1 /* $NetBSD: uvm_map.c,v 1.263 2008/07/29 00:03:06 matt 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. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by Charles D. Cranor, 23 * Washington University, the University of California, Berkeley and 24 * its contributors. 25 * 4. Neither the name of the University nor the names of its contributors 26 * may be used to endorse or promote products derived from this software 27 * without specific prior written permission. 28 * 29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 39 * SUCH DAMAGE. 40 * 41 * @(#)vm_map.c 8.3 (Berkeley) 1/12/94 42 * from: Id: uvm_map.c,v 1.1.2.27 1998/02/07 01:16:54 chs Exp 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Permission to use, copy, modify and distribute this software and 49 * its documentation is hereby granted, provided that both the copyright 50 * notice and this permission notice appear in all copies of the 51 * software, derivative works or modified versions, and any portions 52 * thereof, and that both notices appear in supporting documentation. 53 * 54 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 55 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 56 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 57 * 58 * Carnegie Mellon requests users of this software to return to 59 * 60 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 61 * School of Computer Science 62 * Carnegie Mellon University 63 * Pittsburgh PA 15213-3890 64 * 65 * any improvements or extensions that they make and grant Carnegie the 66 * rights to redistribute these changes. 67 */ 68 69 /* 70 * uvm_map.c: uvm map operations 71 */ 72 73 #include <sys/cdefs.h> 74 __KERNEL_RCSID(0, "$NetBSD: uvm_map.c,v 1.263 2008/07/29 00:03:06 matt Exp $"); 75 76 #include "opt_ddb.h" 77 #include "opt_uvmhist.h" 78 #include "opt_uvm.h" 79 #include "opt_sysv.h" 80 81 #include <sys/param.h> 82 #include <sys/systm.h> 83 #include <sys/mman.h> 84 #include <sys/proc.h> 85 #include <sys/malloc.h> 86 #include <sys/pool.h> 87 #include <sys/kernel.h> 88 #include <sys/mount.h> 89 #include <sys/vnode.h> 90 #include <sys/lockdebug.h> 91 #include <sys/atomic.h> 92 93 #ifdef SYSVSHM 94 #include <sys/shm.h> 95 #endif 96 97 #include <uvm/uvm.h> 98 99 #ifdef DDB 100 #include <uvm/uvm_ddb.h> 101 #endif 102 103 #if !defined(UVMMAP_COUNTERS) 104 105 #define UVMMAP_EVCNT_DEFINE(name) /* nothing */ 106 #define UVMMAP_EVCNT_INCR(ev) /* nothing */ 107 #define UVMMAP_EVCNT_DECR(ev) /* nothing */ 108 109 #else /* defined(UVMMAP_NOCOUNTERS) */ 110 111 #include <sys/evcnt.h> 112 #define UVMMAP_EVCNT_DEFINE(name) \ 113 struct evcnt uvmmap_evcnt_##name = EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, \ 114 "uvmmap", #name); \ 115 EVCNT_ATTACH_STATIC(uvmmap_evcnt_##name); 116 #define UVMMAP_EVCNT_INCR(ev) uvmmap_evcnt_##ev.ev_count++ 117 #define UVMMAP_EVCNT_DECR(ev) uvmmap_evcnt_##ev.ev_count-- 118 119 #endif /* defined(UVMMAP_NOCOUNTERS) */ 120 121 UVMMAP_EVCNT_DEFINE(ubackmerge) 122 UVMMAP_EVCNT_DEFINE(uforwmerge) 123 UVMMAP_EVCNT_DEFINE(ubimerge) 124 UVMMAP_EVCNT_DEFINE(unomerge) 125 UVMMAP_EVCNT_DEFINE(kbackmerge) 126 UVMMAP_EVCNT_DEFINE(kforwmerge) 127 UVMMAP_EVCNT_DEFINE(kbimerge) 128 UVMMAP_EVCNT_DEFINE(knomerge) 129 UVMMAP_EVCNT_DEFINE(map_call) 130 UVMMAP_EVCNT_DEFINE(mlk_call) 131 UVMMAP_EVCNT_DEFINE(mlk_hint) 132 UVMMAP_EVCNT_DEFINE(mlk_list) 133 UVMMAP_EVCNT_DEFINE(mlk_tree) 134 UVMMAP_EVCNT_DEFINE(mlk_treeloop) 135 UVMMAP_EVCNT_DEFINE(mlk_listloop) 136 137 UVMMAP_EVCNT_DEFINE(uke_alloc) 138 UVMMAP_EVCNT_DEFINE(uke_free) 139 UVMMAP_EVCNT_DEFINE(ukh_alloc) 140 UVMMAP_EVCNT_DEFINE(ukh_free) 141 142 const char vmmapbsy[] = "vmmapbsy"; 143 144 /* 145 * cache for vmspace structures. 146 */ 147 148 static struct pool_cache uvm_vmspace_cache; 149 150 /* 151 * cache for dynamically-allocated map entries. 152 */ 153 154 static struct pool_cache uvm_map_entry_cache; 155 156 MALLOC_DEFINE(M_VMMAP, "VM map", "VM map structures"); 157 MALLOC_DEFINE(M_VMPMAP, "VM pmap", "VM pmap"); 158 159 #ifdef PMAP_GROWKERNEL 160 /* 161 * This global represents the end of the kernel virtual address 162 * space. If we want to exceed this, we must grow the kernel 163 * virtual address space dynamically. 164 * 165 * Note, this variable is locked by kernel_map's lock. 166 */ 167 vaddr_t uvm_maxkaddr; 168 #endif 169 170 /* 171 * macros 172 */ 173 174 /* 175 * VM_MAP_USE_KMAPENT: determine if uvm_kmapent_alloc/free is used 176 * for the vm_map. 177 */ 178 extern struct vm_map *pager_map; /* XXX */ 179 #define VM_MAP_USE_KMAPENT_FLAGS(flags) \ 180 (((flags) & VM_MAP_INTRSAFE) != 0) 181 #define VM_MAP_USE_KMAPENT(map) \ 182 (VM_MAP_USE_KMAPENT_FLAGS((map)->flags) || (map) == kernel_map) 183 184 /* 185 * UVM_ET_ISCOMPATIBLE: check some requirements for map entry merging 186 */ 187 188 #define UVM_ET_ISCOMPATIBLE(ent, type, uobj, meflags, \ 189 prot, maxprot, inh, adv, wire) \ 190 ((ent)->etype == (type) && \ 191 (((ent)->flags ^ (meflags)) & (UVM_MAP_NOMERGE | UVM_MAP_QUANTUM)) \ 192 == 0 && \ 193 (ent)->object.uvm_obj == (uobj) && \ 194 (ent)->protection == (prot) && \ 195 (ent)->max_protection == (maxprot) && \ 196 (ent)->inheritance == (inh) && \ 197 (ent)->advice == (adv) && \ 198 (ent)->wired_count == (wire)) 199 200 /* 201 * uvm_map_entry_link: insert entry into a map 202 * 203 * => map must be locked 204 */ 205 #define uvm_map_entry_link(map, after_where, entry) do { \ 206 uvm_mapent_check(entry); \ 207 (map)->nentries++; \ 208 (entry)->prev = (after_where); \ 209 (entry)->next = (after_where)->next; \ 210 (entry)->prev->next = (entry); \ 211 (entry)->next->prev = (entry); \ 212 uvm_rb_insert((map), (entry)); \ 213 } while (/*CONSTCOND*/ 0) 214 215 /* 216 * uvm_map_entry_unlink: remove entry from a map 217 * 218 * => map must be locked 219 */ 220 #define uvm_map_entry_unlink(map, entry) do { \ 221 KASSERT((entry) != (map)->first_free); \ 222 KASSERT((entry) != (map)->hint); \ 223 uvm_mapent_check(entry); \ 224 (map)->nentries--; \ 225 (entry)->next->prev = (entry)->prev; \ 226 (entry)->prev->next = (entry)->next; \ 227 uvm_rb_remove((map), (entry)); \ 228 } while (/*CONSTCOND*/ 0) 229 230 /* 231 * SAVE_HINT: saves the specified entry as the hint for future lookups. 232 * 233 * => map need not be locked. 234 */ 235 #define SAVE_HINT(map, check, value) do { \ 236 if ((map)->hint == (check)) \ 237 (map)->hint = (value); \ 238 } while (/*CONSTCOND*/ 0) 239 240 /* 241 * clear_hints: ensure that hints don't point to the entry. 242 * 243 * => map must be write-locked. 244 */ 245 static void 246 clear_hints(struct vm_map *map, struct vm_map_entry *ent) 247 { 248 249 SAVE_HINT(map, ent, ent->prev); 250 if (map->first_free == ent) { 251 map->first_free = ent->prev; 252 } 253 } 254 255 /* 256 * VM_MAP_RANGE_CHECK: check and correct range 257 * 258 * => map must at least be read locked 259 */ 260 261 #define VM_MAP_RANGE_CHECK(map, start, end) do { \ 262 if (start < vm_map_min(map)) \ 263 start = vm_map_min(map); \ 264 if (end > vm_map_max(map)) \ 265 end = vm_map_max(map); \ 266 if (start > end) \ 267 start = end; \ 268 } while (/*CONSTCOND*/ 0) 269 270 /* 271 * local prototypes 272 */ 273 274 static struct vm_map_entry * 275 uvm_mapent_alloc(struct vm_map *, int); 276 static struct vm_map_entry * 277 uvm_mapent_alloc_split(struct vm_map *, 278 const struct vm_map_entry *, int, 279 struct uvm_mapent_reservation *); 280 static void uvm_mapent_copy(struct vm_map_entry *, struct vm_map_entry *); 281 static void uvm_mapent_free(struct vm_map_entry *); 282 #if defined(DEBUG) 283 static void _uvm_mapent_check(const struct vm_map_entry *, const char *, 284 int); 285 #define uvm_mapent_check(map) _uvm_mapent_check(map, __FILE__, __LINE__) 286 #else /* defined(DEBUG) */ 287 #define uvm_mapent_check(e) /* nothing */ 288 #endif /* defined(DEBUG) */ 289 static struct vm_map_entry * 290 uvm_kmapent_alloc(struct vm_map *, int); 291 static void uvm_kmapent_free(struct vm_map_entry *); 292 static vsize_t uvm_kmapent_overhead(vsize_t); 293 294 static void uvm_map_entry_unwire(struct vm_map *, struct vm_map_entry *); 295 static void uvm_map_reference_amap(struct vm_map_entry *, int); 296 static int uvm_map_space_avail(vaddr_t *, vsize_t, voff_t, vsize_t, int, 297 struct vm_map_entry *); 298 static void uvm_map_unreference_amap(struct vm_map_entry *, int); 299 300 int _uvm_map_sanity(struct vm_map *); 301 int _uvm_tree_sanity(struct vm_map *); 302 static vsize_t uvm_rb_maxgap(const struct vm_map_entry *); 303 304 CTASSERT(offsetof(struct vm_map_entry, rb_node) == 0); 305 #define ROOT_ENTRY(map) ((struct vm_map_entry *)(map)->rb_tree.rbt_root) 306 #define LEFT_ENTRY(entry) ((struct vm_map_entry *)(entry)->rb_node.rb_left) 307 #define RIGHT_ENTRY(entry) ((struct vm_map_entry *)(entry)->rb_node.rb_right) 308 #define PARENT_ENTRY(map, entry) \ 309 (ROOT_ENTRY(map) == (entry) \ 310 ? NULL \ 311 : (struct vm_map_entry *)RB_FATHER(&(entry)->rb_node)) 312 313 static int 314 uvm_map_compare_nodes(const struct rb_node *nparent, 315 const struct rb_node *nkey) 316 { 317 const struct vm_map_entry *eparent = (const void *) nparent; 318 const struct vm_map_entry *ekey = (const void *) nkey; 319 320 KASSERT(eparent->start < ekey->start || eparent->start >= ekey->end); 321 KASSERT(ekey->start < eparent->start || ekey->start >= eparent->end); 322 323 if (ekey->start < eparent->start) 324 return -1; 325 if (ekey->start >= eparent->end) 326 return 1; 327 return 0; 328 } 329 330 static int 331 uvm_map_compare_key(const struct rb_node *nparent, const void *vkey) 332 { 333 const struct vm_map_entry *eparent = (const void *) nparent; 334 const vaddr_t va = *(const vaddr_t *) vkey; 335 336 if (va < eparent->start) 337 return -1; 338 if (va >= eparent->end) 339 return 1; 340 return 0; 341 } 342 343 static const struct rb_tree_ops uvm_map_tree_ops = { 344 .rbto_compare_nodes = uvm_map_compare_nodes, 345 .rbto_compare_key = uvm_map_compare_key, 346 }; 347 348 static inline vsize_t 349 uvm_rb_gap(const struct vm_map_entry *entry) 350 { 351 KASSERT(entry->next != NULL); 352 return entry->next->start - entry->end; 353 } 354 355 static vsize_t 356 uvm_rb_maxgap(const struct vm_map_entry *entry) 357 { 358 struct vm_map_entry *child; 359 vsize_t maxgap = entry->gap; 360 361 /* 362 * We need maxgap to be the largest gap of us or any of our 363 * descendents. Since each of our children's maxgap is the 364 * cached value of their largest gap of themselves or their 365 * descendents, we can just use that value and avoid recursing 366 * down the tree to calculate it. 367 */ 368 if ((child = LEFT_ENTRY(entry)) != NULL && maxgap < child->maxgap) 369 maxgap = child->maxgap; 370 371 if ((child = RIGHT_ENTRY(entry)) != NULL && maxgap < child->maxgap) 372 maxgap = child->maxgap; 373 374 return maxgap; 375 } 376 377 static void 378 uvm_rb_fixup(struct vm_map *map, struct vm_map_entry *entry) 379 { 380 struct vm_map_entry *parent; 381 382 KASSERT(entry->gap == uvm_rb_gap(entry)); 383 entry->maxgap = uvm_rb_maxgap(entry); 384 385 while ((parent = PARENT_ENTRY(map, entry)) != NULL) { 386 struct vm_map_entry *brother; 387 vsize_t maxgap = parent->gap; 388 389 KDASSERT(parent->gap == uvm_rb_gap(parent)); 390 if (maxgap < entry->maxgap) 391 maxgap = entry->maxgap; 392 /* 393 * Since we work our towards the root, we know entry's maxgap 394 * value is ok but its brothers may now be out-of-date due 395 * rebalancing. So refresh it. 396 */ 397 brother = (struct vm_map_entry *)parent->rb_node.rb_nodes[RB_POSITION(&entry->rb_node) ^ RB_DIR_OTHER]; 398 if (brother != NULL) { 399 KDASSERT(brother->gap == uvm_rb_gap(brother)); 400 brother->maxgap = uvm_rb_maxgap(brother); 401 if (maxgap < brother->maxgap) 402 maxgap = brother->maxgap; 403 } 404 405 parent->maxgap = maxgap; 406 entry = parent; 407 } 408 } 409 410 static void 411 uvm_rb_insert(struct vm_map *map, struct vm_map_entry *entry) 412 { 413 entry->gap = entry->maxgap = uvm_rb_gap(entry); 414 if (entry->prev != &map->header) 415 entry->prev->gap = uvm_rb_gap(entry->prev); 416 417 if (!rb_tree_insert_node(&map->rb_tree, &entry->rb_node)) 418 panic("uvm_rb_insert: map %p: duplicate entry?", map); 419 420 /* 421 * If the previous entry is not our immediate left child, then it's an 422 * ancestor and will be fixed up on the way to the root. We don't 423 * have to check entry->prev against &map->header since &map->header 424 * will never be in the tree. 425 */ 426 uvm_rb_fixup(map, 427 LEFT_ENTRY(entry) == entry->prev ? entry->prev : entry); 428 } 429 430 static void 431 uvm_rb_remove(struct vm_map *map, struct vm_map_entry *entry) 432 { 433 struct vm_map_entry *prev_parent = NULL, *next_parent = NULL; 434 435 /* 436 * If we are removing an interior node, then an adjacent node will 437 * be used to replace its position in the tree. Therefore we will 438 * need to fixup the tree starting at the parent of the replacement 439 * node. So record their parents for later use. 440 */ 441 if (entry->prev != &map->header) 442 prev_parent = PARENT_ENTRY(map, entry->prev); 443 if (entry->next != &map->header) 444 next_parent = PARENT_ENTRY(map, entry->next); 445 446 rb_tree_remove_node(&map->rb_tree, &entry->rb_node); 447 448 /* 449 * If the previous node has a new parent, fixup the tree starting 450 * at the previous node's old parent. 451 */ 452 if (entry->prev != &map->header) { 453 /* 454 * Update the previous entry's gap due to our absence. 455 */ 456 entry->prev->gap = uvm_rb_gap(entry->prev); 457 uvm_rb_fixup(map, entry->prev); 458 if (prev_parent != NULL 459 && prev_parent != entry 460 && prev_parent != PARENT_ENTRY(map, entry->prev)) 461 uvm_rb_fixup(map, prev_parent); 462 } 463 464 /* 465 * If the next node has a new parent, fixup the tree starting 466 * at the next node's old parent. 467 */ 468 if (entry->next != &map->header) { 469 uvm_rb_fixup(map, entry->next); 470 if (next_parent != NULL 471 && next_parent != entry 472 && next_parent != PARENT_ENTRY(map, entry->next)) 473 uvm_rb_fixup(map, next_parent); 474 } 475 } 476 477 #if defined(DEBUG) 478 int uvm_debug_check_map = 0; 479 int uvm_debug_check_rbtree = 0; 480 #define uvm_map_check(map, name) \ 481 _uvm_map_check((map), (name), __FILE__, __LINE__) 482 static void 483 _uvm_map_check(struct vm_map *map, const char *name, 484 const char *file, int line) 485 { 486 487 if ((uvm_debug_check_map && _uvm_map_sanity(map)) || 488 (uvm_debug_check_rbtree && _uvm_tree_sanity(map))) { 489 panic("uvm_map_check failed: \"%s\" map=%p (%s:%d)", 490 name, map, file, line); 491 } 492 } 493 #else /* defined(DEBUG) */ 494 #define uvm_map_check(map, name) /* nothing */ 495 #endif /* defined(DEBUG) */ 496 497 #if defined(DEBUG) || defined(DDB) 498 int 499 _uvm_map_sanity(struct vm_map *map) 500 { 501 bool first_free_found = false; 502 bool hint_found = false; 503 const struct vm_map_entry *e; 504 505 e = &map->header; 506 for (;;) { 507 if (map->first_free == e) { 508 first_free_found = true; 509 } else if (!first_free_found && e->next->start > e->end) { 510 printf("first_free %p should be %p\n", 511 map->first_free, e); 512 return -1; 513 } 514 if (map->hint == e) { 515 hint_found = true; 516 } 517 518 e = e->next; 519 if (e == &map->header) { 520 break; 521 } 522 } 523 if (!first_free_found) { 524 printf("stale first_free\n"); 525 return -1; 526 } 527 if (!hint_found) { 528 printf("stale hint\n"); 529 return -1; 530 } 531 return 0; 532 } 533 534 int 535 _uvm_tree_sanity(struct vm_map *map) 536 { 537 struct vm_map_entry *tmp, *trtmp; 538 int n = 0, i = 1; 539 540 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 541 if (tmp->gap != uvm_rb_gap(tmp)) { 542 printf("%d/%d gap %lx != %lx %s\n", 543 n + 1, map->nentries, 544 (ulong)tmp->gap, (ulong)uvm_rb_gap(tmp), 545 tmp->next == &map->header ? "(last)" : ""); 546 goto error; 547 } 548 /* 549 * If any entries are out of order, tmp->gap will be unsigned 550 * and will likely exceed the size of the map. 551 */ 552 KASSERT(tmp->gap < map->size); 553 n++; 554 } 555 556 if (n != map->nentries) { 557 printf("nentries: %d vs %d\n", n, map->nentries); 558 goto error; 559 } 560 561 trtmp = NULL; 562 for (tmp = map->header.next; tmp != &map->header; tmp = tmp->next) { 563 if (tmp->maxgap != uvm_rb_maxgap(tmp)) { 564 printf("maxgap %lx != %lx\n", 565 (ulong)tmp->maxgap, 566 (ulong)uvm_rb_maxgap(tmp)); 567 goto error; 568 } 569 if (trtmp != NULL && trtmp->start >= tmp->start) { 570 printf("corrupt: 0x%lx >= 0x%lx\n", 571 trtmp->start, tmp->start); 572 goto error; 573 } 574 575 trtmp = tmp; 576 } 577 578 for (tmp = map->header.next; tmp != &map->header; 579 tmp = tmp->next, i++) { 580 trtmp = (void *) rb_tree_iterate(&map->rb_tree, &tmp->rb_node, 581 RB_DIR_LEFT); 582 if (trtmp == NULL) 583 trtmp = &map->header; 584 if (tmp->prev != trtmp) { 585 printf("lookup: %d: %p->prev=%p: %p\n", 586 i, tmp, tmp->prev, trtmp); 587 goto error; 588 } 589 trtmp = (void *) rb_tree_iterate(&map->rb_tree, &tmp->rb_node, 590 RB_DIR_RIGHT); 591 if (trtmp == NULL) 592 trtmp = &map->header; 593 if (tmp->next != trtmp) { 594 printf("lookup: %d: %p->next=%p: %p\n", 595 i, tmp, tmp->next, trtmp); 596 goto error; 597 } 598 trtmp = (void *)rb_tree_find_node(&map->rb_tree, &tmp->start); 599 if (trtmp != tmp) { 600 printf("lookup: %d: %p - %p: %p\n", i, tmp, trtmp, 601 PARENT_ENTRY(map, tmp)); 602 goto error; 603 } 604 } 605 606 return (0); 607 error: 608 return (-1); 609 } 610 #endif /* defined(DEBUG) || defined(DDB) */ 611 612 #ifdef DIAGNOSTIC 613 static struct vm_map *uvm_kmapent_map(struct vm_map_entry *); 614 #endif 615 616 /* 617 * vm_map_lock: acquire an exclusive (write) lock on a map. 618 * 619 * => Note that "intrsafe" maps use only exclusive, spin locks. 620 * 621 * => The locking protocol provides for guaranteed upgrade from shared -> 622 * exclusive by whichever thread currently has the map marked busy. 623 * See "LOCKING PROTOCOL NOTES" in uvm_map.h. This is horrible; among 624 * other problems, it defeats any fairness guarantees provided by RW 625 * locks. 626 */ 627 628 void 629 vm_map_lock(struct vm_map *map) 630 { 631 632 if ((map->flags & VM_MAP_INTRSAFE) != 0) { 633 mutex_spin_enter(&map->mutex); 634 return; 635 } 636 637 for (;;) { 638 rw_enter(&map->lock, RW_WRITER); 639 if (map->busy == NULL) 640 break; 641 if (map->busy == curlwp) 642 break; 643 mutex_enter(&map->misc_lock); 644 rw_exit(&map->lock); 645 if (map->busy != NULL) 646 cv_wait(&map->cv, &map->misc_lock); 647 mutex_exit(&map->misc_lock); 648 } 649 650 map->timestamp++; 651 } 652 653 /* 654 * vm_map_lock_try: try to lock a map, failing if it is already locked. 655 */ 656 657 bool 658 vm_map_lock_try(struct vm_map *map) 659 { 660 661 if ((map->flags & VM_MAP_INTRSAFE) != 0) 662 return mutex_tryenter(&map->mutex); 663 if (!rw_tryenter(&map->lock, RW_WRITER)) 664 return false; 665 if (map->busy != NULL) { 666 rw_exit(&map->lock); 667 return false; 668 } 669 670 map->timestamp++; 671 return true; 672 } 673 674 /* 675 * vm_map_unlock: release an exclusive lock on a map. 676 */ 677 678 void 679 vm_map_unlock(struct vm_map *map) 680 { 681 682 if ((map->flags & VM_MAP_INTRSAFE) != 0) 683 mutex_spin_exit(&map->mutex); 684 else { 685 KASSERT(rw_write_held(&map->lock)); 686 KASSERT(map->busy == NULL || map->busy == curlwp); 687 rw_exit(&map->lock); 688 } 689 } 690 691 /* 692 * vm_map_unbusy: mark the map as unbusy, and wake any waiters that 693 * want an exclusive lock. 694 */ 695 696 void 697 vm_map_unbusy(struct vm_map *map) 698 { 699 700 KASSERT(map->busy == curlwp); 701 702 /* 703 * Safe to clear 'busy' and 'waiters' with only a read lock held: 704 * 705 * o they can only be set with a write lock held 706 * o writers are blocked out with a read or write hold 707 * o at any time, only one thread owns the set of values 708 */ 709 mutex_enter(&map->misc_lock); 710 map->busy = NULL; 711 cv_broadcast(&map->cv); 712 mutex_exit(&map->misc_lock); 713 } 714 715 /* 716 * vm_map_lock_read: acquire a shared (read) lock on a map. 717 */ 718 719 void 720 vm_map_lock_read(struct vm_map *map) 721 { 722 723 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 724 725 rw_enter(&map->lock, RW_READER); 726 } 727 728 /* 729 * vm_map_unlock_read: release a shared lock on a map. 730 */ 731 732 void 733 vm_map_unlock_read(struct vm_map *map) 734 { 735 736 KASSERT((map->flags & VM_MAP_INTRSAFE) == 0); 737 738 rw_exit(&map->lock); 739 } 740 741 /* 742 * vm_map_busy: mark a map as busy. 743 * 744 * => the caller must hold the map write locked 745 */ 746 747 void 748 vm_map_busy(struct vm_map *map) 749 { 750 751 KASSERT(rw_write_held(&map->lock)); 752 KASSERT(map->busy == NULL); 753 754 map->busy = curlwp; 755 } 756 757 /* 758 * vm_map_locked_p: return true if the map is write locked. 759 */ 760 761 bool 762 vm_map_locked_p(struct vm_map *map) 763 { 764 765 if ((map->flags & VM_MAP_INTRSAFE) != 0) { 766 return mutex_owned(&map->mutex); 767 } else { 768 return rw_write_held(&map->lock); 769 } 770 } 771 772 /* 773 * uvm_mapent_alloc: allocate a map entry 774 */ 775 776 static struct vm_map_entry * 777 uvm_mapent_alloc(struct vm_map *map, int flags) 778 { 779 struct vm_map_entry *me; 780 int pflags = (flags & UVM_FLAG_NOWAIT) ? PR_NOWAIT : PR_WAITOK; 781 UVMHIST_FUNC("uvm_mapent_alloc"); UVMHIST_CALLED(maphist); 782 783 if (VM_MAP_USE_KMAPENT(map)) { 784 me = uvm_kmapent_alloc(map, flags); 785 } else { 786 me = pool_cache_get(&uvm_map_entry_cache, pflags); 787 if (__predict_false(me == NULL)) 788 return NULL; 789 me->flags = 0; 790 } 791 792 UVMHIST_LOG(maphist, "<- new entry=0x%x [kentry=%d]", me, 793 ((map->flags & VM_MAP_INTRSAFE) != 0 || map == kernel_map), 0, 0); 794 return (me); 795 } 796 797 /* 798 * uvm_mapent_alloc_split: allocate a map entry for clipping. 799 * 800 * => map must be locked by caller if UVM_MAP_QUANTUM is set. 801 */ 802 803 static struct vm_map_entry * 804 uvm_mapent_alloc_split(struct vm_map *map, 805 const struct vm_map_entry *old_entry, int flags, 806 struct uvm_mapent_reservation *umr) 807 { 808 struct vm_map_entry *me; 809 810 KASSERT(!VM_MAP_USE_KMAPENT(map) || 811 (old_entry->flags & UVM_MAP_QUANTUM) || !UMR_EMPTY(umr)); 812 813 if (old_entry->flags & UVM_MAP_QUANTUM) { 814 struct vm_map_kernel *vmk = vm_map_to_kernel(map); 815 816 KASSERT(vm_map_locked_p(map)); 817 me = vmk->vmk_merged_entries; 818 KASSERT(me); 819 vmk->vmk_merged_entries = me->next; 820 KASSERT(me->flags & UVM_MAP_QUANTUM); 821 } else { 822 me = uvm_mapent_alloc(map, flags); 823 } 824 825 return me; 826 } 827 828 /* 829 * uvm_mapent_free: free map entry 830 */ 831 832 static void 833 uvm_mapent_free(struct vm_map_entry *me) 834 { 835 UVMHIST_FUNC("uvm_mapent_free"); UVMHIST_CALLED(maphist); 836 837 UVMHIST_LOG(maphist,"<- freeing map entry=0x%x [flags=%d]", 838 me, me->flags, 0, 0); 839 if (me->flags & UVM_MAP_KERNEL) { 840 uvm_kmapent_free(me); 841 } else { 842 pool_cache_put(&uvm_map_entry_cache, me); 843 } 844 } 845 846 /* 847 * uvm_mapent_free_merged: free merged map entry 848 * 849 * => keep the entry if needed. 850 * => caller shouldn't hold map locked if VM_MAP_USE_KMAPENT(map) is true. 851 * => map should be locked if UVM_MAP_QUANTUM is set. 852 */ 853 854 static void 855 uvm_mapent_free_merged(struct vm_map *map, struct vm_map_entry *me) 856 { 857 858 KASSERT(!(me->flags & UVM_MAP_KERNEL) || uvm_kmapent_map(me) == map); 859 860 if (me->flags & UVM_MAP_QUANTUM) { 861 /* 862 * keep this entry for later splitting. 863 */ 864 struct vm_map_kernel *vmk; 865 866 KASSERT(vm_map_locked_p(map)); 867 KASSERT(VM_MAP_IS_KERNEL(map)); 868 KASSERT(!VM_MAP_USE_KMAPENT(map) || 869 (me->flags & UVM_MAP_KERNEL)); 870 871 vmk = vm_map_to_kernel(map); 872 me->next = vmk->vmk_merged_entries; 873 vmk->vmk_merged_entries = me; 874 } else { 875 uvm_mapent_free(me); 876 } 877 } 878 879 /* 880 * uvm_mapent_copy: copy a map entry, preserving flags 881 */ 882 883 static inline void 884 uvm_mapent_copy(struct vm_map_entry *src, struct vm_map_entry *dst) 885 { 886 887 memcpy(dst, src, ((char *)&src->uvm_map_entry_stop_copy) - 888 ((char *)src)); 889 } 890 891 /* 892 * uvm_mapent_overhead: calculate maximum kva overhead necessary for 893 * map entries. 894 * 895 * => size and flags are the same as uvm_km_suballoc's ones. 896 */ 897 898 vsize_t 899 uvm_mapent_overhead(vsize_t size, int flags) 900 { 901 902 if (VM_MAP_USE_KMAPENT_FLAGS(flags)) { 903 return uvm_kmapent_overhead(size); 904 } 905 return 0; 906 } 907 908 #if defined(DEBUG) 909 static void 910 _uvm_mapent_check(const struct vm_map_entry *entry, const char *file, int line) 911 { 912 913 if (entry->start >= entry->end) { 914 goto bad; 915 } 916 if (UVM_ET_ISOBJ(entry)) { 917 if (entry->object.uvm_obj == NULL) { 918 goto bad; 919 } 920 } else if (UVM_ET_ISSUBMAP(entry)) { 921 if (entry->object.sub_map == NULL) { 922 goto bad; 923 } 924 } else { 925 if (entry->object.uvm_obj != NULL || 926 entry->object.sub_map != NULL) { 927 goto bad; 928 } 929 } 930 if (!UVM_ET_ISOBJ(entry)) { 931 if (entry->offset != 0) { 932 goto bad; 933 } 934 } 935 936 return; 937 938 bad: 939 panic("%s: bad entry %p (%s:%d)", __func__, entry, file, line); 940 } 941 #endif /* defined(DEBUG) */ 942 943 /* 944 * uvm_map_entry_unwire: unwire a map entry 945 * 946 * => map should be locked by caller 947 */ 948 949 static inline void 950 uvm_map_entry_unwire(struct vm_map *map, struct vm_map_entry *entry) 951 { 952 953 entry->wired_count = 0; 954 uvm_fault_unwire_locked(map, entry->start, entry->end); 955 } 956 957 958 /* 959 * wrapper for calling amap_ref() 960 */ 961 static inline void 962 uvm_map_reference_amap(struct vm_map_entry *entry, int flags) 963 { 964 965 amap_ref(entry->aref.ar_amap, entry->aref.ar_pageoff, 966 (entry->end - entry->start) >> PAGE_SHIFT, flags); 967 } 968 969 970 /* 971 * wrapper for calling amap_unref() 972 */ 973 static inline void 974 uvm_map_unreference_amap(struct vm_map_entry *entry, int flags) 975 { 976 977 amap_unref(entry->aref.ar_amap, entry->aref.ar_pageoff, 978 (entry->end - entry->start) >> PAGE_SHIFT, flags); 979 } 980 981 982 /* 983 * uvm_map_init: init mapping system at boot time. 984 */ 985 986 void 987 uvm_map_init(void) 988 { 989 #if defined(UVMHIST) 990 static struct uvm_history_ent maphistbuf[100]; 991 static struct uvm_history_ent pdhistbuf[100]; 992 #endif 993 994 /* 995 * first, init logging system. 996 */ 997 998 UVMHIST_FUNC("uvm_map_init"); 999 UVMHIST_INIT_STATIC(maphist, maphistbuf); 1000 UVMHIST_INIT_STATIC(pdhist, pdhistbuf); 1001 UVMHIST_CALLED(maphist); 1002 UVMHIST_LOG(maphist,"<starting uvm map system>", 0, 0, 0, 0); 1003 1004 /* 1005 * initialize the global lock for kernel map entry. 1006 */ 1007 1008 mutex_init(&uvm_kentry_lock, MUTEX_DRIVER, IPL_VM); 1009 1010 /* 1011 * initialize caches. 1012 */ 1013 1014 pool_cache_bootstrap(&uvm_map_entry_cache, sizeof(struct vm_map_entry), 1015 0, 0, 0, "vmmpepl", NULL, IPL_NONE, NULL, NULL, NULL); 1016 pool_cache_bootstrap(&uvm_vmspace_cache, sizeof(struct vmspace), 1017 0, 0, 0, "vmsppl", NULL, IPL_NONE, NULL, NULL, NULL); 1018 } 1019 1020 /* 1021 * clippers 1022 */ 1023 1024 /* 1025 * uvm_mapent_splitadj: adjust map entries for splitting, after uvm_mapent_copy. 1026 */ 1027 1028 static void 1029 uvm_mapent_splitadj(struct vm_map_entry *entry1, struct vm_map_entry *entry2, 1030 vaddr_t splitat) 1031 { 1032 vaddr_t adj; 1033 1034 KASSERT(entry1->start < splitat); 1035 KASSERT(splitat < entry1->end); 1036 1037 adj = splitat - entry1->start; 1038 entry1->end = entry2->start = splitat; 1039 1040 if (entry1->aref.ar_amap) { 1041 amap_splitref(&entry1->aref, &entry2->aref, adj); 1042 } 1043 if (UVM_ET_ISSUBMAP(entry1)) { 1044 /* ... unlikely to happen, but play it safe */ 1045 uvm_map_reference(entry1->object.sub_map); 1046 } else if (UVM_ET_ISOBJ(entry1)) { 1047 KASSERT(entry1->object.uvm_obj != NULL); /* suppress coverity */ 1048 entry2->offset += adj; 1049 if (entry1->object.uvm_obj->pgops && 1050 entry1->object.uvm_obj->pgops->pgo_reference) 1051 entry1->object.uvm_obj->pgops->pgo_reference( 1052 entry1->object.uvm_obj); 1053 } 1054 } 1055 1056 /* 1057 * uvm_map_clip_start: ensure that the entry begins at or after 1058 * the starting address, if it doesn't we split the entry. 1059 * 1060 * => caller should use UVM_MAP_CLIP_START macro rather than calling 1061 * this directly 1062 * => map must be locked by caller 1063 */ 1064 1065 void 1066 uvm_map_clip_start(struct vm_map *map, struct vm_map_entry *entry, 1067 vaddr_t start, struct uvm_mapent_reservation *umr) 1068 { 1069 struct vm_map_entry *new_entry; 1070 1071 /* uvm_map_simplify_entry(map, entry); */ /* XXX */ 1072 1073 uvm_map_check(map, "clip_start entry"); 1074 uvm_mapent_check(entry); 1075 1076 /* 1077 * Split off the front portion. note that we must insert the new 1078 * entry BEFORE this one, so that this entry has the specified 1079 * starting address. 1080 */ 1081 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 1082 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1083 uvm_mapent_splitadj(new_entry, entry, start); 1084 uvm_map_entry_link(map, entry->prev, new_entry); 1085 1086 uvm_map_check(map, "clip_start leave"); 1087 } 1088 1089 /* 1090 * uvm_map_clip_end: ensure that the entry ends at or before 1091 * the ending address, if it does't we split the reference 1092 * 1093 * => caller should use UVM_MAP_CLIP_END macro rather than calling 1094 * this directly 1095 * => map must be locked by caller 1096 */ 1097 1098 void 1099 uvm_map_clip_end(struct vm_map *map, struct vm_map_entry *entry, vaddr_t end, 1100 struct uvm_mapent_reservation *umr) 1101 { 1102 struct vm_map_entry *new_entry; 1103 1104 uvm_map_check(map, "clip_end entry"); 1105 uvm_mapent_check(entry); 1106 1107 /* 1108 * Create a new entry and insert it 1109 * AFTER the specified entry 1110 */ 1111 new_entry = uvm_mapent_alloc_split(map, entry, 0, umr); 1112 uvm_mapent_copy(entry, new_entry); /* entry -> new_entry */ 1113 uvm_mapent_splitadj(entry, new_entry, end); 1114 uvm_map_entry_link(map, entry, new_entry); 1115 1116 uvm_map_check(map, "clip_end leave"); 1117 } 1118 1119 static void 1120 vm_map_drain(struct vm_map *map, uvm_flag_t flags) 1121 { 1122 1123 if (!VM_MAP_IS_KERNEL(map)) { 1124 return; 1125 } 1126 1127 uvm_km_va_drain(map, flags); 1128 } 1129 1130 /* 1131 * M A P - m a i n e n t r y p o i n t 1132 */ 1133 /* 1134 * uvm_map: establish a valid mapping in a map 1135 * 1136 * => assume startp is page aligned. 1137 * => assume size is a multiple of PAGE_SIZE. 1138 * => assume sys_mmap provides enough of a "hint" to have us skip 1139 * over text/data/bss area. 1140 * => map must be unlocked (we will lock it) 1141 * => <uobj,uoffset> value meanings (4 cases): 1142 * [1] <NULL,uoffset> == uoffset is a hint for PMAP_PREFER 1143 * [2] <NULL,UVM_UNKNOWN_OFFSET> == don't PMAP_PREFER 1144 * [3] <uobj,uoffset> == normal mapping 1145 * [4] <uobj,UVM_UNKNOWN_OFFSET> == uvm_map finds offset based on VA 1146 * 1147 * case [4] is for kernel mappings where we don't know the offset until 1148 * we've found a virtual address. note that kernel object offsets are 1149 * always relative to vm_map_min(kernel_map). 1150 * 1151 * => if `align' is non-zero, we align the virtual address to the specified 1152 * alignment. 1153 * this is provided as a mechanism for large pages. 1154 * 1155 * => XXXCDC: need way to map in external amap? 1156 */ 1157 1158 int 1159 uvm_map(struct vm_map *map, vaddr_t *startp /* IN/OUT */, vsize_t size, 1160 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags) 1161 { 1162 struct uvm_map_args args; 1163 struct vm_map_entry *new_entry; 1164 int error; 1165 1166 KASSERT((flags & UVM_FLAG_QUANTUM) == 0 || VM_MAP_IS_KERNEL(map)); 1167 KASSERT((size & PAGE_MASK) == 0); 1168 1169 /* 1170 * for pager_map, allocate the new entry first to avoid sleeping 1171 * for memory while we have the map locked. 1172 * 1173 * Also, because we allocate entries for in-kernel maps 1174 * a bit differently (cf. uvm_kmapent_alloc/free), we need to 1175 * allocate them before locking the map. 1176 */ 1177 1178 new_entry = NULL; 1179 if (VM_MAP_USE_KMAPENT(map) || (flags & UVM_FLAG_QUANTUM) || 1180 map == pager_map) { 1181 new_entry = uvm_mapent_alloc(map, (flags & UVM_FLAG_NOWAIT)); 1182 if (__predict_false(new_entry == NULL)) 1183 return ENOMEM; 1184 if (flags & UVM_FLAG_QUANTUM) 1185 new_entry->flags |= UVM_MAP_QUANTUM; 1186 } 1187 if (map == pager_map) 1188 flags |= UVM_FLAG_NOMERGE; 1189 1190 error = uvm_map_prepare(map, *startp, size, uobj, uoffset, align, 1191 flags, &args); 1192 if (!error) { 1193 error = uvm_map_enter(map, &args, new_entry); 1194 *startp = args.uma_start; 1195 } else if (new_entry) { 1196 uvm_mapent_free(new_entry); 1197 } 1198 1199 #if defined(DEBUG) 1200 if (!error && VM_MAP_IS_KERNEL(map)) { 1201 uvm_km_check_empty(*startp, *startp + size, 1202 (map->flags & VM_MAP_INTRSAFE) != 0); 1203 } 1204 #endif /* defined(DEBUG) */ 1205 1206 return error; 1207 } 1208 1209 int 1210 uvm_map_prepare(struct vm_map *map, vaddr_t start, vsize_t size, 1211 struct uvm_object *uobj, voff_t uoffset, vsize_t align, uvm_flag_t flags, 1212 struct uvm_map_args *args) 1213 { 1214 struct vm_map_entry *prev_entry; 1215 vm_prot_t prot = UVM_PROTECTION(flags); 1216 vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1217 1218 UVMHIST_FUNC("uvm_map_prepare"); 1219 UVMHIST_CALLED(maphist); 1220 1221 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1222 map, start, size, flags); 1223 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1224 1225 /* 1226 * detect a popular device driver bug. 1227 */ 1228 1229 KASSERT(doing_shutdown || curlwp != NULL || 1230 (map->flags & VM_MAP_INTRSAFE)); 1231 1232 /* 1233 * zero-sized mapping doesn't make any sense. 1234 */ 1235 KASSERT(size > 0); 1236 1237 KASSERT((~flags & (UVM_FLAG_NOWAIT | UVM_FLAG_WAITVA)) != 0); 1238 1239 uvm_map_check(map, "map entry"); 1240 1241 /* 1242 * check sanity of protection code 1243 */ 1244 1245 if ((prot & maxprot) != prot) { 1246 UVMHIST_LOG(maphist, "<- prot. failure: prot=0x%x, max=0x%x", 1247 prot, maxprot,0,0); 1248 return EACCES; 1249 } 1250 1251 /* 1252 * figure out where to put new VM range 1253 */ 1254 1255 retry: 1256 if (vm_map_lock_try(map) == false) { 1257 if ((flags & UVM_FLAG_TRYLOCK) != 0 && 1258 (map->flags & VM_MAP_INTRSAFE) == 0) { 1259 return EAGAIN; 1260 } 1261 vm_map_lock(map); /* could sleep here */ 1262 } 1263 prev_entry = uvm_map_findspace(map, start, size, &start, 1264 uobj, uoffset, align, flags); 1265 if (prev_entry == NULL) { 1266 unsigned int timestamp; 1267 1268 timestamp = map->timestamp; 1269 UVMHIST_LOG(maphist,"waiting va timestamp=0x%x", 1270 timestamp,0,0,0); 1271 map->flags |= VM_MAP_WANTVA; 1272 vm_map_unlock(map); 1273 1274 /* 1275 * try to reclaim kva and wait until someone does unmap. 1276 * fragile locking here, so we awaken every second to 1277 * recheck the condition. 1278 */ 1279 1280 vm_map_drain(map, flags); 1281 1282 mutex_enter(&map->misc_lock); 1283 while ((map->flags & VM_MAP_WANTVA) != 0 && 1284 map->timestamp == timestamp) { 1285 if ((flags & UVM_FLAG_WAITVA) == 0) { 1286 mutex_exit(&map->misc_lock); 1287 UVMHIST_LOG(maphist, 1288 "<- uvm_map_findspace failed!", 0,0,0,0); 1289 return ENOMEM; 1290 } else { 1291 cv_timedwait(&map->cv, &map->misc_lock, hz); 1292 } 1293 } 1294 mutex_exit(&map->misc_lock); 1295 goto retry; 1296 } 1297 1298 #ifdef PMAP_GROWKERNEL 1299 /* 1300 * If the kernel pmap can't map the requested space, 1301 * then allocate more resources for it. 1302 */ 1303 if (map == kernel_map && uvm_maxkaddr < (start + size)) 1304 uvm_maxkaddr = pmap_growkernel(start + size); 1305 #endif 1306 1307 UVMMAP_EVCNT_INCR(map_call); 1308 1309 /* 1310 * if uobj is null, then uoffset is either a VAC hint for PMAP_PREFER 1311 * [typically from uvm_map_reserve] or it is UVM_UNKNOWN_OFFSET. in 1312 * either case we want to zero it before storing it in the map entry 1313 * (because it looks strange and confusing when debugging...) 1314 * 1315 * if uobj is not null 1316 * if uoffset is not UVM_UNKNOWN_OFFSET then we have a normal mapping 1317 * and we do not need to change uoffset. 1318 * if uoffset is UVM_UNKNOWN_OFFSET then we need to find the offset 1319 * now (based on the starting address of the map). this case is 1320 * for kernel object mappings where we don't know the offset until 1321 * the virtual address is found (with uvm_map_findspace). the 1322 * offset is the distance we are from the start of the map. 1323 */ 1324 1325 if (uobj == NULL) { 1326 uoffset = 0; 1327 } else { 1328 if (uoffset == UVM_UNKNOWN_OFFSET) { 1329 KASSERT(UVM_OBJ_IS_KERN_OBJECT(uobj)); 1330 uoffset = start - vm_map_min(kernel_map); 1331 } 1332 } 1333 1334 args->uma_flags = flags; 1335 args->uma_prev = prev_entry; 1336 args->uma_start = start; 1337 args->uma_size = size; 1338 args->uma_uobj = uobj; 1339 args->uma_uoffset = uoffset; 1340 1341 return 0; 1342 } 1343 1344 int 1345 uvm_map_enter(struct vm_map *map, const struct uvm_map_args *args, 1346 struct vm_map_entry *new_entry) 1347 { 1348 struct vm_map_entry *prev_entry = args->uma_prev; 1349 struct vm_map_entry *dead = NULL; 1350 1351 const uvm_flag_t flags = args->uma_flags; 1352 const vm_prot_t prot = UVM_PROTECTION(flags); 1353 const vm_prot_t maxprot = UVM_MAXPROTECTION(flags); 1354 const vm_inherit_t inherit = UVM_INHERIT(flags); 1355 const int amapwaitflag = (flags & UVM_FLAG_NOWAIT) ? 1356 AMAP_EXTEND_NOWAIT : 0; 1357 const int advice = UVM_ADVICE(flags); 1358 const int meflagval = (flags & UVM_FLAG_QUANTUM) ? 1359 UVM_MAP_QUANTUM : 0; 1360 1361 vaddr_t start = args->uma_start; 1362 vsize_t size = args->uma_size; 1363 struct uvm_object *uobj = args->uma_uobj; 1364 voff_t uoffset = args->uma_uoffset; 1365 1366 const int kmap = (vm_map_pmap(map) == pmap_kernel()); 1367 int merged = 0; 1368 int error; 1369 int newetype; 1370 1371 UVMHIST_FUNC("uvm_map_enter"); 1372 UVMHIST_CALLED(maphist); 1373 1374 UVMHIST_LOG(maphist, "(map=0x%x, start=0x%x, size=%d, flags=0x%x)", 1375 map, start, size, flags); 1376 UVMHIST_LOG(maphist, " uobj/offset 0x%x/%d", uobj, uoffset,0,0); 1377 1378 KASSERT(map->hint == prev_entry); /* bimerge case assumes this */ 1379 1380 if (flags & UVM_FLAG_QUANTUM) { 1381 KASSERT(new_entry); 1382 KASSERT(new_entry->flags & UVM_MAP_QUANTUM); 1383 } 1384 1385 if (uobj) 1386 newetype = UVM_ET_OBJ; 1387 else 1388 newetype = 0; 1389 1390 if (flags & UVM_FLAG_COPYONW) { 1391 newetype |= UVM_ET_COPYONWRITE; 1392 if ((flags & UVM_FLAG_OVERLAY) == 0) 1393 newetype |= UVM_ET_NEEDSCOPY; 1394 } 1395 1396 /* 1397 * try and insert in map by extending previous entry, if possible. 1398 * XXX: we don't try and pull back the next entry. might be useful 1399 * for a stack, but we are currently allocating our stack in advance. 1400 */ 1401 1402 if (flags & UVM_FLAG_NOMERGE) 1403 goto nomerge; 1404 1405 if (prev_entry->end == start && 1406 prev_entry != &map->header && 1407 UVM_ET_ISCOMPATIBLE(prev_entry, newetype, uobj, meflagval, 1408 prot, maxprot, inherit, advice, 0)) { 1409 1410 if (uobj && prev_entry->offset + 1411 (prev_entry->end - prev_entry->start) != uoffset) 1412 goto forwardmerge; 1413 1414 /* 1415 * can't extend a shared amap. note: no need to lock amap to 1416 * look at refs since we don't care about its exact value. 1417 * if it is one (i.e. we have only reference) it will stay there 1418 */ 1419 1420 if (prev_entry->aref.ar_amap && 1421 amap_refs(prev_entry->aref.ar_amap) != 1) { 1422 goto forwardmerge; 1423 } 1424 1425 if (prev_entry->aref.ar_amap) { 1426 error = amap_extend(prev_entry, size, 1427 amapwaitflag | AMAP_EXTEND_FORWARDS); 1428 if (error) 1429 goto nomerge; 1430 } 1431 1432 if (kmap) { 1433 UVMMAP_EVCNT_INCR(kbackmerge); 1434 } else { 1435 UVMMAP_EVCNT_INCR(ubackmerge); 1436 } 1437 UVMHIST_LOG(maphist," starting back merge", 0, 0, 0, 0); 1438 1439 /* 1440 * drop our reference to uobj since we are extending a reference 1441 * that we already have (the ref count can not drop to zero). 1442 */ 1443 1444 if (uobj && uobj->pgops->pgo_detach) 1445 uobj->pgops->pgo_detach(uobj); 1446 1447 /* 1448 * Now that we've merged the entries, note that we've grown 1449 * and our gap has shrunk. Then fix the tree. 1450 */ 1451 prev_entry->end += size; 1452 prev_entry->gap -= size; 1453 uvm_rb_fixup(map, prev_entry); 1454 1455 uvm_map_check(map, "map backmerged"); 1456 1457 UVMHIST_LOG(maphist,"<- done (via backmerge)!", 0, 0, 0, 0); 1458 merged++; 1459 } 1460 1461 forwardmerge: 1462 if (prev_entry->next->start == (start + size) && 1463 prev_entry->next != &map->header && 1464 UVM_ET_ISCOMPATIBLE(prev_entry->next, newetype, uobj, meflagval, 1465 prot, maxprot, inherit, advice, 0)) { 1466 1467 if (uobj && prev_entry->next->offset != uoffset + size) 1468 goto nomerge; 1469 1470 /* 1471 * can't extend a shared amap. note: no need to lock amap to 1472 * look at refs since we don't care about its exact value. 1473 * if it is one (i.e. we have only reference) it will stay there. 1474 * 1475 * note that we also can't merge two amaps, so if we 1476 * merged with the previous entry which has an amap, 1477 * and the next entry also has an amap, we give up. 1478 * 1479 * Interesting cases: 1480 * amap, new, amap -> give up second merge (single fwd extend) 1481 * amap, new, none -> double forward extend (extend again here) 1482 * none, new, amap -> double backward extend (done here) 1483 * uobj, new, amap -> single backward extend (done here) 1484 * 1485 * XXX should we attempt to deal with someone refilling 1486 * the deallocated region between two entries that are 1487 * backed by the same amap (ie, arefs is 2, "prev" and 1488 * "next" refer to it, and adding this allocation will 1489 * close the hole, thus restoring arefs to 1 and 1490 * deallocating the "next" vm_map_entry)? -- @@@ 1491 */ 1492 1493 if (prev_entry->next->aref.ar_amap && 1494 (amap_refs(prev_entry->next->aref.ar_amap) != 1 || 1495 (merged && prev_entry->aref.ar_amap))) { 1496 goto nomerge; 1497 } 1498 1499 if (merged) { 1500 /* 1501 * Try to extend the amap of the previous entry to 1502 * cover the next entry as well. If it doesn't work 1503 * just skip on, don't actually give up, since we've 1504 * already completed the back merge. 1505 */ 1506 if (prev_entry->aref.ar_amap) { 1507 if (amap_extend(prev_entry, 1508 prev_entry->next->end - 1509 prev_entry->next->start, 1510 amapwaitflag | AMAP_EXTEND_FORWARDS)) 1511 goto nomerge; 1512 } 1513 1514 /* 1515 * Try to extend the amap of the *next* entry 1516 * back to cover the new allocation *and* the 1517 * previous entry as well (the previous merge 1518 * didn't have an amap already otherwise we 1519 * wouldn't be checking here for an amap). If 1520 * it doesn't work just skip on, again, don't 1521 * actually give up, since we've already 1522 * completed the back merge. 1523 */ 1524 else if (prev_entry->next->aref.ar_amap) { 1525 if (amap_extend(prev_entry->next, 1526 prev_entry->end - 1527 prev_entry->start, 1528 amapwaitflag | AMAP_EXTEND_BACKWARDS)) 1529 goto nomerge; 1530 } 1531 } else { 1532 /* 1533 * Pull the next entry's amap backwards to cover this 1534 * new allocation. 1535 */ 1536 if (prev_entry->next->aref.ar_amap) { 1537 error = amap_extend(prev_entry->next, size, 1538 amapwaitflag | AMAP_EXTEND_BACKWARDS); 1539 if (error) 1540 goto nomerge; 1541 } 1542 } 1543 1544 if (merged) { 1545 if (kmap) { 1546 UVMMAP_EVCNT_DECR(kbackmerge); 1547 UVMMAP_EVCNT_INCR(kbimerge); 1548 } else { 1549 UVMMAP_EVCNT_DECR(ubackmerge); 1550 UVMMAP_EVCNT_INCR(ubimerge); 1551 } 1552 } else { 1553 if (kmap) { 1554 UVMMAP_EVCNT_INCR(kforwmerge); 1555 } else { 1556 UVMMAP_EVCNT_INCR(uforwmerge); 1557 } 1558 } 1559 UVMHIST_LOG(maphist," starting forward merge", 0, 0, 0, 0); 1560 1561 /* 1562 * drop our reference to uobj since we are extending a reference 1563 * that we already have (the ref count can not drop to zero). 1564 * (if merged, we've already detached) 1565 */ 1566 if (uobj && uobj->pgops->pgo_detach && !merged) 1567 uobj->pgops->pgo_detach(uobj); 1568 1569 if (merged) { 1570 dead = prev_entry->next; 1571 prev_entry->end = dead->end; 1572 uvm_map_entry_unlink(map, dead); 1573 if (dead->aref.ar_amap != NULL) { 1574 prev_entry->aref = dead->aref; 1575 dead->aref.ar_amap = NULL; 1576 } 1577 } else { 1578 prev_entry->next->start -= size; 1579 if (prev_entry != &map->header) { 1580 prev_entry->gap -= size; 1581 KASSERT(prev_entry->gap == uvm_rb_gap(prev_entry)); 1582 uvm_rb_fixup(map, prev_entry); 1583 } 1584 if (uobj) 1585 prev_entry->next->offset = uoffset; 1586 } 1587 1588 uvm_map_check(map, "map forwardmerged"); 1589 1590 UVMHIST_LOG(maphist,"<- done forwardmerge", 0, 0, 0, 0); 1591 merged++; 1592 } 1593 1594 nomerge: 1595 if (!merged) { 1596 UVMHIST_LOG(maphist," allocating new map entry", 0, 0, 0, 0); 1597 if (kmap) { 1598 UVMMAP_EVCNT_INCR(knomerge); 1599 } else { 1600 UVMMAP_EVCNT_INCR(unomerge); 1601 } 1602 1603 /* 1604 * allocate new entry and link it in. 1605 */ 1606 1607 if (new_entry == NULL) { 1608 new_entry = uvm_mapent_alloc(map, 1609 (flags & UVM_FLAG_NOWAIT)); 1610 if (__predict_false(new_entry == NULL)) { 1611 error = ENOMEM; 1612 goto done; 1613 } 1614 } 1615 new_entry->start = start; 1616 new_entry->end = new_entry->start + size; 1617 new_entry->object.uvm_obj = uobj; 1618 new_entry->offset = uoffset; 1619 1620 new_entry->etype = newetype; 1621 1622 if (flags & UVM_FLAG_NOMERGE) { 1623 new_entry->flags |= UVM_MAP_NOMERGE; 1624 } 1625 1626 new_entry->protection = prot; 1627 new_entry->max_protection = maxprot; 1628 new_entry->inheritance = inherit; 1629 new_entry->wired_count = 0; 1630 new_entry->advice = advice; 1631 if (flags & UVM_FLAG_OVERLAY) { 1632 1633 /* 1634 * to_add: for BSS we overallocate a little since we 1635 * are likely to extend 1636 */ 1637 1638 vaddr_t to_add = (flags & UVM_FLAG_AMAPPAD) ? 1639 UVM_AMAP_CHUNK << PAGE_SHIFT : 0; 1640 struct vm_amap *amap = amap_alloc(size, to_add, 1641 (flags & UVM_FLAG_NOWAIT)); 1642 if (__predict_false(amap == NULL)) { 1643 error = ENOMEM; 1644 goto done; 1645 } 1646 new_entry->aref.ar_pageoff = 0; 1647 new_entry->aref.ar_amap = amap; 1648 } else { 1649 new_entry->aref.ar_pageoff = 0; 1650 new_entry->aref.ar_amap = NULL; 1651 } 1652 uvm_map_entry_link(map, prev_entry, new_entry); 1653 1654 /* 1655 * Update the free space hint 1656 */ 1657 1658 if ((map->first_free == prev_entry) && 1659 (prev_entry->end >= new_entry->start)) 1660 map->first_free = new_entry; 1661 1662 new_entry = NULL; 1663 } 1664 1665 map->size += size; 1666 1667 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 1668 1669 error = 0; 1670 done: 1671 if ((flags & UVM_FLAG_QUANTUM) == 0) { 1672 /* 1673 * vmk_merged_entries is locked by the map's lock. 1674 */ 1675 vm_map_unlock(map); 1676 } 1677 if (new_entry && error == 0) { 1678 KDASSERT(merged); 1679 uvm_mapent_free_merged(map, new_entry); 1680 new_entry = NULL; 1681 } 1682 if (dead) { 1683 KDASSERT(merged); 1684 uvm_mapent_free_merged(map, dead); 1685 } 1686 if ((flags & UVM_FLAG_QUANTUM) != 0) { 1687 vm_map_unlock(map); 1688 } 1689 if (new_entry != NULL) { 1690 uvm_mapent_free(new_entry); 1691 } 1692 return error; 1693 } 1694 1695 /* 1696 * uvm_map_lookup_entry_bytree: lookup an entry in tree 1697 */ 1698 1699 static inline bool 1700 uvm_map_lookup_entry_bytree(struct vm_map *map, vaddr_t address, 1701 struct vm_map_entry **entry /* OUT */) 1702 { 1703 struct vm_map_entry *prev = &map->header; 1704 struct vm_map_entry *cur = ROOT_ENTRY(map); 1705 1706 while (cur) { 1707 UVMMAP_EVCNT_INCR(mlk_treeloop); 1708 if (address >= cur->start) { 1709 if (address < cur->end) { 1710 *entry = cur; 1711 return true; 1712 } 1713 prev = cur; 1714 cur = RIGHT_ENTRY(cur); 1715 } else 1716 cur = LEFT_ENTRY(cur); 1717 } 1718 *entry = prev; 1719 return false; 1720 } 1721 1722 /* 1723 * uvm_map_lookup_entry: find map entry at or before an address 1724 * 1725 * => map must at least be read-locked by caller 1726 * => entry is returned in "entry" 1727 * => return value is true if address is in the returned entry 1728 */ 1729 1730 bool 1731 uvm_map_lookup_entry(struct vm_map *map, vaddr_t address, 1732 struct vm_map_entry **entry /* OUT */) 1733 { 1734 struct vm_map_entry *cur; 1735 bool use_tree = false; 1736 UVMHIST_FUNC("uvm_map_lookup_entry"); 1737 UVMHIST_CALLED(maphist); 1738 1739 UVMHIST_LOG(maphist,"(map=0x%x,addr=0x%x,ent=0x%x)", 1740 map, address, entry, 0); 1741 1742 /* 1743 * start looking either from the head of the 1744 * list, or from the hint. 1745 */ 1746 1747 cur = map->hint; 1748 1749 if (cur == &map->header) 1750 cur = cur->next; 1751 1752 UVMMAP_EVCNT_INCR(mlk_call); 1753 if (address >= cur->start) { 1754 1755 /* 1756 * go from hint to end of list. 1757 * 1758 * but first, make a quick check to see if 1759 * we are already looking at the entry we 1760 * want (which is usually the case). 1761 * note also that we don't need to save the hint 1762 * here... it is the same hint (unless we are 1763 * at the header, in which case the hint didn't 1764 * buy us anything anyway). 1765 */ 1766 1767 if (cur != &map->header && cur->end > address) { 1768 UVMMAP_EVCNT_INCR(mlk_hint); 1769 *entry = cur; 1770 UVMHIST_LOG(maphist,"<- got it via hint (0x%x)", 1771 cur, 0, 0, 0); 1772 uvm_mapent_check(*entry); 1773 return (true); 1774 } 1775 1776 if (map->nentries > 15) 1777 use_tree = true; 1778 } else { 1779 1780 /* 1781 * invalid hint. use tree. 1782 */ 1783 use_tree = true; 1784 } 1785 1786 uvm_map_check(map, __func__); 1787 1788 if (use_tree) { 1789 /* 1790 * Simple lookup in the tree. Happens when the hint is 1791 * invalid, or nentries reach a threshold. 1792 */ 1793 UVMMAP_EVCNT_INCR(mlk_tree); 1794 if (uvm_map_lookup_entry_bytree(map, address, entry)) { 1795 goto got; 1796 } else { 1797 goto failed; 1798 } 1799 } 1800 1801 /* 1802 * search linearly 1803 */ 1804 1805 UVMMAP_EVCNT_INCR(mlk_list); 1806 while (cur != &map->header) { 1807 UVMMAP_EVCNT_INCR(mlk_listloop); 1808 if (cur->end > address) { 1809 if (address >= cur->start) { 1810 /* 1811 * save this lookup for future 1812 * hints, and return 1813 */ 1814 1815 *entry = cur; 1816 got: 1817 SAVE_HINT(map, map->hint, *entry); 1818 UVMHIST_LOG(maphist,"<- search got it (0x%x)", 1819 cur, 0, 0, 0); 1820 KDASSERT((*entry)->start <= address); 1821 KDASSERT(address < (*entry)->end); 1822 uvm_mapent_check(*entry); 1823 return (true); 1824 } 1825 break; 1826 } 1827 cur = cur->next; 1828 } 1829 *entry = cur->prev; 1830 failed: 1831 SAVE_HINT(map, map->hint, *entry); 1832 UVMHIST_LOG(maphist,"<- failed!",0,0,0,0); 1833 KDASSERT((*entry) == &map->header || (*entry)->end <= address); 1834 KDASSERT((*entry)->next == &map->header || 1835 address < (*entry)->next->start); 1836 return (false); 1837 } 1838 1839 /* 1840 * See if the range between start and start + length fits in the gap 1841 * entry->next->start and entry->end. Returns 1 if fits, 0 if doesn't 1842 * fit, and -1 address wraps around. 1843 */ 1844 static int 1845 uvm_map_space_avail(vaddr_t *start, vsize_t length, voff_t uoffset, 1846 vsize_t align, int topdown, struct vm_map_entry *entry) 1847 { 1848 vaddr_t end; 1849 1850 #ifdef PMAP_PREFER 1851 /* 1852 * push start address forward as needed to avoid VAC alias problems. 1853 * we only do this if a valid offset is specified. 1854 */ 1855 1856 if (uoffset != UVM_UNKNOWN_OFFSET) 1857 PMAP_PREFER(uoffset, start, length, topdown); 1858 #endif 1859 if (align != 0) { 1860 if ((*start & (align - 1)) != 0) { 1861 if (topdown) 1862 *start &= ~(align - 1); 1863 else 1864 *start = roundup(*start, align); 1865 } 1866 /* 1867 * XXX Should we PMAP_PREFER() here again? 1868 * eh...i think we're okay 1869 */ 1870 } 1871 1872 /* 1873 * Find the end of the proposed new region. Be sure we didn't 1874 * wrap around the address; if so, we lose. Otherwise, if the 1875 * proposed new region fits before the next entry, we win. 1876 */ 1877 1878 end = *start + length; 1879 if (end < *start) 1880 return (-1); 1881 1882 if (entry->next->start >= end && *start >= entry->end) 1883 return (1); 1884 1885 return (0); 1886 } 1887 1888 /* 1889 * uvm_map_findspace: find "length" sized space in "map". 1890 * 1891 * => "hint" is a hint about where we want it, unless UVM_FLAG_FIXED is 1892 * set in "flags" (in which case we insist on using "hint"). 1893 * => "result" is VA returned 1894 * => uobj/uoffset are to be used to handle VAC alignment, if required 1895 * => if "align" is non-zero, we attempt to align to that value. 1896 * => caller must at least have read-locked map 1897 * => returns NULL on failure, or pointer to prev. map entry if success 1898 * => note this is a cross between the old vm_map_findspace and vm_map_find 1899 */ 1900 1901 struct vm_map_entry * 1902 uvm_map_findspace(struct vm_map *map, vaddr_t hint, vsize_t length, 1903 vaddr_t *result /* OUT */, struct uvm_object *uobj, voff_t uoffset, 1904 vsize_t align, int flags) 1905 { 1906 struct vm_map_entry *entry; 1907 struct vm_map_entry *child, *prev, *tmp; 1908 vaddr_t orig_hint; 1909 const int topdown = map->flags & VM_MAP_TOPDOWN; 1910 UVMHIST_FUNC("uvm_map_findspace"); 1911 UVMHIST_CALLED(maphist); 1912 1913 UVMHIST_LOG(maphist, "(map=0x%x, hint=0x%x, len=%d, flags=0x%x)", 1914 map, hint, length, flags); 1915 KASSERT((align & (align - 1)) == 0); 1916 KASSERT((flags & UVM_FLAG_FIXED) == 0 || align == 0); 1917 1918 uvm_map_check(map, "map_findspace entry"); 1919 1920 /* 1921 * remember the original hint. if we are aligning, then we 1922 * may have to try again with no alignment constraint if 1923 * we fail the first time. 1924 */ 1925 1926 orig_hint = hint; 1927 if (hint < vm_map_min(map)) { /* check ranges ... */ 1928 if (flags & UVM_FLAG_FIXED) { 1929 UVMHIST_LOG(maphist,"<- VA below map range",0,0,0,0); 1930 return (NULL); 1931 } 1932 hint = vm_map_min(map); 1933 } 1934 if (hint > vm_map_max(map)) { 1935 UVMHIST_LOG(maphist,"<- VA 0x%x > range [0x%x->0x%x]", 1936 hint, vm_map_min(map), vm_map_max(map), 0); 1937 return (NULL); 1938 } 1939 1940 /* 1941 * Look for the first possible address; if there's already 1942 * something at this address, we have to start after it. 1943 */ 1944 1945 /* 1946 * @@@: there are four, no, eight cases to consider. 1947 * 1948 * 0: found, fixed, bottom up -> fail 1949 * 1: found, fixed, top down -> fail 1950 * 2: found, not fixed, bottom up -> start after entry->end, 1951 * loop up 1952 * 3: found, not fixed, top down -> start before entry->start, 1953 * loop down 1954 * 4: not found, fixed, bottom up -> check entry->next->start, fail 1955 * 5: not found, fixed, top down -> check entry->next->start, fail 1956 * 6: not found, not fixed, bottom up -> check entry->next->start, 1957 * loop up 1958 * 7: not found, not fixed, top down -> check entry->next->start, 1959 * loop down 1960 * 1961 * as you can see, it reduces to roughly five cases, and that 1962 * adding top down mapping only adds one unique case (without 1963 * it, there would be four cases). 1964 */ 1965 1966 if ((flags & UVM_FLAG_FIXED) == 0 && hint == vm_map_min(map)) { 1967 entry = map->first_free; 1968 } else { 1969 if (uvm_map_lookup_entry(map, hint, &entry)) { 1970 /* "hint" address already in use ... */ 1971 if (flags & UVM_FLAG_FIXED) { 1972 UVMHIST_LOG(maphist, "<- fixed & VA in use", 1973 0, 0, 0, 0); 1974 return (NULL); 1975 } 1976 if (topdown) 1977 /* Start from lower gap. */ 1978 entry = entry->prev; 1979 } else if (flags & UVM_FLAG_FIXED) { 1980 if (entry->next->start >= hint + length && 1981 hint + length > hint) 1982 goto found; 1983 1984 /* "hint" address is gap but too small */ 1985 UVMHIST_LOG(maphist, "<- fixed mapping failed", 1986 0, 0, 0, 0); 1987 return (NULL); /* only one shot at it ... */ 1988 } else { 1989 /* 1990 * See if given hint fits in this gap. 1991 */ 1992 switch (uvm_map_space_avail(&hint, length, 1993 uoffset, align, topdown, entry)) { 1994 case 1: 1995 goto found; 1996 case -1: 1997 goto wraparound; 1998 } 1999 2000 if (topdown) { 2001 /* 2002 * Still there is a chance to fit 2003 * if hint > entry->end. 2004 */ 2005 } else { 2006 /* Start from higher gap. */ 2007 entry = entry->next; 2008 if (entry == &map->header) 2009 goto notfound; 2010 goto nextgap; 2011 } 2012 } 2013 } 2014 2015 /* 2016 * Note that all UVM_FLAGS_FIXED case is already handled. 2017 */ 2018 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2019 2020 /* Try to find the space in the red-black tree */ 2021 2022 /* Check slot before any entry */ 2023 hint = topdown ? entry->next->start - length : entry->end; 2024 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2025 topdown, entry)) { 2026 case 1: 2027 goto found; 2028 case -1: 2029 goto wraparound; 2030 } 2031 2032 nextgap: 2033 KDASSERT((flags & UVM_FLAG_FIXED) == 0); 2034 /* If there is not enough space in the whole tree, we fail */ 2035 tmp = ROOT_ENTRY(map); 2036 if (tmp == NULL || tmp->maxgap < length) 2037 goto notfound; 2038 2039 prev = NULL; /* previous candidate */ 2040 2041 /* Find an entry close to hint that has enough space */ 2042 for (; tmp;) { 2043 KASSERT(tmp->next->start == tmp->end + tmp->gap); 2044 if (topdown) { 2045 if (tmp->next->start < hint + length && 2046 (prev == NULL || tmp->end > prev->end)) { 2047 if (tmp->gap >= length) 2048 prev = tmp; 2049 else if ((child = LEFT_ENTRY(tmp)) != NULL 2050 && child->maxgap >= length) 2051 prev = tmp; 2052 } 2053 } else { 2054 if (tmp->end >= hint && 2055 (prev == NULL || tmp->end < prev->end)) { 2056 if (tmp->gap >= length) 2057 prev = tmp; 2058 else if ((child = RIGHT_ENTRY(tmp)) != NULL 2059 && child->maxgap >= length) 2060 prev = tmp; 2061 } 2062 } 2063 if (tmp->next->start < hint + length) 2064 child = RIGHT_ENTRY(tmp); 2065 else if (tmp->end > hint) 2066 child = LEFT_ENTRY(tmp); 2067 else { 2068 if (tmp->gap >= length) 2069 break; 2070 if (topdown) 2071 child = LEFT_ENTRY(tmp); 2072 else 2073 child = RIGHT_ENTRY(tmp); 2074 } 2075 if (child == NULL || child->maxgap < length) 2076 break; 2077 tmp = child; 2078 } 2079 2080 if (tmp != NULL && tmp->start < hint && hint < tmp->next->start) { 2081 /* 2082 * Check if the entry that we found satifies the 2083 * space requirement 2084 */ 2085 if (topdown) { 2086 if (hint > tmp->next->start - length) 2087 hint = tmp->next->start - length; 2088 } else { 2089 if (hint < tmp->end) 2090 hint = tmp->end; 2091 } 2092 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2093 topdown, tmp)) { 2094 case 1: 2095 entry = tmp; 2096 goto found; 2097 case -1: 2098 goto wraparound; 2099 } 2100 if (tmp->gap >= length) 2101 goto listsearch; 2102 } 2103 if (prev == NULL) 2104 goto notfound; 2105 2106 if (topdown) { 2107 KASSERT(orig_hint >= prev->next->start - length || 2108 prev->next->start - length > prev->next->start); 2109 hint = prev->next->start - length; 2110 } else { 2111 KASSERT(orig_hint <= prev->end); 2112 hint = prev->end; 2113 } 2114 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2115 topdown, prev)) { 2116 case 1: 2117 entry = prev; 2118 goto found; 2119 case -1: 2120 goto wraparound; 2121 } 2122 if (prev->gap >= length) 2123 goto listsearch; 2124 2125 if (topdown) 2126 tmp = LEFT_ENTRY(prev); 2127 else 2128 tmp = RIGHT_ENTRY(prev); 2129 for (;;) { 2130 KASSERT(tmp && tmp->maxgap >= length); 2131 if (topdown) 2132 child = RIGHT_ENTRY(tmp); 2133 else 2134 child = LEFT_ENTRY(tmp); 2135 if (child && child->maxgap >= length) { 2136 tmp = child; 2137 continue; 2138 } 2139 if (tmp->gap >= length) 2140 break; 2141 if (topdown) 2142 tmp = LEFT_ENTRY(tmp); 2143 else 2144 tmp = RIGHT_ENTRY(tmp); 2145 } 2146 2147 if (topdown) { 2148 KASSERT(orig_hint >= tmp->next->start - length || 2149 tmp->next->start - length > tmp->next->start); 2150 hint = tmp->next->start - length; 2151 } else { 2152 KASSERT(orig_hint <= tmp->end); 2153 hint = tmp->end; 2154 } 2155 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2156 topdown, tmp)) { 2157 case 1: 2158 entry = tmp; 2159 goto found; 2160 case -1: 2161 goto wraparound; 2162 } 2163 2164 /* 2165 * The tree fails to find an entry because of offset or alignment 2166 * restrictions. Search the list instead. 2167 */ 2168 listsearch: 2169 /* 2170 * Look through the rest of the map, trying to fit a new region in 2171 * the gap between existing regions, or after the very last region. 2172 * note: entry->end = base VA of current gap, 2173 * entry->next->start = VA of end of current gap 2174 */ 2175 2176 for (;;) { 2177 /* Update hint for current gap. */ 2178 hint = topdown ? entry->next->start - length : entry->end; 2179 2180 /* See if it fits. */ 2181 switch (uvm_map_space_avail(&hint, length, uoffset, align, 2182 topdown, entry)) { 2183 case 1: 2184 goto found; 2185 case -1: 2186 goto wraparound; 2187 } 2188 2189 /* Advance to next/previous gap */ 2190 if (topdown) { 2191 if (entry == &map->header) { 2192 UVMHIST_LOG(maphist, "<- failed (off start)", 2193 0,0,0,0); 2194 goto notfound; 2195 } 2196 entry = entry->prev; 2197 } else { 2198 entry = entry->next; 2199 if (entry == &map->header) { 2200 UVMHIST_LOG(maphist, "<- failed (off end)", 2201 0,0,0,0); 2202 goto notfound; 2203 } 2204 } 2205 } 2206 2207 found: 2208 SAVE_HINT(map, map->hint, entry); 2209 *result = hint; 2210 UVMHIST_LOG(maphist,"<- got it! (result=0x%x)", hint, 0,0,0); 2211 KASSERT( topdown || hint >= orig_hint); 2212 KASSERT(!topdown || hint <= orig_hint); 2213 KASSERT(entry->end <= hint); 2214 KASSERT(hint + length <= entry->next->start); 2215 return (entry); 2216 2217 wraparound: 2218 UVMHIST_LOG(maphist, "<- failed (wrap around)", 0,0,0,0); 2219 2220 return (NULL); 2221 2222 notfound: 2223 UVMHIST_LOG(maphist, "<- failed (notfound)", 0,0,0,0); 2224 2225 return (NULL); 2226 } 2227 2228 /* 2229 * U N M A P - m a i n h e l p e r f u n c t i o n s 2230 */ 2231 2232 /* 2233 * uvm_unmap_remove: remove mappings from a vm_map (from "start" up to "stop") 2234 * 2235 * => caller must check alignment and size 2236 * => map must be locked by caller 2237 * => we return a list of map entries that we've remove from the map 2238 * in "entry_list" 2239 */ 2240 2241 void 2242 uvm_unmap_remove(struct vm_map *map, vaddr_t start, vaddr_t end, 2243 struct vm_map_entry **entry_list /* OUT */, 2244 struct uvm_mapent_reservation *umr, int flags) 2245 { 2246 struct vm_map_entry *entry, *first_entry, *next; 2247 vaddr_t len; 2248 UVMHIST_FUNC("uvm_unmap_remove"); UVMHIST_CALLED(maphist); 2249 2250 UVMHIST_LOG(maphist,"(map=0x%x, start=0x%x, end=0x%x)", 2251 map, start, end, 0); 2252 VM_MAP_RANGE_CHECK(map, start, end); 2253 2254 uvm_map_check(map, "unmap_remove entry"); 2255 2256 /* 2257 * find first entry 2258 */ 2259 2260 if (uvm_map_lookup_entry(map, start, &first_entry) == true) { 2261 /* clip and go... */ 2262 entry = first_entry; 2263 UVM_MAP_CLIP_START(map, entry, start, umr); 2264 /* critical! prevents stale hint */ 2265 SAVE_HINT(map, entry, entry->prev); 2266 } else { 2267 entry = first_entry->next; 2268 } 2269 2270 /* 2271 * Save the free space hint 2272 */ 2273 2274 if (map->first_free != &map->header && map->first_free->start >= start) 2275 map->first_free = entry->prev; 2276 2277 /* 2278 * note: we now re-use first_entry for a different task. we remove 2279 * a number of map entries from the map and save them in a linked 2280 * list headed by "first_entry". once we remove them from the map 2281 * the caller should unlock the map and drop the references to the 2282 * backing objects [c.f. uvm_unmap_detach]. the object is to 2283 * separate unmapping from reference dropping. why? 2284 * [1] the map has to be locked for unmapping 2285 * [2] the map need not be locked for reference dropping 2286 * [3] dropping references may trigger pager I/O, and if we hit 2287 * a pager that does synchronous I/O we may have to wait for it. 2288 * [4] we would like all waiting for I/O to occur with maps unlocked 2289 * so that we don't block other threads. 2290 */ 2291 2292 first_entry = NULL; 2293 *entry_list = NULL; 2294 2295 /* 2296 * break up the area into map entry sized regions and unmap. note 2297 * that all mappings have to be removed before we can even consider 2298 * dropping references to amaps or VM objects (otherwise we could end 2299 * up with a mapping to a page on the free list which would be very bad) 2300 */ 2301 2302 while ((entry != &map->header) && (entry->start < end)) { 2303 KASSERT((entry->flags & UVM_MAP_FIRST) == 0); 2304 2305 UVM_MAP_CLIP_END(map, entry, end, umr); 2306 next = entry->next; 2307 len = entry->end - entry->start; 2308 2309 /* 2310 * unwire before removing addresses from the pmap; otherwise 2311 * unwiring will put the entries back into the pmap (XXX). 2312 */ 2313 2314 if (VM_MAPENT_ISWIRED(entry)) { 2315 uvm_map_entry_unwire(map, entry); 2316 } 2317 if (flags & UVM_FLAG_VAONLY) { 2318 2319 /* nothing */ 2320 2321 } else if ((map->flags & VM_MAP_PAGEABLE) == 0) { 2322 2323 /* 2324 * if the map is non-pageable, any pages mapped there 2325 * must be wired and entered with pmap_kenter_pa(), 2326 * and we should free any such pages immediately. 2327 * this is mostly used for kmem_map and mb_map. 2328 */ 2329 2330 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2331 uvm_km_pgremove_intrsafe(entry->start, 2332 entry->end); 2333 pmap_kremove(entry->start, len); 2334 } 2335 } else if (UVM_ET_ISOBJ(entry) && 2336 UVM_OBJ_IS_KERN_OBJECT(entry->object.uvm_obj)) { 2337 KASSERT(vm_map_pmap(map) == pmap_kernel()); 2338 2339 /* 2340 * note: kernel object mappings are currently used in 2341 * two ways: 2342 * [1] "normal" mappings of pages in the kernel object 2343 * [2] uvm_km_valloc'd allocations in which we 2344 * pmap_enter in some non-kernel-object page 2345 * (e.g. vmapbuf). 2346 * 2347 * for case [1], we need to remove the mapping from 2348 * the pmap and then remove the page from the kernel 2349 * object (because, once pages in a kernel object are 2350 * unmapped they are no longer needed, unlike, say, 2351 * a vnode where you might want the data to persist 2352 * until flushed out of a queue). 2353 * 2354 * for case [2], we need to remove the mapping from 2355 * the pmap. there shouldn't be any pages at the 2356 * specified offset in the kernel object [but it 2357 * doesn't hurt to call uvm_km_pgremove just to be 2358 * safe?] 2359 * 2360 * uvm_km_pgremove currently does the following: 2361 * for pages in the kernel object in range: 2362 * - drops the swap slot 2363 * - uvm_pagefree the page 2364 */ 2365 2366 /* 2367 * remove mappings from pmap and drop the pages 2368 * from the object. offsets are always relative 2369 * to vm_map_min(kernel_map). 2370 */ 2371 2372 pmap_remove(pmap_kernel(), entry->start, 2373 entry->start + len); 2374 uvm_km_pgremove(entry->start, entry->end); 2375 2376 /* 2377 * null out kernel_object reference, we've just 2378 * dropped it 2379 */ 2380 2381 entry->etype &= ~UVM_ET_OBJ; 2382 entry->object.uvm_obj = NULL; 2383 } else if (UVM_ET_ISOBJ(entry) || entry->aref.ar_amap) { 2384 2385 /* 2386 * remove mappings the standard way. 2387 */ 2388 2389 pmap_remove(map->pmap, entry->start, entry->end); 2390 } 2391 2392 #if defined(DEBUG) 2393 if ((entry->flags & UVM_MAP_KMAPENT) == 0) { 2394 2395 /* 2396 * check if there's remaining mapping, 2397 * which is a bug in caller. 2398 */ 2399 2400 vaddr_t va; 2401 for (va = entry->start; va < entry->end; 2402 va += PAGE_SIZE) { 2403 if (pmap_extract(vm_map_pmap(map), va, NULL)) { 2404 panic("uvm_unmap_remove: has mapping"); 2405 } 2406 } 2407 2408 if (VM_MAP_IS_KERNEL(map)) { 2409 uvm_km_check_empty(entry->start, entry->end, 2410 (map->flags & VM_MAP_INTRSAFE) != 0); 2411 } 2412 } 2413 #endif /* defined(DEBUG) */ 2414 2415 /* 2416 * remove entry from map and put it on our list of entries 2417 * that we've nuked. then go to next entry. 2418 */ 2419 2420 UVMHIST_LOG(maphist, " removed map entry 0x%x", entry, 0, 0,0); 2421 2422 /* critical! prevents stale hint */ 2423 SAVE_HINT(map, entry, entry->prev); 2424 2425 uvm_map_entry_unlink(map, entry); 2426 KASSERT(map->size >= len); 2427 map->size -= len; 2428 entry->prev = NULL; 2429 entry->next = first_entry; 2430 first_entry = entry; 2431 entry = next; 2432 } 2433 if ((map->flags & VM_MAP_DYING) == 0) { 2434 pmap_update(vm_map_pmap(map)); 2435 } 2436 2437 uvm_map_check(map, "unmap_remove leave"); 2438 2439 /* 2440 * now we've cleaned up the map and are ready for the caller to drop 2441 * references to the mapped objects. 2442 */ 2443 2444 *entry_list = first_entry; 2445 UVMHIST_LOG(maphist,"<- done!", 0, 0, 0, 0); 2446 2447 if (map->flags & VM_MAP_WANTVA) { 2448 mutex_enter(&map->misc_lock); 2449 map->flags &= ~VM_MAP_WANTVA; 2450 cv_broadcast(&map->cv); 2451 mutex_exit(&map->misc_lock); 2452 } 2453 } 2454 2455 /* 2456 * uvm_unmap_detach: drop references in a chain of map entries 2457 * 2458 * => we will free the map entries as we traverse the list. 2459 */ 2460 2461 void 2462 uvm_unmap_detach(struct vm_map_entry *first_entry, int flags) 2463 { 2464 struct vm_map_entry *next_entry; 2465 UVMHIST_FUNC("uvm_unmap_detach"); UVMHIST_CALLED(maphist); 2466 2467 while (first_entry) { 2468 KASSERT(!VM_MAPENT_ISWIRED(first_entry)); 2469 UVMHIST_LOG(maphist, 2470 " detach 0x%x: amap=0x%x, obj=0x%x, submap?=%d", 2471 first_entry, first_entry->aref.ar_amap, 2472 first_entry->object.uvm_obj, 2473 UVM_ET_ISSUBMAP(first_entry)); 2474 2475 /* 2476 * drop reference to amap, if we've got one 2477 */ 2478 2479 if (first_entry->aref.ar_amap) 2480 uvm_map_unreference_amap(first_entry, flags); 2481 2482 /* 2483 * drop reference to our backing object, if we've got one 2484 */ 2485 2486 KASSERT(!UVM_ET_ISSUBMAP(first_entry)); 2487 if (UVM_ET_ISOBJ(first_entry) && 2488 first_entry->object.uvm_obj->pgops->pgo_detach) { 2489 (*first_entry->object.uvm_obj->pgops->pgo_detach) 2490 (first_entry->object.uvm_obj); 2491 } 2492 next_entry = first_entry->next; 2493 uvm_mapent_free(first_entry); 2494 first_entry = next_entry; 2495 } 2496 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 2497 } 2498 2499 /* 2500 * E X T R A C T I O N F U N C T I O N S 2501 */ 2502 2503 /* 2504 * uvm_map_reserve: reserve space in a vm_map for future use. 2505 * 2506 * => we reserve space in a map by putting a dummy map entry in the 2507 * map (dummy means obj=NULL, amap=NULL, prot=VM_PROT_NONE) 2508 * => map should be unlocked (we will write lock it) 2509 * => we return true if we were able to reserve space 2510 * => XXXCDC: should be inline? 2511 */ 2512 2513 int 2514 uvm_map_reserve(struct vm_map *map, vsize_t size, 2515 vaddr_t offset /* hint for pmap_prefer */, 2516 vsize_t align /* alignment */, 2517 vaddr_t *raddr /* IN:hint, OUT: reserved VA */, 2518 uvm_flag_t flags /* UVM_FLAG_FIXED or 0 */) 2519 { 2520 UVMHIST_FUNC("uvm_map_reserve"); UVMHIST_CALLED(maphist); 2521 2522 UVMHIST_LOG(maphist, "(map=0x%x, size=0x%x, offset=0x%x,addr=0x%x)", 2523 map,size,offset,raddr); 2524 2525 size = round_page(size); 2526 2527 /* 2528 * reserve some virtual space. 2529 */ 2530 2531 if (uvm_map(map, raddr, size, NULL, offset, align, 2532 UVM_MAPFLAG(UVM_PROT_NONE, UVM_PROT_NONE, UVM_INH_NONE, 2533 UVM_ADV_RANDOM, UVM_FLAG_NOMERGE|flags)) != 0) { 2534 UVMHIST_LOG(maphist, "<- done (no VM)", 0,0,0,0); 2535 return (false); 2536 } 2537 2538 UVMHIST_LOG(maphist, "<- done (*raddr=0x%x)", *raddr,0,0,0); 2539 return (true); 2540 } 2541 2542 /* 2543 * uvm_map_replace: replace a reserved (blank) area of memory with 2544 * real mappings. 2545 * 2546 * => caller must WRITE-LOCK the map 2547 * => we return true if replacement was a success 2548 * => we expect the newents chain to have nnewents entrys on it and 2549 * we expect newents->prev to point to the last entry on the list 2550 * => note newents is allowed to be NULL 2551 */ 2552 2553 int 2554 uvm_map_replace(struct vm_map *map, vaddr_t start, vaddr_t end, 2555 struct vm_map_entry *newents, int nnewents, struct vm_map_entry **oldentryp) 2556 { 2557 struct vm_map_entry *oldent, *last; 2558 2559 uvm_map_check(map, "map_replace entry"); 2560 2561 /* 2562 * first find the blank map entry at the specified address 2563 */ 2564 2565 if (!uvm_map_lookup_entry(map, start, &oldent)) { 2566 return (false); 2567 } 2568 2569 /* 2570 * check to make sure we have a proper blank entry 2571 */ 2572 2573 if (end < oldent->end && !VM_MAP_USE_KMAPENT(map)) { 2574 UVM_MAP_CLIP_END(map, oldent, end, NULL); 2575 } 2576 if (oldent->start != start || oldent->end != end || 2577 oldent->object.uvm_obj != NULL || oldent->aref.ar_amap != NULL) { 2578 return (false); 2579 } 2580 2581 #ifdef DIAGNOSTIC 2582 2583 /* 2584 * sanity check the newents chain 2585 */ 2586 2587 { 2588 struct vm_map_entry *tmpent = newents; 2589 int nent = 0; 2590 vaddr_t cur = start; 2591 2592 while (tmpent) { 2593 nent++; 2594 if (tmpent->start < cur) 2595 panic("uvm_map_replace1"); 2596 if (tmpent->start > tmpent->end || tmpent->end > end) { 2597 printf("tmpent->start=0x%lx, tmpent->end=0x%lx, end=0x%lx\n", 2598 tmpent->start, tmpent->end, end); 2599 panic("uvm_map_replace2"); 2600 } 2601 cur = tmpent->end; 2602 if (tmpent->next) { 2603 if (tmpent->next->prev != tmpent) 2604 panic("uvm_map_replace3"); 2605 } else { 2606 if (newents->prev != tmpent) 2607 panic("uvm_map_replace4"); 2608 } 2609 tmpent = tmpent->next; 2610 } 2611 if (nent != nnewents) 2612 panic("uvm_map_replace5"); 2613 } 2614 #endif 2615 2616 /* 2617 * map entry is a valid blank! replace it. (this does all the 2618 * work of map entry link/unlink...). 2619 */ 2620 2621 if (newents) { 2622 last = newents->prev; 2623 2624 /* critical: flush stale hints out of map */ 2625 SAVE_HINT(map, map->hint, newents); 2626 if (map->first_free == oldent) 2627 map->first_free = last; 2628 2629 last->next = oldent->next; 2630 last->next->prev = last; 2631 2632 /* Fix RB tree */ 2633 uvm_rb_remove(map, oldent); 2634 2635 newents->prev = oldent->prev; 2636 newents->prev->next = newents; 2637 map->nentries = map->nentries + (nnewents - 1); 2638 2639 /* Fixup the RB tree */ 2640 { 2641 int i; 2642 struct vm_map_entry *tmp; 2643 2644 tmp = newents; 2645 for (i = 0; i < nnewents && tmp; i++) { 2646 uvm_rb_insert(map, tmp); 2647 tmp = tmp->next; 2648 } 2649 } 2650 } else { 2651 /* NULL list of new entries: just remove the old one */ 2652 clear_hints(map, oldent); 2653 uvm_map_entry_unlink(map, oldent); 2654 } 2655 2656 uvm_map_check(map, "map_replace leave"); 2657 2658 /* 2659 * now we can free the old blank entry and return. 2660 */ 2661 2662 *oldentryp = oldent; 2663 return (true); 2664 } 2665 2666 /* 2667 * uvm_map_extract: extract a mapping from a map and put it somewhere 2668 * (maybe removing the old mapping) 2669 * 2670 * => maps should be unlocked (we will write lock them) 2671 * => returns 0 on success, error code otherwise 2672 * => start must be page aligned 2673 * => len must be page sized 2674 * => flags: 2675 * UVM_EXTRACT_REMOVE: remove mappings from srcmap 2676 * UVM_EXTRACT_CONTIG: abort if unmapped area (advisory only) 2677 * UVM_EXTRACT_QREF: for a temporary extraction do quick obj refs 2678 * UVM_EXTRACT_FIXPROT: set prot to maxprot as we go 2679 * >>>NOTE: if you set REMOVE, you are not allowed to use CONTIG or QREF!<<< 2680 * >>>NOTE: QREF's must be unmapped via the QREF path, thus should only 2681 * be used from within the kernel in a kernel level map <<< 2682 */ 2683 2684 int 2685 uvm_map_extract(struct vm_map *srcmap, vaddr_t start, vsize_t len, 2686 struct vm_map *dstmap, vaddr_t *dstaddrp, int flags) 2687 { 2688 vaddr_t dstaddr, end, newend, oldoffset, fudge, orig_fudge; 2689 struct vm_map_entry *chain, *endchain, *entry, *orig_entry, *newentry, 2690 *deadentry, *oldentry; 2691 struct vm_map_entry *resentry = NULL; /* a dummy reservation entry */ 2692 vsize_t elen; 2693 int nchain, error, copy_ok; 2694 UVMHIST_FUNC("uvm_map_extract"); UVMHIST_CALLED(maphist); 2695 2696 UVMHIST_LOG(maphist,"(srcmap=0x%x,start=0x%x, len=0x%x", srcmap, start, 2697 len,0); 2698 UVMHIST_LOG(maphist," ...,dstmap=0x%x, flags=0x%x)", dstmap,flags,0,0); 2699 2700 uvm_map_check(srcmap, "map_extract src enter"); 2701 uvm_map_check(dstmap, "map_extract dst enter"); 2702 2703 /* 2704 * step 0: sanity check: start must be on a page boundary, length 2705 * must be page sized. can't ask for CONTIG/QREF if you asked for 2706 * REMOVE. 2707 */ 2708 2709 KASSERT((start & PAGE_MASK) == 0 && (len & PAGE_MASK) == 0); 2710 KASSERT((flags & UVM_EXTRACT_REMOVE) == 0 || 2711 (flags & (UVM_EXTRACT_CONTIG|UVM_EXTRACT_QREF)) == 0); 2712 2713 /* 2714 * step 1: reserve space in the target map for the extracted area 2715 */ 2716 2717 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 2718 dstaddr = vm_map_min(dstmap); 2719 if (!uvm_map_reserve(dstmap, len, start, 0, &dstaddr, 0)) 2720 return (ENOMEM); 2721 *dstaddrp = dstaddr; /* pass address back to caller */ 2722 UVMHIST_LOG(maphist, " dstaddr=0x%x", dstaddr,0,0,0); 2723 } else { 2724 dstaddr = *dstaddrp; 2725 } 2726 2727 /* 2728 * step 2: setup for the extraction process loop by init'ing the 2729 * map entry chain, locking src map, and looking up the first useful 2730 * entry in the map. 2731 */ 2732 2733 end = start + len; 2734 newend = dstaddr + len; 2735 chain = endchain = NULL; 2736 nchain = 0; 2737 vm_map_lock(srcmap); 2738 2739 if (uvm_map_lookup_entry(srcmap, start, &entry)) { 2740 2741 /* "start" is within an entry */ 2742 if (flags & UVM_EXTRACT_QREF) { 2743 2744 /* 2745 * for quick references we don't clip the entry, so 2746 * the entry may map space "before" the starting 2747 * virtual address... this is the "fudge" factor 2748 * (which can be non-zero only the first time 2749 * through the "while" loop in step 3). 2750 */ 2751 2752 fudge = start - entry->start; 2753 } else { 2754 2755 /* 2756 * normal reference: we clip the map to fit (thus 2757 * fudge is zero) 2758 */ 2759 2760 UVM_MAP_CLIP_START(srcmap, entry, start, NULL); 2761 SAVE_HINT(srcmap, srcmap->hint, entry->prev); 2762 fudge = 0; 2763 } 2764 } else { 2765 2766 /* "start" is not within an entry ... skip to next entry */ 2767 if (flags & UVM_EXTRACT_CONTIG) { 2768 error = EINVAL; 2769 goto bad; /* definite hole here ... */ 2770 } 2771 2772 entry = entry->next; 2773 fudge = 0; 2774 } 2775 2776 /* save values from srcmap for step 6 */ 2777 orig_entry = entry; 2778 orig_fudge = fudge; 2779 2780 /* 2781 * step 3: now start looping through the map entries, extracting 2782 * as we go. 2783 */ 2784 2785 while (entry->start < end && entry != &srcmap->header) { 2786 2787 /* if we are not doing a quick reference, clip it */ 2788 if ((flags & UVM_EXTRACT_QREF) == 0) 2789 UVM_MAP_CLIP_END(srcmap, entry, end, NULL); 2790 2791 /* clear needs_copy (allow chunking) */ 2792 if (UVM_ET_ISNEEDSCOPY(entry)) { 2793 amap_copy(srcmap, entry, 2794 AMAP_COPY_NOWAIT|AMAP_COPY_NOMERGE, start, end); 2795 if (UVM_ET_ISNEEDSCOPY(entry)) { /* failed? */ 2796 error = ENOMEM; 2797 goto bad; 2798 } 2799 2800 /* amap_copy could clip (during chunk)! update fudge */ 2801 if (fudge) { 2802 fudge = start - entry->start; 2803 orig_fudge = fudge; 2804 } 2805 } 2806 2807 /* calculate the offset of this from "start" */ 2808 oldoffset = (entry->start + fudge) - start; 2809 2810 /* allocate a new map entry */ 2811 newentry = uvm_mapent_alloc(dstmap, 0); 2812 if (newentry == NULL) { 2813 error = ENOMEM; 2814 goto bad; 2815 } 2816 2817 /* set up new map entry */ 2818 newentry->next = NULL; 2819 newentry->prev = endchain; 2820 newentry->start = dstaddr + oldoffset; 2821 newentry->end = 2822 newentry->start + (entry->end - (entry->start + fudge)); 2823 if (newentry->end > newend || newentry->end < newentry->start) 2824 newentry->end = newend; 2825 newentry->object.uvm_obj = entry->object.uvm_obj; 2826 if (newentry->object.uvm_obj) { 2827 if (newentry->object.uvm_obj->pgops->pgo_reference) 2828 newentry->object.uvm_obj->pgops-> 2829 pgo_reference(newentry->object.uvm_obj); 2830 newentry->offset = entry->offset + fudge; 2831 } else { 2832 newentry->offset = 0; 2833 } 2834 newentry->etype = entry->etype; 2835 newentry->protection = (flags & UVM_EXTRACT_FIXPROT) ? 2836 entry->max_protection : entry->protection; 2837 newentry->max_protection = entry->max_protection; 2838 newentry->inheritance = entry->inheritance; 2839 newentry->wired_count = 0; 2840 newentry->aref.ar_amap = entry->aref.ar_amap; 2841 if (newentry->aref.ar_amap) { 2842 newentry->aref.ar_pageoff = 2843 entry->aref.ar_pageoff + (fudge >> PAGE_SHIFT); 2844 uvm_map_reference_amap(newentry, AMAP_SHARED | 2845 ((flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0)); 2846 } else { 2847 newentry->aref.ar_pageoff = 0; 2848 } 2849 newentry->advice = entry->advice; 2850 if ((flags & UVM_EXTRACT_QREF) != 0) { 2851 newentry->flags |= UVM_MAP_NOMERGE; 2852 } 2853 2854 /* now link it on the chain */ 2855 nchain++; 2856 if (endchain == NULL) { 2857 chain = endchain = newentry; 2858 } else { 2859 endchain->next = newentry; 2860 endchain = newentry; 2861 } 2862 2863 /* end of 'while' loop! */ 2864 if ((flags & UVM_EXTRACT_CONTIG) && entry->end < end && 2865 (entry->next == &srcmap->header || 2866 entry->next->start != entry->end)) { 2867 error = EINVAL; 2868 goto bad; 2869 } 2870 entry = entry->next; 2871 fudge = 0; 2872 } 2873 2874 /* 2875 * step 4: close off chain (in format expected by uvm_map_replace) 2876 */ 2877 2878 if (chain) 2879 chain->prev = endchain; 2880 2881 /* 2882 * step 5: attempt to lock the dest map so we can pmap_copy. 2883 * note usage of copy_ok: 2884 * 1 => dstmap locked, pmap_copy ok, and we "replace" here (step 5) 2885 * 0 => dstmap unlocked, NO pmap_copy, and we will "replace" in step 7 2886 */ 2887 2888 if (srcmap == dstmap || vm_map_lock_try(dstmap) == true) { 2889 copy_ok = 1; 2890 if (!uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2891 nchain, &resentry)) { 2892 if (srcmap != dstmap) 2893 vm_map_unlock(dstmap); 2894 error = EIO; 2895 goto bad; 2896 } 2897 } else { 2898 copy_ok = 0; 2899 /* replace defered until step 7 */ 2900 } 2901 2902 /* 2903 * step 6: traverse the srcmap a second time to do the following: 2904 * - if we got a lock on the dstmap do pmap_copy 2905 * - if UVM_EXTRACT_REMOVE remove the entries 2906 * we make use of orig_entry and orig_fudge (saved in step 2) 2907 */ 2908 2909 if (copy_ok || (flags & UVM_EXTRACT_REMOVE)) { 2910 2911 /* purge possible stale hints from srcmap */ 2912 if (flags & UVM_EXTRACT_REMOVE) { 2913 SAVE_HINT(srcmap, srcmap->hint, orig_entry->prev); 2914 if (srcmap->first_free != &srcmap->header && 2915 srcmap->first_free->start >= start) 2916 srcmap->first_free = orig_entry->prev; 2917 } 2918 2919 entry = orig_entry; 2920 fudge = orig_fudge; 2921 deadentry = NULL; /* for UVM_EXTRACT_REMOVE */ 2922 2923 while (entry->start < end && entry != &srcmap->header) { 2924 if (copy_ok) { 2925 oldoffset = (entry->start + fudge) - start; 2926 elen = MIN(end, entry->end) - 2927 (entry->start + fudge); 2928 pmap_copy(dstmap->pmap, srcmap->pmap, 2929 dstaddr + oldoffset, elen, 2930 entry->start + fudge); 2931 } 2932 2933 /* we advance "entry" in the following if statement */ 2934 if (flags & UVM_EXTRACT_REMOVE) { 2935 pmap_remove(srcmap->pmap, entry->start, 2936 entry->end); 2937 oldentry = entry; /* save entry */ 2938 entry = entry->next; /* advance */ 2939 uvm_map_entry_unlink(srcmap, oldentry); 2940 /* add to dead list */ 2941 oldentry->next = deadentry; 2942 deadentry = oldentry; 2943 } else { 2944 entry = entry->next; /* advance */ 2945 } 2946 2947 /* end of 'while' loop */ 2948 fudge = 0; 2949 } 2950 pmap_update(srcmap->pmap); 2951 2952 /* 2953 * unlock dstmap. we will dispose of deadentry in 2954 * step 7 if needed 2955 */ 2956 2957 if (copy_ok && srcmap != dstmap) 2958 vm_map_unlock(dstmap); 2959 2960 } else { 2961 deadentry = NULL; 2962 } 2963 2964 /* 2965 * step 7: we are done with the source map, unlock. if copy_ok 2966 * is 0 then we have not replaced the dummy mapping in dstmap yet 2967 * and we need to do so now. 2968 */ 2969 2970 vm_map_unlock(srcmap); 2971 if ((flags & UVM_EXTRACT_REMOVE) && deadentry) 2972 uvm_unmap_detach(deadentry, 0); /* dispose of old entries */ 2973 2974 /* now do the replacement if we didn't do it in step 5 */ 2975 if (copy_ok == 0) { 2976 vm_map_lock(dstmap); 2977 error = uvm_map_replace(dstmap, dstaddr, dstaddr+len, chain, 2978 nchain, &resentry); 2979 vm_map_unlock(dstmap); 2980 2981 if (error == false) { 2982 error = EIO; 2983 goto bad2; 2984 } 2985 } 2986 2987 if (resentry != NULL) 2988 uvm_mapent_free(resentry); 2989 2990 uvm_map_check(srcmap, "map_extract src leave"); 2991 uvm_map_check(dstmap, "map_extract dst leave"); 2992 2993 return (0); 2994 2995 /* 2996 * bad: failure recovery 2997 */ 2998 bad: 2999 vm_map_unlock(srcmap); 3000 bad2: /* src already unlocked */ 3001 if (chain) 3002 uvm_unmap_detach(chain, 3003 (flags & UVM_EXTRACT_QREF) ? AMAP_REFALL : 0); 3004 3005 if (resentry != NULL) 3006 uvm_mapent_free(resentry); 3007 3008 uvm_map_check(srcmap, "map_extract src err leave"); 3009 uvm_map_check(dstmap, "map_extract dst err leave"); 3010 3011 if ((flags & UVM_EXTRACT_RESERVED) == 0) { 3012 uvm_unmap(dstmap, dstaddr, dstaddr+len); /* ??? */ 3013 } 3014 return (error); 3015 } 3016 3017 /* end of extraction functions */ 3018 3019 /* 3020 * uvm_map_submap: punch down part of a map into a submap 3021 * 3022 * => only the kernel_map is allowed to be submapped 3023 * => the purpose of submapping is to break up the locking granularity 3024 * of a larger map 3025 * => the range specified must have been mapped previously with a uvm_map() 3026 * call [with uobj==NULL] to create a blank map entry in the main map. 3027 * [And it had better still be blank!] 3028 * => maps which contain submaps should never be copied or forked. 3029 * => to remove a submap, use uvm_unmap() on the main map 3030 * and then uvm_map_deallocate() the submap. 3031 * => main map must be unlocked. 3032 * => submap must have been init'd and have a zero reference count. 3033 * [need not be locked as we don't actually reference it] 3034 */ 3035 3036 int 3037 uvm_map_submap(struct vm_map *map, vaddr_t start, vaddr_t end, 3038 struct vm_map *submap) 3039 { 3040 struct vm_map_entry *entry; 3041 struct uvm_mapent_reservation umr; 3042 int error; 3043 3044 uvm_mapent_reserve(map, &umr, 2, 0); 3045 3046 vm_map_lock(map); 3047 VM_MAP_RANGE_CHECK(map, start, end); 3048 3049 if (uvm_map_lookup_entry(map, start, &entry)) { 3050 UVM_MAP_CLIP_START(map, entry, start, &umr); 3051 UVM_MAP_CLIP_END(map, entry, end, &umr); /* to be safe */ 3052 } else { 3053 entry = NULL; 3054 } 3055 3056 if (entry != NULL && 3057 entry->start == start && entry->end == end && 3058 entry->object.uvm_obj == NULL && entry->aref.ar_amap == NULL && 3059 !UVM_ET_ISCOPYONWRITE(entry) && !UVM_ET_ISNEEDSCOPY(entry)) { 3060 entry->etype |= UVM_ET_SUBMAP; 3061 entry->object.sub_map = submap; 3062 entry->offset = 0; 3063 uvm_map_reference(submap); 3064 error = 0; 3065 } else { 3066 error = EINVAL; 3067 } 3068 vm_map_unlock(map); 3069 3070 uvm_mapent_unreserve(map, &umr); 3071 3072 return error; 3073 } 3074 3075 /* 3076 * uvm_map_setup_kernel: init in-kernel map 3077 * 3078 * => map must not be in service yet. 3079 */ 3080 3081 void 3082 uvm_map_setup_kernel(struct vm_map_kernel *map, 3083 vaddr_t vmin, vaddr_t vmax, int flags) 3084 { 3085 3086 uvm_map_setup(&map->vmk_map, vmin, vmax, flags); 3087 callback_head_init(&map->vmk_reclaim_callback, IPL_VM); 3088 LIST_INIT(&map->vmk_kentry_free); 3089 map->vmk_merged_entries = NULL; 3090 } 3091 3092 3093 /* 3094 * uvm_map_protect: change map protection 3095 * 3096 * => set_max means set max_protection. 3097 * => map must be unlocked. 3098 */ 3099 3100 #define MASK(entry) (UVM_ET_ISCOPYONWRITE(entry) ? \ 3101 ~VM_PROT_WRITE : VM_PROT_ALL) 3102 3103 int 3104 uvm_map_protect(struct vm_map *map, vaddr_t start, vaddr_t end, 3105 vm_prot_t new_prot, bool set_max) 3106 { 3107 struct vm_map_entry *current, *entry; 3108 int error = 0; 3109 UVMHIST_FUNC("uvm_map_protect"); UVMHIST_CALLED(maphist); 3110 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_prot=0x%x)", 3111 map, start, end, new_prot); 3112 3113 vm_map_lock(map); 3114 VM_MAP_RANGE_CHECK(map, start, end); 3115 if (uvm_map_lookup_entry(map, start, &entry)) { 3116 UVM_MAP_CLIP_START(map, entry, start, NULL); 3117 } else { 3118 entry = entry->next; 3119 } 3120 3121 /* 3122 * make a first pass to check for protection violations. 3123 */ 3124 3125 current = entry; 3126 while ((current != &map->header) && (current->start < end)) { 3127 if (UVM_ET_ISSUBMAP(current)) { 3128 error = EINVAL; 3129 goto out; 3130 } 3131 if ((new_prot & current->max_protection) != new_prot) { 3132 error = EACCES; 3133 goto out; 3134 } 3135 /* 3136 * Don't allow VM_PROT_EXECUTE to be set on entries that 3137 * point to vnodes that are associated with a NOEXEC file 3138 * system. 3139 */ 3140 if (UVM_ET_ISOBJ(current) && 3141 UVM_OBJ_IS_VNODE(current->object.uvm_obj)) { 3142 struct vnode *vp = 3143 (struct vnode *) current->object.uvm_obj; 3144 3145 if ((new_prot & VM_PROT_EXECUTE) != 0 && 3146 (vp->v_mount->mnt_flag & MNT_NOEXEC) != 0) { 3147 error = EACCES; 3148 goto out; 3149 } 3150 } 3151 3152 current = current->next; 3153 } 3154 3155 /* go back and fix up protections (no need to clip this time). */ 3156 3157 current = entry; 3158 while ((current != &map->header) && (current->start < end)) { 3159 vm_prot_t old_prot; 3160 3161 UVM_MAP_CLIP_END(map, current, end, NULL); 3162 old_prot = current->protection; 3163 if (set_max) 3164 current->protection = 3165 (current->max_protection = new_prot) & old_prot; 3166 else 3167 current->protection = new_prot; 3168 3169 /* 3170 * update physical map if necessary. worry about copy-on-write 3171 * here -- CHECK THIS XXX 3172 */ 3173 3174 if (current->protection != old_prot) { 3175 /* update pmap! */ 3176 pmap_protect(map->pmap, current->start, current->end, 3177 current->protection & MASK(entry)); 3178 3179 /* 3180 * If this entry points at a vnode, and the 3181 * protection includes VM_PROT_EXECUTE, mark 3182 * the vnode as VEXECMAP. 3183 */ 3184 if (UVM_ET_ISOBJ(current)) { 3185 struct uvm_object *uobj = 3186 current->object.uvm_obj; 3187 3188 if (UVM_OBJ_IS_VNODE(uobj) && 3189 (current->protection & VM_PROT_EXECUTE)) { 3190 vn_markexec((struct vnode *) uobj); 3191 } 3192 } 3193 } 3194 3195 /* 3196 * If the map is configured to lock any future mappings, 3197 * wire this entry now if the old protection was VM_PROT_NONE 3198 * and the new protection is not VM_PROT_NONE. 3199 */ 3200 3201 if ((map->flags & VM_MAP_WIREFUTURE) != 0 && 3202 VM_MAPENT_ISWIRED(entry) == 0 && 3203 old_prot == VM_PROT_NONE && 3204 new_prot != VM_PROT_NONE) { 3205 if (uvm_map_pageable(map, entry->start, 3206 entry->end, false, 3207 UVM_LK_ENTER|UVM_LK_EXIT) != 0) { 3208 3209 /* 3210 * If locking the entry fails, remember the 3211 * error if it's the first one. Note we 3212 * still continue setting the protection in 3213 * the map, but will return the error 3214 * condition regardless. 3215 * 3216 * XXX Ignore what the actual error is, 3217 * XXX just call it a resource shortage 3218 * XXX so that it doesn't get confused 3219 * XXX what uvm_map_protect() itself would 3220 * XXX normally return. 3221 */ 3222 3223 error = ENOMEM; 3224 } 3225 } 3226 current = current->next; 3227 } 3228 pmap_update(map->pmap); 3229 3230 out: 3231 vm_map_unlock(map); 3232 3233 UVMHIST_LOG(maphist, "<- done, error=%d",error,0,0,0); 3234 return error; 3235 } 3236 3237 #undef MASK 3238 3239 /* 3240 * uvm_map_inherit: set inheritance code for range of addrs in map. 3241 * 3242 * => map must be unlocked 3243 * => note that the inherit code is used during a "fork". see fork 3244 * code for details. 3245 */ 3246 3247 int 3248 uvm_map_inherit(struct vm_map *map, vaddr_t start, vaddr_t end, 3249 vm_inherit_t new_inheritance) 3250 { 3251 struct vm_map_entry *entry, *temp_entry; 3252 UVMHIST_FUNC("uvm_map_inherit"); UVMHIST_CALLED(maphist); 3253 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_inh=0x%x)", 3254 map, start, end, new_inheritance); 3255 3256 switch (new_inheritance) { 3257 case MAP_INHERIT_NONE: 3258 case MAP_INHERIT_COPY: 3259 case MAP_INHERIT_SHARE: 3260 break; 3261 default: 3262 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3263 return EINVAL; 3264 } 3265 3266 vm_map_lock(map); 3267 VM_MAP_RANGE_CHECK(map, start, end); 3268 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3269 entry = temp_entry; 3270 UVM_MAP_CLIP_START(map, entry, start, NULL); 3271 } else { 3272 entry = temp_entry->next; 3273 } 3274 while ((entry != &map->header) && (entry->start < end)) { 3275 UVM_MAP_CLIP_END(map, entry, end, NULL); 3276 entry->inheritance = new_inheritance; 3277 entry = entry->next; 3278 } 3279 vm_map_unlock(map); 3280 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3281 return 0; 3282 } 3283 3284 /* 3285 * uvm_map_advice: set advice code for range of addrs in map. 3286 * 3287 * => map must be unlocked 3288 */ 3289 3290 int 3291 uvm_map_advice(struct vm_map *map, vaddr_t start, vaddr_t end, int new_advice) 3292 { 3293 struct vm_map_entry *entry, *temp_entry; 3294 UVMHIST_FUNC("uvm_map_advice"); UVMHIST_CALLED(maphist); 3295 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_adv=0x%x)", 3296 map, start, end, new_advice); 3297 3298 vm_map_lock(map); 3299 VM_MAP_RANGE_CHECK(map, start, end); 3300 if (uvm_map_lookup_entry(map, start, &temp_entry)) { 3301 entry = temp_entry; 3302 UVM_MAP_CLIP_START(map, entry, start, NULL); 3303 } else { 3304 entry = temp_entry->next; 3305 } 3306 3307 /* 3308 * XXXJRT: disallow holes? 3309 */ 3310 3311 while ((entry != &map->header) && (entry->start < end)) { 3312 UVM_MAP_CLIP_END(map, entry, end, NULL); 3313 3314 switch (new_advice) { 3315 case MADV_NORMAL: 3316 case MADV_RANDOM: 3317 case MADV_SEQUENTIAL: 3318 /* nothing special here */ 3319 break; 3320 3321 default: 3322 vm_map_unlock(map); 3323 UVMHIST_LOG(maphist,"<- done (INVALID ARG)",0,0,0,0); 3324 return EINVAL; 3325 } 3326 entry->advice = new_advice; 3327 entry = entry->next; 3328 } 3329 3330 vm_map_unlock(map); 3331 UVMHIST_LOG(maphist,"<- done (OK)",0,0,0,0); 3332 return 0; 3333 } 3334 3335 /* 3336 * uvm_map_pageable: sets the pageability of a range in a map. 3337 * 3338 * => wires map entries. should not be used for transient page locking. 3339 * for that, use uvm_fault_wire()/uvm_fault_unwire() (see uvm_vslock()). 3340 * => regions specified as not pageable require lock-down (wired) memory 3341 * and page tables. 3342 * => map must never be read-locked 3343 * => if islocked is true, map is already write-locked 3344 * => we always unlock the map, since we must downgrade to a read-lock 3345 * to call uvm_fault_wire() 3346 * => XXXCDC: check this and try and clean it up. 3347 */ 3348 3349 int 3350 uvm_map_pageable(struct vm_map *map, vaddr_t start, vaddr_t end, 3351 bool new_pageable, int lockflags) 3352 { 3353 struct vm_map_entry *entry, *start_entry, *failed_entry; 3354 int rv; 3355 #ifdef DIAGNOSTIC 3356 u_int timestamp_save; 3357 #endif 3358 UVMHIST_FUNC("uvm_map_pageable"); UVMHIST_CALLED(maphist); 3359 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,new_pageable=0x%x)", 3360 map, start, end, new_pageable); 3361 KASSERT(map->flags & VM_MAP_PAGEABLE); 3362 3363 if ((lockflags & UVM_LK_ENTER) == 0) 3364 vm_map_lock(map); 3365 VM_MAP_RANGE_CHECK(map, start, end); 3366 3367 /* 3368 * only one pageability change may take place at one time, since 3369 * uvm_fault_wire assumes it will be called only once for each 3370 * wiring/unwiring. therefore, we have to make sure we're actually 3371 * changing the pageability for the entire region. we do so before 3372 * making any changes. 3373 */ 3374 3375 if (uvm_map_lookup_entry(map, start, &start_entry) == false) { 3376 if ((lockflags & UVM_LK_EXIT) == 0) 3377 vm_map_unlock(map); 3378 3379 UVMHIST_LOG(maphist,"<- done (fault)",0,0,0,0); 3380 return EFAULT; 3381 } 3382 entry = start_entry; 3383 3384 /* 3385 * handle wiring and unwiring separately. 3386 */ 3387 3388 if (new_pageable) { /* unwire */ 3389 UVM_MAP_CLIP_START(map, entry, start, NULL); 3390 3391 /* 3392 * unwiring. first ensure that the range to be unwired is 3393 * really wired down and that there are no holes. 3394 */ 3395 3396 while ((entry != &map->header) && (entry->start < end)) { 3397 if (entry->wired_count == 0 || 3398 (entry->end < end && 3399 (entry->next == &map->header || 3400 entry->next->start > entry->end))) { 3401 if ((lockflags & UVM_LK_EXIT) == 0) 3402 vm_map_unlock(map); 3403 UVMHIST_LOG(maphist, "<- done (INVAL)",0,0,0,0); 3404 return EINVAL; 3405 } 3406 entry = entry->next; 3407 } 3408 3409 /* 3410 * POSIX 1003.1b - a single munlock call unlocks a region, 3411 * regardless of the number of mlock calls made on that 3412 * region. 3413 */ 3414 3415 entry = start_entry; 3416 while ((entry != &map->header) && (entry->start < end)) { 3417 UVM_MAP_CLIP_END(map, entry, end, NULL); 3418 if (VM_MAPENT_ISWIRED(entry)) 3419 uvm_map_entry_unwire(map, entry); 3420 entry = entry->next; 3421 } 3422 if ((lockflags & UVM_LK_EXIT) == 0) 3423 vm_map_unlock(map); 3424 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3425 return 0; 3426 } 3427 3428 /* 3429 * wire case: in two passes [XXXCDC: ugly block of code here] 3430 * 3431 * 1: holding the write lock, we create any anonymous maps that need 3432 * to be created. then we clip each map entry to the region to 3433 * be wired and increment its wiring count. 3434 * 3435 * 2: we downgrade to a read lock, and call uvm_fault_wire to fault 3436 * in the pages for any newly wired area (wired_count == 1). 3437 * 3438 * downgrading to a read lock for uvm_fault_wire avoids a possible 3439 * deadlock with another thread that may have faulted on one of 3440 * the pages to be wired (it would mark the page busy, blocking 3441 * us, then in turn block on the map lock that we hold). because 3442 * of problems in the recursive lock package, we cannot upgrade 3443 * to a write lock in vm_map_lookup. thus, any actions that 3444 * require the write lock must be done beforehand. because we 3445 * keep the read lock on the map, the copy-on-write status of the 3446 * entries we modify here cannot change. 3447 */ 3448 3449 while ((entry != &map->header) && (entry->start < end)) { 3450 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3451 3452 /* 3453 * perform actions of vm_map_lookup that need the 3454 * write lock on the map: create an anonymous map 3455 * for a copy-on-write region, or an anonymous map 3456 * for a zero-fill region. (XXXCDC: submap case 3457 * ok?) 3458 */ 3459 3460 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3461 if (UVM_ET_ISNEEDSCOPY(entry) && 3462 ((entry->max_protection & VM_PROT_WRITE) || 3463 (entry->object.uvm_obj == NULL))) { 3464 amap_copy(map, entry, 0, start, end); 3465 /* XXXCDC: wait OK? */ 3466 } 3467 } 3468 } 3469 UVM_MAP_CLIP_START(map, entry, start, NULL); 3470 UVM_MAP_CLIP_END(map, entry, end, NULL); 3471 entry->wired_count++; 3472 3473 /* 3474 * Check for holes 3475 */ 3476 3477 if (entry->protection == VM_PROT_NONE || 3478 (entry->end < end && 3479 (entry->next == &map->header || 3480 entry->next->start > entry->end))) { 3481 3482 /* 3483 * found one. amap creation actions do not need to 3484 * be undone, but the wired counts need to be restored. 3485 */ 3486 3487 while (entry != &map->header && entry->end > start) { 3488 entry->wired_count--; 3489 entry = entry->prev; 3490 } 3491 if ((lockflags & UVM_LK_EXIT) == 0) 3492 vm_map_unlock(map); 3493 UVMHIST_LOG(maphist,"<- done (INVALID WIRE)",0,0,0,0); 3494 return EINVAL; 3495 } 3496 entry = entry->next; 3497 } 3498 3499 /* 3500 * Pass 2. 3501 */ 3502 3503 #ifdef DIAGNOSTIC 3504 timestamp_save = map->timestamp; 3505 #endif 3506 vm_map_busy(map); 3507 vm_map_unlock(map); 3508 3509 rv = 0; 3510 entry = start_entry; 3511 while (entry != &map->header && entry->start < end) { 3512 if (entry->wired_count == 1) { 3513 rv = uvm_fault_wire(map, entry->start, entry->end, 3514 entry->max_protection, 1); 3515 if (rv) { 3516 3517 /* 3518 * wiring failed. break out of the loop. 3519 * we'll clean up the map below, once we 3520 * have a write lock again. 3521 */ 3522 3523 break; 3524 } 3525 } 3526 entry = entry->next; 3527 } 3528 3529 if (rv) { /* failed? */ 3530 3531 /* 3532 * Get back to an exclusive (write) lock. 3533 */ 3534 3535 vm_map_lock(map); 3536 vm_map_unbusy(map); 3537 3538 #ifdef DIAGNOSTIC 3539 if (timestamp_save + 1 != map->timestamp) 3540 panic("uvm_map_pageable: stale map"); 3541 #endif 3542 3543 /* 3544 * first drop the wiring count on all the entries 3545 * which haven't actually been wired yet. 3546 */ 3547 3548 failed_entry = entry; 3549 while (entry != &map->header && entry->start < end) { 3550 entry->wired_count--; 3551 entry = entry->next; 3552 } 3553 3554 /* 3555 * now, unwire all the entries that were successfully 3556 * wired above. 3557 */ 3558 3559 entry = start_entry; 3560 while (entry != failed_entry) { 3561 entry->wired_count--; 3562 if (VM_MAPENT_ISWIRED(entry) == 0) 3563 uvm_map_entry_unwire(map, entry); 3564 entry = entry->next; 3565 } 3566 if ((lockflags & UVM_LK_EXIT) == 0) 3567 vm_map_unlock(map); 3568 UVMHIST_LOG(maphist, "<- done (RV=%d)", rv,0,0,0); 3569 return (rv); 3570 } 3571 3572 if ((lockflags & UVM_LK_EXIT) == 0) { 3573 vm_map_unbusy(map); 3574 } else { 3575 3576 /* 3577 * Get back to an exclusive (write) lock. 3578 */ 3579 3580 vm_map_lock(map); 3581 vm_map_unbusy(map); 3582 } 3583 3584 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3585 return 0; 3586 } 3587 3588 /* 3589 * uvm_map_pageable_all: special case of uvm_map_pageable - affects 3590 * all mapped regions. 3591 * 3592 * => map must not be locked. 3593 * => if no flags are specified, all regions are unwired. 3594 * => XXXJRT: has some of the same problems as uvm_map_pageable() above. 3595 */ 3596 3597 int 3598 uvm_map_pageable_all(struct vm_map *map, int flags, vsize_t limit) 3599 { 3600 struct vm_map_entry *entry, *failed_entry; 3601 vsize_t size; 3602 int rv; 3603 #ifdef DIAGNOSTIC 3604 u_int timestamp_save; 3605 #endif 3606 UVMHIST_FUNC("uvm_map_pageable_all"); UVMHIST_CALLED(maphist); 3607 UVMHIST_LOG(maphist,"(map=0x%x,flags=0x%x)", map, flags, 0, 0); 3608 3609 KASSERT(map->flags & VM_MAP_PAGEABLE); 3610 3611 vm_map_lock(map); 3612 3613 /* 3614 * handle wiring and unwiring separately. 3615 */ 3616 3617 if (flags == 0) { /* unwire */ 3618 3619 /* 3620 * POSIX 1003.1b -- munlockall unlocks all regions, 3621 * regardless of how many times mlockall has been called. 3622 */ 3623 3624 for (entry = map->header.next; entry != &map->header; 3625 entry = entry->next) { 3626 if (VM_MAPENT_ISWIRED(entry)) 3627 uvm_map_entry_unwire(map, entry); 3628 } 3629 map->flags &= ~VM_MAP_WIREFUTURE; 3630 vm_map_unlock(map); 3631 UVMHIST_LOG(maphist,"<- done (OK UNWIRE)",0,0,0,0); 3632 return 0; 3633 } 3634 3635 if (flags & MCL_FUTURE) { 3636 3637 /* 3638 * must wire all future mappings; remember this. 3639 */ 3640 3641 map->flags |= VM_MAP_WIREFUTURE; 3642 } 3643 3644 if ((flags & MCL_CURRENT) == 0) { 3645 3646 /* 3647 * no more work to do! 3648 */ 3649 3650 UVMHIST_LOG(maphist,"<- done (OK no wire)",0,0,0,0); 3651 vm_map_unlock(map); 3652 return 0; 3653 } 3654 3655 /* 3656 * wire case: in three passes [XXXCDC: ugly block of code here] 3657 * 3658 * 1: holding the write lock, count all pages mapped by non-wired 3659 * entries. if this would cause us to go over our limit, we fail. 3660 * 3661 * 2: still holding the write lock, we create any anonymous maps that 3662 * need to be created. then we increment its wiring count. 3663 * 3664 * 3: we downgrade to a read lock, and call uvm_fault_wire to fault 3665 * in the pages for any newly wired area (wired_count == 1). 3666 * 3667 * downgrading to a read lock for uvm_fault_wire avoids a possible 3668 * deadlock with another thread that may have faulted on one of 3669 * the pages to be wired (it would mark the page busy, blocking 3670 * us, then in turn block on the map lock that we hold). because 3671 * of problems in the recursive lock package, we cannot upgrade 3672 * to a write lock in vm_map_lookup. thus, any actions that 3673 * require the write lock must be done beforehand. because we 3674 * keep the read lock on the map, the copy-on-write status of the 3675 * entries we modify here cannot change. 3676 */ 3677 3678 for (size = 0, entry = map->header.next; entry != &map->header; 3679 entry = entry->next) { 3680 if (entry->protection != VM_PROT_NONE && 3681 VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3682 size += entry->end - entry->start; 3683 } 3684 } 3685 3686 if (atop(size) + uvmexp.wired > uvmexp.wiredmax) { 3687 vm_map_unlock(map); 3688 return ENOMEM; 3689 } 3690 3691 if (limit != 0 && 3692 (size + ptoa(pmap_wired_count(vm_map_pmap(map))) > limit)) { 3693 vm_map_unlock(map); 3694 return ENOMEM; 3695 } 3696 3697 /* 3698 * Pass 2. 3699 */ 3700 3701 for (entry = map->header.next; entry != &map->header; 3702 entry = entry->next) { 3703 if (entry->protection == VM_PROT_NONE) 3704 continue; 3705 if (VM_MAPENT_ISWIRED(entry) == 0) { /* not already wired? */ 3706 3707 /* 3708 * perform actions of vm_map_lookup that need the 3709 * write lock on the map: create an anonymous map 3710 * for a copy-on-write region, or an anonymous map 3711 * for a zero-fill region. (XXXCDC: submap case 3712 * ok?) 3713 */ 3714 3715 if (!UVM_ET_ISSUBMAP(entry)) { /* not submap */ 3716 if (UVM_ET_ISNEEDSCOPY(entry) && 3717 ((entry->max_protection & VM_PROT_WRITE) || 3718 (entry->object.uvm_obj == NULL))) { 3719 amap_copy(map, entry, 0, entry->start, 3720 entry->end); 3721 /* XXXCDC: wait OK? */ 3722 } 3723 } 3724 } 3725 entry->wired_count++; 3726 } 3727 3728 /* 3729 * Pass 3. 3730 */ 3731 3732 #ifdef DIAGNOSTIC 3733 timestamp_save = map->timestamp; 3734 #endif 3735 vm_map_busy(map); 3736 vm_map_unlock(map); 3737 3738 rv = 0; 3739 for (entry = map->header.next; entry != &map->header; 3740 entry = entry->next) { 3741 if (entry->wired_count == 1) { 3742 rv = uvm_fault_wire(map, entry->start, entry->end, 3743 entry->max_protection, 1); 3744 if (rv) { 3745 3746 /* 3747 * wiring failed. break out of the loop. 3748 * we'll clean up the map below, once we 3749 * have a write lock again. 3750 */ 3751 3752 break; 3753 } 3754 } 3755 } 3756 3757 if (rv) { 3758 3759 /* 3760 * Get back an exclusive (write) lock. 3761 */ 3762 3763 vm_map_lock(map); 3764 vm_map_unbusy(map); 3765 3766 #ifdef DIAGNOSTIC 3767 if (timestamp_save + 1 != map->timestamp) 3768 panic("uvm_map_pageable_all: stale map"); 3769 #endif 3770 3771 /* 3772 * first drop the wiring count on all the entries 3773 * which haven't actually been wired yet. 3774 * 3775 * Skip VM_PROT_NONE entries like we did above. 3776 */ 3777 3778 failed_entry = entry; 3779 for (/* nothing */; entry != &map->header; 3780 entry = entry->next) { 3781 if (entry->protection == VM_PROT_NONE) 3782 continue; 3783 entry->wired_count--; 3784 } 3785 3786 /* 3787 * now, unwire all the entries that were successfully 3788 * wired above. 3789 * 3790 * Skip VM_PROT_NONE entries like we did above. 3791 */ 3792 3793 for (entry = map->header.next; entry != failed_entry; 3794 entry = entry->next) { 3795 if (entry->protection == VM_PROT_NONE) 3796 continue; 3797 entry->wired_count--; 3798 if (VM_MAPENT_ISWIRED(entry)) 3799 uvm_map_entry_unwire(map, entry); 3800 } 3801 vm_map_unlock(map); 3802 UVMHIST_LOG(maphist,"<- done (RV=%d)", rv,0,0,0); 3803 return (rv); 3804 } 3805 3806 vm_map_unbusy(map); 3807 3808 UVMHIST_LOG(maphist,"<- done (OK WIRE)",0,0,0,0); 3809 return 0; 3810 } 3811 3812 /* 3813 * uvm_map_clean: clean out a map range 3814 * 3815 * => valid flags: 3816 * if (flags & PGO_CLEANIT): dirty pages are cleaned first 3817 * if (flags & PGO_SYNCIO): dirty pages are written synchronously 3818 * if (flags & PGO_DEACTIVATE): any cached pages are deactivated after clean 3819 * if (flags & PGO_FREE): any cached pages are freed after clean 3820 * => returns an error if any part of the specified range isn't mapped 3821 * => never a need to flush amap layer since the anonymous memory has 3822 * no permanent home, but may deactivate pages there 3823 * => called from sys_msync() and sys_madvise() 3824 * => caller must not write-lock map (read OK). 3825 * => we may sleep while cleaning if SYNCIO [with map read-locked] 3826 */ 3827 3828 int 3829 uvm_map_clean(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 3830 { 3831 struct vm_map_entry *current, *entry; 3832 struct uvm_object *uobj; 3833 struct vm_amap *amap; 3834 struct vm_anon *anon; 3835 struct vm_page *pg; 3836 vaddr_t offset; 3837 vsize_t size; 3838 voff_t uoff; 3839 int error, refs; 3840 UVMHIST_FUNC("uvm_map_clean"); UVMHIST_CALLED(maphist); 3841 3842 UVMHIST_LOG(maphist,"(map=0x%x,start=0x%x,end=0x%x,flags=0x%x)", 3843 map, start, end, flags); 3844 KASSERT((flags & (PGO_FREE|PGO_DEACTIVATE)) != 3845 (PGO_FREE|PGO_DEACTIVATE)); 3846 3847 vm_map_lock_read(map); 3848 VM_MAP_RANGE_CHECK(map, start, end); 3849 if (uvm_map_lookup_entry(map, start, &entry) == false) { 3850 vm_map_unlock_read(map); 3851 return EFAULT; 3852 } 3853 3854 /* 3855 * Make a first pass to check for holes and wiring problems. 3856 */ 3857 3858 for (current = entry; current->start < end; current = current->next) { 3859 if (UVM_ET_ISSUBMAP(current)) { 3860 vm_map_unlock_read(map); 3861 return EINVAL; 3862 } 3863 if ((flags & PGO_FREE) != 0 && VM_MAPENT_ISWIRED(entry)) { 3864 vm_map_unlock_read(map); 3865 return EBUSY; 3866 } 3867 if (end <= current->end) { 3868 break; 3869 } 3870 if (current->end != current->next->start) { 3871 vm_map_unlock_read(map); 3872 return EFAULT; 3873 } 3874 } 3875 3876 error = 0; 3877 for (current = entry; start < end; current = current->next) { 3878 amap = current->aref.ar_amap; /* top layer */ 3879 uobj = current->object.uvm_obj; /* bottom layer */ 3880 KASSERT(start >= current->start); 3881 3882 /* 3883 * No amap cleaning necessary if: 3884 * 3885 * (1) There's no amap. 3886 * 3887 * (2) We're not deactivating or freeing pages. 3888 */ 3889 3890 if (amap == NULL || (flags & (PGO_DEACTIVATE|PGO_FREE)) == 0) 3891 goto flush_object; 3892 3893 amap_lock(amap); 3894 offset = start - current->start; 3895 size = MIN(end, current->end) - start; 3896 for ( ; size != 0; size -= PAGE_SIZE, offset += PAGE_SIZE) { 3897 anon = amap_lookup(¤t->aref, offset); 3898 if (anon == NULL) 3899 continue; 3900 3901 mutex_enter(&anon->an_lock); 3902 pg = anon->an_page; 3903 if (pg == NULL) { 3904 mutex_exit(&anon->an_lock); 3905 continue; 3906 } 3907 3908 switch (flags & (PGO_CLEANIT|PGO_FREE|PGO_DEACTIVATE)) { 3909 3910 /* 3911 * In these first 3 cases, we just deactivate the page. 3912 */ 3913 3914 case PGO_CLEANIT|PGO_FREE: 3915 case PGO_CLEANIT|PGO_DEACTIVATE: 3916 case PGO_DEACTIVATE: 3917 deactivate_it: 3918 /* 3919 * skip the page if it's loaned or wired, 3920 * since it shouldn't be on a paging queue 3921 * at all in these cases. 3922 */ 3923 3924 mutex_enter(&uvm_pageqlock); 3925 if (pg->loan_count != 0 || 3926 pg->wire_count != 0) { 3927 mutex_exit(&uvm_pageqlock); 3928 mutex_exit(&anon->an_lock); 3929 continue; 3930 } 3931 KASSERT(pg->uanon == anon); 3932 uvm_pagedeactivate(pg); 3933 mutex_exit(&uvm_pageqlock); 3934 mutex_exit(&anon->an_lock); 3935 continue; 3936 3937 case PGO_FREE: 3938 3939 /* 3940 * If there are multiple references to 3941 * the amap, just deactivate the page. 3942 */ 3943 3944 if (amap_refs(amap) > 1) 3945 goto deactivate_it; 3946 3947 /* skip the page if it's wired */ 3948 if (pg->wire_count != 0) { 3949 mutex_exit(&anon->an_lock); 3950 continue; 3951 } 3952 amap_unadd(¤t->aref, offset); 3953 refs = --anon->an_ref; 3954 mutex_exit(&anon->an_lock); 3955 if (refs == 0) 3956 uvm_anfree(anon); 3957 continue; 3958 } 3959 } 3960 amap_unlock(amap); 3961 3962 flush_object: 3963 /* 3964 * flush pages if we've got a valid backing object. 3965 * note that we must always clean object pages before 3966 * freeing them since otherwise we could reveal stale 3967 * data from files. 3968 */ 3969 3970 uoff = current->offset + (start - current->start); 3971 size = MIN(end, current->end) - start; 3972 if (uobj != NULL) { 3973 mutex_enter(&uobj->vmobjlock); 3974 if (uobj->pgops->pgo_put != NULL) 3975 error = (uobj->pgops->pgo_put)(uobj, uoff, 3976 uoff + size, flags | PGO_CLEANIT); 3977 else 3978 error = 0; 3979 } 3980 start += size; 3981 } 3982 vm_map_unlock_read(map); 3983 return (error); 3984 } 3985 3986 3987 /* 3988 * uvm_map_checkprot: check protection in map 3989 * 3990 * => must allow specified protection in a fully allocated region. 3991 * => map must be read or write locked by caller. 3992 */ 3993 3994 bool 3995 uvm_map_checkprot(struct vm_map *map, vaddr_t start, vaddr_t end, 3996 vm_prot_t protection) 3997 { 3998 struct vm_map_entry *entry; 3999 struct vm_map_entry *tmp_entry; 4000 4001 if (!uvm_map_lookup_entry(map, start, &tmp_entry)) { 4002 return (false); 4003 } 4004 entry = tmp_entry; 4005 while (start < end) { 4006 if (entry == &map->header) { 4007 return (false); 4008 } 4009 4010 /* 4011 * no holes allowed 4012 */ 4013 4014 if (start < entry->start) { 4015 return (false); 4016 } 4017 4018 /* 4019 * check protection associated with entry 4020 */ 4021 4022 if ((entry->protection & protection) != protection) { 4023 return (false); 4024 } 4025 start = entry->end; 4026 entry = entry->next; 4027 } 4028 return (true); 4029 } 4030 4031 /* 4032 * uvmspace_alloc: allocate a vmspace structure. 4033 * 4034 * - structure includes vm_map and pmap 4035 * - XXX: no locking on this structure 4036 * - refcnt set to 1, rest must be init'd by caller 4037 */ 4038 struct vmspace * 4039 uvmspace_alloc(vaddr_t vmin, vaddr_t vmax) 4040 { 4041 struct vmspace *vm; 4042 UVMHIST_FUNC("uvmspace_alloc"); UVMHIST_CALLED(maphist); 4043 4044 vm = pool_cache_get(&uvm_vmspace_cache, PR_WAITOK); 4045 uvmspace_init(vm, NULL, vmin, vmax); 4046 UVMHIST_LOG(maphist,"<- done (vm=0x%x)", vm,0,0,0); 4047 return (vm); 4048 } 4049 4050 /* 4051 * uvmspace_init: initialize a vmspace structure. 4052 * 4053 * - XXX: no locking on this structure 4054 * - refcnt set to 1, rest must be init'd by caller 4055 */ 4056 void 4057 uvmspace_init(struct vmspace *vm, struct pmap *pmap, vaddr_t vmin, vaddr_t vmax) 4058 { 4059 UVMHIST_FUNC("uvmspace_init"); UVMHIST_CALLED(maphist); 4060 4061 memset(vm, 0, sizeof(*vm)); 4062 uvm_map_setup(&vm->vm_map, vmin, vmax, VM_MAP_PAGEABLE 4063 #ifdef __USING_TOPDOWN_VM 4064 | VM_MAP_TOPDOWN 4065 #endif 4066 ); 4067 if (pmap) 4068 pmap_reference(pmap); 4069 else 4070 pmap = pmap_create(); 4071 vm->vm_map.pmap = pmap; 4072 vm->vm_refcnt = 1; 4073 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4074 } 4075 4076 /* 4077 * uvmspace_share: share a vmspace between two processes 4078 * 4079 * - used for vfork, threads(?) 4080 */ 4081 4082 void 4083 uvmspace_share(struct proc *p1, struct proc *p2) 4084 { 4085 4086 uvmspace_addref(p1->p_vmspace); 4087 p2->p_vmspace = p1->p_vmspace; 4088 } 4089 4090 /* 4091 * uvmspace_unshare: ensure that process "p" has its own, unshared, vmspace 4092 * 4093 * - XXX: no locking on vmspace 4094 */ 4095 4096 void 4097 uvmspace_unshare(struct lwp *l) 4098 { 4099 struct proc *p = l->l_proc; 4100 struct vmspace *nvm, *ovm = p->p_vmspace; 4101 4102 if (ovm->vm_refcnt == 1) 4103 /* nothing to do: vmspace isn't shared in the first place */ 4104 return; 4105 4106 /* make a new vmspace, still holding old one */ 4107 nvm = uvmspace_fork(ovm); 4108 4109 kpreempt_disable(); 4110 pmap_deactivate(l); /* unbind old vmspace */ 4111 p->p_vmspace = nvm; 4112 pmap_activate(l); /* switch to new vmspace */ 4113 kpreempt_enable(); 4114 4115 uvmspace_free(ovm); /* drop reference to old vmspace */ 4116 } 4117 4118 /* 4119 * uvmspace_exec: the process wants to exec a new program 4120 */ 4121 4122 void 4123 uvmspace_exec(struct lwp *l, vaddr_t start, vaddr_t end) 4124 { 4125 struct proc *p = l->l_proc; 4126 struct vmspace *nvm, *ovm = p->p_vmspace; 4127 struct vm_map *map = &ovm->vm_map; 4128 4129 #ifdef __sparc__ 4130 /* XXX cgd 960926: the sparc #ifdef should be a MD hook */ 4131 kill_user_windows(l); /* before stack addresses go away */ 4132 #endif 4133 4134 /* 4135 * see if more than one process is using this vmspace... 4136 */ 4137 4138 if (ovm->vm_refcnt == 1) { 4139 4140 /* 4141 * if p is the only process using its vmspace then we can safely 4142 * recycle that vmspace for the program that is being exec'd. 4143 */ 4144 4145 #ifdef SYSVSHM 4146 /* 4147 * SYSV SHM semantics require us to kill all segments on an exec 4148 */ 4149 4150 if (ovm->vm_shm) 4151 shmexit(ovm); 4152 #endif 4153 4154 /* 4155 * POSIX 1003.1b -- "lock future mappings" is revoked 4156 * when a process execs another program image. 4157 */ 4158 4159 map->flags &= ~VM_MAP_WIREFUTURE; 4160 4161 /* 4162 * now unmap the old program 4163 */ 4164 4165 pmap_remove_all(map->pmap); 4166 uvm_unmap(map, vm_map_min(map), vm_map_max(map)); 4167 KASSERT(map->header.prev == &map->header); 4168 KASSERT(map->nentries == 0); 4169 4170 /* 4171 * resize the map 4172 */ 4173 4174 vm_map_setmin(map, start); 4175 vm_map_setmax(map, end); 4176 } else { 4177 4178 /* 4179 * p's vmspace is being shared, so we can't reuse it for p since 4180 * it is still being used for others. allocate a new vmspace 4181 * for p 4182 */ 4183 4184 nvm = uvmspace_alloc(start, end); 4185 4186 /* 4187 * install new vmspace and drop our ref to the old one. 4188 */ 4189 4190 kpreempt_disable(); 4191 pmap_deactivate(l); 4192 p->p_vmspace = nvm; 4193 pmap_activate(l); 4194 kpreempt_enable(); 4195 4196 uvmspace_free(ovm); 4197 } 4198 } 4199 4200 /* 4201 * uvmspace_addref: add a referece to a vmspace. 4202 */ 4203 4204 void 4205 uvmspace_addref(struct vmspace *vm) 4206 { 4207 struct vm_map *map = &vm->vm_map; 4208 4209 KASSERT((map->flags & VM_MAP_DYING) == 0); 4210 4211 mutex_enter(&map->misc_lock); 4212 KASSERT(vm->vm_refcnt > 0); 4213 vm->vm_refcnt++; 4214 mutex_exit(&map->misc_lock); 4215 } 4216 4217 /* 4218 * uvmspace_free: free a vmspace data structure 4219 */ 4220 4221 void 4222 uvmspace_free(struct vmspace *vm) 4223 { 4224 struct vm_map_entry *dead_entries; 4225 struct vm_map *map = &vm->vm_map; 4226 int n; 4227 4228 UVMHIST_FUNC("uvmspace_free"); UVMHIST_CALLED(maphist); 4229 4230 UVMHIST_LOG(maphist,"(vm=0x%x) ref=%d", vm, vm->vm_refcnt,0,0); 4231 mutex_enter(&map->misc_lock); 4232 n = --vm->vm_refcnt; 4233 mutex_exit(&map->misc_lock); 4234 if (n > 0) 4235 return; 4236 4237 /* 4238 * at this point, there should be no other references to the map. 4239 * delete all of the mappings, then destroy the pmap. 4240 */ 4241 4242 map->flags |= VM_MAP_DYING; 4243 pmap_remove_all(map->pmap); 4244 #ifdef SYSVSHM 4245 /* Get rid of any SYSV shared memory segments. */ 4246 if (vm->vm_shm != NULL) 4247 shmexit(vm); 4248 #endif 4249 if (map->nentries) { 4250 uvm_unmap_remove(map, vm_map_min(map), vm_map_max(map), 4251 &dead_entries, NULL, 0); 4252 if (dead_entries != NULL) 4253 uvm_unmap_detach(dead_entries, 0); 4254 } 4255 KASSERT(map->nentries == 0); 4256 KASSERT(map->size == 0); 4257 mutex_destroy(&map->misc_lock); 4258 mutex_destroy(&map->mutex); 4259 rw_destroy(&map->lock); 4260 cv_destroy(&map->cv); 4261 pmap_destroy(map->pmap); 4262 pool_cache_put(&uvm_vmspace_cache, vm); 4263 } 4264 4265 /* 4266 * F O R K - m a i n e n t r y p o i n t 4267 */ 4268 /* 4269 * uvmspace_fork: fork a process' main map 4270 * 4271 * => create a new vmspace for child process from parent. 4272 * => parent's map must not be locked. 4273 */ 4274 4275 struct vmspace * 4276 uvmspace_fork(struct vmspace *vm1) 4277 { 4278 struct vmspace *vm2; 4279 struct vm_map *old_map = &vm1->vm_map; 4280 struct vm_map *new_map; 4281 struct vm_map_entry *old_entry; 4282 struct vm_map_entry *new_entry; 4283 UVMHIST_FUNC("uvmspace_fork"); UVMHIST_CALLED(maphist); 4284 4285 vm_map_lock(old_map); 4286 4287 vm2 = uvmspace_alloc(vm_map_min(old_map), vm_map_max(old_map)); 4288 memcpy(&vm2->vm_startcopy, &vm1->vm_startcopy, 4289 (char *) (vm1 + 1) - (char *) &vm1->vm_startcopy); 4290 new_map = &vm2->vm_map; /* XXX */ 4291 4292 old_entry = old_map->header.next; 4293 new_map->size = old_map->size; 4294 4295 /* 4296 * go entry-by-entry 4297 */ 4298 4299 while (old_entry != &old_map->header) { 4300 4301 /* 4302 * first, some sanity checks on the old entry 4303 */ 4304 4305 KASSERT(!UVM_ET_ISSUBMAP(old_entry)); 4306 KASSERT(UVM_ET_ISCOPYONWRITE(old_entry) || 4307 !UVM_ET_ISNEEDSCOPY(old_entry)); 4308 4309 switch (old_entry->inheritance) { 4310 case MAP_INHERIT_NONE: 4311 4312 /* 4313 * drop the mapping, modify size 4314 */ 4315 new_map->size -= old_entry->end - old_entry->start; 4316 break; 4317 4318 case MAP_INHERIT_SHARE: 4319 4320 /* 4321 * share the mapping: this means we want the old and 4322 * new entries to share amaps and backing objects. 4323 */ 4324 /* 4325 * if the old_entry needs a new amap (due to prev fork) 4326 * then we need to allocate it now so that we have 4327 * something we own to share with the new_entry. [in 4328 * other words, we need to clear needs_copy] 4329 */ 4330 4331 if (UVM_ET_ISNEEDSCOPY(old_entry)) { 4332 /* get our own amap, clears needs_copy */ 4333 amap_copy(old_map, old_entry, AMAP_COPY_NOCHUNK, 4334 0, 0); 4335 /* XXXCDC: WAITOK??? */ 4336 } 4337 4338 new_entry = uvm_mapent_alloc(new_map, 0); 4339 /* old_entry -> new_entry */ 4340 uvm_mapent_copy(old_entry, new_entry); 4341 4342 /* new pmap has nothing wired in it */ 4343 new_entry->wired_count = 0; 4344 4345 /* 4346 * gain reference to object backing the map (can't 4347 * be a submap, already checked this case). 4348 */ 4349 4350 if (new_entry->aref.ar_amap) 4351 uvm_map_reference_amap(new_entry, AMAP_SHARED); 4352 4353 if (new_entry->object.uvm_obj && 4354 new_entry->object.uvm_obj->pgops->pgo_reference) 4355 new_entry->object.uvm_obj-> 4356 pgops->pgo_reference( 4357 new_entry->object.uvm_obj); 4358 4359 /* insert entry at end of new_map's entry list */ 4360 uvm_map_entry_link(new_map, new_map->header.prev, 4361 new_entry); 4362 4363 break; 4364 4365 case MAP_INHERIT_COPY: 4366 4367 /* 4368 * copy-on-write the mapping (using mmap's 4369 * MAP_PRIVATE semantics) 4370 * 4371 * allocate new_entry, adjust reference counts. 4372 * (note that new references are read-only). 4373 */ 4374 4375 new_entry = uvm_mapent_alloc(new_map, 0); 4376 /* old_entry -> new_entry */ 4377 uvm_mapent_copy(old_entry, new_entry); 4378 4379 if (new_entry->aref.ar_amap) 4380 uvm_map_reference_amap(new_entry, 0); 4381 4382 if (new_entry->object.uvm_obj && 4383 new_entry->object.uvm_obj->pgops->pgo_reference) 4384 new_entry->object.uvm_obj->pgops->pgo_reference 4385 (new_entry->object.uvm_obj); 4386 4387 /* new pmap has nothing wired in it */ 4388 new_entry->wired_count = 0; 4389 4390 new_entry->etype |= 4391 (UVM_ET_COPYONWRITE|UVM_ET_NEEDSCOPY); 4392 uvm_map_entry_link(new_map, new_map->header.prev, 4393 new_entry); 4394 4395 /* 4396 * the new entry will need an amap. it will either 4397 * need to be copied from the old entry or created 4398 * from scratch (if the old entry does not have an 4399 * amap). can we defer this process until later 4400 * (by setting "needs_copy") or do we need to copy 4401 * the amap now? 4402 * 4403 * we must copy the amap now if any of the following 4404 * conditions hold: 4405 * 1. the old entry has an amap and that amap is 4406 * being shared. this means that the old (parent) 4407 * process is sharing the amap with another 4408 * process. if we do not clear needs_copy here 4409 * we will end up in a situation where both the 4410 * parent and child process are refering to the 4411 * same amap with "needs_copy" set. if the 4412 * parent write-faults, the fault routine will 4413 * clear "needs_copy" in the parent by allocating 4414 * a new amap. this is wrong because the 4415 * parent is supposed to be sharing the old amap 4416 * and the new amap will break that. 4417 * 4418 * 2. if the old entry has an amap and a non-zero 4419 * wire count then we are going to have to call 4420 * amap_cow_now to avoid page faults in the 4421 * parent process. since amap_cow_now requires 4422 * "needs_copy" to be clear we might as well 4423 * clear it here as well. 4424 * 4425 */ 4426 4427 if (old_entry->aref.ar_amap != NULL) { 4428 if ((amap_flags(old_entry->aref.ar_amap) & 4429 AMAP_SHARED) != 0 || 4430 VM_MAPENT_ISWIRED(old_entry)) { 4431 4432 amap_copy(new_map, new_entry, 4433 AMAP_COPY_NOCHUNK, 0, 0); 4434 /* XXXCDC: M_WAITOK ... ok? */ 4435 } 4436 } 4437 4438 /* 4439 * if the parent's entry is wired down, then the 4440 * parent process does not want page faults on 4441 * access to that memory. this means that we 4442 * cannot do copy-on-write because we can't write 4443 * protect the old entry. in this case we 4444 * resolve all copy-on-write faults now, using 4445 * amap_cow_now. note that we have already 4446 * allocated any needed amap (above). 4447 */ 4448 4449 if (VM_MAPENT_ISWIRED(old_entry)) { 4450 4451 /* 4452 * resolve all copy-on-write faults now 4453 * (note that there is nothing to do if 4454 * the old mapping does not have an amap). 4455 */ 4456 if (old_entry->aref.ar_amap) 4457 amap_cow_now(new_map, new_entry); 4458 4459 } else { 4460 4461 /* 4462 * setup mappings to trigger copy-on-write faults 4463 * we must write-protect the parent if it has 4464 * an amap and it is not already "needs_copy"... 4465 * if it is already "needs_copy" then the parent 4466 * has already been write-protected by a previous 4467 * fork operation. 4468 */ 4469 4470 if (old_entry->aref.ar_amap && 4471 !UVM_ET_ISNEEDSCOPY(old_entry)) { 4472 if (old_entry->max_protection & VM_PROT_WRITE) { 4473 pmap_protect(old_map->pmap, 4474 old_entry->start, 4475 old_entry->end, 4476 old_entry->protection & 4477 ~VM_PROT_WRITE); 4478 pmap_update(old_map->pmap); 4479 } 4480 old_entry->etype |= UVM_ET_NEEDSCOPY; 4481 } 4482 } 4483 break; 4484 } /* end of switch statement */ 4485 old_entry = old_entry->next; 4486 } 4487 4488 vm_map_unlock(old_map); 4489 4490 #ifdef SYSVSHM 4491 if (vm1->vm_shm) 4492 shmfork(vm1, vm2); 4493 #endif 4494 4495 #ifdef PMAP_FORK 4496 pmap_fork(vm1->vm_map.pmap, vm2->vm_map.pmap); 4497 #endif 4498 4499 UVMHIST_LOG(maphist,"<- done",0,0,0,0); 4500 return (vm2); 4501 } 4502 4503 4504 /* 4505 * in-kernel map entry allocation. 4506 */ 4507 4508 struct uvm_kmapent_hdr { 4509 LIST_ENTRY(uvm_kmapent_hdr) ukh_listq; 4510 int ukh_nused; 4511 struct vm_map_entry *ukh_freelist; 4512 struct vm_map *ukh_map; 4513 struct vm_map_entry ukh_entries[0]; 4514 }; 4515 4516 #define UVM_KMAPENT_CHUNK \ 4517 ((PAGE_SIZE - sizeof(struct uvm_kmapent_hdr)) \ 4518 / sizeof(struct vm_map_entry)) 4519 4520 #define UVM_KHDR_FIND(entry) \ 4521 ((struct uvm_kmapent_hdr *)(((vaddr_t)entry) & ~PAGE_MASK)) 4522 4523 4524 #ifdef DIAGNOSTIC 4525 static struct vm_map * 4526 uvm_kmapent_map(struct vm_map_entry *entry) 4527 { 4528 const struct uvm_kmapent_hdr *ukh; 4529 4530 ukh = UVM_KHDR_FIND(entry); 4531 return ukh->ukh_map; 4532 } 4533 #endif 4534 4535 static inline struct vm_map_entry * 4536 uvm_kmapent_get(struct uvm_kmapent_hdr *ukh) 4537 { 4538 struct vm_map_entry *entry; 4539 4540 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4541 KASSERT(ukh->ukh_nused >= 0); 4542 4543 entry = ukh->ukh_freelist; 4544 if (entry) { 4545 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4546 == UVM_MAP_KERNEL); 4547 ukh->ukh_freelist = entry->next; 4548 ukh->ukh_nused++; 4549 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4550 } else { 4551 KASSERT(ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4552 } 4553 4554 return entry; 4555 } 4556 4557 static inline void 4558 uvm_kmapent_put(struct uvm_kmapent_hdr *ukh, struct vm_map_entry *entry) 4559 { 4560 4561 KASSERT((entry->flags & (UVM_MAP_KERNEL | UVM_MAP_KMAPENT)) 4562 == UVM_MAP_KERNEL); 4563 KASSERT(ukh->ukh_nused <= UVM_KMAPENT_CHUNK); 4564 KASSERT(ukh->ukh_nused > 0); 4565 KASSERT(ukh->ukh_freelist != NULL || 4566 ukh->ukh_nused == UVM_KMAPENT_CHUNK); 4567 KASSERT(ukh->ukh_freelist == NULL || 4568 ukh->ukh_nused < UVM_KMAPENT_CHUNK); 4569 4570 ukh->ukh_nused--; 4571 entry->next = ukh->ukh_freelist; 4572 ukh->ukh_freelist = entry; 4573 } 4574 4575 /* 4576 * uvm_kmapent_alloc: allocate a map entry for in-kernel map 4577 */ 4578 4579 static struct vm_map_entry * 4580 uvm_kmapent_alloc(struct vm_map *map, int flags) 4581 { 4582 struct vm_page *pg; 4583 struct uvm_map_args args; 4584 struct uvm_kmapent_hdr *ukh; 4585 struct vm_map_entry *entry; 4586 uvm_flag_t mapflags = UVM_MAPFLAG(UVM_PROT_ALL, UVM_PROT_ALL, 4587 UVM_INH_NONE, UVM_ADV_RANDOM, flags | UVM_FLAG_NOMERGE); 4588 vaddr_t va; 4589 int error; 4590 int i; 4591 4592 KDASSERT(UVM_KMAPENT_CHUNK > 2); 4593 KDASSERT(kernel_map != NULL); 4594 KASSERT(vm_map_pmap(map) == pmap_kernel()); 4595 4596 UVMMAP_EVCNT_INCR(uke_alloc); 4597 entry = NULL; 4598 again: 4599 /* 4600 * try to grab an entry from freelist. 4601 */ 4602 mutex_spin_enter(&uvm_kentry_lock); 4603 ukh = LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free); 4604 if (ukh) { 4605 entry = uvm_kmapent_get(ukh); 4606 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK) 4607 LIST_REMOVE(ukh, ukh_listq); 4608 } 4609 mutex_spin_exit(&uvm_kentry_lock); 4610 4611 if (entry) 4612 return entry; 4613 4614 /* 4615 * there's no free entry for this vm_map. 4616 * now we need to allocate some vm_map_entry. 4617 * for simplicity, always allocate one page chunk of them at once. 4618 */ 4619 4620 pg = uvm_pagealloc(NULL, 0, NULL, 0); 4621 if (__predict_false(pg == NULL)) { 4622 if (flags & UVM_FLAG_NOWAIT) 4623 return NULL; 4624 uvm_wait("kme_alloc"); 4625 goto again; 4626 } 4627 4628 error = uvm_map_prepare(map, 0, PAGE_SIZE, NULL, UVM_UNKNOWN_OFFSET, 4629 0, mapflags, &args); 4630 if (error) { 4631 uvm_pagefree(pg); 4632 return NULL; 4633 } 4634 4635 va = args.uma_start; 4636 4637 pmap_kenter_pa(va, VM_PAGE_TO_PHYS(pg), 4638 VM_PROT_READ|VM_PROT_WRITE|PMAP_KMPAGE); 4639 pmap_update(vm_map_pmap(map)); 4640 4641 ukh = (void *)va; 4642 4643 /* 4644 * use the first entry for ukh itsself. 4645 */ 4646 4647 entry = &ukh->ukh_entries[0]; 4648 entry->flags = UVM_MAP_KERNEL | UVM_MAP_KMAPENT; 4649 error = uvm_map_enter(map, &args, entry); 4650 KASSERT(error == 0); 4651 4652 ukh->ukh_nused = UVM_KMAPENT_CHUNK; 4653 ukh->ukh_map = map; 4654 ukh->ukh_freelist = NULL; 4655 for (i = UVM_KMAPENT_CHUNK - 1; i >= 2; i--) { 4656 struct vm_map_entry *xentry = &ukh->ukh_entries[i]; 4657 4658 xentry->flags = UVM_MAP_KERNEL; 4659 uvm_kmapent_put(ukh, xentry); 4660 } 4661 KASSERT(ukh->ukh_nused == 2); 4662 4663 mutex_spin_enter(&uvm_kentry_lock); 4664 LIST_INSERT_HEAD(&vm_map_to_kernel(map)->vmk_kentry_free, 4665 ukh, ukh_listq); 4666 mutex_spin_exit(&uvm_kentry_lock); 4667 4668 /* 4669 * return second entry. 4670 */ 4671 4672 entry = &ukh->ukh_entries[1]; 4673 entry->flags = UVM_MAP_KERNEL; 4674 UVMMAP_EVCNT_INCR(ukh_alloc); 4675 return entry; 4676 } 4677 4678 /* 4679 * uvm_mapent_free: free map entry for in-kernel map 4680 */ 4681 4682 static void 4683 uvm_kmapent_free(struct vm_map_entry *entry) 4684 { 4685 struct uvm_kmapent_hdr *ukh; 4686 struct vm_page *pg; 4687 struct vm_map *map; 4688 struct pmap *pmap; 4689 vaddr_t va; 4690 paddr_t pa; 4691 struct vm_map_entry *deadentry; 4692 4693 UVMMAP_EVCNT_INCR(uke_free); 4694 ukh = UVM_KHDR_FIND(entry); 4695 map = ukh->ukh_map; 4696 4697 mutex_spin_enter(&uvm_kentry_lock); 4698 uvm_kmapent_put(ukh, entry); 4699 if (ukh->ukh_nused > 1) { 4700 if (ukh->ukh_nused == UVM_KMAPENT_CHUNK - 1) 4701 LIST_INSERT_HEAD( 4702 &vm_map_to_kernel(map)->vmk_kentry_free, 4703 ukh, ukh_listq); 4704 mutex_spin_exit(&uvm_kentry_lock); 4705 return; 4706 } 4707 4708 /* 4709 * now we can free this ukh. 4710 * 4711 * however, keep an empty ukh to avoid ping-pong. 4712 */ 4713 4714 if (LIST_FIRST(&vm_map_to_kernel(map)->vmk_kentry_free) == ukh && 4715 LIST_NEXT(ukh, ukh_listq) == NULL) { 4716 mutex_spin_exit(&uvm_kentry_lock); 4717 return; 4718 } 4719 LIST_REMOVE(ukh, ukh_listq); 4720 mutex_spin_exit(&uvm_kentry_lock); 4721 4722 KASSERT(ukh->ukh_nused == 1); 4723 4724 /* 4725 * remove map entry for ukh itsself. 4726 */ 4727 4728 va = (vaddr_t)ukh; 4729 KASSERT((va & PAGE_MASK) == 0); 4730 vm_map_lock(map); 4731 uvm_unmap_remove(map, va, va + PAGE_SIZE, &deadentry, NULL, 0); 4732 KASSERT(deadentry->flags & UVM_MAP_KERNEL); 4733 KASSERT(deadentry->flags & UVM_MAP_KMAPENT); 4734 KASSERT(deadentry->next == NULL); 4735 KASSERT(deadentry == &ukh->ukh_entries[0]); 4736 4737 /* 4738 * unmap the page from pmap and free it. 4739 */ 4740 4741 pmap = vm_map_pmap(map); 4742 KASSERT(pmap == pmap_kernel()); 4743 if (!pmap_extract(pmap, va, &pa)) 4744 panic("%s: no mapping", __func__); 4745 pmap_kremove(va, PAGE_SIZE); 4746 pmap_update(vm_map_pmap(map)); 4747 vm_map_unlock(map); 4748 pg = PHYS_TO_VM_PAGE(pa); 4749 uvm_pagefree(pg); 4750 UVMMAP_EVCNT_INCR(ukh_free); 4751 } 4752 4753 static vsize_t 4754 uvm_kmapent_overhead(vsize_t size) 4755 { 4756 4757 /* 4758 * - the max number of unmerged entries is howmany(size, PAGE_SIZE) 4759 * as the min allocation unit is PAGE_SIZE. 4760 * - UVM_KMAPENT_CHUNK "kmapent"s are allocated from a page. 4761 * one of them are used to map the page itself. 4762 */ 4763 4764 return howmany(howmany(size, PAGE_SIZE), (UVM_KMAPENT_CHUNK - 1)) * 4765 PAGE_SIZE; 4766 } 4767 4768 /* 4769 * map entry reservation 4770 */ 4771 4772 /* 4773 * uvm_mapent_reserve: reserve map entries for clipping before locking map. 4774 * 4775 * => needed when unmapping entries allocated without UVM_FLAG_QUANTUM. 4776 * => caller shouldn't hold map locked. 4777 */ 4778 int 4779 uvm_mapent_reserve(struct vm_map *map, struct uvm_mapent_reservation *umr, 4780 int nentries, int flags) 4781 { 4782 4783 umr->umr_nentries = 0; 4784 4785 if ((flags & UVM_FLAG_QUANTUM) != 0) 4786 return 0; 4787 4788 if (!VM_MAP_USE_KMAPENT(map)) 4789 return 0; 4790 4791 while (nentries--) { 4792 struct vm_map_entry *ent; 4793 ent = uvm_kmapent_alloc(map, flags); 4794 if (!ent) { 4795 uvm_mapent_unreserve(map, umr); 4796 return ENOMEM; 4797 } 4798 UMR_PUTENTRY(umr, ent); 4799 } 4800 4801 return 0; 4802 } 4803 4804 /* 4805 * uvm_mapent_unreserve: 4806 * 4807 * => caller shouldn't hold map locked. 4808 * => never fail or sleep. 4809 */ 4810 void 4811 uvm_mapent_unreserve(struct vm_map *map, struct uvm_mapent_reservation *umr) 4812 { 4813 4814 while (!UMR_EMPTY(umr)) 4815 uvm_kmapent_free(UMR_GETENTRY(umr)); 4816 } 4817 4818 /* 4819 * uvm_mapent_trymerge: try to merge an entry with its neighbors. 4820 * 4821 * => called with map locked. 4822 * => return non zero if successfully merged. 4823 */ 4824 4825 int 4826 uvm_mapent_trymerge(struct vm_map *map, struct vm_map_entry *entry, int flags) 4827 { 4828 struct uvm_object *uobj; 4829 struct vm_map_entry *next; 4830 struct vm_map_entry *prev; 4831 vsize_t size; 4832 int merged = 0; 4833 bool copying; 4834 int newetype; 4835 4836 if (VM_MAP_USE_KMAPENT(map)) { 4837 return 0; 4838 } 4839 if (entry->aref.ar_amap != NULL) { 4840 return 0; 4841 } 4842 if ((entry->flags & UVM_MAP_NOMERGE) != 0) { 4843 return 0; 4844 } 4845 4846 uobj = entry->object.uvm_obj; 4847 size = entry->end - entry->start; 4848 copying = (flags & UVM_MERGE_COPYING) != 0; 4849 newetype = copying ? (entry->etype & ~UVM_ET_NEEDSCOPY) : entry->etype; 4850 4851 next = entry->next; 4852 if (next != &map->header && 4853 next->start == entry->end && 4854 ((copying && next->aref.ar_amap != NULL && 4855 amap_refs(next->aref.ar_amap) == 1) || 4856 (!copying && next->aref.ar_amap == NULL)) && 4857 UVM_ET_ISCOMPATIBLE(next, newetype, 4858 uobj, entry->flags, entry->protection, 4859 entry->max_protection, entry->inheritance, entry->advice, 4860 entry->wired_count) && 4861 (uobj == NULL || entry->offset + size == next->offset)) { 4862 int error; 4863 4864 if (copying) { 4865 error = amap_extend(next, size, 4866 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_BACKWARDS); 4867 } else { 4868 error = 0; 4869 } 4870 if (error == 0) { 4871 if (uobj) { 4872 if (uobj->pgops->pgo_detach) { 4873 uobj->pgops->pgo_detach(uobj); 4874 } 4875 } 4876 4877 entry->end = next->end; 4878 clear_hints(map, next); 4879 uvm_map_entry_unlink(map, next); 4880 if (copying) { 4881 entry->aref = next->aref; 4882 entry->etype &= ~UVM_ET_NEEDSCOPY; 4883 } 4884 uvm_map_check(map, "trymerge forwardmerge"); 4885 uvm_mapent_free_merged(map, next); 4886 merged++; 4887 } 4888 } 4889 4890 prev = entry->prev; 4891 if (prev != &map->header && 4892 prev->end == entry->start && 4893 ((copying && !merged && prev->aref.ar_amap != NULL && 4894 amap_refs(prev->aref.ar_amap) == 1) || 4895 (!copying && prev->aref.ar_amap == NULL)) && 4896 UVM_ET_ISCOMPATIBLE(prev, newetype, 4897 uobj, entry->flags, entry->protection, 4898 entry->max_protection, entry->inheritance, entry->advice, 4899 entry->wired_count) && 4900 (uobj == NULL || 4901 prev->offset + prev->end - prev->start == entry->offset)) { 4902 int error; 4903 4904 if (copying) { 4905 error = amap_extend(prev, size, 4906 AMAP_EXTEND_NOWAIT|AMAP_EXTEND_FORWARDS); 4907 } else { 4908 error = 0; 4909 } 4910 if (error == 0) { 4911 if (uobj) { 4912 if (uobj->pgops->pgo_detach) { 4913 uobj->pgops->pgo_detach(uobj); 4914 } 4915 entry->offset = prev->offset; 4916 } 4917 4918 entry->start = prev->start; 4919 clear_hints(map, prev); 4920 uvm_map_entry_unlink(map, prev); 4921 if (copying) { 4922 entry->aref = prev->aref; 4923 entry->etype &= ~UVM_ET_NEEDSCOPY; 4924 } 4925 uvm_map_check(map, "trymerge backmerge"); 4926 uvm_mapent_free_merged(map, prev); 4927 merged++; 4928 } 4929 } 4930 4931 return merged; 4932 } 4933 4934 #if defined(DDB) 4935 4936 /* 4937 * DDB hooks 4938 */ 4939 4940 /* 4941 * uvm_map_printit: actually prints the map 4942 */ 4943 4944 void 4945 uvm_map_printit(struct vm_map *map, bool full, 4946 void (*pr)(const char *, ...)) 4947 { 4948 struct vm_map_entry *entry; 4949 4950 (*pr)("MAP %p: [0x%lx->0x%lx]\n", map, vm_map_min(map), 4951 vm_map_max(map)); 4952 (*pr)("\t#ent=%d, sz=%d, ref=%d, version=%d, flags=0x%x\n", 4953 map->nentries, map->size, map->ref_count, map->timestamp, 4954 map->flags); 4955 (*pr)("\tpmap=%p(resident=%ld, wired=%ld)\n", map->pmap, 4956 pmap_resident_count(map->pmap), pmap_wired_count(map->pmap)); 4957 if (!full) 4958 return; 4959 for (entry = map->header.next; entry != &map->header; 4960 entry = entry->next) { 4961 (*pr)(" - %p: 0x%lx->0x%lx: obj=%p/0x%llx, amap=%p/%d\n", 4962 entry, entry->start, entry->end, entry->object.uvm_obj, 4963 (long long)entry->offset, entry->aref.ar_amap, 4964 entry->aref.ar_pageoff); 4965 (*pr)( 4966 "\tsubmap=%c, cow=%c, nc=%c, prot(max)=%d/%d, inh=%d, " 4967 "wc=%d, adv=%d\n", 4968 (entry->etype & UVM_ET_SUBMAP) ? 'T' : 'F', 4969 (entry->etype & UVM_ET_COPYONWRITE) ? 'T' : 'F', 4970 (entry->etype & UVM_ET_NEEDSCOPY) ? 'T' : 'F', 4971 entry->protection, entry->max_protection, 4972 entry->inheritance, entry->wired_count, entry->advice); 4973 } 4974 } 4975 4976 /* 4977 * uvm_object_printit: actually prints the object 4978 */ 4979 4980 void 4981 uvm_object_printit(struct uvm_object *uobj, bool full, 4982 void (*pr)(const char *, ...)) 4983 { 4984 struct vm_page *pg; 4985 int cnt = 0; 4986 4987 (*pr)("OBJECT %p: locked=%d, pgops=%p, npages=%d, ", 4988 uobj, mutex_owned(&uobj->vmobjlock), uobj->pgops, uobj->uo_npages); 4989 if (UVM_OBJ_IS_KERN_OBJECT(uobj)) 4990 (*pr)("refs=<SYSTEM>\n"); 4991 else 4992 (*pr)("refs=%d\n", uobj->uo_refs); 4993 4994 if (!full) { 4995 return; 4996 } 4997 (*pr)(" PAGES <pg,offset>:\n "); 4998 TAILQ_FOREACH(pg, &uobj->memq, listq.queue) { 4999 cnt++; 5000 (*pr)("<%p,0x%llx> ", pg, (long long)pg->offset); 5001 if ((cnt % 3) == 0) { 5002 (*pr)("\n "); 5003 } 5004 } 5005 if ((cnt % 3) != 0) { 5006 (*pr)("\n"); 5007 } 5008 } 5009 5010 /* 5011 * uvm_page_printit: actually print the page 5012 */ 5013 5014 static const char page_flagbits[] = UVM_PGFLAGBITS; 5015 static const char page_pqflagbits[] = UVM_PQFLAGBITS; 5016 5017 void 5018 uvm_page_printit(struct vm_page *pg, bool full, 5019 void (*pr)(const char *, ...)) 5020 { 5021 struct vm_page *tpg; 5022 struct uvm_object *uobj; 5023 struct pgflist *pgl; 5024 char pgbuf[128]; 5025 char pqbuf[128]; 5026 5027 (*pr)("PAGE %p:\n", pg); 5028 bitmask_snprintf(pg->flags, page_flagbits, pgbuf, sizeof(pgbuf)); 5029 bitmask_snprintf(pg->pqflags, page_pqflagbits, pqbuf, sizeof(pqbuf)); 5030 (*pr)(" flags=%s, pqflags=%s, wire_count=%d, pa=0x%lx\n", 5031 pgbuf, pqbuf, pg->wire_count, (long)VM_PAGE_TO_PHYS(pg)); 5032 (*pr)(" uobject=%p, uanon=%p, offset=0x%llx loan_count=%d\n", 5033 pg->uobject, pg->uanon, (long long)pg->offset, pg->loan_count); 5034 #if defined(UVM_PAGE_TRKOWN) 5035 if (pg->flags & PG_BUSY) 5036 (*pr)(" owning process = %d, tag=%s\n", 5037 pg->owner, pg->owner_tag); 5038 else 5039 (*pr)(" page not busy, no owner\n"); 5040 #else 5041 (*pr)(" [page ownership tracking disabled]\n"); 5042 #endif 5043 5044 if (!full) 5045 return; 5046 5047 /* cross-verify object/anon */ 5048 if ((pg->pqflags & PQ_FREE) == 0) { 5049 if (pg->pqflags & PQ_ANON) { 5050 if (pg->uanon == NULL || pg->uanon->an_page != pg) 5051 (*pr)(" >>> ANON DOES NOT POINT HERE <<< (%p)\n", 5052 (pg->uanon) ? pg->uanon->an_page : NULL); 5053 else 5054 (*pr)(" anon backpointer is OK\n"); 5055 } else { 5056 uobj = pg->uobject; 5057 if (uobj) { 5058 (*pr)(" checking object list\n"); 5059 TAILQ_FOREACH(tpg, &uobj->memq, listq.queue) { 5060 if (tpg == pg) { 5061 break; 5062 } 5063 } 5064 if (tpg) 5065 (*pr)(" page found on object list\n"); 5066 else 5067 (*pr)(" >>> PAGE NOT FOUND ON OBJECT LIST! <<<\n"); 5068 } 5069 } 5070 } 5071 5072 /* cross-verify page queue */ 5073 if (pg->pqflags & PQ_FREE) { 5074 int fl = uvm_page_lookup_freelist(pg); 5075 int color = VM_PGCOLOR_BUCKET(pg); 5076 pgl = &uvm.page_free[fl].pgfl_buckets[color].pgfl_queues[ 5077 ((pg)->flags & PG_ZERO) ? PGFL_ZEROS : PGFL_UNKNOWN]; 5078 } else { 5079 pgl = NULL; 5080 } 5081 5082 if (pgl) { 5083 (*pr)(" checking pageq list\n"); 5084 LIST_FOREACH(tpg, pgl, pageq.list) { 5085 if (tpg == pg) { 5086 break; 5087 } 5088 } 5089 if (tpg) 5090 (*pr)(" page found on pageq list\n"); 5091 else 5092 (*pr)(" >>> PAGE NOT FOUND ON PAGEQ LIST! <<<\n"); 5093 } 5094 } 5095 5096 /* 5097 * uvm_pages_printthem - print a summary of all managed pages 5098 */ 5099 5100 void 5101 uvm_page_printall(void (*pr)(const char *, ...)) 5102 { 5103 unsigned i; 5104 struct vm_page *pg; 5105 5106 (*pr)("%18s %4s %4s %18s %18s" 5107 #ifdef UVM_PAGE_TRKOWN 5108 " OWNER" 5109 #endif 5110 "\n", "PAGE", "FLAG", "PQ", "UOBJECT", "UANON"); 5111 for (i = 0; i < vm_nphysseg; i++) { 5112 for (pg = vm_physmem[i].pgs; pg <= vm_physmem[i].lastpg; pg++) { 5113 (*pr)("%18p %04x %04x %18p %18p", 5114 pg, pg->flags, pg->pqflags, pg->uobject, 5115 pg->uanon); 5116 #ifdef UVM_PAGE_TRKOWN 5117 if (pg->flags & PG_BUSY) 5118 (*pr)(" %d [%s]", pg->owner, pg->owner_tag); 5119 #endif 5120 (*pr)("\n"); 5121 } 5122 } 5123 } 5124 5125 #endif 5126 5127 /* 5128 * uvm_map_create: create map 5129 */ 5130 5131 struct vm_map * 5132 uvm_map_create(pmap_t pmap, vaddr_t vmin, vaddr_t vmax, int flags) 5133 { 5134 struct vm_map *result; 5135 5136 MALLOC(result, struct vm_map *, sizeof(struct vm_map), 5137 M_VMMAP, M_WAITOK); 5138 uvm_map_setup(result, vmin, vmax, flags); 5139 result->pmap = pmap; 5140 return(result); 5141 } 5142 5143 /* 5144 * uvm_map_setup: init map 5145 * 5146 * => map must not be in service yet. 5147 */ 5148 5149 void 5150 uvm_map_setup(struct vm_map *map, vaddr_t vmin, vaddr_t vmax, int flags) 5151 { 5152 int ipl; 5153 5154 rb_tree_init(&map->rb_tree, &uvm_map_tree_ops); 5155 map->header.next = map->header.prev = &map->header; 5156 map->nentries = 0; 5157 map->size = 0; 5158 map->ref_count = 1; 5159 vm_map_setmin(map, vmin); 5160 vm_map_setmax(map, vmax); 5161 map->flags = flags; 5162 map->first_free = &map->header; 5163 map->hint = &map->header; 5164 map->timestamp = 0; 5165 map->busy = NULL; 5166 5167 if ((flags & VM_MAP_INTRSAFE) != 0) { 5168 ipl = IPL_VM; 5169 } else { 5170 ipl = IPL_NONE; 5171 } 5172 5173 rw_init(&map->lock); 5174 cv_init(&map->cv, "vm_map"); 5175 mutex_init(&map->misc_lock, MUTEX_DRIVER, ipl); 5176 mutex_init(&map->mutex, MUTEX_DRIVER, ipl); 5177 } 5178 5179 5180 /* 5181 * U N M A P - m a i n e n t r y p o i n t 5182 */ 5183 5184 /* 5185 * uvm_unmap1: remove mappings from a vm_map (from "start" up to "stop") 5186 * 5187 * => caller must check alignment and size 5188 * => map must be unlocked (we will lock it) 5189 * => flags is UVM_FLAG_QUANTUM or 0. 5190 */ 5191 5192 void 5193 uvm_unmap1(struct vm_map *map, vaddr_t start, vaddr_t end, int flags) 5194 { 5195 struct vm_map_entry *dead_entries; 5196 struct uvm_mapent_reservation umr; 5197 UVMHIST_FUNC("uvm_unmap"); UVMHIST_CALLED(maphist); 5198 5199 UVMHIST_LOG(maphist, " (map=0x%x, start=0x%x, end=0x%x)", 5200 map, start, end, 0); 5201 if (map == kernel_map) { 5202 LOCKDEBUG_MEM_CHECK((void *)start, end - start); 5203 } 5204 /* 5205 * work now done by helper functions. wipe the pmap's and then 5206 * detach from the dead entries... 5207 */ 5208 uvm_mapent_reserve(map, &umr, 2, flags); 5209 vm_map_lock(map); 5210 uvm_unmap_remove(map, start, end, &dead_entries, &umr, flags); 5211 vm_map_unlock(map); 5212 uvm_mapent_unreserve(map, &umr); 5213 5214 if (dead_entries != NULL) 5215 uvm_unmap_detach(dead_entries, 0); 5216 5217 UVMHIST_LOG(maphist, "<- done", 0,0,0,0); 5218 } 5219 5220 5221 /* 5222 * uvm_map_reference: add reference to a map 5223 * 5224 * => map need not be locked (we use misc_lock). 5225 */ 5226 5227 void 5228 uvm_map_reference(struct vm_map *map) 5229 { 5230 mutex_enter(&map->misc_lock); 5231 map->ref_count++; 5232 mutex_exit(&map->misc_lock); 5233 } 5234 5235 struct vm_map_kernel * 5236 vm_map_to_kernel(struct vm_map *map) 5237 { 5238 5239 KASSERT(VM_MAP_IS_KERNEL(map)); 5240 5241 return (struct vm_map_kernel *)map; 5242 } 5243 5244 bool 5245 vm_map_starved_p(struct vm_map *map) 5246 { 5247 5248 if ((map->flags & VM_MAP_WANTVA) != 0) { 5249 return true; 5250 } 5251 /* XXX */ 5252 if ((vm_map_max(map) - vm_map_min(map)) / 16 * 15 < map->size) { 5253 return true; 5254 } 5255 return false; 5256 } 5257 5258 #if defined(DDB) 5259 void 5260 uvm_whatis(uintptr_t addr, void (*pr)(const char *, ...)) 5261 { 5262 struct vm_map *map; 5263 5264 for (map = kernel_map;;) { 5265 struct vm_map_entry *entry; 5266 5267 if (!uvm_map_lookup_entry_bytree(map, (vaddr_t)addr, &entry)) { 5268 break; 5269 } 5270 (*pr)("%p is %p+%zu from VMMAP %p\n", 5271 (void *)addr, (void *)entry->start, 5272 (size_t)(addr - (uintptr_t)entry->start), map); 5273 if (!UVM_ET_ISSUBMAP(entry)) { 5274 break; 5275 } 5276 map = entry->object.sub_map; 5277 } 5278 } 5279 #endif /* defined(DDB) */ 5280