 3453 if (uvm_map_lookup_entry(map, start, &start_entry) == false) { 3454 if ((lockflags & UVM_LK_EXIT) == 0) 3455 vm_map_unlock(map); 3456 3457 UVMHIST_LOG(maphist,"<- done (fault)",0,0,0,0); 3458 return EFAULT; 3459 } 3460 entry = start_entry; 3461 3462 /* 3463 * handle wiring and unwiring separately. 3464 */ 3465 3466 if (new_pageable) { /* unwire */ 3467 UVM_MAP_CLIP_START(map, entry, start, NULL); 3468 3469 /* 3470 * unwiring. first ensure that the range to be unwired is 3471 * really wired down and that there are no holes. 3472 */ 3473 3474 while ((entry != &map->header) && (entry->start < end)) { 3475 if (entry->wired_count == 0 || 3476 (entry->end < end && 3477 (entry->next == &map->header || 3478 entry->next->start > entry->end))) { 3479 if ((lockflags & UVM_LK_EXIT) == 0) 3480 vm_map_unlock(map); 3481 UVMHIST_LOG(maphist, "<- done (INVAL)",0,0,0,0); 3482 return EINVAL; 3483 } 3484 entry = entry->next; 3485 } 3486 3487 /* 3488 * POSIX 1003.1b - a single munlock call unlocks a region, 3489 * regardless of the number of mlock calls made on that 3490 * region. 3491 */ 3492 3493 entry = start_entry; 3494 while ((entry != &map->header) && (entry->start < end)) { 3495 UVM_MAP_CLIP_END(map, entry, end, NULL); 3496 if (VM_MAPENT_ISWIRED(entry)) 3497 uvm_map_entry_unwire(map, entry); 3498 entry = entry->next; 3499 } 3500 if ((lockflags & UVM_LK_EXIT) == 0) 3501 vm_map_unlock(map); 3502 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3503 return 0; 3504 } 3505 3506 /* 3507 * wire case: in two passes [XXXCDC: ugly block of code here] 3508 * 3509 * 1: holding the write lock, we create any anonymous maps that need 3510 * to be created. then we clip each map entry to the region to 3511 * be wired and increment its wiring count. 3512 * 3513 * 2: we downgrade to a read lock, and call uvm_fault_wire to fault 3514 * in the pages for any newly wired area (wired_count == 1). 3515 * 3516 * downgrading to a read lock for uvm_fault_wire avoids a possible 3517 * deadlock with another thread that may have faulted on one of 3518 * the pages to be wired (it would mark the page busy, blocking 3519 * us, then in turn block on the map lock that we hold). because 3520 * of problems in the recursive lock package, we cannot upgrade 3521 * to a write lock in vm_map_lookup. thus, any actions that 3522 * require the write lock must be done beforehand. because we 3523 * keep the read lock on the map, the copy-on-write status of the 3524 * entries we modify here cannot change. 3525 */ 3526 3527 while ((entry != &map->header) && (entry->start < end)) { 3528 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3529 3530 /* 3531 * perform actions of vm_map_lookup that need the 3532 * write lock on the map: create an anonymous map 3533 * for a copy-on-write region, or an anonymous map 3534 * for a zero-fill region. (XXXCDC: submap case 3535 * ok?) 3536 */ 3537 3538 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3539 if (UVM_ET_ISNEEDSCOPY(entry) && 3540 ((entry->max_protection & VM_PROT_WRITE) || 3541 (entry->object.uvm_obj == NULL))) { 3542 amap_copy(map, entry, 0, start, end); 3543 /* XXXCDC: wait OK? */ 3544 } 3545 } 3546 } 3547 UVM_MAP_CLIP_START(map, entry, start, NULL); 3548 UVM_MAP_CLIP_END(map, entry, end, NULL); 3549 entry->wired_count++; 3550 3551 /* 3552 * Check for holes 3553 */ 3554 3555 if (entry->protection == VM_PROT_NONE || 3556 (entry->end < end && 3557 (entry->next == &map->header || 3558 entry->next->start > entry->end))) { 3559 3560 /* 3561 * found one. amap creation actions do not need to 3562 * be undone, but the wired counts need to be restored. 3563 */ 3564 3565 while (entry != &map->header && entry->end > start) { 3566 entry->wired_count--; 3567 entry = entry->prev; 3568 } 3569 if ((lockflags & UVM_LK_EXIT) == 0) 3570 vm_map_unlock(map); 3571 UVMHIST_LOG(maphist,"<- done (INVALID WIRE)",0,0,0,0); 3572 return EINVAL; 3573 } 3574 entry = entry->next; 3575 } 3576 3577 /* 3578 * Pass 2. 3579 */ 3580 3581 #ifdef DIAGNOSTIC 3582 timestamp_save = map->timestamp; 3583 #endif 3584 vm_map_busy(map); 3585 vm_map_unlock(map); 3586 3587 rv = 0; 3588 entry = start_entry; 3589 while (entry != &map->header && entry->start < end) { 3590 if (entry->wired_count == 1) { 3591 rv = uvm_fault_wire(map, entry->start, entry->end, 3592 entry->max_protection, 1); 3593 if (rv) { 3594 3595 /* 3596 * wiring failed. break out of the loop. 3597 * we'll clean up the map below, once we 3598 * have a write lock again. 3599 */ 3600 3601 break; 3602 } 3603 } 3604 entry = entry->next; 3605 } 3606 3607 if (rv) { /* failed? */ 3608 3609 /* 3610 * Get back to an exclusive (write) lock. 3611 */ 3612 3613 vm_map_lock(map); 3614 vm_map_unbusy(map); 3615 3616 #ifdef DIAGNOSTIC 3617 if (timestamp_save + 1 != map->timestamp) 3618 panic("uvm_map_pageable: stale map"); 3619 #endif 3620 3621 /* 3622 * first drop the wiring count on all the entries 3623 * which haven't actually been wired yet. 3624 */ 3625 3626 failed_entry = entry; 3627 while (entry != &map->header && entry->start < end) { 3628 entry->wired_count--; 3629 entry = entry->next; 3630 } 3631 3632 /* 3633 * now, unwire all the entries that were successfully 3634 * wired above. 3635 */ 3636 3637 entry = start_entry; 3638 while (entry != failed_entry) { 3639 entry->wired_count--; 3640 if (VM_MAPENT_ISWIRED(entry) == 0) 3641 uvm_map_entry_unwire(map, entry); 3642 entry = entry->next; 3643 } 3644 if ((lockflags & UVM_LK_EXIT) == 0) 3645 vm_map_unlock(map); 3646 UVMHIST_LOG(maphist, "<- done (RV=%d)", rv,0,0,0); 3647 return (rv); 3648 } 3649 3650 if ((lockflags & UVM_LK_EXIT) == 0) { 3651 vm_map_unbusy(map); 3652 } else { 3653 3654 /* 3655 * Get back to an exclusive (write) lock. 3656 */ 3657 3658 vm_map_lock(map); 3659 vm_map_unbusy(map); 3660 } 3661 3662 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3663 return 0; 3664 } 3665 3666 /* 3667 * uvm_map_pageable_all: special case of uvm_map_pageable - affects 3668 * all mapped regions. 3669 * 3670 * => map must not be locked. 3671 * => if no flags are specified, all regions are unwired. 3672 * => XXXJRT: has some of the same problems as uvm_map_pageable() above. 3673 */ 3674 3675 int 3676 uvm_map_pageable_all(struct vm_map *map, int flags, vsize_t limit) 3677 { 3678 struct vm_map_entry *entry, *failed_entry; 3679 vsize_t size; 3680 int rv; 3681 #ifdef DIAGNOSTIC 3682 u_int timestamp_save; 3683 #endif 3684 UVMHIST_FUNC("uvm_map_pageable_all"); UVMHIST_CALLED(maphist); 3685 UVMHIST_LOG(maphist,"(map=0x%x,flags=0x%x)", map, flags, 0, 0); 3686 3687 KASSERT(map->flags & VM_MAP_PAGEABLE); 3688 3689 vm_map_lock(map); 3690 3691 /* 3692 * handle wiring and unwiring separately. 3693 */ 3694 3695 if (flags == 0) { /* unwire */ 3696 3697 /* 3698 * POSIX 1003.1b -- munlockall unlocks all regions, 3699 * regardless of how many times mlockall has been called. 3700 */ 3701 3702 for (entry = map->header.next; entry != &map->header; 3703 entry = entry->next) { 3704 if (VM_MAPENT_ISWIRED(entry)) 3705 uvm_map_entry_unwire(map, entry); 3706 } 3707 map->flags &= ~VM_MAP_WIREFUTURE; 3708 vm_map_unlock(map); 3709 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3710 return 0; 3711 } 3712 3713 if (flags & MCL_FUTURE) { 3714 3715 /* 3716 * must wire all future mappings; remember this. 3717 */ 3718 3719 map->flags |= VM_MAP_WIREFUTURE; 3720 } 3721 3722 if ((flags & MCL_CURRENT) == 0) { 3723 3724 /* 3725 * no more work to do! 3726 */ 3727 3728 UVMHIST_LOG(maphist,"<- done (OK no wire)",0,0,0,0); 3729 vm_map_unlock(map); 3730 return 0; 3731 } 3732 3733 /* 3734 * wire case: in three passes [XXXCDC: ugly block of code here] 3735 * 3736 * 1: holding the write lock, count all pages mapped by non-wired 3737 * entries. if this would cause us to go over our limit, we fail. 3738 * 3739 * 2: still holding the write lock, we create any anonymous maps that 3740 * need to be created. then we increment its wiring count. 3741 * 3742 * 3: we downgrade to a read lock, and call uvm_fault_wire to fault 3743 * in the pages for any newly wired area (wired_count == 1). 3744 * 3745 * downgrading to a read lock for uvm_fault_wire avoids a possible 3746 * deadlock with another thread that may have faulted on one of 3747 * the pages to be wired (it would mark the page busy, blocking 3748 * us, then in turn block on the map lock that we hold). because 3749 * of problems in the recursive lock package, we cannot upgrade 3750 * to a write lock in vm_map_lookup. thus, any actions that 3751 * require the write lock must be done beforehand. because we 3752 * keep the read lock on the map, the copy-on-write status of the 3753 * entries we modify here cannot change. 3754 */ 3755 3756 for (size = 0, entry = map->header.next; entry != &map->header; 3757 entry = entry->next) { 3758 if (entry->protection != VM_PROT_NONE && 3759 VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3760 size += entry->end - entry->start; 3761 } 3762 } 3763 3764 if (atop(size) + uvmexp.wired > uvmexp.wiredmax) { 3765 vm_map_unlock(map); 3766 return ENOMEM; 3767 } 3768 3769 if (limit != 0 && 3770 (size + ptoa(pmap_wired_count(vm_map_pmap(map))) > limit)) { 3771 vm_map_unlock(map); 3772 return ENOMEM; 3773 } 3774 3775 /* 3776 * Pass 2. 3777 */ 3778 3779 for (entry = map->header.next; entry != &map->header; 3780 entry = entry->next) { 3781 if (entry->protection == VM_PROT_NONE) 3782 continue; 3783 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3784 3785 /* 3786 * perform actions of vm_map_lookup that need the 3787 * write lock on the map: create an anonymous map 3788 * for a copy-on-write region, or an anonymous map 3789 * for a zero-fill region. (XXXCDC: submap case 3790 * ok?) 3791 */ 3792 3793 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3794 if (UVM_ET_ISNEEDSCOPY(entry) && 3795 ((entry->max_protection & VM_PROT_WRITE) || 3796 (entry->object.uvm_obj == NULL))) { 3797 amap_copy(map, entry, 0, entry->start, 3798 entry->end); 3799 /* XXXCDC: wait OK? */ 3800 } 3801 } 3802 } 3803 entry->wired_count++; 3804 } 3805 3806 /* 3807 * Pass 3. 3808 */ 3809 3810 #ifdef DIAGNOSTIC 3811 timestamp_save = map->timestamp; 3812 #endif 3813 vm_map_busy(map); 3814 vm_map_unlock(map); 3815 3816 rv = 0; 3817 for (entry = map->header.next; entry != &map->header; 3818 entry = entry->next) { 3819 if (entry->wired_count == 1) { 3820 rv = uvm_fault_wire(map, entry->start, entry->end, 3821 entry->max_protection, 1); 3822 if (rv) { 3823 3824 /* 3825 * wiring failed. break out of the loop. 3826 * we'll clean up the map below, once we 3827 * have a write lock again. 3828 */ 3829 3830 break; 3831 } 3832 } 3833 } 3834 3835 if (rv) { 3836 3837 /* 3838 * Get back an exclusive (write) lock. 3839 */ 3840 3841 vm_map_lock(map); 3842 vm_map_unbusy(map); 3843 3844 #ifdef DIAGNOSTIC 3845 if (timestamp_save + 1 != map->timestamp) 3846 panic("uvm_map_pageable_all: stale map"); 3847 #endif 3848 3849 /* 3850 * first drop the wiring count on all the entries 3851 * which haven't actually been wired yet. 3852 * 3853 * Skip VM_PROT_NONE entries like we did above. 3854 */ 3855 3856 failed_entry = entry; 3857 for (/* nothing */; entry != &map->header; 3858 entry = entry->next) { 3859 if (entry->protection == VM_PROT_NONE) 3860 continue; 3861 entry->wired_count--; 3862 } 3863 3864 /* 3865 * now, unwire all the entries that were successfully 3866 * wired above. 3867 * 3868 * Skip VM_PROT_NONE entries like we did above. 3869 */ 3870 3871 for (entry = map->header.next; entry != failed_entry; 3872 entry = entry->next) { 3873 if (entry->protection == VM_PROT_NONE) 3874 continue; 3875 entry->wired_count--; 3876 if (VM_MAPENT_ISWIRED(entry)) 3877 uvm_map_entry_unwire(map, entry); 3878 } 3879 vm_map_unlock(map); 3880 UVMHIST_LOG(maphist,"<- done (RV=%d)", rv,0,0,0); 3881 return (rv); 3882 } 3883 3884 vm_map_unbusy(map); 3885 3886 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3887 return 0; 3888 } 3889 3890 /* 3891 * uvm_map_clean: clean out a map range 3892 * 3893 * => valid flags: 3894 * if (flags & PGO_CLEANIT): dirty pages are cleaned first 3895 * if (flags & PGO_SYNCIO): dirty pages are written synchronously 3896 * if (flags & PGO_DEACTIVATE): any cached pages are deactivated after clean 3897 * if (flags & PGO_FREE): any cached pages are freed after clean 3898 * => returns an error if any part of the specified range isn't mapped 3899 * => never a need to flush amap layer since the anonymous memory has 3900 * no permanent home, but may deactivate pages there 3901 * => called from sys_msync() and sys_madvise() 3902 * => caller must not write-lock map (read OK). 3903 * => we may sleep while cleaning if SYNCIO [with map read-locked] 3904 */ 3905 3906 int 3907 uvm_map_clean(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 3908 { 3909 struct vm_map_entry *current, *entry; 3910 struct uvm_object *uobj; 3911 struct vm_amap *amap; 3912 struct vm_anon *anon, *anon_tofree; 3913 struct vm_page *pg; 3914 vaddr_t offset; 3915 vsize_t size; 3916 voff_t uoff; 3917 int error, refs; 3918 UVMHIST_FUNC("uvm_map_clean"); UVMHIST_CALLED(maphist); 3919 3920 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,flags=0x%x)", 3921 map, start, end, flags); 3922 KASSERT((flags & (PGO_FREE|PGO_DEACTIVATE)) != 3923 (PGO_FREE|PGO_DEACTIVATE)); 3924 3925 vm_map_lock_read(map); 3926 VM_MAP_RANGE_CHECK(map, start, end); 3927 if (uvm_map_lookup_entry(map, start, &entry) == false) { 3928 vm_map_unlock_read(map); 3929 return EFAULT; 3930 } 3931 3932 /* 3933 * Make a first pass to check for holes and wiring problems. 3934 */ 3935 3936 for (current = entry; current->start < end; current = current->next) { 3937 if (UVM_ET_ISSUBMAP(current)) { 3938 vm_map_unlock_read(map); 3939 return EINVAL; 3940 } 3941 if ((flags & PGO_FREE) != 0 && VM_MAPENT_ISWIRED(entry)) { 3942 vm_map_unlock_read(map); 3943 return EBUSY; 3944 } 3945 if (end <= current->end) { 3946 break; 3947 } 3948 if (current->end != current->next->start) { 3949 vm_map_unlock_read(map); 3950 return EFAULT; 3951 } 3952 } 3953 3954 error = 0; 3955 for (current = entry; start < end; current = current->next) { 3956 amap = current->aref.ar_amap; /* upper layer */ 3957 uobj = current->object.uvm_obj; /* lower layer */ 3958 KASSERT(start >= current->start); 3959 3960 /* 3961 * No amap cleaning necessary if: 3962 * 3963 * (1) There's no amap. 3964 * 3965 * (2) We're not deactivating or freeing pages. 3966 */ 3967 3968 if (amap == NULL || (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) 3969 goto flush_object; 3970 3971 offset = start - current->start; 3972 size = MIN(end, current->end) - start; 3973 anon_tofree = NULL; 3974 3975 amap_lock(amap); 3976 for ( ; size != 0; size -= PAGE_SIZE, offset += PAGE_SIZE) { 3977 anon = amap_lookup(¤t->aref, offset); 3978 if (anon == NULL) 3979 continue; 3980 3981 KASSERT(anon->an_lock == amap->am_lock); 3982 pg = anon->an_page; 3983 if (pg == NULL) { 3984 continue; 3985 } 3986 3987 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 3988 3989 /* 3990 * In these first 3 cases, we just deactivate the page. 3991 */ 3992 3993 case PGO_CLEANIT|PGO_FREE: 3994 case PGO_CLEANIT|PGO_DEACTIVATE: 3995 case PGO_DEACTIVATE: 3996 deactivate_it: 3997 /* 3998 * skip the page if it's loaned or wired, 3999 * since it shouldn't be on a paging queue 4000 * at all in these cases. 4001 */ 4002 4003 mutex_enter(&uvm_pageqlock); 4004 if (pg->loan_count != 0 || 4005 pg->wire_count != 0) { 4006 mutex_exit(&uvm_pageqlock); 4007 continue; 4008 } 4009 KASSERT(pg->uanon == anon); 4010 uvm_pagedeactivate(pg); 4011 mutex_exit(&uvm_pageqlock); 4012 continue; 4013 4014 case PGO_FREE: 4015 4016 /* 4017 * If there are multiple references to 4018 * the amap, just deactivate the page. 4019 */ 4020 4021 if (amap_refs(amap) > 1) 4022 goto deactivate_it; 4023 4024 /* skip the page if it's wired */ 4025 if (pg->wire_count != 0) { 4026 continue; 4027 } 4028 amap_unadd(¤t->aref, offset); 4029 refs = --anon->an_ref; 4030 if (refs == 0) { 4031 anon->an_link = anon_tofree; 4032 anon_tofree = anon; 4033 } 4034 continue; 4035 } 4036 } 4037 uvm_anon_freelst(amap, anon_tofree); 4038 4039 flush_object: 4040 /* 4041 * flush pages if we've got a valid backing object. 4042 * note that we must always clean object pages before 4043 * freeing them since otherwise we could reveal stale 4044 * data from files. 4045 */ 4046 4047 uoff = current->offset + (start - current->start); 4048 size = MIN(end, current->end) - start; 4049 if (uobj != NULL) { 4050 mutex_enter(uobj->vmobjlock); 4051 if (uobj->pgops->pgo_put != NULL) 4052 error = (uobj->pgops->pgo_put)(uobj, uoff, 4053 uoff + size, flags | PGO_CLEANIT); 4054 else 4055 error = 0; 4056 } 4057 start += size; 4058 } 4059 vm_map_unlock_read(map); 4060 return (error); 4061 } 4062 4063 4064 /* 4065 * uvm_map_checkprot: check protection in map 4066 * 4067 * => must allow specified protection in a fully allocated region. 4068 * => map must be read or write locked by caller. 4069 */ 4070 4071 bool 4072 uvm_map_checkprot(struct vm_map *map, vaddr_t start, vaddr_t end, 4073 vm_prot_t protection) 4074 { 4075 struct vm_map_entry *entry; 4076 struct vm_map_entry *tmp_entry; 4077 4078 if (!uvm_map_lookup_entry(map, start, &tmp_entry)) { 4079 return (false); 4080 } 4081 entry = tmp_entry; 4082 while (start < end) { 4083 if (entry == &map->header) { 4084 return (false); 4085 } 4086 4087 /* 4088 * no holes allowed 4089 */ 4090 4091 if (start < entry->start) { 4092 return (false); 4093 } 4094 4095 /* 4096 * check protection associated with entry 4097 */ 4098 4099 if ((entry->protection & protection) != protection) { 4100 return (false); 4101 } 4102 start = entry->end; 4103 entry = entry->next; 4104 } 4105 return (true); 4106 } 4107 4108 /* 4109 * uvmspace_alloc: allocate a vmspace structure. 4110 * 4111 * - structure includes vm_map and pmap 4112 * - XXX: no locking on this structure 4113 * - refcnt set to 1, rest must be init'd by caller 4114 */ 4115 struct vmspace * 4116 uvmspace_alloc(vaddr_t vmin, vaddr_t vmax) 4117 { 4118 struct vmspace *vm; 4119 UVMHIST_FUNC("uvmspace_alloc"); UVMHIST_CALLED(maphist); 4120 4121 vm = pool_cache_get(&uvm_vmspace_cache, PR_WAITOK); 4122 uvmspace_init(vm, NULL, vmin, vmax); 4123 UVMHIST_LOG(maphist,"<- done (vm=0x%x)", vm,0,0,0); 4124 return (vm); 4125 } 4126 4127 /* 4128 * uvmspace_init: initialize a vmspace structure. 4129 * 4130 * - XXX: no locking on this structure 4131 * - refcnt set to 1, rest must be init'd by caller 4132 */ 4133 void 4134 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t vmin, vaddr_t vmax) 4135 { 4136 UVMHIST_FUNC("uvmspace_init"); UVMHIST_CALLED(maphist); 4137 4138 memset(vm, 0, sizeof(*vm)); 4139 uvm_map_setup(&vm->vm_map, vmin, vmax, VM_MAP_PAGEABLE 4140 #ifdef __USING_TOPDOWN_VM 4141 | VM_MAP_TOPDOWN 4142 #endif 4143 ); 4144 if (pmap) 4145 pmap_reference(pmap); 4146 else 4147 pmap = pmap_create(); 4148 vm->vm_map.pmap = pmap; 4149 vm->vm_refcnt = 1; 4150 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4151 } 4152 4153 /* 4154 * uvmspace_share: share a vmspace between two processes 4155 * 4156 * - used for vfork, threads(?) 4157 */ 4158 4159 void 4160 uvmspace_share(struct proc *p1, struct proc *p2) 4161 { 4162 4163 uvmspace_addref(p1->p_vmspace); 4164 p2->p_vmspace = p1->p_vmspace; 4165 } 4166 4167 #if 0 4168 4169 /* 4170 * uvmspace_unshare: ensure that process "p" has its own, unshared, vmspace 4171 * 4172 * - XXX: no locking on vmspace 4173 */ 4174 4175 void 4176 uvmspace_unshare(struct lwp *l) 4177 { 4178 struct proc *p = l->l_proc; 4179 struct vmspace *nvm, *ovm = p->p_vmspace; 4180 4181 if (ovm->vm_refcnt == 1) 4182 /* nothing to do: vmspace isn't shared in the first place */ 4183 return; 4184 4185 /* make a new vmspace, still holding old one */ 4186 nvm = uvmspace_fork(ovm); 4187 4188 kpreempt_disable(); 4189 pmap_deactivate(l); /* unbind old vmspace */ 4190 p->p_vmspace = nvm; 4191 pmap_activate(l); /* switch to new vmspace */ 4192 kpreempt_enable(); 4193 4194 uvmspace_free(ovm); /* drop reference to old vmspace */ 4195 } 4196 4197 #endif 4198 4199 /* 4200 * uvmspace_exec: the process wants to exec a new program 4201 */ 4202 4203 void 4204 uvmspace_exec(struct lwp *l, vaddr_t start, vaddr_t end) 4205 { 4206 struct proc *p = l->l_proc; 4207 struct vmspace *nvm, *ovm = p->p_vmspace; 4208 struct vm_map *map; 4209 4210 #ifdef __HAVE_CPU_VMSPACE_EXEC 4211 cpu_vmspace_exec(l, start, end); 4212 #endif 4213 4214 /* 4215 * Special case: no vmspace yet (see posix_spawn) - 4216 * no races possible in this case. 4217 */ 4218 if (ovm == NULL) { 4219 p->p_vmspace = uvmspace_alloc(start, end); 4220 pmap_activate(l); 4221 return; 4222 } 4223 4224 map = &ovm->vm_map; 4225 /* 4226 * see if more than one process is using this vmspace... 4227 */ 4228 4229 if (ovm->vm_refcnt == 1) { 4230 4231 /* 4232 * if p is the only process using its vmspace then we can safely 4233 * recycle that vmspace for the program that is being exec'd. 4234 */ 4235 4236 #ifdef SYSVSHM 4237 /* 4238 * SYSV SHM semantics require us to kill all segments on an exec 4239 */ 4240 4241 if (ovm->vm_shm) 4242 shmexit(ovm); 4243 #endif 4244 4245 /* 4246 * POSIX 1003.1b -- "lock future mappings" is revoked 4247 * when a process execs another program image. 4248 */ 4249 4250 map->flags &= ~VM_MAP_WIREFUTURE; 4251 4252 /* 4253 * now unmap the old program 4254 */ 4255 4256 pmap_remove_all(map->pmap); 4257 uvm_unmap(map, vm_map_min(map), vm_map_max(map)); 4258 KASSERT(map->header.prev == &map->header); 4259 KASSERT(map->nentries == 0); 4260 4261 /* 4262 * resize the map 4263 */ 4264 4265 vm_map_setmin(map, start); 4266 vm_map_setmax(map, end); 4267 } else { 4268 4269 /* 4270 * p's vmspace is being shared, so we can't reuse it for p since 4271 * it is still being used for others. allocate a new vmspace 4272 * for p 4273 */ 4274 4275 nvm = uvmspace_alloc(start, end); 4276 4277 /* 4278 * install new vmspace and drop our ref to the old one. 4279 */ 4280 4281 kpreempt_disable(); 4282 pmap_deactivate(l); 4283 p->p_vmspace = nvm; 4284 pmap_activate(l); 4285 kpreempt_enable(); 4286 4287 uvmspace_free(ovm); 4288 } 4289 } 4290 4291 /* 4292 * uvmspace_addref: add a referece to a vmspace. 4293 */ 4294 4295 void 4296 uvmspace_addref(struct vmspace *vm) 4297 { 4298 struct vm_map *map = &vm->vm_map; 4299 4300 KASSERT((map->flags & VM_MAP_DYING) == 0); 4301 4302 mutex_enter(&map->misc_lock); 4303 KASSERT(vm->vm_refcnt > 0); 4304 vm->vm_refcnt++; 4305 mutex_exit(&map->misc_lock); 4306 } 4307 4308 /* 4309 * uvmspace_free: free a vmspace data structure 4310 */ 4311 4312 void 4313 uvmspace_free(struct vmspace *vm) 4314 { 4315 struct vm_map_entry *dead_entries; 4316 struct vm_map *map = &vm->vm_map; 4317 int n; 4318 4319 UVMHIST_FUNC("uvmspace_free"); UVMHIST_CALLED(maphist); 4320 4321 UVMHIST_LOG(maphist,"(vm=0x%x) ref=%d", vm, vm->vm_refcnt,0,0); 4322 mutex_enter(&map->misc_lock); 4323 n = --vm->vm_refcnt; 4324 mutex_exit(&map->misc_lock); 4325 if (n > 0) 4326 return; 4327 4328 /* 4329 * at this point, there should be no other references to the map. 4330 * delete all of the mappings, then destroy the pmap. 4331 */ 4332 4333 map->flags |= VM_MAP_DYING; 4334 pmap_remove_all(map->pmap); 4335 #ifdef SYSVSHM 4336 /* Get rid of any SYSV shared memory segments. */ 4337 if (vm->vm_shm != NULL) 4338 shmexit(vm); 4339 #endif 4340 if (map->nentries) { 4341 uvm_unmap_remove(map, vm_map_min(map), vm_map_max(map), 4342 &dead_entries, NULL, 0); 4343 if (dead_entries != NULL) 4344 uvm_unmap_detach(dead_entries, 0); 4345 } 4346 KASSERT(map->nentries == 0); 4347 KASSERT(map->size == 0); 4348 mutex_destroy(&map->misc_lock); 4349 mutex_destroy(&map->mutex); 4350 rw_destroy(&map->lock); 4351 cv_destroy(&map->cv); 4352 pmap_destroy(map->pmap); 4353 pool_cache_put(&uvm_vmspace_cache, vm); 4354 } 4355 4356 /* 4357 * F O R K - m a i n e n t r y p o i n t 4358 */ 4359 /* 4360 * uvmspace_fork: fork a process' main map 4361 * 4362 * => create a new vmspace for child process from parent. 4363 * => parent's map must not be locked. 4364 */ 4365 4366 struct vmspace * 4367 uvmspace_fork(struct vmspace *vm1) 4368 { 4369 struct vmspace *vm2; 4370 struct vm_map *old_map = &vm1->vm_map; 4371 struct vm_map *new_map; 4372 struct vm_map_entry *old_entry; 4373 struct vm_map_entry *new_entry; 4374 UVMHIST_FUNC("uvmspace_fork"); UVMHIST_CALLED(maphist); 4375 4376 vm_map_lock(old_map); 4377 4378 vm2 = uvmspace_alloc(vm_map_min(old_map), vm_map_max(old_map)); 4379 memcpy(&vm2->vm_startcopy, &vm1->vm_startcopy, 4380 (char *) (vm1 + 1) - (char *) &vm1->vm_startcopy); 4381 new_map = &vm2->vm_map; /* XXX */ 4382 4383 old_entry = old_map->header.next; 4384 new_map->size = old_map->size; 4385 4386 /* 4387 * go entry-by-entry 4388 */ 4389 4390 while (old_entry != &old_map->header) { 4391 4392 /* 4393 * first, some sanity checks on the old entry 4394 */ 4395 4396 KASSERT(!UVM_ET_ISSUBMAP(old_entry)); 4397 KASSERT(UVM_ET_ISCOPYONWRITE(old_entry) || 4398 !UVM_ET_ISNEEDSCOPY(old_entry)); 4399 4400 switch (old_entry->inheritance) { 4401 case MAP_INHERIT_NONE: 4402 4403 /* 4404 * drop the mapping, modify size 4405 */ 4406 new_map->size -= old_entry->end - old_entry->start; 4407 break; 4408 4409 case MAP_INHERIT_SHARE: 4410 4411 /* 4412 * share the mapping: this means we want the old and 4413 * new entries to share amaps and backing objects. 4414 */ 4415 /* 4416 * if the old_entry needs a new amap (due to prev fork) 4417 * then we need to allocate it now so that we have 4418 * something we own to share with the new_entry. [in 4419 * other words, we need to clear needs_copy] 4420 */ 4421 4422 if (UVM_ET_ISNEEDSCOPY(old_entry)) { 4423 /* get our own amap, clears needs_copy */ 4424 amap_copy(old_map, old_entry, AMAP_COPY_NOCHUNK, 4425 0, 0); 4426 /* XXXCDC: WAITOK??? */ 4427 } 4428 4429 new_entry = uvm_mapent_alloc(new_map, 0); 4430 /* old_entry -> new_entry */ 4431 uvm_mapent_copy(old_entry, new_entry); 4432 4433 /* new pmap has nothing wired in it */ 4434 new_entry->wired_count = 0; 4435 4436 /* 4437 * gain reference to object backing the map (can't 4438 * be a submap, already checked this case). 4439 */ 4440 4441 if (new_entry->aref.ar_amap) 4442 uvm_map_reference_amap(new_entry, AMAP_SHARED); 4443 4444 if (new_entry->object.uvm_obj && 4445 new_entry->object.uvm_obj->pgops->pgo_reference) 4446 new_entry->object.uvm_obj-> 4447 pgops->pgo_reference( 4448 new_entry->object.uvm_obj); 4449 4450 /* insert entry at end of new_map's entry list */ 4451 uvm_map_entry_link(new_map, new_map->header.prev, 4452 new_entry); 4453 4454 break; 4455 4456 case MAP_INHERIT_COPY: 4457 4458 /* 4459 * copy-on-write the mapping (using mmap's 4460 * MAP_PRIVATE semantics) 4461 * 4462 * allocate new_entry, adjust reference counts. 4463 * (note that new references are read-only). 4464 */ 4465 4466 new_entry = uvm_mapent_alloc(new_map, 0); 4467 /* old_entry -> new_entry */ 4468 uvm_mapent_copy(old_entry, new_entry); 4469 4470 if (new_entry->aref.ar_amap) 4471 uvm_map_reference_amap(new_entry, 0); 4472 4473 if (new_entry->object.uvm_obj && 4474 new_entry->object.uvm_obj->pgops->pgo_reference) 4475 new_entry->object.uvm_obj->pgops->pgo_reference 4476 (new_entry->object.uvm_obj); 4477 4478 /* new pmap has nothing wired in it */ 4479 new_entry->wired_count = 0; 4480 4481 new_entry->etype |= 4482 (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY); 4483 uvm_map_entry_link(new_map, new_map->header.prev, 4484 new_entry); 4485 4486 /* 4487 * the new entry will need an amap. it will either 4488 * need to be copied from the old entry or created 4489 * from scratch (if the old entry does not have an 4490 * amap). can we defer this process until later 4491 * (by setting "needs_copy") or do we need to copy 4492 * the amap now? 4493 * 4494 * we must copy the amap now if any of the following 4495 * conditions hold: 4496 * 1. the old entry has an amap and that amap is 4497 * being shared. this means that the old (parent) 4498 * process is sharing the amap with another 4499 * process. if we do not clear needs_copy here 4500 * we will end up in a situation where both the 4501 * parent and child process are refering to the 4502 * same amap with "needs_copy" set. if the 4503 * parent write-faults, the fault routine will 4504 * clear "needs_copy" in the parent by allocating 4505 * a new amap. this is wrong because the 4506 * parent is supposed to be sharing the old amap 4507 * and the new amap will break that. 4508 * 4509 * 2. if the old entry has an amap and a non-zero 4510 * wire count then we are going to have to call 4511 * amap_cow_now to avoid page faults in the 4512 * parent process. since amap_cow_now requires 4513 * "needs_copy" to be clear we might as well 4514 * clear it here as well. 4515 * 4516 */ 4517 4518 if (old_entry->aref.ar_amap != NULL) { 4519 if ((amap_flags(old_entry->aref.ar_amap) & 4520 AMAP_SHARED) != 0 || 4521 VM_MAPENT_ISWIRED(old_entry)) { 4522 4523 amap_copy(new_map, new_entry, 4524 AMAP_COPY_NOCHUNK, 0, 0); 4525 /* XXXCDC: M_WAITOK ... ok? */ 4526 } 4527 } 4528 4529 /* 4530 * if the parent's entry is wired down, then the 4531 * parent process does not want page faults on 4532 * access to that memory. this means that we 4533 * cannot do copy-on-write because we can't write 4534 * protect the old entry. in this case we 4535 * resolve all copy-on-write faults now, using 4536 * amap_cow_now. note that we have already 4537 * allocated any needed amap (above). 4538 */ 4539 4540 if (VM_MAPENT_ISWIRED(old_entry)) { 4541 4542 /* 4543 * resolve all copy-on-write faults now 4544 * (note that there is nothing to do if 4545 * the old mapping does not have an amap). 4546 */ 4547 if (old_entry->aref.ar_amap) 4548 amap_cow_now(new_map, new_entry); 4549 4550 } else { 4551 4552 /* 4553 * setup mappings to trigger copy-on-write faults 4554 * we must write-protect the parent if it has 4555 * an amap and it is not already "needs_copy"... 4556 * if it is already "needs_copy" then the parent 4557 * has already been write-protected by a previous 4558 * fork operation. 4559 */ 4560 4561 if (old_entry->aref.ar_amap && 4562 !UVM_ET_ISNEEDSCOPY(old_entry)) { 4563 if (old_entry->max_protection & VM_PROT_WRITE) { 4564 pmap_protect(old_map->pmap, 4565 old_entry->start, 4566 old_entry->end, 4567 old_entry->protection & 4568 ~VM_PROT_WRITE); 4569 } 4570 old_entry->etype |= UVM_ET_NEEDSCOPY; 4571 } 4572 } 4573 break; 4574 } /* end of switch statement */ 4575 old_entry = old_entry->next; 4576 } 4577 4578 pmap_update(old_map->pmap); 4579 vm_map_unlock(old_map); 4580 4581 #ifdef SYSVSHM 4582 if (vm1->vm_shm) 4583 shmfork(vm1, vm2); 4584 #endif 4585 4586 #ifdef PMAP_FORK 4587 pmap_fork(vm1->vm_map.pmap, vm2->vm_map.pmap); 4588 #endif 4589 4590 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4591 return (vm2); 4592 } 4593 4594 4595 /* 4596 * in-kernel map entry allocation. 4597 */ 4598 4599 struct uvm_kmapent_hdr { 4600 LIST_ENTRY(uvm_kmapent_hdr) ukh_listq; 4601 int ukh_nused; 4602 struct vm_map_entry *ukh_freelist; 4603 struct vm_map *ukh_map; 4604 struct vm_map_entry ukh_entries[0]; 4605 }; 4606 4607 #define UVM_KMAPENT_CHUNK \ 4608 ((PAGE_SIZE - sizeof(struct uvm_kmapent_hdr)) \ 4609 / sizeof(struct vm_map_entry)) 4610 4611 #define UVM_KHDR_FIND(entry) \ 4612 ((struct uvm_kmapent_hdr *)(((vaddr_t)entry) & ~PAGE_MASK)) 4613 4614 4615 #ifdef DIAGNOSTIC 4616 static struct vm_map * 4617 uvm_kmapent_map(struct vm_map_entry *entry) 4618 { 4619 const struct uvm_kmapent_hdr *ukh; 4620 4621 ukh = UVM_KHDR_FIND(entry); 4622 return ukh->ukh_map; 4623 } 4624 #endif 4625 4626 static inline struct vm_map_entry * 4627 uvm_kmapent_get(struct uvm_kmapent_hdr *ukh) 4628 { 4629 struct vm_map_entry *entry; 4630 4631 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4632 KASSERT(ukh->ukh_nused >= 0); 4633 4634 entry = ukh->ukh_freelist; 4635 if (entry) { 4636 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4637 == UVM_MAP_KERNEL); 4638 ukh->ukh_freelist = entry->next; 4639 ukh->ukh_nused++; 4640 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4641 } else { 4642 KASSERT(ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4643 } 4644 4645 return entry; 4646 } 4647 4648 static inline void 4649 uvm_kmapent_put(struct uvm_kmapent_hdr *ukh, struct vm_map_entry *entry) 4650 { 4651 4652 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4653 == UVM_MAP_KERNEL); 4654 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4655 KASSERT(ukh->ukh_nused > 0); 4656 KASSERT(ukh->ukh_freelist != NULL || 4657 ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4658 KASSERT(ukh->ukh_freelist == NULL || 4659 ukh->ukh_nused < UVM_KMAPENT_CHUNK); 4660 4661 ukh->ukh_nused--; 4662 entry->next = ukh->ukh_freelist; 4663 ukh->ukh_freelist = entry; 4664 } 4665 4666 /* 4667 * uvm_kmapent_alloc: allocate a map entry for in-kernel map 4668 */ 4669 4670 static struct vm_map_entry * 4671 uvm_kmapent_alloc(struct vm_map *map, int flags) 4672 { 4673 struct vm_page *pg; 4674 struct uvm_kmapent_hdr *ukh; 4675 struct vm_map_entry *entry; 4676 #ifndef PMAP_MAP_POOLPAGE 4677 struct uvm_map_args args; 4678 uvm_flag_t mapflags = UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, 4679 UVM_INH_NONE, UVM_ADV_RANDOM, flags | UVM_FLAG_NOMERGE); 4680 int error; 4681 #endif 4682 vaddr_t va; 4683 int i; 4684 4685 KDASSERT(UVM_KMAPENT_CHUNK > 2); 4686 KDASSERT(kernel_map != NULL); 4687 KASSERT(vm_map_pmap(map) == pmap_kernel()); 4688 4689 UVMMAP_EVCNT_INCR(uke_alloc); 4690 entry = NULL; 4691 again: 4692 /* 4693 * try to grab an entry from freelist. 4694 */ 4695 mutex_spin_enter(&uvm_kentry_lock); 4696 ukh = LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free); 4697 if (ukh) { 4698 entry = uvm_kmapent_get(ukh); 4699 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK) 4700 LIST_REMOVE(ukh, ukh_listq); 4701 } 4702 mutex_spin_exit(&uvm_kentry_lock); 4703 4704 if (entry) 4705 return entry; 4706 4707 /* 4708 * there's no free entry for this vm_map. 4709 * now we need to allocate some vm_map_entry. 4710 * for simplicity, always allocate one page chunk of them at once. 4711 */ 4712 4713 pg = uvm_pagealloc(NULL, 0, NULL, 4714 (flags & UVM_KMF_NOWAIT) != 0 ? UVM_PGA_USERESERVE : 0); 4715 if (__predict_false(pg == NULL)) { 4716 if (flags & UVM_FLAG_NOWAIT) 4717 return NULL; 4718 uvm_wait("kme_alloc"); 4719 goto again; 4720 } 4721 4722 #ifdef PMAP_MAP_POOLPAGE 4723 va = PMAP_MAP_POOLPAGE(VM_PAGE_TO_PHYS(pg)); 4724 KASSERT(va != 0); 4725 #else 4726 error = uvm_map_prepare(map, 0, PAGE_SIZE, NULL, UVM_UNKNOWN_OFFSET, 4727 VM_PGCOLOR_BUCKET(pg), mapflags | UVM_FLAG_COLORMATCH, &args); 4728 if (error) { 4729 uvm_pagefree(pg); 4730 return NULL; 4731 } 4732 4733 va = args.uma_start; 4734 4735 pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), 4736 VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE); 4737 pmap_update(vm_map_pmap(map)); 4738 4739 #endif 4740 ukh = (void *)va; 4741 4742 /* 4743 * use the last entry for ukh itsself. 4744 */ 4745 4746 i = UVM_KMAPENT_CHUNK - 1; 4747 #ifndef PMAP_MAP_POOLPAGE 4748 entry = &ukh->ukh_entries[i--]; 4749 entry->flags = UVM_MAP_KERNEL | UVM_MAP_KMAPENT; 4750 error = uvm_map_enter(map, &args, entry); 4751 KASSERT(error == 0); 4752 #endif 4753 4754 ukh->ukh_nused = UVM_KMAPENT_CHUNK; 4755 ukh->ukh_map = map; 4756 ukh->ukh_freelist = NULL; 4757 for (; i >= 1; i--) { 4758 struct vm_map_entry *xentry = &ukh->ukh_entries[i]; 4759 4760 xentry->flags = UVM_MAP_KERNEL; 4761 uvm_kmapent_put(ukh, xentry); 4762 } 4763 #ifdef PMAP_MAP_POOLPAGE 4764 KASSERT(ukh->ukh_nused == 1); 4765 #else 4766 KASSERT(ukh->ukh_nused == 2); 4767 #endif 4768 4769 mutex_spin_enter(&uvm_kentry_lock); 4770 LIST_INSERT_HEAD(&vm_map_to_kernel(map)->vmk_kentry_free, 4771 ukh, ukh_listq); 4772 mutex_spin_exit(&uvm_kentry_lock); 4773 4774 /* 4775 * return first entry. 4776 */ 4777 4778 entry = &ukh->ukh_entries[0]; 4779 entry->flags = UVM_MAP_KERNEL; 4780 UVMMAP_EVCNT_INCR(ukh_alloc); 4781 4782 return entry; 4783 } 4784 4785 /* 4786 * uvm_mapent_free: free map entry for in-kernel map 4787 */ 4788 4789 static void 4790 uvm_kmapent_free(struct vm_map_entry *entry) 4791 { 4792 struct uvm_kmapent_hdr *ukh; 4793 struct vm_page *pg; 4794 struct vm_map *map; 4795 #ifndef PMAP_UNMAP_POOLPAGE 4796 struct pmap *pmap; 4797 struct vm_map_entry *deadentry; 4798 #endif 4799 vaddr_t va; 4800 paddr_t pa; 4801 4802 UVMMAP_EVCNT_INCR(uke_free); 4803 ukh = UVM_KHDR_FIND(entry); 4804 map = ukh->ukh_map; 4805 4806 mutex_spin_enter(&uvm_kentry_lock); 4807 uvm_kmapent_put(ukh, entry); 4808 #ifdef PMAP_UNMAP_POOLPAGE 4809 if (ukh->ukh_nused > 0) { 4810 #else 4811 if (ukh->ukh_nused > 1) { 4812 #endif 4813 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK - 1) 4814 LIST_INSERT_HEAD( 4815 &vm_map_to_kernel(map)->vmk_kentry_free, 4816 ukh, ukh_listq); 4817 mutex_spin_exit(&uvm_kentry_lock); 4818 return; 4819 } 4820 4821 /* 4822 * now we can free this ukh. 4823 * 4824 * however, keep an empty ukh to avoid ping-pong. 4825 */ 4826 4827 if (LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free) == ukh && 4828 LIST_NEXT(ukh, ukh_listq) == NULL) { 4829 mutex_spin_exit(&uvm_kentry_lock); 4830 return; 4831 } 4832 LIST_REMOVE(ukh, ukh_listq); 4833 mutex_spin_exit(&uvm_kentry_lock); 4834 4835 va = (vaddr_t)ukh; 4836 4837 #ifdef PMAP_UNMAP_POOLPAGE 4838 KASSERT(ukh->ukh_nused == 0); 4839 pa = PMAP_UNMAP_POOLPAGE(va); 4840 KASSERT(pa != 0); 4841 #else 4842 KASSERT(ukh->ukh_nused == 1); 4843 4844 /* 4845 * remove map entry for ukh itsself. 4846 */ 4847 4848 KASSERT((va & PAGE_MASK) == 0); 4849 vm_map_lock(map); 4850 uvm_unmap_remove(map, va, va + PAGE_SIZE, &deadentry, NULL, 0); 4851 KASSERT(deadentry->flags & UVM_MAP_KERNEL); 4852 KASSERT(deadentry->flags & UVM_MAP_KMAPENT); 4853 KASSERT(deadentry->next == NULL); 4854 KASSERT(deadentry == &ukh->ukh_entries[UVM_KMAPENT_CHUNK - 1]); 4855 4856 /* 4857 * unmap the page from pmap and free it. 4858 */ 4859 4860 pmap = vm_map_pmap(map); 4861 KASSERT(pmap == pmap_kernel()); 4862 if (!pmap_extract(pmap, va, &pa)) 4863 panic("%s: no mapping", __func__); 4864 pmap_kremove(va, PAGE_SIZE); 4865 pmap_update(vm_map_pmap(map)); 4866 vm_map_unlock(map); 4867 #endif /* !PMAP_UNMAP_POOLPAGE */ 4868 pg = PHYS_TO_VM_PAGE(pa); 4869 uvm_pagefree(pg); 4870 UVMMAP_EVCNT_INCR(ukh_free); 4871 } 4872 4873 static vsize_t 4874 uvm_kmapent_overhead(vsize_t size) 4875 { 4876 4877 /* 4878 * - the max number of unmerged entries is howmany(size, PAGE_SIZE) 4879 * as the min allocation unit is PAGE_SIZE. 4880 * - UVM_KMAPENT_CHUNK "kmapent"s are allocated from a page. 4881 * one of them are used to map the page itself. 4882 */ 4883 4884 return howmany(howmany(size, PAGE_SIZE), (UVM_KMAPENT_CHUNK - 1)) * 4885 PAGE_SIZE; 4886 } 4887 4888 /* 4889 * map entry reservation 4890 */ 4891 4892 /* 4893 * uvm_mapent_reserve: reserve map entries for clipping before locking map. 4894 * 4895 * => needed when unmapping entries allocated without UVM_FLAG_QUANTUM. 4896 * => caller shouldn't hold map locked. 4897 */ 4898 int 4899 uvm_mapent_reserve(struct vm_map *map, struct uvm_mapent_reservation *umr, 4900 int nentries, int flags) 4901 { 4902 4903 umr->umr_nentries = 0; 4904 4905 if ((flags & UVM_FLAG_QUANTUM) != 0) 4906 return 0; 4907 4908 if (!VM_MAP_USE_KMAPENT(map)) 4909 return 0; 4910 4911 while (nentries--) { 4912 struct vm_map_entry *ent; 4913 ent = uvm_kmapent_alloc(map, flags); 4914 if (!ent) { 4915 uvm_mapent_unreserve(map, umr); 4916 return ENOMEM; 4917 } 4918 UMR_PUTENTRY(umr, ent); 4919 } 4920 4921 return 0; 4922 } 4923 4924 /* 4925 * uvm_mapent_unreserve: 4926 * 4927 * => caller shouldn't hold map locked. 4928 * => never fail or sleep. 4929 */ 4930 void 4931 uvm_mapent_unreserve(struct vm_map *map, struct uvm_mapent_reservation *umr) 4932 { 4933 4934 while (!UMR_EMPTY(umr)) 4935 uvm_kmapent_free(UMR_GETENTRY(umr)); 4936 } 4937 4938 /* 4939 * uvm_mapent_trymerge: try to merge an entry with its neighbors. 4940 * 4941 * => called with map locked. 4942 * => return non zero if successfully merged. 4943 */ 4944 4945 int 4946 uvm_mapent_trymerge(struct vm_map *map, struct vm_map_entry *entry, int flags) 4947 { 4948 struct uvm_object *uobj; 4949 struct vm_map_entry *next; 4950 struct vm_map_entry *prev; 4951 vsize_t size; 4952 int merged = 0; 4953 bool copying; 4954 int newetype; 4955 4956 if (VM_MAP_USE_KMAPENT(map)) { 4957 return 0; 4958 } 4959 if (entry->aref.ar_amap != NULL) { 4960 return 0; 4961 } 4962 if ((entry->flags & UVM_MAP_NOMERGE) != 0) { 4963 return 0; 4964 } 4965 4966 uobj = entry->object.uvm_obj; 4967 size = entry->end - entry->start; 4968 copying = (flags & UVM_MERGE_COPYING) != 0; 4969 newetype = copying ? (entry->etype & ~UVM_ET_NEEDSCOPY) : entry->etype; 4970 4971 next = entry->next; 4972 if (next != &map->header && 4973 next->start == entry->end && 4974 ((copying && next->aref.ar_amap != NULL && 4975 amap_refs(next->aref.ar_amap) == 1) || 4976 (!copying && next->aref.ar_amap == NULL)) && 4977 UVM_ET_ISCOMPATIBLE(next, newetype, 4978 uobj, entry->flags, entry->protection, 4979 entry->max_protection, entry->inheritance, entry->advice, 4980 entry->wired_count) && 4981 (uobj == NULL || entry->offset + size == next->offset)) { 4982 int error; 4983 4984 if (copying) { 4985 error = amap_extend(next, size, 4986 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_BACKWARDS); 4987 } else { 4988 error = 0; 4989 } 4990 if (error == 0) { 4991 if (uobj) { 4992 if (uobj->pgops->pgo_detach) { 4993 uobj->pgops->pgo_detach(uobj); 4994 } 4995 } 4996 4997 entry->end = next->end; 4998 clear_hints(map, next); 4999 uvm_map_entry_unlink(map, next); 5000 if (copying) { 5001 entry->aref = next->aref; 5002 entry->etype &= ~UVM_ET_NEEDSCOPY; 5003 } 5004 uvm_map_check(map, "trymerge forwardmerge"); 5005 uvm_mapent_free_merged(map, next); 5006 merged++; 5007 } 5008 } 5009 5010 prev = entry->prev; 5011 if (prev != &map->header && 5012 prev->end == entry->start && 5013 ((copying && !merged && prev->aref.ar_amap != NULL && 5014 amap_refs(prev->aref.ar_amap) == 1) || 5015 (!copying && prev->aref.ar_amap == NULL)) && 5016 UVM_ET_ISCOMPATIBLE(prev, newetype, 5017 uobj, entry->flags, entry->protection, 5018 entry->max_protection, entry->inheritance, entry->advice, 5019 entry->wired_count) && 5020 (uobj == NULL || 5021 prev->offset + prev->end - prev->start == entry->offset)) { 5022 int error; 5023 5024 if (copying) { 5025 error = amap_extend(prev, size, 5026 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_FORWARDS); 5027 } else { 5028 error = 0; 5029 } 5030 if (error == 0) { 5031 if (uobj) { 5032 if (uobj->pgops->pgo_detach) { 5033 uobj->pgops->pgo_detach(uobj); 5034 } 5035 entry->offset = prev->offset; 5036 } 5037 5038 entry->start = prev->start; 5039 clear_hints(map, prev); 5040 uvm_map_entry_unlink(map, prev); 5041 if (copying) { 5042 entry->aref = prev->aref; 5043 entry->etype &= ~UVM_ET_NEEDSCOPY; 5044 } 5045 uvm_map_check(map, "trymerge backmerge"); 5046 uvm_mapent_free_merged(map, prev); 5047 merged++; 5048 } 5049 } 5050 5051 return merged; 5052 } 5053 5054 /* 5055 * uvm_map_create: create map 5056 */ 5057 5058 struct vm_map * 5059 uvm_map_create(pmap_t pmap, vaddr_t vmin, vaddr_t vmax, int flags) 5060 { 5061 struct vm_map *result; 5062 5063 result = malloc(sizeof(struct vm_map), M_VMMAP, M_WAITOK); 5064 uvm_map_setup(result, vmin, vmax, flags); 5065 result->pmap = pmap; 5066 return(result); 5067 } 5068 5069 /* 5070 * uvm_map_setup: init map 5071 * 5072 * => map must not be in service yet. 5073 */ 5074 5075 void 5076 uvm_map_setup(struct vm_map *map, vaddr_t vmin, vaddr_t vmax, int flags) 5077 { 5078 int ipl; 5079 5080 rb_tree_init(&map->rb_tree, &uvm_map_tree_ops); 5081 map->header.next = map->header.prev = &map->header; 5082 map->nentries = 0; 5083 map->size = 0; 5084 map->ref_count = 1; 5085 vm_map_setmin(map, vmin); 5086 vm_map_setmax(map, vmax); 5087 map->flags = flags; 5088 map->first_free = &map->header; 5089 map->hint = &map->header; 5090 map->timestamp = 0; 5091 map->busy = NULL; 5092 5093 if ((flags & VM_MAP_INTRSAFE) != 0) { 5094 ipl = IPL_VM; 5095 } else { 5096 ipl = IPL_NONE; 5097 } 5098 5099 rw_init(&map->lock); 5100 cv_init(&map->cv, "vm_map"); 5101 mutex_init(&map->misc_lock, MUTEX_DRIVER, ipl); 5102 mutex_init(&map->mutex, MUTEX_DRIVER, ipl); 5103 } 5104 5105 5106 /* 5107 * U N M A P - m a i n e n t r y p o i n t 5108 */ 5109 5110 /* 5111 * uvm_unmap1: remove mappings from a vm_map (from "start" up to "stop") 5112 * 5113 * => caller must check alignment and size 5114 * => map must be unlocked (we will lock it) 5115 * => flags is UVM_FLAG_QUANTUM or 0. 5116 */ 5117 5118 void 5119 uvm_unmap1(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 5120 { 5121 struct vm_map_entry *dead_entries; 5122 struct uvm_mapent_reservation umr; 5123 UVMHIST_FUNC("uvm_unmap"); UVMHIST_CALLED(maphist); 5124 5125 UVMHIST_LOG(maphist, " (map=0x%x, start=0x%x, end=0x%x)", 5126 map, start, end, 0); 5127 if (map == kernel_map) { 5128 LOCKDEBUG_MEM_CHECK((void *)start, end - start); 5129 } 5130 /* 5131 * work now done by helper functions. wipe the pmap's and then 5132 * detach from the dead entries... 5133 */ 5134 uvm_mapent_reserve(map, &umr, 2, flags); 5135 vm_map_lock(map); 5136 uvm_unmap_remove(map, start, end, &dead_entries, &umr, flags); 5137 vm_map_unlock(map); 5138 uvm_mapent_unreserve(map, &umr); 5139 5140 if (dead_entries != NULL) 5141 uvm_unmap_detach(dead_entries, 0); 5142 5143 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 5144 } 5145 5146 5147 /* 5148 * uvm_map_reference: add reference to a map 5149 * 5150 * => map need not be locked (we use misc_lock). 5151 */ 5152 5153 void 5154 uvm_map_reference(struct vm_map *map) 5155 { 5156 mutex_enter(&map->misc_lock); 5157 map->ref_count++; 5158 mutex_exit(&map->misc_lock); 5159 } 5160 5161 struct vm_map_kernel * 5162 vm_map_to_kernel(struct vm_map *map) 5163 { 5164 5165 KASSERT(VM_MAP_IS_KERNEL(map)); 5166 5167 return (struct vm_map_kernel *)map; 5168 } 5169 5170 bool 5171 vm_map_starved_p(struct vm_map *map) 5172 { 5173 5174 if ((map->flags & VM_MAP_WANTVA) != 0) { 5175 return true; 5176 } 5177 /* XXX */ 5178 if ((vm_map_max(map) - vm_map_min(map)) / 16 * 15 < map->size) { 5179 return true; 5180 } 5181 return false; 5182 } 5183 5184 void 5185 uvm_map_lock_entry(struct vm_map_entry *entry) 5186 { 5187 5188 if (entry->aref.ar_amap != NULL) { 5189 amap_lock(entry->aref.ar_amap); 5190 } 5191 if (UVM_ET_ISOBJ(entry)) { 5192 mutex_enter(entry->object.uvm_obj->vmobjlock); 5193 } 5194 } 5195 5196 void 5197 uvm_map_unlock_entry(struct vm_map_entry *entry) 5198 { 5199 5200 if (UVM_ET_ISOBJ(entry)) { 5201 mutex_exit(entry->object.uvm_obj->vmobjlock); 5202 } 5203 if (entry->aref.ar_amap != NULL) { 5204 amap_unlock(entry->aref.ar_amap); 5205 } 5206 } 5207 5208 #if defined(DDB) || defined(DEBUGPRINT) 5209 5210 /* 5211 * uvm_map_printit: actually prints the map 5212 */ 5213 5214 void 5215 uvm_map_printit(struct vm_map *map, bool full, 5216 void (*pr)(const char *, ...)) 5217 { 5218 struct vm_map_entry *entry; 5219 5220 (*pr)("MAP %p: [0x%lx->0x%lx]\n", map, vm_map_min(map), 5221 vm_map_max(map)); 5222 (*pr)("\t#ent=%d, sz=%d, ref=%d, version=%d, flags=0x%x\n", 5223 map->nentries, map->size, map->ref_count, map->timestamp, 5224 map->flags); 5225 (*pr)("\tpmap=%p(resident=%ld, wired=%ld)\n", map->pmap, 5226 pmap_resident_count(map->pmap), pmap_wired_count(map->pmap)); 5227 if (!full) 5228 return; 5229 for (entry = map->header.next; entry != &map->header; 5230 entry = entry->next) { 5231 (*pr)(" - %p: 0x%lx->0x%lx: obj=%p/0x%llx, amap=%p/%d\n", 5232 entry, entry->start, entry->end, entry->object.uvm_obj, 5233 (long long)entry->offset, entry->aref.ar_amap, 5234 entry->aref.ar_pageoff); 5235 (*pr)( 5236 "\tsubmap=%c, cow=%c, nc=%c, prot(max)=%d/%d, inh=%d, " 5237 "wc=%d, adv=%d\n", 5238 (entry->etype & UVM_ET_SUBMAP) ? 'T' : 'F', 5239 (entry->etype & UVM_ET_COPYONWRITE) ? 'T' : 'F', 5240 (entry->etype & UVM_ET_NEEDSCOPY) ? 'T' : 'F', 5241 entry->protection, entry->max_protection, 5242 entry->inheritance, entry->wired_count, entry->advice); 5243 } 5244 } 5245 5246 void 5247 uvm_whatis(uintptr_t addr, void (*pr)(const char *, ...)) 5248 { 5249 struct vm_map *map; 5250 5251 for (map = kernel_map;;) { 5252 struct vm_map_entry *entry; 5253 5254 if (!uvm_map_lookup_entry_bytree(map, (vaddr_t)addr, &entry)) { 5255 break; 5256 } 5257 (*pr)("%p is %p+%zu from VMMAP %p\n", 5258 (void *)addr, (void *)entry->start, 5259 (size_t)(addr - (uintptr_t)entry->start), map); 5260 if (!UVM_ET_ISSUBMAP(entry)) { 5261 break; 5262 } 5263 map = entry->object.sub_map; 5264 } 5265 } 5266 5267 #endif /* DDB || DEBUGPRINT */ 5268 5269 #ifndef __USER_VA0_IS_SAFE 5270 static int 5271 sysctl_user_va0_disable(SYSCTLFN_ARGS) 5272 { 5273 struct sysctlnode node; 5274 int t, error; 5275 5276 node = *rnode; 5277 node.sysctl_data = &t; 5278 t = user_va0_disable; 5279 error = sysctl_lookup(SYSCTLFN_CALL(&node)); 5280 if (error || newp == NULL) 5281 return (error); 5282 5283 /* lower only at securelevel < 1 */ 5284 if (!t && user_va0_disable && 5285 kauth_authorize_system(l->l_cred, 5286 KAUTH_SYSTEM_CHSYSFLAGS /* XXX */, 0, 5287 NULL, NULL, NULL)) 5288 return EPERM; 5289 5290 user_va0_disable = !!t; 5291 return 0; 5292 } 5293 5294 SYSCTL_SETUP(sysctl_uvmmap_setup, "sysctl uvmmap setup") 5295 { 5296 5297 sysctl_createv(clog, 0, NULL, NULL, 5298 CTLFLAG_PERMANENT|CTLFLAG_READWRITE, 5299 CTLTYPE_INT, "user_va0_disable", 5300 SYSCTL_DESCR("Disable VA 0"), 5301 sysctl_user_va0_disable, 0, &user_va0_disable, 0, 5302 CTL_VM, CTL_CREATE, CTL_EOL); 5303 } 5304 #endif 5305